Skip to content

Release v0.16.0 — Loopmaxxing, Google apps through Kortix, desktop tabs and a calmer session view - #12852

Merged
markokraemer merged 163 commits into
prodfrom
release/v0.16.0
Nov 18, 2026
Merged

markokraemer merged 163 commits into
prodfrom
release/v0.16.0

Conversation

@github-actions

@github-actions github-actions Bot commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Loopmaxxing, Google apps through Kortix, desktop tabs and a calmer session view

New

  • Loopmaxxing. A new developer landing at loopmaxxing.sh: one prompt in Claude Code, Codex, Cursor, OpenCode or pi adds a Software Factory to your repository. The page shows real runs that integrations start (a Sentry alert, a pull request, a Slack message), the plans and the cloud computer pricing, and a full-screen guide after you copy the prompt, with a send-to-yourself option on phones.
  • Guided setup. The setup guide offers the template that fits your repository, updates an older CLI instead of keeping it, offers Slack so teams can give agents work from Slack, and connects Linear with one sign-in. kortix onboard opens one setup page for every step a person must do and waits for it.
  • Your repository holds your Kortix setup. Setup proves your dev environment by running it, sizes the agents' computer to your stack, and resizes it from measured peaks.
  • Gmail, Google Calendar, Drive, Docs and Sheets connect through Kortix instead of a third-party connector service. The privacy policy states how Kortix handles Google user data.
  • Browser-style tabs in the desktop app.
  • kortix sessions usage shows tokens, cache and paid spend per session.
  • Reduce motion switch in Settings ▸ Preferences.
  • Chat links to artifacts show the artifact's title and details.
  • The model picker shows each subscription in one line with a usage meter.
  • ActionBar: one floating bar for a selection, unsaved edits and paste.
  • Session groups in the sidebar sit below Pinned and gain New group and Collapse others. One promo card covers Kortix in ChatGPT and Claude, your computer, Chrome, and the desktop and phone apps.

Improved

  • Sessions: side panels dock, the results column is compact, and right-side panels keep the page, history and files correct. Chat keeps line breaks, @ finds repository files, toasts read clearly, one agent name shows throughout, and pins stay put.
  • Sign-in screens are simpler: one legal line, cleaner type, and a code step that checks itself.
  • Workspace settings: 22 fixes from an audit of every settings page. Settings pages stop loading hidden app previews, and a rate-limited request retries.
  • Templates use the workspace layout, and a workspace's config gets its own repository. New workspaces no longer get a .kortix folder; the CLI link moves to git config.
  • Software Factory: its agent is now called the engineer, tickets link the real session, and the first ticket starts intake once.
  • Desktop app: the Review and Setup pages open as named tabs.
  • The terminal app splits sessions the same way as the web app. Mobile session search and filters reach the server.
  • Faster: sandbox health reads come from a warm git mirror, and spend per session is aggregated before it is joined.
  • Workspaces groundwork: every record that points at a project now also points at its workspace, and permissions read workspace-level grants. No change to how you work today.

Fixed

  • A Stop no longer loses a queued message to the turn it ends, and prompts survive an extension's own run and a stalled billing check.
  • A canceled plan no longer offers Cancel subscription, and it shows its remaining balance.
  • Managers can refresh login usage.
  • After kortix projects rm, the CLI moves its default off the deleted workspace.
  • An agent call with no matching subagent type runs inline instead of failing.
  • A shell command with a heredoc after cd no longer asks for permission.
  • A git server's "not our ref" refusal is retried instead of failing the fetch.
  • Parked sandbox removal with a dead provider credential answers a retryable error and alerts the team.
  • iOS: the thread scrolls to the end of the last answer, the agent and model sheet shows its tabs, and the sidebar stays still under the page.
  • A gated deployment serves the setup guide and the installer.
  • Fewer false alerts: Android WebView bridge errors, expected not-ready responses, and client-fixable billing refusals no longer page.
  • Get Kortix Usage no longer times out, and a script's calls draw once.
  • The session panel's state words and resize handle are translated.

Merging this PR publishes v0.16.0: tags it, cuts the GitHub Release, retags the tested images, and deploys ECS Fargate. Nothing is published until this merges.


Autofix with [code]smith
Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.

markokraemer and others added 30 commits November 13, 2026 22:14
…prompt and one copy button (#12670)

## Summary

The Loopmaxxing page (`/loopmaxxing`, behind
`KORTIX_PUBLIC_LOOPMAXXING_ENABLED`) now leads with one template. The
headline is fixed: "Add a Software Factory with 1 prompt." The prompt
box holds the Software Factory prompt and one "Copy the prompt" button;
the "Claude Code" and "Codex" marks under the prompt are gone. The
header button copies the same prompt. The template section shows the
Software Factory card and the "Any loop you describe" card; Marketing,
Sales, Support and Recruiting move into that card's example list.
Follow-up to #12644.

## Demo

**Web, `/loopmaxxing`, 1440×900, local stack with the flag on: the
prompt box under the headline.** A video adds nothing to a static
layout.

![prompt
box](https://github.com/user-attachments/assets/6778a357-75ce-428b-8632-e31184ebb937)

## What is affected

- Only the Loopmaxxing page, which is off on staging and prod.
- Not affected: the templates themselves, the CLI, the setup guide. The
other four templates stay in the build.

## Verification

| Check | Result | What it proves |
| --- | --- | --- |
| Screenshots of the hero and the template section on the local stack |
looked at | fixed headline, prompt names `kortix-ai/software-factory`,
one button, 2 cards |
| `pnpm --filter Kortix-Computer-Frontend typecheck` | PASS, 0 errors |
|
| `npx eslint` on the two changed files | PASS, no output | |
| A click on the button and a read of the clipboard | NOT RUN | the copy
handler is unchanged |

## Risks and rollback

The brand audit flags older lines in `factory-home.tsx` (a white label
on the accent button, two dimmed text colours, the headline's custom
size); none are new. Revert the commit to roll back.

<!-- codesmith:footer -->
---
<a
href="https://backend.blacksmith.sh/track/enable-autofix?expires=1797196480&installation_model_id=434224&pr_number=12670&ref=codesmith_pr_footer&repository=kortix-ai%2FCompanyOS&return_to=https%3A%2F%2Fgithub.com%2Fkortix-ai%2FCompanyOS%2Fpull%2F12670&signature=fa27f667eab0bb962a8d73e282e0bdfb85cb4234d2747ea4d1ca27ef7aaa6e62"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-light.svg"><img
alt="Autofix with [code]smith"
src="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark-3x.png"
width="190" height="28"></picture></a>
<sup><a
href="https://app.blacksmith.sh/kortix-ai/codesmith/CompanyOS/pr/12670?autoLogin=true&ref=codesmith_pr_footer">Open
in [code]smith</a> · Need help on this PR? Tag
<code>@codesmith-bot</code> with what you need.</sup>

<!-- codesmith:autofix:disabled -->
<!-- /codesmith:footer -->

---------

Co-authored-by: Dimitrije Glibic <dimitrije.glibic1@gmail.com>
…nies (KRTX-4871) (#12671)

## Review in 60 seconds

- `tests/e2e/helpers/desktop-parity.ts` (3430 lines) is now a 2-line
re-export index. Its code moved verbatim into
`tests/e2e/helpers/desktop-parity/{fixture,session,queue,shell,frame,native}.ts`
(largest 974 lines). Both spec files import the same two names unchanged
(KRTX-4871).
- The five repeated ceremonies now exist once, in `fixture.ts`:
`provisionDesktopScene` (user + account + project + one teardown),
`openNativeDesktop` (profile, launch, close/rm), `desktopDatabaseUrl`,
`resize`, `pinMacPlatform`.
- Every journey keeps its title, timeout, skip condition and assertions;
only the setup/teardown lines around them changed.

No demo video: code-only change (e2e test helpers; no product surface).

**Risk:** medium — the journeys need the Supabase + Electron stack this
sandbox does not have, so their bodies were not executed; titles, order
and counts were pinned with `playwright --list` before and after.

**Verified:** `playwright --list` identical for web/desktop (12) and
native (19); `tsc` on the spec-27 graph has the same 11 inherited errors
as the trunk and no unused locals; `vitest run test-runner-contract
create-auth-user-replay browser-transient-retry` 31 passed; `pnpm
test:unit` fails the same 4 tests at origin/dev. Not run: the journeys
themselves.

suna-skills: worktree, testing, learnings, contributing

ponytail: full · review: no findings left · markers: 0

Self-review: 5 items found and fixed (two `import.meta.url`-relative
paths that broke at the new depth, the project double-dispose in the
queue journey, unused imports, one mis-indented moved line).

LOC: 3430 → 3458 (`git diff --shortstat origin/dev`: 7 files changed,
3458 insertions(+), 3430 deletions(-)). `signal:` files_over_1000 for
this file 1 → 0, largest new file 974.

## Summary

Splits the 3430-line desktop-parity e2e helper by journey group and
collapses the five ceremonies every journey repeated. Behavior is
unchanged: every journey keeps its title, `test.setTimeout`, skip
conditions and assertions; the helpers only own the setup/teardown
around them. Resolves KRTX-4871.

Chosen readings (stated per the brief, where the issue was loose):
- Churn: the file had 15 commits in 30 days; two earlier workers blocked
on that. Intake re-queued with "decide it yourself", so this ships as
one fast PR on a fresh trunk (no other PR touched the file since).
- `queue.ts` alone was 1180 lines, so the six `nativeTest` journeys that
sit inside `defineDesktopParity` live in `shell.ts`;
`defineDesktopParity` itself is in `session.ts` and calls the four
groups in the original order.
- `withNativeDesktop(fn)` became `openNativeDesktop(baseURL, prefix)`
returning `{ profile, launch, dispose }`, so journey bodies keep their
`try/finally` and are not re-indented. `provisionDesktopScene` takes `{
databaseUrl, emailPrefix, name, fund? }` (no `page`/`baseURL`: they were
unused) and returns `{ email, user, session, accounts, project, dispose
}`; without `name` it only creates the user (the frame-back journey has
no project).
- Teardown order is now one: app close, profile rm, then project dispose
+ user delete (it drifted before). The database-URL guard has one
message.
- Two deliberate non-verbatim lines: `new
URL("../../../apps/desktop-electron", import.meta.url)` and the preload
path in `session.ts` gain one `../` for the new directory depth.

## Demo video

No demo video: code-only change.

## Type of change

Refactor (behavior-preserving restructure of e2e test helpers).

## What is affected

`tests/e2e/helpers/desktop-parity*` only (e2e test code). Specs
`27-desktop-parity.spec.ts` and `27-desktop-parity-desktop.spec.ts` are
untouched. No product code, API or SDK surface.

## How was this tested?

```
$ pnpm exec vitest run --config unit/vitest.config.ts test-runner-contract create-auth-user-replay browser-transient-retry
 Test Files  3 passed (3)
      Tests  31 passed (31)
```
```
$ E2E_DESKTOP_NATIVE=0 bun --bun x playwright test -c playwright.config.ts --list 27-desktop-parity   # trunk vs PR, locations normalised
list 0: titles identical (12 tests)
$ E2E_DESKTOP_NATIVE=1 ... same
list 1: titles identical (19 tests)
```
```
$ pnpm exec tsc -p tsconfig.tmp.json   # spec 27 graph, noUnusedLocals (e2e is excluded from the package tsconfig)
trunk: 11 errors (webkitAppRegion, __TAURI__, JSHandle, MessageBoxSync)   PR: same 11, no unused imports/locals
```
```
$ pnpm test:unit   # tests package, PR head
Test Files  5 failed | 137 passed (142)
Tests  4 failed | 1090 passed | 23 skipped (1117)
Pre-existing: pnpm test:unit fails the same at origin/dev (Test Files 5 failed | 137 passed; carry-web-static, cli-commands-doc, deployed-set, discover-flows-idempotent, shard)
```
```
$ grep -c <ceremony> across desktop-parity*.ts   # each is 1, all in fixture.ts
await createAuthUser( = 1 · mkdtemp(join(tmpdir() = 1 · KE2E_DATABASE_URL || process.env.E2E_DATABASE_URL = 1 · setViewportSize = 1 · Object.defineProperty(navigator, "platform" = 1
```
```
$ git diff --shortstat origin/dev
7 files changed, 3458 insertions(+), 3430 deletions(-)    # +0.8 % of the split file
$ git diff --color-moved=zebra --color-moved-ws=allow-indentation-change origin/dev   # 6364 moved-colored lines, 531 plain add/delete
$ wc -l: desktop-parity.ts 2 · fixture 334 · session 662 · queue 623 · shell 567 · frame 296 · native 974   # files_over_1000: this file 1 -> 0
```
Dedupe (step 2 vs the pure split): 290 insertions, 376 deletions. The
helpers (≈145 lines) cost more than the pure ceremony deletions, so the
"removes at least 200 lines" criterion holds for gross deletions (376)
but not net (−86); net vs trunk is +28 lines, inside the +2 % allowance.

Not run: the journeys themselves (need Docker/Supabase/Electron; this
sandbox has none). Their bodies are unchanged except for the ceremony
lines above.

## Security & data review

No secrets, customer data or auth changes; test helpers only, synthetic
users as before.

## Rollout / rollback

Merges with the trunk; revert the PR to restore the single file.

## Risks and rollback

The unexecuted journeys are the risk: a mistake in a helper would show
only in CI's browser lane. Revert the PR to restore the single file.

## Reviewer checklist

- [x] Behavior preserved (titles, timeouts, skips, assertions)
- [x] No product code touched
- [x] Pre-existing unit failures named

## Reviewer focus

1. `desktop-parity/fixture.ts` `provisionDesktopScene` /
`openNativeDesktop`: the failure paths still delete the user and
profile.
2. `queue.ts` mid-journey `project.dispose()` + `projectDisposed` flag
(replaces `project = undefined`).

<details>
<summary>Changed files</summary>

- `tests/e2e/helpers/desktop-parity.ts`: re-export index
- `tests/e2e/helpers/desktop-parity/fixture.ts`: shared fixture, API
client and the five helpers
- `session.ts`, `queue.ts`, `shell.ts`, `frame.ts`: journey groups;
`native.ts`: native-only journeys

</details>

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
## Summary

The v0.15.11 release suite (run 43511979254 on #12672) failed on drift
from the features merged today. This PR fixes every blocking failure. It
is cherry-picked onto `staging` next, then Promote runs again with
`ref=staging`.

No demo video: the change is test, generator, lint and lockfile drift.
No behavior a person sees changes.

## What failed, and the fix

| Failure | Cause | Fix |
|---|---|---|
| core-flows, core-db, deployed smoke: `Duplicate flow id "CHN-35"` |
#12644 gave the Slack join flow `CHN-35`, which already named the
agent-marks flow | The Slack join flow and its spec line are `CHN-36` |
| `cli.test.ts`, `cli-commands-doc.test.ts` | New CLI commands `plan`,
`onboard`, `templates` | MCP denial for each (they need a local
checkout), one CLI-doc row each, regenerated docs region |
| `unit-audit-route-labels`, web `audit-display-helpers` | 4 new routes
had no audit label | Labels in `audit-route-labels.ts`, regenerated
`audit-http-routes.generated.ts` |
| `unit-routes-manifest`, route coverage | Manifest not regenerated; 2
internal gateway routes in no flow | Regenerated
`routes.generated.json`; GW-10 covers
`/internal/gateway/idempotency/{get,put}` |
| `unit-catch-database-error-echo-guard` | The template reconcile catch
echoed any error | `rethrowDatabaseError(err)` first |
| `unit-request-primitives-tripwire` | A UUID regex literal in
`goals/rules.ts` | Match `[0-9a-f-]{36}`, then `isUuid` |
| web i18n strict audit (81 findings) | Unannounced Loopmaxxing copy
(79); 2 duration placeholders | Loopmaxxing files exempt until its flag
turns on, the same as `/launch`; placeholders are named constants |
| web Route Handler coverage | `/loops/[file]` | `/loops/` is a non-page
prefix |
| ui module layers | `composer-prefill-store` re-exported a store from a
higher layer | Re-export removed; the 2 web importers use
`@kortix/ui/composer` |
| starter `kortix-system` budget (1520 > 1500) | A second
prompt-variable table in `scheduling.md` | One table in
`kortix-yaml.md`, with the new rows. The removed copy said every fire
has `fired_at`, but a schedule fire has none. |
| GHSA lockfile check | `GHSA-r3h5-4585-7x3w` (astro, @astrojs/node) and
`GHSA-3v2p-hm54-8qcq` (@opentelemetry/instrumentation-http), both
published 2026-11-13 | Overrides: astro 7.4.7, @astrojs/node 11.1.12,
instrumentation-http 0.222.0 |
| gitleaks | A live-key-shaped literal in `plan.blackbox.test.ts` |
Assembled at runtime; `.gitleaksignore` names the commit that holds the
literal |
| CodeQL (17 new alerts) | New CLI, SDK and daemon code | 2 fixed
(GitHub link from the parsed remote; linear trailing-slash trim). The
rest are dismissed on GitHub with reasons. |

## Verification

| Command | Result |
|---|---|
| `pnpm --filter kortix-api test` cli, audit labels, routes manifest,
primitives tripwire, catch guard | 73 pass, 0 fail |
| `pnpm --filter Kortix-Computer-Frontend test` audit-display-helpers,
middleware-locale-routing, generate-audit-http-routes,
reset-client-state | 83 pass, 0 fail |
| `node --test scripts/audit-i18n.test.mjs` | 3 pass |
| `pnpm --filter @kortix/ui test` module-cycles, composer-prefill-store
| 6 pass |
| `pnpm --filter @kortix/starter test` managed-skills-drift | 6 pass
(1500 lines) |
| `vitest` deployed-set, discover-flows-idempotent, shard,
cli-commands-doc | 42 pass |
| `bun tests/bin/ke2e.ts coverage` | passed, 0 uncovered |
| CLI plan, onboard, templates black-box tests; shared setup-guide test
| all pass |
| `pnpm audit` (pnpm 8.11, CI filter) | 0 advisories |
| `blume build` with astro 7.4.7 | built; `#plan-one-go-onboard` and
`#templates` resolve |
| `gitleaks git` from `d61aa12438^` | no leaks |
| typecheck: kortix-api, @kortix/cli, @kortix/shared, @kortix/ui, web |
0 errors |

## Not fixed here

- `compliance-as-code-action` (Drata) is advisory (`tests-release.yml`
excludes it). The daily scan on `dev` already failed on 2026-11-12 and
2026-11-13.

## Risk and rollback

- The lockfile moves only the astro tree (used by the `blume` docs
build) and instrumentation-http (used by the legacy `@daytonaio/sdk`).
The docs build passed locally. Rollback: revert this PR.

<!-- codesmith:footer -->
---
<a
href="https://backend.blacksmith.sh/track/enable-autofix?expires=1797200383&installation_model_id=434224&pr_number=12675&ref=codesmith_pr_footer&repository=kortix-ai%2FCompanyOS&return_to=https%3A%2F%2Fgithub.com%2Fkortix-ai%2FCompanyOS%2Fpull%2F12675&signature=ae5a63dbbadfeda44831611aeb42ec9e4edbd0989c7f0601ddbde50763ad41b9"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-light.svg"><img
alt="Autofix with [code]smith"
src="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark-3x.png"
width="190" height="28"></picture></a>
<sup><a
href="https://app.blacksmith.sh/kortix-ai/codesmith/CompanyOS/pr/12675?autoLogin=true&ref=codesmith_pr_footer">Open
in [code]smith</a> · Need help on this PR? Tag
<code>@codesmith-bot</code> with what you need.</sup>

<!-- codesmith:autofix:disabled -->
<!-- /codesmith:footer -->
…ion joins (KRTX-4869) (#12661)

## Review in 60 seconds

- `listProjectGatewaySourceSpend` (`GET
/v1/workspaces/:id/gateway/sources`) now aggregates the window's log
rows per session first, then joins `project_sessions` (and its parent
alias) once per distinct session instead of once per request
(`apps/api/src/shared/session-costs.ts`).
- Why: the old plan hash-joined every `project_sessions` row of the
project twice against the 24 h window's rows, then sorted the whole
window by a computed source expression. That cost grows with the
project's lifetime session count, which is why the company workspace's
every-5-min poll slid from ~0.4 s to 1.3-2 s (prod p50 of the 200s,
2026-11-10 → 2026-11-13) with no deploy involved.
- Same response, same columns: a new integration test pins the
per-session regrouping (one session counted once for many requests, a
session with no row stays `unknown`).

No demo video: code-only change

**Risk:** low — a read-only aggregate rewritten to return identical
numbers (sums of per-session sums; cost columns are float sums so a
last-digit float difference is possible). It does not reduce the
window's O(rows) scan, see Reviewer focus.
**Verified:** `bun test
src/shared/gateway-source-spend.integration.test.ts` on real Postgres 17
(4 pass, 0 fail; the new plan test fails on the base query); every other
test file naming the function or its module passes (0 fail); `tsc
--noEmit -p apps/api` exit 0; eslint on both files exit 0. Local
stack/HTTP not run: the route is a thin wrapper over the function and
the function ran against the migrated schema. Acceptance (p95 within
1.5x of 398 ms for 24 h in prod) can only be proven after prod carries
it.
suna-skills: testing, learnings, contributing, create-pr, worktree
(read; the clone is a plain checkout, not a `pnpm worktree`)
ponytail: full · review: nothing to cut: one subquery replaces per-row
joins, no new helper or dependency · markers: 0
Self-review: 2 items found and fixed (reviewer's G9 items: plan
regression test, no non-null assertions) (checks A-H run by hand on the
diff: the no-session group is not counted as a session via
`count(perSession.sessionId)`, the session join keeps the project
predicate, source expression unchanged)

## Summary

KRTX-4869 (factory-key `infra:log:acacd8ef1736`): prod p95 of `GET
/v1/workspaces/:id/gateway/sources` was 1498 ms (121 requests, last 60
min) against 398 ms over the previous 24 h. Root cause, from Better
Stack ClickHouse: not a deploy. Only one workspace (the company's own,
polled every 5 min = 12 requests/h) gets a 200 here and those 200s have
grown steadily slower (p50 385 ms on 2026-11-10 06:00 UTC, 1.3-2.0 s
since 2026-11-11), while hundreds of cheap 403s per hour (a caller
without `workspace.gateway.spend.read`, ~20 ms) from 2026-11-12 18:00
UTC dilute the all-status p95. When the 403 share dipped, the 200s moved
above the 95th percentile. The 200 path's cost is the session joins
described above.

Refs KRTX-4869 (not "Closes": the 403 caller and the O(window rows) scan
remain, see Reviewer focus).

## Demo video

No demo video: code-only change (an API aggregate; the evidence is the
real-Postgres test run and the query plans under How was this tested?).

## Type of change

performance: a read-only aggregate query rewritten with identical
output.

## What is affected

`GET /v1/workspaces/:id/gateway/sources` (and its `/v1/projects/`
alias), the `kortix gateway sources` CLI and anything polling them.
Response shape and numbers unchanged. Not affected: the other gateway
aggregates, the budgets and `gateway_spend_daily`, migrations (none).

## How was this tested?

Local Postgres 17 (docker), full migration chain applied, then the
integration suite; 1.2 M synthetic `gateway_request_logs` rows over 14
days (86k in the 24 h window, ~2.9k distinct sessions), 150k
`project_sessions`:

```
$ bun test --env-file=/dev/null src/shared/gateway-source-spend.integration.test.ts   # TEST_DATABASE_URL set
(pass) listProjectGatewaySourceSpend > rolls spend up to the trigger that started a session or its parent
(pass) listProjectGatewaySourceSpend > counts one session once however many requests it made, and keeps a session with no row as unknown
(pass) listProjectGatewaySourceSpend > aggregates the window per session before it joins project_sessions (KRTX-4869)
(pass) listProjectGatewaySourceSpend > clamps the window to 1..720 hours
 4 pass
 0 fail

$ bun test src/shared/integration-llm-value.test.ts src/shared/integration-moved-project-costs.test.ts src/shared/dead-credential-log.test.ts
 3 pass / 6 pass / 28 pass, 0 fail

$ bun test src/shared/session-costs.test.ts session-costs-service.test.ts cost-rollups.test.ts cost-value-parity.test.ts src/router/routes/usage-{cost-by-project,cost-summary,session-costs}-http.test.ts src/mcp/shape.test.ts src/__tests__/unit-workspace-binding-coverage.test.ts
 14 / 14 / 43 / 2 / 23 / 9 / 33 / 33 / 85 pass, 0 fail

$ NODE_OPTIONS=--max-old-space-size=8192 npx tsc --noEmit -p apps/api   → exit 0
$ npx eslint src/shared/session-costs.ts src/shared/gateway-source-spend.integration.test.ts → exit 0
```

Timing of the real function against that data (4 runs each, ms) and
`EXPLAIN (ANALYZE)`:

```
before: 561 528 528 529   Execution Time 522 ms: two full Seq Scans of project_sessions (150k rows, hash joins) + sort of 86k rows
after : 295 241 256 243   Execution Time 255 ms: one hash join of the 2.9k session groups, parent alias by project_sessions_pkey index scan
```

Index-only scan on `idx_gateway_logs_project_time_covering_v4` (Heap
Fetches: 0) in both plans, so no covering-index change is needed.
Regression test: `aggregates the window per session before it joins
project_sessions` runs `EXPLAIN (GENERIC_PLAN)` on the real query (now
exported as `gatewaySourceSpendQuery`) and requires the per-session
`Group Key: gateway_request_logs.session_id` aggregate. On the base
query it fails (`expect(false).toBe(true)`), on the fix it passes.
Reviewer item 2 fixed: the `!` assertions are replaced by explicit
throws on a missing source.

## Security & data review

No auth, IAM, secret or customer-data change. The capability check
(`PROJECT_GATEWAY_SPEND_READ`) and the project/account row scope
(`projectRowsLlm`) are untouched; the data in this PR is synthetic.

## Rollout / rollback

No migration or flag. Ships with the normal API deploy; revert the one
commit to roll back.

## Risks and rollback

Revert the one commit. No migration, no flag. Float sums: per-session
`float8` sums are added in the outer query, so the last digit of `cost`
can differ from the previous single sum.

## Reviewer checklist

- [x] Tests ran on real Postgres and pass
- [x] Typecheck and eslint clean
- [x] No migration, no public SDK export change

## Reviewer focus

1. `apps/api/src/shared/session-costs.ts` (`perSession` subquery): the
`sessions` column must count non-null session groups only
(`count(perSession.sessionId)`), and the outer join keeps
`projectSessions.projectId = input.projectId`.
2. What this does not fix: the 24 h window's rows are still scanned and
grouped (O(requests in window)); the route stays ~250 ms at 86k rows
here, so prod may still read > 1.5x of 398 ms. A per-source rollup table
would be the next step (a migration, out of this PR). Also the 403 flood
from one agent session without `workspace.gateway.spend.read` is a
caller problem, separate from this latency.

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…-4881) (#12679)

## Review in 60 seconds

- #12675 fixed most of the v0.15.11 release suite (run
[43511979254](https://github.com/kortix-ai/CompanyOS/actions/runs/43511979254),
release PR #12672). Three packages-api tests still fail on `dev`
5418d83, so a re-promoted candidate would go red again. This PR fixes
those three.
- One is a real bug.
`apps/api/src/projects/lib/session-sandbox-env-build.ts` passed
`expectedGovernanceEtag` (from #12660) as a promise derived from
`agentConfigRead` with no rejection handler, so a failed agent-config
read surfaced as an unhandled rejection. It now goes through `started()`
like every other read in `buildSessionSandboxEnvVars`.
- The other two tests follow merged product changes: #12645 makes a
`retryable: false` fire terminal on its first attempt, and #12644 adds
the `github_event` webhook header. The five-attempt retry budget stays
covered by `trigger-execution-store.integration.test.ts`.

No demo video: code-only change

**Risk:** low. The only runtime change adds a rejection handler to one
promise; `publish.ts` still awaits it with its own `.catch(() => null)`.
**Verified:** the 3 test files → 3 fail on `dev` 5418d83, 59 pass, 0
fail on this branch (bun 1.4.4, the CI version) · `tsgo` typecheck of
apps/api → exit 0 · eslint on the changed files → 0 errors
Self-review: 1 item found and fixed (the rewritten trigger test first
used a thrown provider error. That requeues the create command and takes
the overlap-skip path, a different rule, so the test asserts the
ownerless rule #12645 introduced instead)
suna-skills: kortix-release, testing, learnings, worktree, contributing
ponytail: full · review: no over-engineering found; three hunks, one per
failure · markers: 0

## Summary

The v0.15.11 candidate on `staging` carries #12675 (via #12676) but
still fails three packages-api tests. This PR fixes them on `dev`; a
cherry-pick onto `staging` follows, then one re-promote from `staging`.
Closes KRTX-4881.

## Demo video

No demo video: code-only change

## Type of change

- [x] Bug fix: the unhandled `expectedGovernanceEtag` rejection
- [x] Tests that follow merged product changes (#12644, #12645)

## How was this tested?

Red on `dev` 5418d83 (after #12675), bun 1.4.4:

```
$ cd apps/api && bun test --isolate --env-file=scripts/test.env src/__tests__/unit-trigger-fire-paths.test.ts src/projects/lib/sessions.create-characterization.test.ts src/projects/lib/trigger-split-characterization.test.ts
(fail) git-backed triggers — runtime fire paths > a failed attempt the cron retries raises nothing; the fifth dead-letters and raises once
(fail) trigger split characterization > webhook payload keeps parsed body and selected headers
(fail) repository_access follows the agent grant AND launcherMayReadRepository
 55 pass
 3 fail
```

The same command on this branch (red before the fix, above):

```
$ cd apps/api && bun test --isolate --env-file=scripts/test.env src/__tests__/unit-trigger-fire-paths.test.ts src/projects/lib/sessions.create-characterization.test.ts src/projects/lib/trigger-split-characterization.test.ts
 59 pass
 0 fail
Ran 59 tests across 3 files. [4.64s]
```

```
$ cd apps/api && pnpm run typecheck
exit 0
```

Not run here: the Docker-backed lanes (no Docker in this sandbox). The
re-promoted release suite is the full gate.

## Security & data review

No new data path, secret or permission. Synthetic data only.

## Rollout / rollback

Merges into `dev`, then a cherry-pick PR onto `staging`, then one
Promote with `ref=staging`. To roll back, revert the squash commit;
nothing is stateful.

## Reviewer checklist

- [x] Each changed test follows a named product change, or proves the
fix
- [x] No check weakened, skipped or deleted

## Reviewer focus

1. `apps/api/src/projects/lib/session-sandbox-env-build.ts:426`: the
promise guard.
2. `apps/api/src/__tests__/unit-trigger-fire-paths.test.ts`: the test
asserts the #12645 rule.


## Re-check against `dev` ba0e573 (feedback cycle, 2026-11-13 23:05Z)

The merge-gate conflict reported on head 90a9420 is gone: head 8fe42d4
is a rebase onto `dev`, and `git merge-tree --write-tree origin/dev
8fe42d4` is clean against the newer `dev` ba0e573 too. No new commit was
needed. On a local (unpushed) merge of 8fe42d4 with ba0e573:

```
$ cd apps/api && bun test --env-file=scripts/test.env src/__tests__/unit-trigger-fire-paths.test.ts src/projects/lib/trigger-split-characterization.test.ts src/projects/lib/sessions.create-characterization.test.ts
 59 pass
 0 fail
Ran 59 tests across 3 files. [2.47s]

# the bug test with the session-sandbox-env-build.ts hunk reverted
$ bun test --env-file=scripts/test.env src/projects/lib/sessions.create-characterization.test.ts
(fail) repository_access follows the agent grant AND launcherMayReadRepository [1.98ms]
 12 pass
 1 fail

$ tsgo --noEmit -p apps/api   -> exit 0
$ bash apps/api/scripts/test.sh   (whole apps/api unit suite, 18 batches)
-> 37 fail, every one "partial-store tests need git 2.45 or later, as in the API image; this host has git version 2.39.5"
```

Pre-existing: the apps/api partial-store tests
(`config-releases/partial-store`, `projects/git/*`,
`git-proxy/project-snapshot-build`) fail the same at origin/dev ba0e573
in this sandbox (`3 pass, 22 fail` on 4 of those files; this sandbox has
git 2.39.5). Every other batch of the suite passes. The CI image has git
≥ 2.45.

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…d the default agent (KRTX-4876) (#12673)

## Review in 60 seconds

- **Grant membership has one implementation.** `agent-config-v2.ts` and
`session-secret-reach.ts` deleted their private copies (`listAdmits`,
`admits`) and call the delivery gate's own `listAdmits`, now exported
from `apps/api/src/secrets/strategy.ts`. A non-string entry admits
nothing in that one predicate, so `grantSecretToAgentV2` drops its own
string pre-filter.
- **The board reads the three ask tables through one builder.**
`openAskRows(table, decided, sessionIds)` in
`apps/api/src/projects/lib/session-board.ts` replaces three hand-written
selects (questions `answeredAt`, permissions `decidedAt`, colleague
access `decidedAt`). The connector-approval query keeps its own shape.
- **No `'default'` literal in the session reads.** `session-usage.ts`
and `session-message-authors.ts` compare with `DEFAULT_AGENT_SENTINEL`.
Characterization tests pin the behavior before and after: a DB suite for
the board's attention read, the serializer/branch-label name agreement,
and the grant-membership predicate.

No demo video: code-only change

**Risk:** low — a behavior-preserving restructure of server reads. The
one observable difference: a non-string entry in a grant list now admits
nothing, where each old copy threw `TypeError` (`entry.toUpperCase is
not a function`). Every typed caller passes `string[]`, so their answers
are identical.
**Verified:** unit set (`src/projects/lib` + every test naming a changed
symbol) → 2206 pass / 2 fail before, 2212 pass / 2 fail after (the same
2 pre-existing at origin/dev) · DB suites (board attention, message
authors, session usage, session asks) → 4/4 suites, 21 tests PASS ·
review item 1: dropping the `surfaceSessionId` arm of `openAskRows` →
board suite FAIL (1 of 4 tests), restored → PASS; same suite on
origin/dev's `session-board.ts` → PASS 4 tests · unit set rerun after
the trunk merge → origin/dev 2042 pass / 103 fail, head 2048 pass / 103
fail, identical fail list (this sandbox's direct `bun test` run; the 103
fail the same at origin/dev) · board attention suite → PASS on the
unchanged board code first · real HTTP routes (`GET
/v1/projects/:id/sessions`, `…/sessions/:sid`, `…/sessions/:sid/usage`)
→ identical answers at origin/dev and at the head · `pnpm typecheck`
(apps/api) → exit 0 · eslint on the changed files → 0 errors · whole
`apps/api` unit suite → every failure also fails at origin/dev (below)
suna-skills: worktree, testing, learnings, contributing, create-pr,
ponytail, ponytail-review
ponytail: full · review: Lean already. Ship. · markers: 0
Self-review: 4 items found and fixed
Feedback: 1 addressed · 0 answered without a change


## Feedback cycle (G9 review of head 33465f6)

1. **[C] `integration-session-board-attention.test.ts:105` — the
`surfaceSessionId` arm was unpinned.** Fixed in 37a692c: the
hidden-worker test now reads the three surface sessions alone, then the
three workers alone. Before, `boards()` also read w1..w3, whose own rows
marked the surfaces through the `sessionId` arm.

```
$ # mutant: openAskRows filter = inArray(table.sessionId, sessionIds) only (reverted after)
$ pnpm test -- --db-only integration-session-board-attention
[db-suites]   | (fail) readSessionBoards: what a session waits on > a hidden worker's ask of each kind marks the session a person opens for it
[db-suites] FAIL 0/1 suites, 4 tests, 2.2s
$ # unmutated head
[db-suites] PASS apps/api/src/__tests__/integration-session-board-attention.test.ts 4 tests 1.7s
[db-suites] PASS 4/4 suites, 21 tests, 3.3s   # with message authors, session usage, session asks
$ # origin/dev's session-board.ts with the new test
[db-suites] PASS 1/1 suites, 4 tests, 2.3s
```

## Summary

The session read layer kept a second copy of rules that already have an
owner. A change to one rule had to be found and made in every copy. One
comment already said the copies "must agree or the UI and the delivery
gate disagree". This PR gives three of the issue's rules one owner each.
It changes no behavior: tests pin each rule at origin/dev and at the
head.

Root cause: each module wrote its own private copy of a rule instead of
importing the module that owns it.

Closes KRTX-4876

### Findings, one line each

| Finding | Result |
| --- | --- |
| `dedupe-secret-grant-membership` | Done for `strategy.ts`,
`agent-config-v2.ts` and `session-secret-reach.ts`. **Struck for the
`serializers.ts` copy:** high-churn file (37 commits in 30 days; the
playbook skips a file over 10). |
| `dedupe-session-board-attention-queries` | Done: `openAskRows` in
`session-board.ts`, called 3×. |
| `dedupe-session-branch-title` | **Struck:** it needs `serializers.ts`
(37 commits in 30 days). The characterization test that both reads
resolve the same name is added. |
| `dedupe-live-session-predicate` | **Struck:** it needs
`session-list.ts` (12 commits in 30 days). |
| `dedupe-named-agent-predicate` | Done: both sites compare with
`DEFAULT_AGENT_SENTINEL`. |

Churn, `git log origin/dev --since=30.days --oneline -- <file> | wc -l`:
`serializers.ts` 37, `session-list.ts` 12, `agent-config-v2.ts` 10,
`session-secret-reach.ts` 4, `session-board.ts` 4,
`session-branch-labels.ts` 4, `session-message-authors.ts` 3,
`session-title-placeholder.ts` 2, `session-usage.ts` 2, `strategy.ts` 1.
The diff touches no file with more than 10.

**One deviation from `## Proposed change`.** The issue proposed a new
leaf `apps/api/src/secrets/grant-membership.ts`. The existing
`strategy.ts` already is that leaf: it imports only `node:crypto` and
types, and it is the delivery gate's own module. Exporting its
`listAdmits` gives the same single owner with one file fewer. The
untouched `serializers.ts` copy keeps an accurate comment ("mirroring
`listAdmits` in ../../secrets/strategy.ts"), and deduplicating it later
is a one-line import.

### Metrics, before → after

| Metric | Before (origin/dev) | After (head) |
| --- | --- | --- |
| `grant_membership_implementations` (`git grep -n "function
\(listAdmits\|grantAdmits\|admits\)(" apps/api/src`) | 4 | 2
(`strategy.ts`, and `serializers.ts`, which is struck) |
| `open_ask_query_copies` (the `or(inArray(<t>.sessionId),
inArray(<t>.surfaceSessionId))` predicate in `session-board.ts`) | 3 | 1
|
| `session_name_derivations` | 2 | 2 (struck) |
| `live_session_predicate_copies` | 3 | 3 (struck) |
| `default_agent_literal_sites` | 2 | 0 |
| Production LOC, `git diff --shortstat origin/dev...HEAD -- .
':(exclude)*.test.ts'` | — | 6 files changed, 32 insertions(+), 45
deletions(-): net −13 |
| All LOC, `git diff --shortstat origin/dev...HEAD` | — | 9 files
changed, 250 insertions(+), 45 deletions(-) (+218 of characterization
tests) |
| codebase-analysis `api-session-lib` `prod_loc` | 18427 | 18422 |
| codebase-analysis `dup_lines` / `unused_exports` / `import_cycles` | 0
/ 96 / 1 | 0 / 96 / 1 |

The issue's estimate (net −68) assumed the two struck findings. The
three findings done here delete 45 production lines and add 32.

`git diff --color-moved=zebra --color-moved-ws=allow-indentation-change
origin/dev...HEAD -- ':(exclude)*.test.ts'`: 0 moved blocks, 32 added
and 45 removed lines. Nothing is relocated verbatim. Each hunk deletes a
copy and imports the owner, or folds three selects into one builder
call. The reason in `agent-config-v2.ts`'s deleted comment (a mixed-case
list; the UI and the delivery gate must agree) moved onto the owner's
doc comment in `strategy.ts`, next to its original lines. That comment
also names the one remaining mirror, `grantAdmits` in the untouched
`serializers.ts`.

## Demo video

No demo video: code-only change

## Type of change

- [ ] Bug fix
- [ ] New feature
- [x] Refactor (no behavior change)
- [x] Tests

## What is affected

- Server reads only, in `apps/api`: the session board fields
(`board_state`, `attention`) of the session list and session read, the
secret-withheld advisory, the agent-scope secret grant write
(`grantSecretToAgentV2`), the session usage tree's `agent_name`, and
message authors' `agent`.
- The secret delivery gate (`resolveSecretDelivery`) calls the same
`listAdmits` as before. Only the export and the non-string guard are
new.
- Not affected: `@kortix/sdk` (no `packages/sdk` file), the DB schema,
migrations, routes, and every exported name of an `apps/api` module (one
export added, none removed).

## Verification

| Command / scenario | Result | What it proves |
| --- | --- | --- |
| Unit set at origin/dev `eab0879` (167 `src/projects/lib` files + 6
that name a changed symbol) | 2206 pass, 2 fail (pre-existing) |
baseline |
| Same unit set at the head | 2212 pass, 2 fail (the same 2) | behavior
preserved; +6 new tests |
| `pnpm test -- --db-only integration-session-board-attention` on the
unchanged board code | PASS, 4 tests | the characterization pins the old
query |
| `pnpm test -- --db-only` board attention, message authors, session
usage, session asks, at the head | PASS, 4/4 suites, 21 tests | the
refactored SQL on real Postgres |
| Mutation: colleague access filtered on `consumedAt` instead of
`decidedAt` | FAIL, 1 test | the board test detects a wrong decided
column |
| `listAdmits` tests without the `typeof` guard | 2 pass, 1 fail
(`TypeError`) | the non-string test depends on the guard |
| HTTP probe through the real Hono `app` on real Postgres, at origin/dev
and at the head | identical answers | the routes answer the same |
| `pnpm typecheck` in apps/api | exit 0 | types |
| eslint on the 6 changed source files | 0 errors | lint |
| Whole `apps/api` unit suite (`bash scripts/test.sh`) | 47 failing
tests: 46 fail the same at origin/dev; 1 passes alone (timing) | nothing
new breaks |

The change works because each deleted copy had the same body as the
owner it now calls, and the DB suite and the HTTP probe show the same
rows at origin/dev and at the head. Not verified: the browser UI,
because no web file changes. The local `pnpm worktree start` stack could
not start in this sandbox: the shared primary Supabase has no `kortix`
schema, and the API `.env` has no decryption key here. The HTTP probe
drives the same routes in-process instead.

## How was this tested?

All commands ran in the worktree `/workspace/suna-krtx-4876`. Unit runs
use the suite's own hermetic flags (`.
../../scripts/hermetic-test-env.sh; bun test --isolate
--env-file=scripts/test.env`, from `apps/api`).

Baseline: the unit set at origin/dev `eab0879`, before any change. The
two failures are pre-existing at origin/dev:

```
$ bun test --isolate --parallel=2 --env-file=scripts/test.env --timeout=15000 <167 src/projects/lib/*.test.ts> src/secrets/strategy.test.ts src/__tests__/unit-secret-grant-route.test.ts src/__tests__/unit-agent-config-v2.test.ts src/__tests__/unit-agent-config-v3.test.ts src/__tests__/unit-secret-request-withheld-route.test.ts src/setup-links/public-app.test.ts
 2206 pass
 2 fail (pre-existing at origin/dev)
Ran 2208 tests across 173 files. [27.55s]
(fail) repository_access follows the agent grant AND launcherMayReadRepository
(fail) trigger split characterization > webhook payload keeps parsed body and selected headers
```

Pre-existing: both fail the same at origin/dev when run alone:

```
$ bun test --isolate --env-file=scripts/test.env src/projects/lib/sessions.create-characterization.test.ts
(fail) repository_access follows the agent grant AND launcherMayReadRepository
 12 pass
 1 fail (pre-existing at origin/dev: TypeError in grantFromLoadedAgents, loaded.specs undefined)
$ bun test --isolate --env-file=scripts/test.env src/projects/lib/trigger-split-characterization.test.ts
(fail) trigger split characterization > webhook payload keeps parsed body and selected headers
 15 pass
 1 fail (pre-existing at origin/dev: the payload gained a github_event header field)
```

The same unit set at the head, after merging origin/dev `652c972`:

```
$ bun test --isolate --parallel=2 --env-file=scripts/test.env --timeout=15000 <the same 173 files>
 2212 pass
 2 fail (the same 2, pre-existing at origin/dev)
Ran 2214 tests across 173 files. [26.52s]
```

The new tests:

```
$ bun test --isolate --env-file=scripts/test.env src/secrets/strategy.test.ts -t listAdmits
(pass) listAdmits — grant membership > matches an identifier in any case
(pass) listAdmits — grant membership > an empty list admits nothing
(pass) listAdmits — grant membership > a non-string entry admits nothing and does not stop the scan
 3 pass
 0 fail
$ bun test --isolate --env-file=scripts/test.env src/projects/lib/session-branch-labels.test.ts
(pass) the branch label and the session read resolve the same name > a renamed session
(pass) the branch label and the session read resolve the same name > a runtime-titled session
(pass) the branch label and the session read resolve the same name > a placeholder-titled session has no name in either read
 10 pass
 0 fail
```

The non-string test fails without the guard (a red run with `typeof
entry === 'string' &&` removed, then restored):

```
$ bun test --isolate --env-file=scripts/test.env src/secrets/strategy.test.ts -t listAdmits
TypeError: entry.toUpperCase is not a function. (In 'entry.toUpperCase()', 'entry.toUpperCase' is undefined)
(fail) listAdmits — grant membership > a non-string entry admits nothing and does not stop the scan
 2 pass
 1 fail (red run before the guard)
```

The board characterization DB suite, run first on the unchanged board
code:

```
$ pnpm test -- --db-only integration-session-board-attention
[db-suites] PASS apps/api/src/__tests__/integration-session-board-attention.test.ts 4 tests 2.0s
[test] PASS db-suites 6.5s
```

The DB suites that cover the changed modules, at the head after the
merge:

```
$ pnpm test -- --db-only integration-session-board-attention integration-message-authors integration-session-usage integration-session-asks
[db-suites] PASS apps/api/src/__tests__/integration-session-usage.test.ts 6 tests 2.1s
[db-suites] PASS apps/api/src/__tests__/integration-session-asks.test.ts 6 tests 2.1s
[db-suites] PASS apps/api/src/__tests__/integration-message-authors.test.ts 5 tests 2.1s
[db-suites] PASS apps/api/src/__tests__/integration-session-board-attention.test.ts 4 tests 2.2s
[db-suites] PASS 4/4 suites, 21 tests, 3.6s
```

Mutation check: `openAskRows(sessionColleagueAccess,
sessionColleagueAccess.consumedAt, …)` instead of `decidedAt` (reverted
after):

```
$ pnpm test -- --db-only integration-session-board-attention
[db-suites]   | (fail) readSessionBoards: what a session waits on > an answered, decided or expired ask waits on nobody
[db-suites] FAIL 0/1 suites, 4 tests, 2.2s
```

The real routes: a scratch DB suite, not committed, imports `app` from
`src/index.ts`, seeds an account, a project, an owner token and five
sessions, and calls the routes. It ran with the source at origin/dev
(the six source files stashed) and at the head. With the list rows
sorted by session (the seeded UUIDs are random), the outputs are
identical:

```
$ pnpm test -- --db-only integration-zz-krtx4876-http-probe
[db-suites] PASS 1/1 suites, 1 tests, 4.7s
GET /v1/projects/:id/sessions -> 200 {"question":["needs_input","question"],"question+permission":["needs_input","question"],"colleague":["needs_input","approval"],"quiet":["idle",null],"permission":["needs_input","approval"]}
GET /v1/projects/:id/sessions/:sid -> {"question":[200,"needs_input","question"],"permission":[200,"needs_input","approval"],"colleague":[200,"needs_input","approval"],"quiet":[200,"idle",null],"question+permission":[200,"needs_input","question"]}
GET /v1/projects/:id/sessions/:sid/usage (agent 'default') -> 200 agent_name=null
GET /v1/projects/:id/sessions/:sid/usage (agent 'kortix') -> 200 agent_name="kortix"
$ diff <(sorted before) <(sorted after)
IDENTICAL-AFTER-SORT
```

Typecheck and lint:

```
$ cd apps/api && pnpm typecheck
> tsgo --noEmit --incremental --tsBuildInfoFile node_modules/.cache/tsgo.tsbuildinfo
exit 0
$ npx eslint <the 6 changed source files and 3 test files>
✖ 3 problems (0 errors, 3 warnings)   # the 3 warnings: test files are in the config's ignore list
```

The whole `apps/api` unit suite at the head. Every failing file also
fails at origin/dev:

```
$ cd apps/api && bash scripts/test.sh
API unit suite: 1403 files in 18 batches
batch failures: 2, 3, 0, 2, 0, 1, 0, 6, 2, 0, 29, 0, 2, 0, 0, 0, 0, 0 (47 failing tests)
$ bun test --isolate --parallel=2 --env-file=scripts/test.env --timeout=15000 <the 24 files with a failure>   # origin/dev (stashed)
 296 pass
 46 fail (pre-existing at origin/dev: git partial-store, route manifest, audit labels, CLI denial list, trigger fire paths, the 2 above)
$ bun test --isolate --parallel=2 --env-file=scripts/test.env --timeout=15000 <the same 24 files>   # head
 296 pass
 46 fail (pre-existing: the identical 46 tests as origin/dev)
$ bun test --isolate --env-file=scripts/test.env src/__tests__/e2e-preview-proxy.test.ts   # the 47th, run twice
 90 pass	 0 fail
 90 pass	 0 fail
```

The 47th failure ("Preview proxy: forwarding > retries transient project
env sync failures") failed once under full-suite load. It passes in both
targeted runs and twice alone, so it is timing, unrelated to this diff.

## Security & data review

- The grant predicate guards secret delivery. Its answers stay identical
for every `string[]` list. The unit tests of `resolveSecretDelivery`,
`secretDeliveryBlockedReason`, `withheldSecrets` and
`grantSecretToAgentV2`, and `unit-secret-grant-route.test.ts`, pass
unchanged.
- A non-string entry in a grant list now admits nothing (fail-closed)
instead of throwing.
- No data, no migration, no new route. Tests use synthetic values only.

## Rollout / rollback

No flag, migration or config. Revert the squash commit to roll back.

## Reviewer checklist

- [x] Behavior preserved: the characterization tests pass at origin/dev
and at the head.
- [x] No `@kortix/sdk` file and no exported `apps/api` name removed or
renamed.
- [x] No file with more than 10 commits in 30 days is touched.

## Reviewer focus

1. `apps/api/src/projects/lib/session-board.ts` `openAskRows`: each call
pairs a table with its own "decided" column (`answeredAt` for
questions).
2. `apps/api/src/secrets/strategy.ts` `listAdmits`: the `typeof entry
=== 'string'` guard is the only new logic; `agent-config-v2.ts` relies
on it after dropping its pre-filter.
3. The struck findings and their churn counts in the Findings table
above. A fifth case-insensitive membership check, `agentMayUseEnv` in
`apps/api/src/iam/agent-scope.ts` (existing, outside the issue's unit),
is left as it is.

<details>
<summary>Changed files</summary>

- `apps/api/src/secrets/strategy.ts`: `listAdmits` exported and guards
non-strings. Its doc comment names the remaining `serializers.ts` mirror
and why the answers must agree.
- `apps/api/src/projects/lib/agent-config-v2.ts`: its private
`listAdmits` and string pre-filter deleted; imports the owner.
- `apps/api/src/projects/lib/session-secret-reach.ts`: private `admits`
deleted; imports the owner.
- `apps/api/src/projects/lib/session-board.ts`: `openAskRows` builder;
three selects become calls.
- `apps/api/src/projects/lib/session-usage.ts`,
`apps/api/src/projects/lib/session-message-authors.ts`:
`DEFAULT_AGENT_SENTINEL` instead of `'default'`.
- Tests:
`apps/api/src/__tests__/integration-session-board-attention.test.ts`
(new DB suite),
`apps/api/src/projects/lib/session-branch-labels.test.ts` (name
agreement), `apps/api/src/secrets/strategy.test.ts` (grant predicate).

</details>

Routing: model=claude-agent/opus route=visual reason=visual-surface (the
diff turned out code-only).

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…sign helper (#12684)

## Review in 60 seconds

- **What:** the TRG-20 flow's GitHub repository webhook step (added by
#12644) calls `sign()`, which is defined only inside another step's
closure. At run time the step throws `ReferenceError: sign is not
defined`, so `contract / core-flows` fails on the v0.15.11 candidate
once the flow registry loads again. This PR defines `sign` in the step
itself.
- **Why:** #12675 and #12679 fixed the rest of the v0.15.11 suite (run
43511979254, release PR #12672), but `dev` 1f7fbc1 still carries this
one. The Promote Marko dispatched from 1f7fbc1 (run 43524336081)
includes the bug, so the fix goes to `dev` and then onto `staging`. No
check is weakened.
- **How:** `const sign = (body, key = secret) => webhookSignature(key,
body)`. GitHub signs the raw body alone, and `webhookSignature` already
returns `sha256=<hex HMAC>`. The step's five calls (`sign(rawBody)`,
`sign(rawBody, 'wrong-secret')`) type-check against it.

No regression test: the `tests` typecheck (`tsc --noEmit -p tests`)
already names this exact error, and TRG-20 itself is the test; it needs
the local stack, which this sandbox cannot run (no Docker).

**Risk:** low — two lines in one test flow; no product code.
**Verified:** at head f20ecc1 (origin/dev 83fc1ca merged in, no
conflict): `npx tsc --noEmit -p tests`: 5 × `TS2304 Cannot find name
'sign'` at origin/dev 1f7fbc1, 0 errors with the fix; `pnpm --filter
@kortix/tests test -- --exclude '**/carry-web-static*'` 1107 passed, 0
failed (the excluded file fails the same at origin/dev: sandbox
coreutils 9.1 lacks `cp --update=none`); `bun bin/ke2e.ts coverage`
passes; biome on the changed file: 25 diagnostics at origin/dev and at
head, none on the changed lines. Not run: TRG-20 itself (needs the local
stack; no Docker here).
suna-skills: worktree, testing, learnings
Self-review: 1 item found and fixed (the earlier, wider version of this
PR duplicated #12675 and #12679; it is reduced to the one remaining
failure).
ponytail: full · review: no findings (one local two-line helper, reuses
`webhookSignature`) · markers: 0
Feedback: 1 addressed (merge conflict with dev: the branch is rebuilt on
dev, then origin/dev 83fc1ca merged in cleanly; targeted tests rerun
after the merge) · 0 answered without a change
No demo video: code-only change

## Summary

TRG-20 (webhook ingress against a real webhook trigger) gained a
GitHub-scheme step in #12644. The step uses a `sign` helper that exists
only in a later flow's step (`const sign = (ts) => …` at line 2468), so
the reference is unresolved in TRG-20. The flow runs in the release
suite's core-flows lane (no `deployed` tag), so this fails the v0.15.11
suite once `CHN-35` (fixed in #12675) no longer stops the registry from
loading.

Closes KRTX-4886

## Demo video

No demo video: code-only change

## Type of change

- [x] Bug fix (release gate)

## How was this tested?

Red at origin/dev 1f7fbc1:

```
$ npx tsc --noEmit -p tests
src/flows/triggers.flow.ts(1776,29): error TS2304: Cannot find name 'sign'.
src/flows/triggers.flow.ts(1777,29): error TS2304: Cannot find name 'sign'.
src/flows/triggers.flow.ts(1778,89): error TS2304: Cannot find name 'sign'.
src/flows/triggers.flow.ts(1780,29): error TS2304: Cannot find name 'sign'.
src/flows/triggers.flow.ts(1780,68): error TS2304: Cannot find name 'sign'.
```

Green with the fix (rerun at head f20ecc1, after merging origin/dev
83fc1ca):

```
$ npx tsc --noEmit -p tests
(exit 0, no output)
$ pnpm --filter @kortix/tests test -- --exclude '**/carry-web-static*'
 Test Files  141 passed (141)
      Tests  1107 passed | 8 skipped (1115)
$ bun bin/ke2e.ts coverage
95.8% covered · 1687/1761 routes · 74 allowlisted · 0 uncovered
✓ coverage gate passed
```

Every other test the issue names (fixed on dev by #12675 and #12679)
passes at head f20ecc1, so this PR closes the last one:

```
$ cd apps/api && bash scripts/test.sh src/mcp/cli.test.ts src/__tests__/unit-catch-database-error-echo-guard.test.ts src/__tests__/unit-audit-route-labels.test.ts src/__tests__/unit-routes-manifest.test.ts src/__tests__/unit-request-primitives-tripwire.test.ts src/__tests__/unit-trigger-fire-paths.test.ts src/projects/lib/sessions.create-characterization.test.ts src/projects/lib/trigger-split-characterization.test.ts
 132 pass
 0 fail
Ran 132 tests across 8 files. [13.61s]
$ cd apps/web && bun test src/components/iam/audit-display-helpers.test.ts src/middleware-locale-routing.test.ts
 68 pass
 0 fail
$ cd packages/starter && bun test src/__tests__/managed-skills-drift.test.ts
 0 fail
Ran 6 tests across 1 file. [120.00ms]
$ cd packages/ui && bun test src/module-cycles.test.ts
 0 fail
Ran 2 tests across 1 file. [185.00ms]
```

Pre-existing, and the one file excluded above: the two carry-web-static
script tests fail the same at origin/dev in this sandbox (the full `pnpm
--filter @kortix/tests test` run: 1107 passed, those 2 failed): its
`infra/scripts/carry-web-static.sh` calls `cp --update=…`, which this
sandbox's coreutils `cp` rejects (`cp: option '--update' doesn't allow
an argument`). This PR does not touch it.

Not run: TRG-20 itself, which needs the local stack (Docker is not
available in this sandbox). The release suite on `staging` is the full
gate.

## Security & data review

No secrets, no customer data. Test code only; the secret is the flow's
own random `ke2e-hook-<uuid>`.

## Rollout / rollback

Merges into `dev`, then is cherry-picked onto `staging` for the v0.15.11
candidate. Revert the commit to roll back.

## Reviewer checklist

- `tests/src/flows/triggers.flow.ts`: `sign(body, key)` signs the raw
body (GitHub's `X-Hub-Signature-256`), not `<ts>.<body>` (the Kortix
scheme the step's other helper `signed()` covers).

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…box removal hits a dead provider credential (KRTX-4883) (#12687)

## Review in 60 seconds

- Prod `DELETE /v1/account/delete-immediately` answered a bare 500 (`2
parked sandbox(es) of account <uuid> could not be removed`) because the
provider's credential was dead: every parked-box removal failed with
HTTP 401 `Invalid credentials` (Better Stack logs, 2026-11-13
16:34:35-45, 4 attempts of one user, 2 boxes). The aggregate error hid
the cause and filed as an error of the account.
- Now `removeParkedBoxes`
(`apps/api/src/billing/services/account-erasure-stores.ts`) recognises a
provider auth failure (`isProviderAuthFailure`), pages the outage once
per window through the existing `escalateProviderAuthOutage` (names the
key to rotate), and, when every failure is that class, throws
`BillingError` 503 (controlled response, no per-account error group).
The deletion stays retryable: nothing is erased, the Stripe subscription
is untouched. Any other failure still throws the loud plain error.
- Verified by a real-database test that failed before the fix and passes
after.

No demo video: code-only change (API behavior, no UI).

**Risk:** low — Only the all-auth-failure branch changes; mixed or
non-auth failures behave as before. The sandbox reclaim pass
(`account-deletion.ts`) is unchanged and still logs its own per-box
errors. The real fix for the outage itself is rotating the provider key
(KRTX-4792), not code.
**Verified:** `pnpm test -- --db-only integration-account-erasure-stores
integration-account-deletion integration-retention-erasure` PASS 4/4
suites, 35 tests; unit `account-deletion` 41 pass 0 fail; `pnpm
typecheck` (api) exit 0. Not run: the local web stack (no UI; route
mapping `BillingError` -> 503 is the existing `http-errors.ts` branch).
suna-skills: worktree, testing, learnings, contributing
ponytail: full · review: 0 deletions left (one conditional, one existing
helper reused, no new module) · markers: 0

## Summary

KRTX-4883. Deleting an account whose parked sandboxes live on a provider
with a dead credential used to return a generic 500 and a Sentry error
group naming the account. It now returns a 503 "waiting on our sandbox
provider, try again shortly" and pages the actual outage (provider and
remediation) once per 15 min.

Closes KRTX-4883

## Demo video

No demo video: code-only change (API behavior, covered by a
real-database test).

## Type of change

fix (API behavior, no migration, no UI).

## Security & data review

No secrets or customer data logged: the escalation emits provider,
action and numeric status only. The 503 body carries no account id.
Erasure is not weakened: nothing is deleted or marked completed on this
path.

## What is affected

Immediate and scheduled account/person deletion when a parked
(stopped/archived) box fails to remove. Not affected: sandbox reclaim
pass, boxes removed successfully, 404/not-running handling, non-auth
provider failures, the DB schema.

## Verification

| Command / scenario | Result | What it proves |
| --- | --- | --- |
| New DB test before the fix (`--db-only
integration-account-erasure-stores`) | FAIL (8 pass, 1 fail) | 401
failure was a plain 500 Error |
| Same after the fix | PASS | 503 `BillingError`, one
`captureException`, account intact, retry completes |
| `--db-only integration-account-erasure-stores
integration-account-deletion integration-retention-erasure` | PASS 4/4
suites, 35 tests | neighbouring deletion suites hold |
| `pnpm test src/billing/services/account-deletion.test.ts
src/__tests__/billing/account-deletion.test.ts` (apps/api) | 41 pass, 0
fail | unit deletion tests |
| `pnpm typecheck` (apps/api) | exit 0 | types |

## Rollout / rollback

Low; revert the commit. After deploy, pattern `9329304b` should see no
new occurrences while the credential outage lasts; the outage itself
pages as `[daytona] API rejects every call`.

## Reviewer checklist

1. `account-erasure-stores.ts` `removeParkedBoxes`: the `authOutages ===
failed` branch.
2. The two new tests in `integration-account-erasure-stores.test.ts`.

<details>
<summary>Changed files</summary>

- `apps/api/src/billing/services/account-erasure-stores.ts`: classify
and escalate auth failures, 503 when all failures are auth.
-
`apps/api/src/billing/services/integration-account-erasure-stores.test.ts`:
401 and non-auth failure cases.

</details>

## How was this tested?

Before the fix (new test red):

```
$ pnpm test -- --db-only integration-account-erasure-stores
(fail) account erasure empties the stores outside the database > a dead provider credential answers a retryable 503 and pages the outage once, not the account
 8 pass
 1 fail
[db-suites] FAIL 0/1 suites, 9 tests, 4.3s
```

After the fix:

```
$ pnpm test -- --db-only integration-account-erasure-stores integration-account-deletion integration-retention-erasure
[db-suites] PASS apps/api/src/billing/services/integration-account-erasure-stores.test.ts 9 tests 4.2s
[db-suites] PASS apps/api/src/billing/services/integration-account-deletion.test.ts 18 tests 8.1s
[db-suites] PASS 4/4 suites, 35 tests, 9.4s

$ pnpm test src/billing/services/account-deletion.test.ts src/__tests__/billing/account-deletion.test.ts
 41 pass
 0 fail
Ran 41 tests across 2 files.

$ pnpm typecheck   (apps/api)
exit 0
```

Self-review: 1 items found and fixed (a new non-null assertion in my
test, replaced by `?.`; the 5 biome errors in the touched files exist at
origin/dev too).

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…12652)

## Review in 60 seconds

- `GET /v1/workspaces/:id/sandbox-health` (the sidebar poll) now calls
`allowStaleMirrorReads()`, as its sibling page-view routes `/snapshots`
and `/sandbox-templates` already do. Before, a cold memo ran the
template listing's manifest sync through a blocking `git fetch`.
- Prod p95 was 5575 ms over 194 requests (baseline 2029 ms) against a 22
ms median: about 15% of polls, on every API version since 0.15.3 (not a
deploy regression), took 1.1–8.6 s, right after a project opens. The
route lacked the opt-in; the 10 s SWR memo only hides the cost after the
first poll per project per API process.
- A regression test pins the opt-in through the real handler (red before
the fix, green after).

No demo video: code-only change

**Risk:** low — a poll may read a mirror up to 5 min old
(`KORTIX_GIT_VIEW_MAX_STALE_MS`) and never after a Kortix-made base move
(those invalidate the marker). The alert reads template state that
already tolerated a 10 s TTL. A cold clone of a never-fetched mirror is
unchanged, so the slow tail is reduced, not proven gone (the fetch share
of the 5 s is inferred from code and logs, not profiled).
**Verified:** `bun test` new test red before / green after; 48 related
API tests `48 pass; 0 fail`; `tsc --noEmit -p apps/api` exit 0. Not run:
local stack (no Docker in this sandbox); prod p95 (needs the fix live).
suna-skills: testing, contributing, learnings
ponytail: full · review: nothing to cut (one call, one test) · markers:
0
Self-review: 0 items found and fixed

## Summary

Prod p95 for `GET /v1/workspaces/:id/sandbox-health` was 5575 ms for an
hour (baseline 2029 ms); the median is 22 ms. The slow polls come from a
cold per-process memo, whose loader reaches `listTemplatesForProject` →
`syncManifestTemplatesForProject` → a manifest read through the git
mirror that blocks on `git fetch` unless the request opted in with
`allowStaleMirrorReads()`. This route did not. Closes KRTX-4867.

## Demo video

No demo video: code-only change

## Type of change

- [x] Performance fix (API)

## How was this tested?

```
$ bun test src/projects/routes/sandbox-health-stale-mirror.test.ts   # before the fix
error: expect(received).toEqual(expected)
-   "stale-reads",
    "git-project",
(fail) sandbox-health serves the warm mirror before it reads the git project
 0 pass
 1 fail

$ bun test src/projects/routes/sandbox-health-stale-mirror.test.ts   # after the fix, with the health budget, ttl-memo, sandbox-writer-gate, mirror-view-stale and project-resources tests
 48 pass
 0 fail
Ran 48 tests across 6 files.

$ bun test src/projects/routes/   # unit + the DB suite in the dir
 685 pass
 2 fail   # byo-new-workspace.integration.test.ts only: DB suite, ECONNREFUSED, excluded from the unit lane

$ tsc --noEmit -p apps/api
exit 0
```

```
$ bun tests/bin/ke2e.ts coverage
ke2e coverage — spec ↔ flow ↔ route parity
95.8% covered · 1681/1755 routes · 74 allowlisted · 0 uncovered
✓ coverage gate passed
```

Prod proof once deployed (ClickHouse, read-only): hourly
`quantile(0.95)(JSONExtractFloat(raw,'duration'))` over
`remote(t502678_kortix_api_logs)` where `JSONExtractString(raw,'path')
LIKE '%/sandbox-health'`; target within 1.5x of 2029 ms for 24 h.

## Security & data review

No auth, data, secret or customer-data change. Evidence above uses no
customer values.

## Rollout / rollback

Ships with the normal API deploy. Revert the one
`allowStaleMirrorReads()` call.

## Reviewer checklist

- [x] The call sits before the memo in the handler
(`apps/api/src/projects/routes/sandboxes.ts`), so each memo refresh runs
with it
- [x] Regression test fails without the fix


## Update 223c0c1 (G9 item 1)
The regression test no longer mocks `allowStaleMirrorReads`. It runs the
real handler inside a real request context against a real warm git
mirror and a moved local remote: with the opt-in the handler's git read
sees the mirror as it was (no fetch); with the opt-in removed it fetches
(test red: tip mismatch). Ran: `bun test --isolate
--env-file=scripts/test.env` on the new test + mirror-view-stale +
session-scope → `22 pass; 0 fail`.
Self-review: 1 items found and fixed

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…e timing out (KRTX-4877) (#12683)

## Review in 60 seconds

- **SCRIPT calls drawn once.** When a codemode script called a platform
tool (`kortix_schedules`, `kortix_usage`, …), the transcript listed the
call inside the `Ran script` row and then drew it a second time as its
own card below the script. `segmentTurn`
(`packages/sdk/src/core/turns/segments/segment-turn.ts`) now skips a
call whose id is `<script callID>/<n>` (pi's nested id) when that script
is in the same turn. A deliverable (`show`, `image_gen`), a sub-agent
(`agent`/`task`, whose card opens the child session), a call waiting on
a permission, a connector sign-in (its Connect card draws only under its
own row), and a direct call keep their own card. Web and mobile share
this segmentation.
- **Get Kortix Usage `summary` no longer times out.** Root cause: the
tool's `summary` read `GET /v1/usage/cost-summary`, which also builds a
daily series, a model breakdown and the project list in one GROUPING
SETS scan of every gateway row in the window. The card shows none of
that. On the largest workspace (~8M gateway rows per 30 days) that scan
hit the API's 25 s deadline, and the SDK's two 503 retries made it 76 s,
then an error. `kortix_usage`
(`apps/kortix-sandbox-agent-server/src/services/tools/platform/kortix_usage.ts`)
now reads `kortix billing costs --by workspace` for the window and the
window before, in parallel: the same totals at about a quarter of the DB
time (table below).
- **The CLI gives each workspace row its credits**
(`apps/cli/src/commands/billing-costs.ts`), as the summary JSON already
did, so the tool still never converts a price. The CLI ships in the same
sandbox image as the tool (`apps/sandbox/Dockerfile`, `cli-builder`,
existing).


https://github.com/user-attachments/assets/20b2059c-a895-4601-8164-2579fcc8adcd

**Walkthrough** (`/debug/session-fixture` on this worktree's web app,
with one local-only fixture turn shaped like the report: a script that
calls `kortix_schedules` twice and `kortix_usage` twice; that fixture
turn is not committed)
1. Before (fix stashed): `Ran script`, then the same four calls again as
four cards. ![Step
1](https://github.com/user-attachments/assets/0723ff75-7cea-466a-be41-4b9d1c51a165)
2. Before, script row open: the row already lists each call with its
status, time and error, plus the script output. ![Step
2](https://github.com/user-attachments/assets/bc529141-3c51-4d52-b424-0fdac311301d)
3. After: `Ran script · 4 tool calls`, then the answer text. No
duplicate cards. ![Step
3](https://github.com/user-attachments/assets/3068eef7-ce8e-48ad-9fcb-34fbc02a8429)
4. After, script row open: every call, the code and the output are still
there. ![Step
4](https://github.com/user-attachments/assets/c249dcd3-8581-48a7-aa0c-4d04225d13cf)

**Risk:** medium. A script's platform-tool card used to carry its own
detail, such as the list of people who can grant a refused permission.
Now that detail lives only in the script row's per-call line (`denied`)
and the script output. A call that waits on a permission still draws its
card, so the person can still answer it. The usage `summary` makes two
reads where it made one, but each is about 4× cheaper.
**Feedback:** G9 item 1 (a script's `kortix_connectors` connect lost its
Connect card) is fixed in e727b2a. `scriptChildIsListed` keeps a child
that `partStartsConnectorSignIn`. A refused connect (`ok: false`)
started no sign-in, so it stays listed in the script row, the same rule
`work-split.ts` already applies. Two new tests (v1 parts and the v2
`nested_part` reload) go red without the fix and green with it. The demo
video is unchanged: its scenario, a script that calls `kortix_schedules`
and `kortix_usage`, renders the same; the connect case now matches
origin/dev.
**Verified:** SDK `bash scripts/test.sh`: 7315 pass, 0 fail. CLI `pnpm
test`: 2808 pass, 0 fail. Agent server `pnpm test`: 1938 pass, 2 fail;
the same 2 fail at origin/dev (pre-existing). The new tests fail without
the fix (5 SDK, 1 CLI, 2 tool). `tsgo` typecheck: exit 0 for all 3
packages. ESLint (agent server) and Biome (sdk, cli): no new findings.
On prod, the old 30-day `cost-summary` read took 24.9–25.0 s (503), and
the new reads took 5.4–7.3 s. The real tool, run end to end against prod
with this branch's CLI, took 6.7 s for 30 days. Not run: the full
worktree stack (`pnpm worktree start` refuses: the shared local Supabase
has no `kortix` schema), so the demo runs on the worktree's web app
alone; the page needs no API.
suna-skills: worktree, testing, learnings, contributing, sdk,
kortix-design-system, kortix-brand, ponytail, ponytail-review
ponytail: full · review: Lean already. Ship. · markers: 0
Self-review: 5 items found and fixed

## Summary

Two fixes from one #kortix-triage report.

**Duplicate script rows.** pi publishes a script's call to a native-part
tool (`NATIVE_PART_TOOLS` in `harness/pi/runtime.ts`, existing) as its
own tool part with callID `<script callID>/<n>`. The session log stores
it as a `nested_part` block with `parent_call_id`, which reads back to
the same part. Since #11240 that includes every `kortix_*` platform
tool, so each such call drew twice: once in the script row, which lists
`metadata.calls`, and once as a standalone card. The new rule lives in
`segmentTurn` and uses that parent-child id. It covers the live stream
and a reloaded session, on web and mobile, with no harness change. A
child whose script is not in the turn is kept, so a partly loaded turn
never loses a call. A child that started a connector sign-in is kept
too: the web and mobile Connect card draws only under that call's own
row (`tool-part-renderer.tsx`), never inside the script row.

**Get Kortix Usage.** The `summary` action now reads the per-workspace
roll-up (`GET /v1/usage/cost-by-project`, existing) twice: once for
`[now − days, now)` and once for `[now − 2·days, now − days)`. The
current and previous totals are the same figures as before. On the same
1-day window, `cost-summary` and `cost-by-project` returned equal LLM
and storage totals; the compute totals differed by $0.0003 of live
billing between the two calls. An empty window shows `0 credits`.

Closes KRTX-4877. Out of scope: `top_sessions` (`GET
/v1/usage/session-costs`) and the web cost explorer's own `cost-summary`
read are unchanged. They aggregate the same raw rows per session, and on
the largest workspace they can still approach the deadline (measured
5.4–25 s for 30 days). Speeding them up needs a prod EXPLAIN and likely
a daily credit roll-up; that work is filed as KRTX-4882 with these
measurements.

## Demo video


https://github.com/user-attachments/assets/20b2059c-a895-4601-8164-2579fcc8adcd

Recorded with agent-browser on
`http://localhost:13000/debug/session-fixture`, the worktree's web app.
The first half is before (the SDK fix stashed): four duplicate cards
under `Ran script`. The second half is after: one script row and the
answer. Synthetic fixture data only.

## Type of change

- [x] Bug fix
- [ ] New feature
- [ ] Refactor
- [ ] Breaking change

## What is affected

- Session transcript (web and mobile, live and reloaded): a codemode
script's calls to `kortix_*`, messaging (`send_message`, `list_agents`),
team-task and `task_stop` tools draw only inside the script row. Not
affected: direct calls of the same tools; a script's `show`,
`image_gen`, `agent` and `task` calls; any call with a pending
permission; a script's connector sign-in (`kortix_connectors` connect,
or a shell `kortix connectors connect`), which keeps its Connect card;
other harnesses (their call ids have no script parent in the turn).
- My Kortix `kortix_usage` tool, action `summary`. `top_sessions` is
unchanged.
- CLI `kortix billing costs --by workspace --json`: each row gains
`credits: { total, llm, compute }` (additive). Human output and the
other modes are unchanged.
- No API, database, migration, auth or billing change.

## How was this tested?

Red before the fix: the new tests fail with the fix stashed.

```
$ cd packages/sdk && bun test src/core/turns/segments/segment-turn.test.ts src/core/session-log/turns/segments.test.ts   # fix stashed
(fail) a script's own calls > a platform or messaging call the script made draws only in the script row; its text and a direct call stay
(fail) a script's own calls > a deliverable, a sub-agent and a call waiting on a permission keep their own card
(fail) a codemode script's nested call on v2 (a pi `nested_part` block) > a nested platform call draws only in its script row, at rest as live (KRTX-4877)
 42 pass
 3 fail
```

The G9 fix (e727b2a): its two tests fail with only `segment-turn.ts`
stashed.

```
$ cd packages/sdk && bun test src/core/turns/segments/segment-turn.test.ts   # segment-turn.ts stashed
(fail) a script's own calls > a connector sign-in the script started keeps its own row: its Connect card draws there
 23 pass
 1 fail
$ cd packages/sdk && bun test src/core/session-log/turns/segments.test.ts   # segment-turn.ts stashed
(fail) a codemode script's nested call on v2 (a pi `nested_part` block) > a nested connector sign-in keeps its own row, where its Connect card draws (KRTX-4877)
 22 pass
 1 fail
```

```
$ cd apps/cli && bun test src/__tests__/billing-command.test.ts -t "credits, as the summary"   # fix not applied
error: expect(received).toEqual(expected)
 0 pass
 1 fail
```

```
$ cd apps/kortix-sandbox-agent-server && bun test src/__tests__/platform-tools.test.ts -t kortix_usage   # kortix_usage.ts stashed
(fail) kortix_usage > summary reads this workspace's credits for the window and the one before, never the account summary, and never converts a price itself
(fail) kortix_usage > summary of a workspace with no spend says 0, not an error
 0 pass
 2 fail
```

Green with the fix, head after merging origin/dev:

```
$ cd packages/sdk && bun test src/core/turns/segments/segment-turn.test.ts src/core/session-log/turns/segments.test.ts
 47 pass
 0 fail
$ cd packages/sdk && bun test src/core/turns/segments/ src/core/session-log/turns/   # head b766794, after merging origin/dev
 583 pass
 0 fail
$ cd apps/cli && bun test src/__tests__/billing-command.test.ts
 24 pass
 0 fail
$ cd apps/kortix-sandbox-agent-server && bun test src/__tests__/platform-tools.test.ts
 56 pass
 0 fail
```

Whole unit suite of each touched package, after merging origin/dev:

```
$ cd packages/sdk && bash scripts/test.sh          # summed over its isolated per-file runs, with e727b2a
7315 pass
0 fail
$ cd apps/cli && pnpm test
 2808 pass
 0 fail
Ran 2808 tests across 223 files. [461.09s]
$ cd apps/kortix-sandbox-agent-server && pnpm test
 1938 pass
 2 fail
Ran 1940 tests across 123 files.
```

The 2 agent-server failures are pre-existing: they fail the same at
origin/dev, with this PR's files stashed.

```
$ cd apps/kortix-sandbox-agent-server && bun test --isolate src/__tests__/pi-claude-agent.test.ts src/__tests__/harness-boundary.test.ts   # at origin/dev
(fail) Claude Code on a box whose image lacks it > refuses a download whose checksum is not the pin, and installs nothing
(fail) harness ownership boundary > host controllers call a different resolved service and retain its extra fields and native features
 21 pass
 2 fail
```

The consumers of the segmentation (web and mobile renderers):

```
$ cd packages/ui && bun test --isolate src/session/turn/ src/session/thread/
 791 pass
 0 fail
Ran 791 tests across 59 files.
$ cd packages/ui && bun test --isolate src/session/thread/image-gen-turn.test.tsx src/session/tool/tool-part-renderer.test.tsx
 23 pass
 0 fail
$ cd apps/mobile && bun test --isolate lib/session/turn-body.test.ts components/session/tool/tool-part-renderer.test.tsx
 94 pass
 0 fail
```

Static checks:

```
$ pnpm typecheck   # tsgo, in packages/sdk, apps/cli, apps/kortix-sandbox-agent-server
exit 0 (each)
$ cd apps/kortix-sandbox-agent-server && npx eslint src/__tests__/platform-tools.test.ts src/services/tools/platform/kortix_usage.ts
exit 0
$ npx biome check <each changed file>   # sdk and cli lint with Biome; counts equal origin/dev, segment-turn.ts clean
segment-turn.ts 0 (base 0) · segment-turn.test.ts 2 (base 2) · segments.test.ts 19 (base 19) · billing-costs.ts 1 (base 1) · billing-command.test.ts 1 (base 1)
$ biome lint packages/sdk/src/core/turns/segments/segment-turn.ts packages/sdk/src/core/turns/segments/segment-turn.test.ts packages/sdk/src/core/session-log/turns/segments.test.ts   # e727b2a vs 5e24841
Found 17 errors. (5e24841: Found 17 errors.; all pre-existing noNonNullAssertion lines in segments.test.ts)
$ cd packages/sdk && pnpm typecheck   # with e727b2a
exit 0
```

Reproduction and timing on prod, read-only. Each row is one call with
this session's workspace-bound token to
`https://api.kortix.com/v1/usage/*` on the largest workspace; server
`Server-Timing` `db;dur`, same fixed window, alternating order:

```
# the reported call, before: the tool's old argv
$ time kortix billing costs --project <workspace> --since <now-30d> --json
  ✗  HTTP 503: Request exceeded the 25s server processing deadline
real 1m16.480s

# same 1-day window, 3 rounds: old summary read vs the per-workspace read
cost-summary 200 db;dur=4700 | cost-by-project 200 db;dur=509
cost-summary 200 db;dur=2980 | cost-by-project 200 db;dur=613
cost-summary 200 db;dur=3279 | cost-by-project 200 db;dur=637

# 7 and 30 days: OLD = cost-summary [S,E); NEW = cost-by-project [S,E) and [S2,S), which the tool runs in parallel
d=7  r=1 OLD 200 db;dur=12112 | NEW cur 200 db;dur=2934 | prev 200 db;dur=1109
d=7  r=2 OLD 200 db;dur=12266 | NEW cur 200 db;dur=3051 | prev 200 db;dur=1098
d=30 r=1 OLD 200 db;dur=24895 | NEW cur 200 db;dur=7308 | prev 200 db;dur=2463
d=30 r=2 OLD 200 db;dur=24527 | NEW cur 200 db;dur=5412 | prev 200 db;dur=2219

# the changed tool end to end (runTool → this branch's CLI → prod)
new tool summary days=1:  5.8s ok=true
new tool summary days=7:  13.4s ok=true
new tool summary days=30: 6.7s ok=true   (the old argv, same minute: 24.5s)
```

Prod timings swing with database load: the same read varied 2–4× between
minutes. The interleaved rows compare the two reads under the same load.
The tool's end-to-end times include `bun` compiling the CLI from source
on each call; the image ships a compiled binary.

Not run: Docker-backed suites. None covers a changed module: there is no
API, DB or migration change.

## Security & data review

No auth, permission, data-model or secret change. The tool reads through
the workspace-bound token it already used. `/v1/usage/cost-by-project`
has the same `OWN_WORKSPACE` binding as `/v1/usage/cost-summary`
(`apps/api/src/middleware/workspace-binding.ts`, existing). The demo
uses the synthetic session fixture only. The prod timings above were
read-only calls on our own internal workspace, and this PR records no
customer data, amounts or ids.

## Rollout / rollback

Ships with the next sandbox image (agent server and CLI together) and
the next web and mobile builds. A sandbox image that predates the CLI
change cannot run the new tool, because both come from this commit.
Rollback: revert this PR. Nothing persistent changes.

## Reviewer checklist

- [x] Every hunk traces to KRTX-4877
- [x] New tests fail without the fix and pass with it
- [x] Whole unit suite of each touched package run; failures shown as
pre-existing at origin/dev
- [x] Platform-visible behavior recorded (before and after) with
synthetic data
- [x] No customer data in code, tests, screenshots or this body

Reviewer focus: `segment-turn.ts` `scriptChildIsListed` (which children
keep a card), and `kortix_usage.ts` `summary` (the two windows are
contiguous and the same length).

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
## Review in 60 seconds

- **What:** #12675, #12679 and #12644-follow-ups fixed most of the
v0.15.11 release-suite drift on `dev`. One failure remains: the
`../agents` stub in `sessions.create-characterization.test.ts` returns
neither `specs` nor `errors` (#12660 reads `loaded.specs`).
`tests/src/flows/triggers.flow.ts` (existing, now identical to dev)
needs no change.
- **Why it is safe:** tests only. One stub line.
- **How it was checked:** the three api unit files run green (59 pass, 0
fail) after merging `origin/dev` (which already holds the `started()`
fix and the trigger-fire-path tests from #12679).
**Risk:** low — test-only.
**Verified:** `bash apps/api/scripts/test.sh <3 files>`: 59 pass, 0
fail. Not re-run: tests typecheck (unchanged vs. the last verified
head), full release suite (gate after re-promote).
No demo video: code-only change
suna-skills: worktree, testing, learnings (from AGENTS.md)
Self-review: 0 items found and fixed
ponytail: full · review: clean · markers: 0

## Summary

Closes KRTX-4884

| Failure | Cause | Fix |
| --- | --- | --- |
| `sessions.create-characterization` `repository_access` | #12660 reads
`loaded.specs` | The `../agents` stub returns `specs: []`, `errors: []`
|

## Demo video

No demo video: code-only change

## Type of change

- Bug fix: tests follow merged behavior.

## How was this tested?

```
$ bash apps/api/scripts/test.sh src/__tests__/unit-trigger-fire-paths.test.ts src/projects/lib/sessions.create-characterization.test.ts src/projects/lib/trigger-split-characterization.test.ts
 59 pass
 0 fail
Ran 59 tests across 3 files. (re-run at head 5093d43: sessions.create-characterization 13 pass, unit-trigger-fire-paths 29 pass, unit-trigger-spec-characterization 4 pass; 0 fail)
```

```
$ bun test --env-file=scripts/test.env --isolate src/__tests__/unit-*.test.ts src/projects/lib/*.test.ts (integration files excluded, they need a DB)
 6638 pass
 0 fail
Ran 6639 tests across 532 files.
```

```
$ cd apps/api && bun test --env-file=scripts/test.env --isolate $(ls src/projects/lib/*.test.ts | grep -v integration)   # head 5093d43
 1869 pass
 0 fail
Ran 1869 tests across 154 files.
```
The `*.integration.test.ts` files in that folder need a DB and are
excluded: unrelated to this PR, which changes no file they cover.

## Rollout / rollback

Test-only change; no rollout. Rollback is a revert.

## Reviewer checklist

- [x] Tests only, no runtime code changed
- [x] The three api unit files pass

## Security & data review

No auth, data or secret handling changed; test files only.

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
… git failure (KRTX-4890) (#12688)

## Review in 60 seconds

- A git object fetch the remote refuses (`fatal: remote error:
upload-pack: not our ref <id>`, an object a force-push or deleted branch
left unreachable) was thrown as a plain `Error`: an unhandled 500 that
paged Better Stack (pattern `a1dfeecd`, application 2346961, 1
occurrence / 1 user).
- `fetchMissing` now throws a typed `ObjectNotOnRemoteError`;
`ensureStoreObjects` wraps it in a `GitOperationError`, and `not our
ref` joins the transient mirror pattern, so the global handler answers
the existing retryable 503 without paging. The next refresh drops the
stale ref.
- Every other fetch failure stays a plain `Error` and keeps paging; the
regression test fails without the change.
No demo video: code-only change

**Risk:** low — a permanently unreachable object now answers a retryable
503 instead of a 500 (neither can succeed; the 503 no longer pages).
**Verified:** `bun test src/projects/git/object-not-on-remote.test.ts` 4
pass 0 fail (the ensureStoreObjects test goes red when the wrap is
removed); `tsc --noEmit` clean; eslint clean on changed sources. Not
run: the 29 `partial-store` fixture tests need git >= 2.45 (sandbox has
2.39.5; they fail the same at origin/dev with "partial-store tests need
git 2.45 or later"); local stack (no Docker).
suna-skills: pull_request_template + hermetic test runner only (repo
AGENTS.md and skills were not read in full); ponytail: full · review: no
findings · markers: 0
Self-review: 1 item found and fixed (G9 item C: ensureStoreObjects wrap
now tested end to end)

## Summary

Fixes KRTX-4890 (factory-key
`betterstack:2346961:a1dfeecd03d5e389cdfd28b4828d6b231a8b51ddc6bc9c05311b6a88b6c7c4a7`):
`git fetch of 1 objects failed (exit 128): … not our ref <id>` at
`object-store.ts:fetchMissing` is now a typed, retryable, non-paging
outcome.

## Demo video

No demo video: code-only change

## Type of change

- [x] Bug fix

## How was this tested?

```
$ bun --env-file=scripts/test.env --isolate test src/projects/git/object-not-on-remote.test.ts
 4 pass
 0 fail
$ (same test with the source change stashed)
 0 pass
 1 fail   (ObjectNotOnRemoteError export missing)
$ bun --env-file=scripts/test.env --isolate test src/projects/git/
 229 pass
 29 fail  (all partial-store fixtures: "partial-store tests need git 2.45 or later ... this host has git version 2.39.5"; Pre-existing: fails the same at origin/dev)
$ NODE_OPTIONS=--max-old-space-size=6000 npx tsc --noEmit -p .   # exit clean
$ npx eslint --no-warn-ignored src/projects/git/mirror.ts src/projects/git/object-store.ts   # clean
```

Confirm after deploy: Better Stack pattern `a1dfeecd` (application
2346961) shows 0 new occurrences for 24 h.

## Security & data review

No new input, secret or customer data. The error text carries only an
object id.

## Rollout / rollback

Ships with the API deploy. Revert the PR to restore the old 500.

## Reviewer checklist

- [x] Regression test fails without the fix
- [x] Other fetch failures still page

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…00) (#12691)

## Review in 60 seconds

- On iOS the Sessions page sent every list request with **no query
string**. Search, the All / Mine / Shared / Automated chips, the pinned
and archived views and paging were all ignored by the server, so the
list showed every session, automated runs included, whatever was typed.
- Cause: `@kortix/sdk` added the query only when `URLSearchParams#size >
0`. The app installs `react-native-url-polyfill/auto` (existing
`apps/mobile/api/supabase.ts`), whose URLSearchParams
(whatwg-url-without-unicode 8.0.0-3) has no `size`. So `undefined > 0`
is false and the query is dropped.
- Fix: one helper, `querySuffix()` in
`packages/sdk/src/core/http/url.ts`, reads `toString()` instead. All
nine SDK sites that checked `size` use it, and session-costs' local copy
is deleted. Regression tests fail without the fix, under a size-less
URLSearchParams in the SDK and under the real polyfill class in the
mobile app.

No demo video: the iOS app cannot run in this sandbox (no simulator or
device). The proof is the real SDK under the device's URLSearchParams
against the API, below.

**Risk:** low — the query string each call builds is unchanged where
`size` exists (web, Node, Bun); only runtimes without `size` (the mobile
app) now send it. JS only.
**Verified:** device-polyfill run against the API: trunk sends no query
(50 mixed rows); fix sends `?parent=root&started_by=humans&q=…` (0 rows)
· `bun test` new tests red without fix, green with it · `@kortix/sdk`
full unit suite 475 files, 7316 pass, 0 fail · SDK typecheck, tripwire,
`smoke:install` pass · mobile typecheck adds no error
suna-skills: sdk, contributing, create-pr, testing, worktree, ponytail
ponytail: full · review: Lean already. Ship. · markers: 0
Self-review: 1 item found and fixed (three single-use `qs` locals
inlined)

## Summary

On the iOS app, typing in the Sessions page search left the full list on
screen, and "All" listed automated runs. The app never sent its filters:
every SDK request builder that guarded its query with
`URLSearchParams#size` sent none under the app's URL polyfill. After
this change the search, the scope chips, the pinned and archived views
and the paging cursor reach the server, which already filters correctly.
Also fixed by the same cause: message-history paging (`limit`,
`before`), the live session log, the asks list, cost reads, token list
and revoke, and the session start options on mobile.

Closes KRTX-4900

## Demo video

No demo video: the iOS app cannot run in this sandbox (no simulator or
device). See "How was this tested?" for the real SDK, under the device's
URLSearchParams, against the API.

## Type of change

`fix(sdk)`: a client transport bug. No API, DB, migration or public SDK
surface change (`querySuffix` is internal to the SDK; the public-surface
tests pass). The mobile version is bumped to 1.6.37 as
`apps/mobile/AGENTS.md` requires for a shared-package change that alters
mobile behavior.

## What is affected

- Mobile app (iOS and Android): Sessions page search, scope chips,
Filter sheet Archived view, paging, the drawer's sections, message
history paging, the session log, asks, session start options.
- Web, desktop, CLI and Node consumers: no behavior change. Their
URLSearchParams has `size`, and the query string they send is
byte-identical.
- Not affected: the API, the database, IAM, billing.

## How was this tested?

The real SDK, under the URLSearchParams the app installs, against the
prod API. Read-only GETs with an agent token; only counts are printed.
On origin/dev f3bd64c:

```
All + search "biomarkers": query sent=(none) rows=50 by_initiator={"trigger":35,"member":15} child_rows=44
Automated chip: query sent=(none) rows=50 by_initiator={"trigger":35,"member":15} child_rows=44
```

The same script on this branch:

```
All + search "biomarkers": query sent=?parent=root&started_by=humans&q=<text> rows=0 by_initiator={} child_rows=0
Automated chip: query sent=?parent=root&started_by=automated rows=50 by_initiator={"trigger":50} child_rows=0
```

The mobile regression with the real polyfill class
(`react-native-url-polyfill/js/URLSearchParams`). Red run before the fix
(SDK source reverted, test kept):

```
cd apps/mobile && bun test --isolate lib/projects/url-polyfill-query.test.ts
- {
-   "cursor": "c-2",
-   "parent": "root",
-   "q": "biomarkers",
-   "started_by": "humans",
- }
+ {}
(fail) the Sessions page query reaches the server under the device URLSearchParams
 0 pass
 1 fail
```

With the fix:

```
cd apps/mobile && bun test --isolate lib/projects/url-polyfill-query.test.ts components/session/sessions-list.test.tsx lib/projects/paged-hooks.test.tsx lib/session/session-tree.test.ts
 57 pass
 0 fail
 160 expect() calls
Ran 57 tests across 4 files.
```

The SDK regression. Red run before the fix:

```
cd packages/sdk && pnpm test src/core/rest/projects-client/sessions.test.ts -- -t "without size"
error: expect(received).toBe(expected)
Expected: "root"
Received: null
(fail) listProjectSessionsPage keeps its query under a URLSearchParams without size
 0 pass
 1 fail
```

With the fix, every SDK test file that covers a changed module, one `bun
test --isolate <file>` each (runner pass and fail lines):

```
src/core/rest/projects-client/asks.test.ts: 4 pass 0 fail
src/core/rest/projects-client/session-costs.test.ts: 21 pass 0 fail
src/core/rest/projects-client/session-sandbox.test.ts: 22 pass 0 fail
src/core/rest/projects-client/sessions.test.ts: 138 pass 0 fail
src/core/rest/projects-client/tokens.test.ts: 16 pass 0 fail
src/core/session/runtime-actions.test.ts: 3 pass 0 fail
src/core/session/runtime-verbs.test.ts: 16 pass 0 fail
src/deprecated-exports.test.ts: 173 pass 0 fail
```

The whole `@kortix/sdk` unit suite (`bash scripts/test.sh` runs one bun
process per file, so it prints one summary per file). It exited 0 across
475 per-file runs; every fail line reads `0 fail`, and the pass lines
sum to 7316:

```
cd packages/sdk && bash scripts/test.sh
 73 pass
 0 fail
 135 expect() calls
Ran 73 tests across 1 file.
(last of 475 per-file summaries; exit 0; sum of pass lines 7316; distinct fail lines: " 0 fail")
```

Gates:

```
pnpm --filter @kortix/sdk typecheck
typecheck_exit=0
pnpm --filter @kortix/sdk run smoke:install
OK: @kortix/sdk imports and constructs from the packed tarball
✔ install smoke test passed
bun test --isolate src/index.isomorphic.test.ts (the framework-free tripwire)
 34 pass
 0 fail
bun test --isolate src/public-surface.test.ts
 1 pass
 0 fail
bun test --isolate src/public-type-surface.test.ts
 1 pass
 0 fail
```

Pre-existing: `cd apps/mobile && pnpm typecheck` fails the same at
origin/dev f3bd64c with the one error
`components/session/ProjectScreen.tsx(999,9): error TS2322 … Property
'openSpringConfig' does not exist on type 'IntrinsicAttributes &
DrawerProps'`. This PR adds no error. ESLint: neither `packages/sdk` nor
`apps/mobile` has an ESLint config.

Not verified: the change on a real device. The iOS and Android apps
cannot run here. Device users get it with the next mobile JS release
(see Rollout).

## Security & data review

No auth, permission or data-model change. The fix makes requests carry
the filters the app already meant to send. The server applies its own
visibility rules to them as before. No customer data in the diff or this
body.

## Rollout / rollback

- JS only. Web and the CLI pick it up on their next release, with no
visible change. The mobile app needs a JS release: the mobile version
moves to 1.6.37, and per `apps/mobile/AGENTS.md` the store build for
that runtime follows the prod promotion. Promotion and store submission
are not part of this PR.
- Rollback: revert the commit.

## Reviewer checklist

- [x] Root cause reproduced with the exact polyfill class and with the
real SDK against the API, before and after.
- [x] Each new test fails without the fix and passes with it.
- [x] No SDK source reads `URLSearchParams#size` to build a query: `grep
-rnE 'params\.size|query\.size|search\.size' packages/sdk/src
--include=*.ts`, tests excluded, finds only the comment on
`querySuffix`.
- [x] Public SDK surface unchanged (`querySuffix` is not exported from
an entry point).

## Reviewer focus

1. `packages/sdk/src/core/http/url.ts`: `querySuffix` and its nine
callers. Check that each call site builds the same string as before on a
runtime that has `size`.
2. `apps/mobile/lib/projects/url-polyfill-query.test.ts`: it swaps in
the real polyfill class and asserts the outgoing URL; nothing under test
is mocked.

<!-- codesmith:footer -->
---
<a
href="https://backend.blacksmith.sh/track/enable-autofix?expires=1797208209&installation_model_id=434224&pr_number=12691&ref=codesmith_pr_footer&repository=kortix-ai%2FCompanyOS&return_to=https%3A%2F%2Fgithub.com%2Fkortix-ai%2FCompanyOS%2Fpull%2F12691&signature=91acdc3ccc73675597a50014f9a597fc00f845745e7d246e812f156e40a08760"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-light.svg"><img
alt="Autofix with [code]smith"
src="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark-3x.png"
width="190" height="28"></picture></a>
<sup><a
href="https://app.blacksmith.sh/kortix-ai/codesmith/CompanyOS/pr/12691?autoLogin=true&ref=codesmith_pr_footer">Open
in [code]smith</a> · Need help on this PR? Tag
<code>@codesmith-bot</code> with what you need.</sup>

<!-- codesmith:autofix:disabled -->
<!-- /codesmith:footer -->

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…t (KRTX-4897) (#12690)

## Review in 60 seconds

- Prod logged 284 `[session-token-grant] stored session grant names an
undeclared agent; healing to the session agent` errors in 1 h, all from
ONE session (`storedAgent: hack-dispatch`), one per connector call. That
session was created with `hack-dispatch`, which the manifest no longer
declares, so the "heal" resolves to the same agent: nothing is healed,
yet an error was logged on every call.
- Fix: log only when the heal changes the agent
(`apps/api/src/projects/lib/session-token-grant.ts`). Grants and
authorization are unchanged.
- Why it matters: one stuck session drowned the error channel and
tripped the new-pattern infra alert, hiding real errors.
No demo video: code-only change

**Risk:** low — log-only change; the authorization path is untouched.
**Verified:** 29 pass; 0 fail on the three session-token-grant test
files (new test failed before the fix); `tsc --noEmit -p apps/api` exit
0; local stack not run (no Docker in sandbox).

suna-skills: worktree, testing, learnings, contributing
ponytail: full · review: nothing to cut · markers: 0
Self-review: 0 items found and fixed

## Summary

Closes KRTX-4897. Root cause from ClickHouse
(`t502678.kortix_api_logs_local`: 284 lines, one sessionId,
`storedAgent: hack-dispatch`): a session whose own agent is undeclared
in the manifest re-logged the "healing" error on every
`/v1/connectors/workspaces/:id/call`. The log now fires only when the
stored agent differs from the session agent (a real heal, as in the
`chief-of-staff` case).

## Demo video

No demo video: code-only change

## Type of change

- [x] Bug fix (noisy error log)

## How was this tested?

```
$ bun test --env-file=scripts/test.env --isolate src/projects/lib/session-token-grant-reconcile.test.ts   (before fix)
(fail) a session created with an agent the manifest no longer declares does not log a heal that changes nothing
 8 pass
 1 fail
$ bun test ... session-token-grant-reconcile / session-token-grant-provenance / unit-my-kortix-grant-resolution   (after fix)
 29 pass
 0 fail
Ran 29 tests across 3 files.
$ bun test ... src/projects   -> 4656 pass, 107 fail
```

The 107 failures are real-PostgreSQL and partial-git-store suites that
need Docker/DB; none touch session-token-grant. `tsc --noEmit -p
apps/api`: exit 0.

## Security & data review

No secrets, customer data or authorization change. Only a console.error
condition.

## Rollout / rollback

Ships with the API; revert the commit to restore the old log.

## Reviewer checklist

- [x] New test fails without the fix
- [x] Real-heal case still logs once

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…ds (KRTX-4905) (#12704)

## Review in 60 seconds

- Root cause: the prod "5xx rise" on `HEAD /v1/p/:id/:id/file/raw` is 6
control-plane `sandbox_not_ready` 503s (6–14 ms each, no dial, no error
line) from one parked sandbox at 2026-11-14 01:14:45 UTC. 1.5 s later
the same box answered `GET .../file/raw` 200 (it had woken). It is the
proxy's designed answer, but `shouldSuppressRequestLog` suppressed it
for `GET` only, so a `HEAD` counted as a route 5xx in the sweep (100 %
of 6 requests).
- Fix: treat `HEAD` as the read it is (the proxy never wakes a box on
GET or HEAD, `wake.ts`), so the same designed control-plane / daemon
boot-phase 503s on a HEAD stop being logged as failures. The wire
response is unchanged.
- Test: new regression test in `request-log-suppression.test.ts` failed
before the fix and passes after; dialled-port failures, unattributed
503s and the give-up 502 on a HEAD stay logged.
No demo video: code-only change

**Risk:** low — Log-line suppression only (the response and the span are
untouched). HEAD on `/global/event`, `/session/status`,
`/session/:id/message` and the startup-probe paths now follows the GET
suppression rules on those paths too.

**Verified:** `bun test request-log-suppression.test.ts` 12 pass;
`request-log-level.test.ts` 10 pass; not run: apps/api typecheck,
eslint, local stack (sandbox has no installed workspace dependencies and
no Docker).

suna-skills: contributing (PR loop; the repo skills directory was not
exercised beyond that), ponytail loaded
ponytail: full · review: no findings; the change is a one-condition
widening plus doc lines · markers: 0
Self-review: 0 items found and fixed

## Summary

Closes KRTX-4905 (factory-key `infra:log:167378867561`). A proxied
`HEAD` answered by a parked sandbox with the designed control-plane
`sandbox_not_ready` 503 (or the daemon's boot-phase 503) no longer logs
a `Request completed ... 503` line, the same as a GET since KRTX-397.
Before, the infra sweep counted each such HEAD as a route 5xx and filed
a "5xx rise" issue with nothing to fix.

## Demo video

No demo video: code-only change (log-line suppression in the API, no
user-visible surface).

## Type of change

Bug fix (non-breaking, API log noise).

## What is affected

`apps/api/src/sandbox-proxy/request-log-suppression.ts`, used only by
the post-request logger in `http-middleware.ts`. HTTP responses, spans
and non-HEAD behavior are unchanged. Mutations stay logged.

## Security & data review

No auth, data or secret change; logging only. Synthetic ids only in this
body.

## Rollout / rollback

Low; revert the commit. A real HEAD outage shaped as control-plane 503
is no longer logged (same trade-off already accepted for GET).

## Reviewer checklist

1. `request-log-suppression.ts:63`: the widened method guard.
2. The new test's negative cases (`upstream_port`, no hop, 502).

<details>
<summary>Changed files</summary>

- `apps/api/src/sandbox-proxy/request-log-suppression.ts`: HEAD treated
as a read
- `apps/api/src/sandbox-proxy/request-log-suppression.test.ts`:
regression test

</details>

## How was this tested?

| Command / scenario | Result | What it proves |
| --- | --- | --- |
| `bun test request-log-suppression.test.ts` before the fix | FAIL 11
pass / 1 fail | the new HEAD test reproduces the bug |
| same, after the fix | PASS 12 pass / 0 fail | fixed |
| `bun test request-log-level.test.ts` | PASS 10 pass / 0 fail |
neighbor suite |
| `bun test .` in `sandbox-proxy/` | 97 pass / 38 fail, identical 38
fail at origin/dev | `Pre-existing`: fail the same at origin/dev
(`Cannot find module '@hono/zod-openapi'`, workspace deps not installed)
|
| typecheck, eslint, local stack | SKIPPED | no installed workspace deps
or Docker in this sandbox |

Not verified: behavior after the prod release. Confirm with ClickHouse:
`status >= 500 AND method = 'HEAD' AND path LIKE '%/file/raw'` per hour
returns to 0 for 24 h.



```
$ bun test request-log-suppression.test.ts   # before the fix
 11 pass
 1 fail
Ran 12 tests across 1 file.
$ bun test request-log-suppression.test.ts   # after the fix
 12 pass
 0 fail
Ran 12 tests across 1 file.
$ bun test request-log-level.test.ts
 10 pass
 0 fail
```

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…ble lockfile (#12701)

## Summary

Follow-ups from the v0.15.11 release (#12672). That release needed 3
release-suite runs, about 1 h each. Its Deploy Prod run then failed once
on a Capture check.

1. **`scripts/pre-promote-tripwires.sh`:** 17 checks the release suite
fails on, run in about 2.5 min with no server and no database. The
kortix-release skill's Step 2 runs it before each Promote dispatch. It
gives the 2026-10-21 ledger rule ("run the release suite's fast
tripwires before every Promote") the enforcer that rule lacked.
2. **Capture Windows MSI check:** it now waits up to 60 s for the
WebView2 process and records the wait. Before, it waited a fixed 10 s.
3. **`pnpm-lock.yaml`:** re-resolved. #12675 left it unstable: every
`pnpm install --lockfile-only` rewrote 46 lines.
4. **`gitleaks-download-verification.test.ts`:** the test installs into
a temp `bin`. Before, it extracted into `/usr/local/bin`, which replaced
a developer's real gitleaks or failed on a binary another user owns.
5. **`scripts/ghsa-audit.mjs`:** the root lockfile advisory filter,
moved out of `ci.yml`. One copy now serves CI and the script.
6. **kortix-release skill:**
- The gateway's version is on `/health/live`. `/health` answers
`{status}` only, by design.
- The staging web is behind Vercel auth (an anonymous request gets
`302`).
   - Deploy Staging runs list the `dev` SHA as their head.
7. **Learnings entry:** supersedes the 2026-10-21 tripwire entry.

No demo video: CI, scripts and docs only.

## Evidence

| Check | Result |
|---|---|
| `bash scripts/pre-promote-tripwires.sh` on this branch | 17 of 17 ok |
| Same script on `eab0879db1` (the `dev` SHA v0.15.11 first promoted) |
9 red: api tripwires, flow registry, route coverage, cli commands doc,
web i18n audit, web route registries, ui layers, starter budgets, GHSA.
These are the first release-suite run's failures, found in about 2.5 min
instead of 1 h. |
| `node scripts/ghsa-audit.mjs` on this branch / on `5418d83309^` | exit
0 / exit 1 naming GHSA-r3h5-4585-7x3w (astro, @astrojs/node) and
GHSA-3v2p-hm54-8qcq |
| `pnpm install --lockfile-only`, run twice | the second run leaves no
diff; `--frozen-lockfile` passes |
| `vitest` gitleaks-download-verification, app-runtime-workflow,
sandbox-workflow, learnings-ledger, no-docs-tree | 49 pass;
`/usr/local/bin/gitleaks` untouched |
| `actionlint` ci.yml, capture-build.yml | clean |

## Capture: why a wait, not a code change

Deploy Prod run 43533482514, attempt 1: the Start menu window opened
(`title: Kortix Capture`). The WebView2 profile existed in
`LocalAppData`, and the window wrote no `ui.log`. The screenshot shows a
blank white window, and `webview_processes` was 0 after 10 s. Attempt 2
on the same build: the full UI and 1 WebView2 process. The job passed,
and the Capture installers were attached. A cold WebView2 first start on
a runner that was recording at the same time took longer than 10 s. A
window that never gets a WebView still fails after 60 s.

## Capture check on a real runner

`capture.yml` dispatched on this branch (run 43542763240): every build
and verify job passed. The Windows MSI evidence reads
`webview_processes: 1, webview_wait_s: 0`.

## Risks and rollback

- `ci.yml`: the root GHSA step now runs the script. Its ignore list and
severity rule are unchanged.
- Revert this PR to roll back. Nothing deploys from it.
…list filters (KRTX-4898) (#12693)

## Review in 60 seconds

- **Root cause (mobile):** the app installs `react-native-url-polyfill`,
whose `URLSearchParams` has no `size`. The SDK built query strings with
`params.size > 0`, so on a phone every session-list request lost its
whole query: the drawer's Sessions, Shared, Automated and Pinned lists
and every Sessions page chip read the same unfiltered list (automated
runs under the person's own sessions; KRTX-4705's `started_by=humans`
never reached the server). The SDK source fix landed on dev first in PR
#12691 (KRTX-4900, `querySuffix()` in existing
`packages/sdk/src/core/http/url.ts`, the same 8 sites this PR fixed).
After merging dev, this PR carries no SDK source change: it takes dev's
source as is.
- **One segmentation everywhere:** the server's `started_by` buckets
(`me` / `others` / `automated`, existing
`apps/api/src/projects/lib/session-list.ts`, unchanged) are the one
rule. Web already reads them and mobile does again with the SDK fix.
This PR makes the TUI sidebar read the same three buckets
(`apps/tui/src/features/sidebar/use-sidebar-sessions.ts`) and show the
viewer's sessions by day, then `Shared` and `Automated`, as web does.
Every bucket polls while one of its rows is live, and every session
action (app `/new`, `/fork`, `/rename`, pin, delete; sidebar `n`, `r`,
`d`) refreshes all session lists at once through one invalidation of
`qk.project.sessionsScope(projectId)`, so the sidebar and the header
title never go stale (G9 items 1 and 2 of head a65b153).
- **Regression tests:**
`apps/mobile/lib/session/session-list-request.test.ts` sends every
drawer list and Sessions page chip request through the SDK under the
app's real polyfill. A sizeless `URLSearchParams` test
(`packages/sdk/src/core/test-fixtures.ts`) covers each of the 6 other
SDK sites #12691 left without one: session start, asks, costs, tokens,
runtime log and messages. `apps/tui/src/lib/session-groups.test.ts`
covers the TUI split, and
`apps/tui/src/features/sidebar/use-sidebar-sessions.test.tsx` mounts the
app's list and the sidebar's buckets together against a real local HTTP
server: a refresh from either side refetches all four lists, and a
bucket with a running row polls. With the `size` check put back into
`querySuffix()`, 7 SDK tests and 2 mobile tests fail. A learnings entry
records the rule.

No demo video: the defect is in the phone's request layer and this Linux
sandbox has no iOS or Android simulator. Expo web cannot show it either,
because `react-native-url-polyfill/auto` skips web. The real surfaces
were exercised instead (before the dev merge, same code paths): the
drawer's four SDK requests under the app's polyfill against the
worktree's local API and against the production API (read-only, counts
only), and the real TUI against the local API in tmux. Their output is
in "How was this tested?".

**Risk:** low. The TUI now holds 4 session-list reads instead of 1: the
app's unfiltered list (header title, `/sessions`) and the sidebar's 3
buckets. A session action refetches every mounted list under
`sessionsScope` (also the Spaces screen's thread list when it is open),
the same prefix the SDK's own mutations invalidate. The rest is tests
and a learnings entry; the SDK, the API and the mobile source are
byte-identical to dev. Revert is the squash commit.
**Verified:** (dev-merge cycle, head 1307b94: merged origin/dev 60fcbcb,
the only overlap was the generated learnings index, regenerated with the
existing learnings index script) TUI `use-sidebar-sessions.test.tsx` +
`session-groups.test.ts` 28 pass, 0 fail · TUI suite 1085 pass, 0 fail ·
TUI typecheck exit 0 · mobile `session-list-request.test.ts` 3 pass, 0
fail · 6 touched SDK test files 88 pass, 0 fail · (review cycle, head
c54f90a) `bun test src/features/sidebar/use-sidebar-sessions.test.tsx` 2
pass, 0 fail; red with each G9 defect put back (sidebar refresh = its
own 3 lists; app refresh = its own list; Shared/Automated without poll):
1 pass, 1 fail each · TUI suite 1085 pass, 0 fail · TUI typecheck exit 0
· biome on the 5 changed TUI files: the same 2 errors + 1 warning as
origin/dev, 0 new · SDK touched tests 110 pass, 0 fail · mobile
`session-list-request.test.ts` 3 pass · not run this cycle: the local
stack (this fresh sandbox's shared primary DB has no `kortix` schema and
the baseline migration needs `basejump`; `pnpm worktree start` stops
with "primary Supabase credentials are unavailable"), so the real TUI in
tmux and `SESS-37` are from the first cycle; the API is unchanged ·
(previous cycle, after merging origin/dev) 7 SDK test files touched, 226
pass, 0 fail (7 fail with `size` restored) · `pnpm --filter @kortix/sdk
test` 7322 pass, 0 fail · SDK typecheck exit 0 · mobile
`session-list-request.test.ts` + `url-polyfill-query.test.ts` 4 pass, 0
fail (2 fail with `size` restored) · mobile suite 3866 pass, 4 fail
(pre-existing at origin/dev) · TUI suite 1083 pass, 0 fail · TUI
typecheck exit 0 · `pnpm test -- --id SESS-37` PASS · `biome check` on
the changed files: the same 14 errors at origin/dev, 0 new
suna-skills: worktree, testing, learnings, contributing, create-pr, sdk,
ponytail, ponytail-review
ponytail: full · review: Lean already. Ship. · markers: 0
Self-review: 11 items found and fixed (1 in the second dev-merge cycle:
git's auto-merge put the learnings index out of order, regenerated with
the existing learnings index script; 4 in the first cycle; 3 in the
dev-merge cycle: a duplicate SDK sessions test dropped in favour of
#12691's, the learnings rule pointed at `querySuffix()`, and stale
PR-body claims rewritten; 3 in this review cycle: two casts in the new
test replaced by a real local server, a fixed-delay wait replaced by
waiting on in-flight fetches, and the sidebar error line now names a
failed Shared or Automated read too)
Feedback: 2 addressed · 0 answered without a change

## Summary

On a phone, every session list asked the API for the same unfiltered
page, because the SDK dropped the query string whenever the global
`URLSearchParams` had no `size` (React Native's URL polyfill). The SDK
fix landed on dev in PR #12691 (KRTX-4900); this PR adds the tests that
pin the mobile session-list requests and the TUI half. With both, the
mobile drawer and the Sessions page now send their `started_by`,
`parent`, `limit` and `pinned` filters again, so the person's own
sessions, Shared and Automated show the same buckets as on web, and
automated runs leave the default list. The TUI sidebar adopts the same
server buckets, so web, TUI and mobile render one segmentation. The
segmentation itself was already computed server-side; the API is
unchanged (reading of the issue: "revamp GET sessions server-side" is
satisfied by the existing `started_by` contract, which every client now
actually uses).

Closes KRTX-4898

## Demo video

No demo video: the defect is in the phone's request layer, and this
sandbox has no iOS or Android simulator (see Review in 60 seconds). The
real-surface output is under "How was this tested?".

## Type of change

- [x] Bug fix (mobile session lists, 7 other SDK reads on mobile)
- [x] Behavior change (TUI sidebar sections)
- [ ] Database migration
- [ ] Public SDK export change (none: no export added, renamed or
removed)

## What is affected

- **Mobile app (iOS, Android):** the drawer's Sessions / Shared /
Automated / Pinned lists, the Sessions page chips (All, Mine, Shared,
Automated, Archived), session search, session start `wait_ms`, asks,
token and cost reads, runtime log and message paging. The mobile fix
itself is dev's #12691 and ships with the mobile binary that PR bumped
(1.6.37); this PR adds tests only for mobile.
- **TUI:** the sidebar's session list shows the viewer's sessions by
day, then `Shared` and `Automated`. The `/sessions` panel still lists
every session.
- **Not affected:** the API (no change); web, desktop and the CLI send
byte-identical URLs wherever `URLSearchParams` has `size` (current
browsers, Node, Bun). A browser older than that support (Chrome 113,
Safari 17) lost its queries the same way and is fixed too.

## How was this tested?

### Second dev merge (head 1307b94)

GitHub reported a conflict with dev on c54f90a. Merging the current
origin/dev (60fcbcb) was clean except the generated
`.agents/skills/learnings/MEMORY.md`, which git merged out of date
order; `bash .agents/skills/learnings/scripts/index.sh` regenerated it.
No source file of this PR changed in the merge.

```
$ cd apps/tui && bun test src/features/sidebar/use-sidebar-sessions.test.tsx src/lib/session-groups.test.ts
 28 pass
 0 fail
Ran 28 tests across 2 files. [349.00ms]
$ cd apps/tui && pnpm test     ->  1085 pass, 0 fail
$ cd apps/tui && pnpm typecheck  ->  exit 0
$ cd apps/mobile && bun test lib/session/session-list-request.test.ts
 3 pass
 0 fail
$ cd packages/sdk && bun test <asks, session-costs, session-sandbox, tokens, runtime-actions, runtime-verbs tests>
 88 pass
 0 fail
Ran 88 tests across 6 files. [47.00ms]
```

### Review cycle (G9 request_changes on a65b153, head c54f90a)

```
$ cd apps/tui && bun test src/features/sidebar/use-sidebar-sessions.test.tsx
(pass) useSidebarSessions (KRTX-4898) > a refresh from the app or the sidebar refetches all four session lists
(pass) useSidebarSessions (KRTX-4898) > every bucket with a live row polls; a settled bucket does not
 2 pass
 0 fail
# A: sidebar refresh = refetch of its own 3 lists (head a65b153)
(fail) useSidebarSessions (KRTX-4898) > a refresh from the app or the sidebar refetches all four session lists
 1 pass
 1 fail
# C: app refresh = refetch of its own unfiltered list (head a65b153)
(fail) useSidebarSessions (KRTX-4898) > a refresh from the app or the sidebar refetches all four session lists
 1 pass
 1 fail
# B: Shared and Automated without refetchInterval (head a65b153)
(fail) useSidebarSessions (KRTX-4898) > every bucket with a live row polls; a settled bucket does not
 1 pass
 1 fail
$ cd apps/tui && pnpm test
 1085 pass
 0 fail
Ran 1085 tests across 91 files.
$ cd apps/tui && pnpm typecheck
exit 0
$ cd packages/sdk && bun test <the 6 touched SDK test files> src/react/use-project-sessions.test.ts
 110 pass
 0 fail
$ cd apps/mobile && bun test lib/session/session-list-request.test.ts
 3 pass
 0 fail
$ biome lint <the 5 changed TUI files>     # the same at origin/dev
Found 2 errors.     (sidebar.tsx createSession deps, app.tsx:572 deps: both present at origin/dev)
Found 1 warning.
```

The poll test drives `pollDueQueries`
(`apps/tui/src/lib/query-polling.ts`), the TUI's own poll driver:
TanStack Query treats Bun as a server and never schedules
`refetchInterval`, so the test calls what `main.tsx` mounts.

### After merging origin/dev (head of this cycle)

The merge conflicted in the 7 SDK sources #12691 changed; the resolution
takes dev's `querySuffix()` version of each and drops this branch's
duplicate sessions-list SDK test (dev's #12691 has one).

```
$ cd packages/sdk && bun test src/core/rest/projects-client/{sessions,asks,session-costs,tokens,session-sandbox}.test.ts src/core/session/runtime-{actions,verbs}.test.ts
 226 pass
 0 fail
Ran 226 tests across 7 files.
# the same run with `querySuffix()` reverted to `params.size > 0 ? ... : ''`:
(fail) readRuntimeLogPage keeps its query under the mobile URLSearchParams polyfill (KRTX-4898)
(fail) messages() and liveLog() keep their query under the mobile URLSearchParams polyfill (KRTX-4898)
(fail) startProjectSession keeps wait_ms under the mobile URLSearchParams polyfill (KRTX-4898)
(fail) listSessionAsks keeps its query under the mobile URLSearchParams polyfill (KRTX-4898)
(fail) listProjectSessionsPage keeps its query under a URLSearchParams without size
(fail) listSessionCosts keeps its query under the mobile URLSearchParams polyfill (KRTX-4898)
(fail) listAccountTokens and revokeAccountToken keep their query under the mobile URLSearchParams polyfill (KRTX-4898)
 219 pass
 7 fail
$ cd packages/sdk && pnpm test        # 475 file runs, summed
 7322 pass
 0 fail
$ cd packages/sdk && pnpm typecheck
exit 0
$ cd apps/mobile && bun test lib/session/session-list-request.test.ts lib/projects/url-polyfill-query.test.ts
 4 pass
 0 fail
# session-list-request.test.ts with `size` restored in querySuffix():
(fail) session list requests under the app URL polyfill > each Sessions page chip asks for its own starter bucket
(fail) session list requests under the app URL polyfill > the drawer's four lists ask for four different buckets
 1 pass
 2 fail
$ cd apps/mobile && pnpm test
 3866 pass
 4 fail
Ran 3870 tests across 352 files.
(fail) react-native-markdown-display AST keys (patched) > ... (4 tests; pre-existing: this branch changes no mobile file but the new test file, and they fail the same at origin/dev)
$ cd apps/tui && pnpm test
 1083 pass
 0 fail
Ran 1083 tests across 90 files.
$ cd apps/tui && pnpm typecheck
exit 0
$ pnpm test -- --id SESS-37
[test] PASS api-cli-flows 126.5s
$ biome check <the 11 modified .ts/.tsx files>    # same at origin/dev
Found 14 errors.
Found 1 warning.
```

Pre-existing: `cd apps/mobile && pnpm typecheck` fails the same at
origin/dev (`components/session/ProjectScreen.tsx(999,9): error
TS2322`).

### First cycle (head 6b0665c, before #12691 landed)


Red before the fix — the 7 new SDK tests at the trunk's SDK source (red
run before the fix):

```
$ cd packages/sdk && bun test src/core/rest/projects-client/{sessions,asks,session-costs,tokens,session-sandbox}.test.ts src/core/session/runtime-{actions,verbs}.test.ts -t "KRTX-4898"
Expected: "http://test.local/workspaces/P1/sessions?limit=20&parent=root&started_by=me"
Received: "http://test.local/workspaces/P1/sessions"
(fail) listProjectSessionsPage keeps its filters under the mobile URLSearchParams polyfill (KRTX-4898)
Expected to contain: "/start?wait_ms=5500"
Received: "http://test.local/v1/workspaces/P1/sessions/S1/start"
 0 pass
 7 fail
Ran 7 tests across 7 files.
```

Green after the fix — the 7 touched SDK test files:

```
$ cd packages/sdk && bun test src/core/rest/projects-client/{sessions,asks,session-costs,tokens,session-sandbox}.test.ts src/core/session/runtime-{actions,verbs}.test.ts
 226 pass
 0 fail
 500 expect() calls
Ran 226 tests across 7 files.
```

The whole SDK suite (baseline at the trunk: 7309 pass, 0 fail, 475 file
runs):

```
$ pnpm --filter @kortix/sdk test
 7316 pass
 0 fail
(475 file runs)
$ pnpm --filter @kortix/sdk typecheck
exit 0
```

Mobile, under the app's real polyfill — red run before the fix (SDK
`sessions.ts` at origin/dev), then green:

```
$ cd apps/mobile && bun test --isolate lib/session/session-list-request.test.ts   # SDK fix stashed
Expected: "?parent=root&started_by=me"
Received: ""
(fail) session list requests under the app URL polyfill > the drawer's four lists ask for four different buckets
 1 pass
 2 fail
$ cd apps/mobile && bun test --isolate lib/session/session-list-request.test.ts   # this branch
 3 pass
 0 fail
Ran 3 tests across 1 file.
```

The whole mobile suite. The 4 failures are pre-existing:
`lib/markdown/markdown-keys.test.ts` fails the same at origin/dev
49de8bc (1 pass, 4 fail there):

```
$ cd apps/mobile && pnpm test
 3865 pass
 4 fail
Ran 3869 tests across 351 files.
(fail) react-native-markdown-display AST keys (patched) > … (4 tests, pre-existing at origin/dev)
```

Pre-existing: `cd apps/mobile && pnpm typecheck` fails the same at
origin/dev with its one error,
`components/session/ProjectScreen.tsx(999,9): error TS2322`
(`openSpringConfig`); this branch adds 0 errors.

TUI — the new grouping test, red run before the fix (the function did
not exist), then the suite:

```
$ cd apps/tui && bun test src/lib/session-groups.test.ts   # before groupSessionsBySection existed
SyntaxError: Export named 'groupSessionsBySection' not found in module '…/apps/tui/src/lib/session-groups.ts'.
 0 pass
 1 fail
$ cd apps/tui && pnpm test
 1083 pass
 0 fail
Ran 1083 tests across 90 files.
$ cd apps/tui && pnpm typecheck
exit 0
```

Server buckets, the targeted flow on the worktree's deterministic stack:

```
$ pnpm test -- --id SESS-37
✓ [1/1] PASS SESS-37 1.0s
results: 1/1 passed · 0 failed · 0 skipped · 0 todo · 1.5s
```

Real surface 1 — the drawer's four requests through `@kortix/sdk` with
the app's `URLSearchParams` polyfill, against the worktree's local API
(http://localhost:13008), synthetic workspace with 3 person-started and
4 trigger-started sessions:

```
### before (origin/dev SDK)
Sessions  query=(none)                                       rows=7 automated=4
Shared    query=(none)                                       rows=7 automated=4
Automated query=(none)                                       rows=7 automated=4
Pinned    query=(none)                                       rows=7 automated=4
### after (this branch)
Sessions  query=?limit=50&parent=root&started_by=me          rows=3 automated=0
Shared    query=?limit=20&parent=root&started_by=others      rows=0 automated=0
Automated query=?limit=20&parent=root&started_by=automated   rows=4 automated=4
Pinned    query=?limit=20&pinned=only                        rows=0 automated=0
```

Real surface 2 — the same script, read-only against the production API
for the reporting workspace (counts only): before, all four lists sent
no query and got the same 50 rows, 35 of them automated; after, Sessions
50 rows / 0 automated, Automated 20 / 20, Pinned 0.

Real surface 3 — the real TUI (`bun run src/index.tsx`, Sidebar layout)
in tmux against the local API, same synthetic workspace:

```
before (origin/dev)        after (this branch)
Today                      Today
 ● Fix login bug            ● Fix login bug
 ● Triage pass run          ● Plan the launch
 ● Nightly sweep run        ● Draft release notes
 ● Hourly watchdog run     Automated
 ● Plan the launch          ● Triage pass run
 ● Draft release notes      ● Nightly sweep run
 ● Weekly report run        ● Hourly watchdog run
                            ● Weekly report run
```



## Security & data review

- No auth, IAM, route or data-model change. The API and its visibility
rules are unchanged; the clients only send the filters they already
intended to send.
- Test and demo data are synthetic (`pr-demo-…@example.test`, invented
session names). The production check printed aggregate counts only.

## Rollout / rollback

- TUI: live with the merge. Mobile: no source change here (the SDK fix
shipped with #12691 and mobile 1.6.37).
- Rollback: revert the squash commit. No migration, flag or data change.

## Reviewer checklist

- [x] Each new test fails without the fix (red runs above)
- [x] Whole suites of the touched packages ran (SDK, mobile, TUI)
- [x] No public SDK export changed (`test-fixtures.ts` is excluded from
the build)
- [x] Synthetic data only

## Reviewer focus

1. `apps/tui/src/features/sidebar/use-sidebar-sessions.ts`: the three
bucket reads, each polling while live, and `useRefreshSessions`, which
`app.tsx` (`act`, `start`) and the sidebar (`n`, `r`, `d`) both call.
`sidebar.tsx`: `reachEnd` pages the sections in render order.
2. The SDK `*.test.ts` additions: one sizeless-`URLSearchParams` test
per SDK site, on top of dev's `querySuffix()`.
3. `apps/mobile/lib/session/session-list-request.test.ts`: the test
loads the polyfill's `js/URLSearchParams` entry (its package index loads
`react-native`).

<details>
<summary>Changed files</summary>

- SDK tests:
`packages/sdk/src/core/rest/projects-client/{asks,session-costs,session-sandbox,tokens}.test.ts`,
`packages/sdk/src/core/session/runtime-{actions,verbs}.test.ts`,
`packages/sdk/src/core/test-fixtures.ts` (no SDK source change: dev's
#12691 has it)
- Mobile: `apps/mobile/lib/session/session-list-request.test.ts`
- TUI:
`apps/tui/src/features/sidebar/{sidebar.tsx,use-sidebar-sessions.ts,use-sidebar-sessions.test.tsx,index.ts}`,
`apps/tui/src/app.tsx` (refresh every list),
`apps/tui/src/lib/session-groups.ts` (+ test), a comment in
`apps/tui/src/commands/panels.tsx`, `apps/tui/SPEC.md`
- Ledger:
`.agents/skills/learnings/entries/2026-11-14T001249Z-never-read-urlsearchparams-size-in-code-the-mobile-app-runs.md`,
`MEMORY.md`

</details>

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…write (KRTX-4910) (#12707)

## Review in 60 seconds

- The v1 sync transcript read (`GET …/transcript?shape=sync` from the
session log) served a `captured_at` 1 ms past the log's real last write
about half the time.
- Cause: the newest message write was selected as an epoch `float8`
(~1.79e12 ms plus a fraction, 16 significant digits) and came back
rounded to the next millisecond; the read now takes `max(updated_at)`
through the column's own Date mapping, truncated to the ms exactly like
`log.updatedAt`.
- A new assertion pins a write at `.405999` and expects `.405Z`: red
before the fix, green after. This is the `contract / core-db lane`
failure on the v0.15.12 release suite (run 43543224012, attempt 3).
No demo video: code-only change
**Risk:** low — one aggregate in one read path; the value only moves
back by under 1 ms to the true write millisecond.
**Verified:** db-suites integration-session-log-read red→green (10
pass), related session-log integration (22 pass) and unit (40 pass)
suites, kortix-api typecheck (tsgo) exit 0, eslint exit 0.
suna-skills: testing, worktree, contributing
Self-review: 0 items found and fixed (diff read as a hostile reviewer:
the pinned timestamps only change ordering-free fields the rest of the
test reads; the green run confirms it)

## Summary

Closes KRTX-4910

The session-log read derived `captured_at` from `(extract(epoch from
max(updated_at)) * 1000)::float8` and `Math.floor`. An epoch in
milliseconds with a microsecond fraction needs 16 significant digits;
the value that reached the API was rounded to the next ms when the
fraction was ≥ .5, so `captured_at` was 1 ms later than the log's last
write. It now selects `max(sessionLogMessages.updatedAt)` with drizzle's
`max()`, which maps through the timestamp column, and takes the later of
that and the session record's `updatedAt`.

## Demo video

No demo video: code-only change

## Type of change

Bug fix (API, release suite failure).

## How was this tested?

Red, before the fix (new pinned assertion, line 185):

```
$ bun --no-env-file ../../tests/bin/db-suites.ts integration-session-log-read
Expected: "2026-11-14T02:16:04.405Z"
Received: "2026-11-14T02:16:04.406Z"
(fail) the v1 reads answer from the log > a session whose threads fromV1 did not build (the pi adapter writes them natively) is served from the log through toV1 [136.30ms]
 9 pass
 1 fail
FAIL 0/1 suites, 10 tests, 11.9s
```

Green, after the fix:

```
$ bun --no-env-file ../../tests/bin/db-suites.ts integration-session-log-read
PASS apps/api/src/__tests__/integration-session-log-read.test.ts 10 tests 6.1s
PASS 1/1 suites, 10 tests, 6.8s
 10 pass; 0 fail
$ bun --no-env-file ../../tests/bin/db-suites.ts integration-session-log-blobs session-log-page-budget.integration session-log-pointer-read.integration
PASS apps/api/src/projects/session-log-page-budget.integration.test.ts 4 tests 5.9s
PASS apps/api/src/__tests__/integration-session-log-blobs.test.ts 16 tests 11.0s
PASS apps/api/src/projects/session-log-pointer-read.integration.test.ts 2 tests 13.8s
PASS 3/3 suites, 22 tests, 14.5s
$ bun test --isolate --env-file=scripts/test.env src/projects/routes/session-log-read-rate-limit.test.ts src/projects/routes/session-open-bundle.test.ts src/projects/session-log-page-budget.test.ts src/projects/session-log-read.test.ts
 40 pass
 0 fail
$ pnpm --filter kortix-api typecheck   # tsgo --noEmit
exit 0
$ npx eslint src/projects/session-log-read.ts
exit 0
```

Not run: the whole kortix-api unit suite (release-train speed rule: the
release suite is the full gate).

## Security & data review

No data, auth or permission change. Synthetic test data only.

## Rollout / rollback

Merges into dev, then is cherry-picked onto staging for the v0.15.12
release (frozen candidate). Revert the commit to roll back.

## Reviewer checklist

- [x] The new assertion fails without the fix (shown above).
- [x] One read path changed; the session record's `updatedAt` is still
part of the maximum.

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…nner (KRTX-4911) (#12708)

## Review in 60 seconds

- The SDK test "eviction drops the memoized rows, not just the store
slices" failed once on the v0.15.12 release suite (run 43543224012,
attempt 2, `contract / core-db lane`) and passed on attempt 3 with the
same code: a flake.
- It checks a WeakRef to the evicted rows is cleared after full GCs. JSC
scans the stack conservatively, so a stale slot can pin the rows through
several collections; the helper now allows up to 20 collections (each on
a fresh job) instead of 5.
- The assertion is unchanged: a real leak keeps the rows through all 20
rounds and still fails. The flake did not reproduce locally (0 failures
in 36 parallel and 42 serial runs at bun 1.4.4), so this widens the
margin rather than proving the race gone.
No demo video: code-only change
**Risk:** low — test-only change in one helper; at most 15 extra
zero-delay ticks on a passing run (the loop stops at the first clear
deref).
**Verified:** sdk sync-store.test.ts 234 pass / 0 fail (3 runs via
scripts/test.sh), @kortix/sdk typecheck (tsgo) exit 0.
suna-skills: testing, worktree, contributing
Self-review: 0 items found and fixed

## Summary

Closes KRTX-4911

`rowsCollectedAfter` in
`packages/sdk/src/browser/stores/sync-store.test.ts` gave the rows five
full collections to disappear. On a loaded CI runner that was not enough
once, and the `contract / core-db lane` of the release suite went red.
The helper now allows twenty. Nothing in the store changes.

## Demo video

No demo video: code-only change

## Type of change

Test fix (flake), release suite.

## How was this tested?

```
$ bash scripts/test.sh src/browser/stores/sync-store.test.ts   # packages/sdk, 3 runs
 234 pass  0 fail
 234 pass  0 fail
 234 pass  0 fail
$ pnpm --filter @kortix/sdk typecheck
exit 0
```

Before the change, the flake did not reproduce here either (36 parallel
`bun test --isolate` runs and 42 serial runs, 0 failures), so this PR
cannot show a red run; the CI failure it answers is run 43543224012
attempt 2, `sync-store.test.ts:2829` (`expect(collected).toBe(true)`,
received false).

## Security & data review

Test-only. No data or permission change.

## Rollout / rollback

Merges into dev, then is cherry-picked onto staging for the v0.15.12
release. Revert the commit to roll back.

## Reviewer checklist

- [x] The assertion is not weakened: `collected` must still be true.

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…write (KRTX-4910) (#12707)

## Review in 60 seconds

- The v1 sync transcript read (`GET …/transcript?shape=sync` from the
session log) served a `captured_at` 1 ms past the log's real last write
about half the time.
- Cause: the newest message write was selected as an epoch `float8`
(~1.79e12 ms plus a fraction, 16 significant digits) and came back
rounded to the next millisecond; the read now takes `max(updated_at)`
through the column's own Date mapping, truncated to the ms exactly like
`log.updatedAt`.
- A new assertion pins a write at `.405999` and expects `.405Z`: red
before the fix, green after. This is the `contract / core-db lane`
failure on the v0.15.12 release suite (run 43543224012, attempt 3).
No demo video: code-only change
**Risk:** low — one aggregate in one read path; the value only moves
back by under 1 ms to the true write millisecond.
**Verified:** db-suites integration-session-log-read red→green (10
pass), related session-log integration (22 pass) and unit (40 pass)
suites, kortix-api typecheck (tsgo) exit 0, eslint exit 0.
suna-skills: testing, worktree, contributing
Self-review: 0 items found and fixed (diff read as a hostile reviewer:
the pinned timestamps only change ordering-free fields the rest of the
test reads; the green run confirms it)

## Summary

Closes KRTX-4910

The session-log read derived `captured_at` from `(extract(epoch from
max(updated_at)) * 1000)::float8` and `Math.floor`. An epoch in
milliseconds with a microsecond fraction needs 16 significant digits;
the value that reached the API was rounded to the next ms when the
fraction was ≥ .5, so `captured_at` was 1 ms later than the log's last
write. It now selects `max(sessionLogMessages.updatedAt)` with drizzle's
`max()`, which maps through the timestamp column, and takes the later of
that and the session record's `updatedAt`.

## Demo video

No demo video: code-only change

## Type of change

Bug fix (API, release suite failure).

## How was this tested?

Red, before the fix (new pinned assertion, line 185):

```
$ bun --no-env-file ../../tests/bin/db-suites.ts integration-session-log-read
Expected: "2026-11-14T02:16:04.405Z"
Received: "2026-11-14T02:16:04.406Z"
(fail) the v1 reads answer from the log > a session whose threads fromV1 did not build (the pi adapter writes them natively) is served from the log through toV1 [136.30ms]
 9 pass
 1 fail
FAIL 0/1 suites, 10 tests, 11.9s
```

Green, after the fix:

```
$ bun --no-env-file ../../tests/bin/db-suites.ts integration-session-log-read
PASS apps/api/src/__tests__/integration-session-log-read.test.ts 10 tests 6.1s
PASS 1/1 suites, 10 tests, 6.8s
 10 pass; 0 fail
$ bun --no-env-file ../../tests/bin/db-suites.ts integration-session-log-blobs session-log-page-budget.integration session-log-pointer-read.integration
PASS apps/api/src/projects/session-log-page-budget.integration.test.ts 4 tests 5.9s
PASS apps/api/src/__tests__/integration-session-log-blobs.test.ts 16 tests 11.0s
PASS apps/api/src/projects/session-log-pointer-read.integration.test.ts 2 tests 13.8s
PASS 3/3 suites, 22 tests, 14.5s
$ bun test --isolate --env-file=scripts/test.env src/projects/routes/session-log-read-rate-limit.test.ts src/projects/routes/session-open-bundle.test.ts src/projects/session-log-page-budget.test.ts src/projects/session-log-read.test.ts
 40 pass
 0 fail
$ pnpm --filter kortix-api typecheck   # tsgo --noEmit
exit 0
$ npx eslint src/projects/session-log-read.ts
exit 0
```

Not run: the whole kortix-api unit suite (release-train speed rule: the
release suite is the full gate).

## Security & data review

No data, auth or permission change. Synthetic test data only.

## Rollout / rollback

Merges into dev, then is cherry-picked onto staging for the v0.15.12
release (frozen candidate). Revert the commit to roll back.

## Reviewer checklist

- [x] The new assertion fails without the fix (shown above).
- [x] One read path changed; the session record's `updatedAt` is still
part of the maximum.

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
(cherry picked from commit 9138e71)
…nner (KRTX-4911) (#12708)

## Review in 60 seconds

- The SDK test "eviction drops the memoized rows, not just the store
slices" failed once on the v0.15.12 release suite (run 43543224012,
attempt 2, `contract / core-db lane`) and passed on attempt 3 with the
same code: a flake.
- It checks a WeakRef to the evicted rows is cleared after full GCs. JSC
scans the stack conservatively, so a stale slot can pin the rows through
several collections; the helper now allows up to 20 collections (each on
a fresh job) instead of 5.
- The assertion is unchanged: a real leak keeps the rows through all 20
rounds and still fails. The flake did not reproduce locally (0 failures
in 36 parallel and 42 serial runs at bun 1.4.4), so this widens the
margin rather than proving the race gone.
No demo video: code-only change
**Risk:** low — test-only change in one helper; at most 15 extra
zero-delay ticks on a passing run (the loop stops at the first clear
deref).
**Verified:** sdk sync-store.test.ts 234 pass / 0 fail (3 runs via
scripts/test.sh), @kortix/sdk typecheck (tsgo) exit 0.
suna-skills: testing, worktree, contributing
Self-review: 0 items found and fixed

## Summary

Closes KRTX-4911

`rowsCollectedAfter` in
`packages/sdk/src/browser/stores/sync-store.test.ts` gave the rows five
full collections to disappear. On a loaded CI runner that was not enough
once, and the `contract / core-db lane` of the release suite went red.
The helper now allows twenty. Nothing in the store changes.

## Demo video

No demo video: code-only change

## Type of change

Test fix (flake), release suite.

## How was this tested?

```
$ bash scripts/test.sh src/browser/stores/sync-store.test.ts   # packages/sdk, 3 runs
 234 pass  0 fail
 234 pass  0 fail
 234 pass  0 fail
$ pnpm --filter @kortix/sdk typecheck
exit 0
```

Before the change, the flake did not reproduce here either (36 parallel
`bun test --isolate` runs and 42 serial runs, 0 failures), so this PR
cannot show a red run; the CI failure it answers is run 43543224012
attempt 2, `sync-store.test.ts:2829` (`expect(collected).toBe(true)`,
received false).

## Security & data review

Test-only. No data or permission change.

## Rollout / rollback

Merges into dev, then is cherry-picked onto staging for the v0.15.12
release. Revert the commit to roll back.

## Reviewer checklist

- [x] The assertion is not weakened: `collected` must still be true.

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
(cherry picked from commit 10af4a4)
…flakes

stage-fix: v0.15.12 release suite flakes (#12707, #12708)
…RTX-4909) (#12706)

## Review in 60 seconds

- **What:** a new cross-session, full-history gateway spend + token
query — API `GET /{workspaceId}/gateway/usage` and `kortix gateway usage
--group-by …`. It aggregates **every** matching `gateway_request_logs`
row (no row ceiling), groups by any subset of
`model,agent,source,session,day`, over an absolute (`--since/--until`)
or rolling (`--days`) window, and keyset-paginates the returned
**groups** via `next_cursor` (`--all` drains the window).
- **Why:** before this, every spend view capped out (top-12 models, 10k
rows, rolling-days only), so a gateway fallback model sat at ~76% of
paid spend for weeks with no way to ask "paid spend by served model
across all sessions". This closes that blind spot.
- **Honesty:** paid Kortix wallet spend, BYOK provider spend and the
total are three separate labelled fields; the list-price usage
**estimate** is kept wholly apart from spend (never called "paid"). All
money math reuses the one definition in `shared/llm-spend.ts`; the
served model reuses one definition now lifted to
`shared/served-model.ts`.
- **Legacy:** `kortix gateway usage` / `--days N` with no new flag is
byte-for-byte unchanged (it still hits `overview`+`breakdown`); the new
endpoint is engaged only when a new flag is passed.

No demo video: code-only change (API + CLI, no user-visible UI).

**Risk:** low — additive read-only endpoint + CLI branch; no schema
change, no migration, legacy paths untouched, gated by the existing
`project.gateway.spend.read` capability.

**Verified:** `tsgo` typecheck clean (api, cli, shared); new parser
(20), route (6), live-DB integration (10, incl. a synthetic >10k-group
cursor drain) and real-CLI-process (6) suites pass;
`session-model-usage` integration still 11/11 after the `servedModelSql`
move; full `@kortix/cli` suite 2814/0; full `kortix-api` suite 16237
pass / 82 pre-existing env-only fails (`git 2.45` guard + missing local
S3 bucket, identical at `origin/dev`); the route manifest + audit label
for the new route are regenerated/added.

suna-skills: worktree, testing, learnings, contributing, ponytail,
ponytail-kortix, ponytail-review

## Summary

Adds a cross-session cost/token observability query to the LLM gateway.
The API gains `GET /{workspaceId}/gateway/usage`, which runs one SQL
aggregate over all matching ledger rows in the window and returns one
row per group (a subset of `model`/`agent`/`source`/`session`/`day`,
combinable) with paid-wallet spend, BYOK provider spend, total,
input/output/cached/cache-write tokens and the list-price usage estimate
as separate fields. Group results are keyset-paginated by total spend
(`next_cursor`), so a dimension with more groups than a page — e.g. by
session across all history — drains fully instead of truncating at a
fixed ceiling. The `kortix gateway usage` CLI gains `--since/--until`,
`--group-by`, `--limit/--cursor` and `--all` that drive this endpoint;
with no new flag its output is unchanged. For an operator, this is the
first honest way to answer "paid spend by served model and by source,
last 7 days, across every session".

## Demo video

No user-visible surface (API route + CLI). A terminal walkthrough adds
nothing a reviewer cannot read from the `How was this tested?` output
below. Code-only change.

## Type of change

- [x] New feature (non-breaking)
- [ ] Bug fix
- [ ] Breaking change
- [ ] Docs only

## How was this tested?

Pure parser + route unit tests (no DB):

```
$ bash scripts/test.sh src/shared/gateway-usage-query.test.ts src/projects/routes/gateway-usage-route.test.ts
 26 pass
 0 fail
Ran 26 tests across 2 files. [527.00ms]
```

Live-DB integration (real local Postgres) — served-model attribution,
payee separation, date boundaries, cache, unknown attribution, caller
visibility, and a **synthetic >10,000-group** cursor drain; plus
`session-model-usage` to prove the `servedModelSql` move is safe:

```
$ pnpm --filter kortix-api test:integration integration-gateway-usage-query integration-session-model-usage
[db-suites] PASS apps/api/src/__tests__/integration-session-model-usage.test.ts 11 tests 6.4s
[db-suites] PASS apps/api/src/__tests__/integration-gateway-usage-query.test.ts 11 tests 6.2s
[db-suites] PASS 2/2 suites, 22 tests
```

Real CLI process runs (`bun index.ts gateway usage …` against a fake
gateway API) — legacy path untouched, flag→query mapping, `--all` cursor
drain, `--json` passthrough, separated money labels:

```
$ bash scripts/test.sh src/commands/gateway-usage.test.ts
 8 pass
 0 fail
Ran 8 tests across 1 file. [3.39s]
```

(The 8 include two regressions added after independent review: `--all`
freezes a rolling window to the absolute bounds page 1 resolved, and
`--all` refuses a repeating cursor instead of truncating silently.)

Full touched-package suites:

```
$ pnpm --filter @kortix/cli test
 2814 pass
 0 fail
Ran 2814 tests across 224 files. [491.62s]

$ pnpm --filter kortix-api test
 16237 pass
 82 fail
```

The 82 `kortix-api` failures are **pre-existing and environmental**,
identical at `origin/dev`: the `partial project store` /
git-object-store suites self-guard with `partial-store tests need git
2.45 or later; this host has "git version 2.39.5"`, and the `config
archive store (S3)` suites fail `NoSuchBucket` because the local
Supabase storage bucket is not provisioned in the sandbox. None are in a
file this PR changes. Route-manifest and audit-label tests pass after
regenerating `tests/spec/routes.generated.json` and adding the
`gateway.usage.query` label:

```
$ bash scripts/test.sh src/__tests__/unit-routes-manifest.test.ts src/__tests__/unit-audit-route-labels.test.ts src/__tests__/unit-workspace-binding-coverage.test.ts src/__tests__/unit-openapi-route-coverage.test.ts
 100 pass
 0 fail
Ran 100 tests across 4 files. [9.04s]
```

Typecheck (`tsgo --noEmit`) is clean for `kortix-api`, `@kortix/cli` and
`@kortix/shared`; ESLint is clean on the changed source files; biome is
format-clean on the changed CLI files.

Second review round (head 75d9db7). The factory reviewer asked for one
change: a `gateway/usage` statement timeout (SQLSTATE 57014) returned an
opaque 500. A factory worker mapped it to 503
(`apps/api/src/http-errors.ts`). A further independent review found two
more items, both fixed:
- `--group` (the alias of `--group-by`) silently fell back to the legacy
ungrouped output. The dispatch check now includes it. The new real-CLI
test is red without the fix:

```
$ bash scripts/test.sh src/commands/gateway-usage.test.ts   # with the gateway.ts fix stashed
(fail) kortix gateway usage (real process) > the --group alias also engages the query path (never a silent legacy fallback)
 8 pass
 1 fail
$ bash scripts/test.sh src/commands/gateway-usage.test.ts   # with the fix
 9 pass
 0 fail
```

- The route's capability gate had no test, because the route test mocks
it. A `gateway usage` row in the leaf-gate HTTP suite now covers it.
Deleting the `assertProjectCapability` call makes that suite fail
(mutation run: `FAIL 0/1 suites, 102 tests`, 1 failed). With the gate in
place:

```
$ pnpm --filter kortix-api test:integration integration-project-read-leaf-gates-http
[db-suites] PASS 1/1 suites, 102 tests, 15.6s
```

Contract coverage: the route is declared in flow GW-9 with a
wire-contract step (400 on a bad group_by, window or cursor; 200 with
separate wallet, provider and total fields and an absolute window). The
flow ran against the local stack:

```
$ bun tests/bin/ke2e.ts coverage
95.8% covered · 1689/1763 routes · 74 allowlisted · 0 uncovered
✓ coverage gate passed
$ pnpm test -- --id GW-9
✓ [1/1] PASS GW-9 0.7s
results: 1/1 passed · 0 failed · 0 skipped · 0 todo · 1.2s
```

Re-run on the new head: timeout-mapping, parser and route tests (35
pass, 0 fail); gateway-usage and session-model-usage integration (2/2
suites, 22 tests); tsgo clean for apps/api, apps/cli and tests.

Self-review: 11 items found and fixed (8 below, plus the --group alias
dispatch, the untested capability gate and the missing GW-9 contract
coverage) (a redundant group-by ordering pass; an `any` in a test mock;
a shared→projects layer violation and a new `(c: any)` route handler —
resolved by lifting `servedModelSql` to the shared layer and typing the
handler; the regenerated route manifest; the missing audit label; a
stray transitive lockfile bump reverted). An independent review then
caught two real pagination-correctness bugs, both fixed with regression
tests: (7) a rolling `--days`/`--since` window drifted across `--all`
pages because each page re-evaluated `now()` — the server now resolves
and echoes the ABSOLUTE window and the drain pins page 1's bounds for
every later page; (8) `--all` silently truncated at 10k pages and
cleared `next_cursor` — it now drains every page and refuses a repeating
cursor loudly instead of reporting partial data as complete.

## Security & data review

- Read-only. No new table, column or migration. Gated by the existing
`project.gateway.spend.read` capability exactly like
`overview`/`breakdown`/`sources` (404 without project read, 403 without
the capability).
- Session-exposing groups (`group_by=session`) apply the same
`projectReadView` visibility filter as the per-session route, so a
non-manager sees only sessions they may open; model/agent/source/day
stay project-wide like the sibling aggregates. Grand totals and groups
are filtered consistently within one response.
- No customer data in tests — every row is synthetic. The CLI never logs
credentials. The list-price estimate is labelled "not a bill" and never
conflated with paid spend.

## Rollout / rollback

- Additive and backward-compatible; no flag, no migration, no data
backfill. Revert is a clean diff revert. The `servedModelSql` move is a
pure relocation of one SQL expression with identical text (pinned by the
passing `session-model-usage` integration suite).

## Reviewer checklist

- [x] Money concepts kept separate and labelled (`shared/llm-spend.ts`,
`shared/llm-value.ts` reused; no fourth rule).
- [x] No 10k/top-N ceiling — groups keyset-paginate; the >10k-group
drain is pinned in `integration-gateway-usage-query.test.ts`.
- [x] Legacy `gateway usage` output unchanged without new flags.
- [x] Route manifest + audit label regenerated/added; OpenAPI coverage
holds.

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…ach, 28 exports dropped (KRTX-4908) (#12712)

## Review in 60 seconds

- One writer for a `project_trigger_runtime` row
(`insertRuntimeTriggerRow`, optional `sessionId`);
`insertRuntimeAgentTrigger` and `insertSessionReminder` delegate to it.
One `connectorJoin()`, one `metaString`, one `oneLine`, one
`stringOrNull`. 28 internal-only `export`s dropped.
- Behavior-preserving restructure of the projects lib; no feature
change, no schema change.
- Two deliberate readings are listed under Summary (writer placement to
avoid an import cycle; one trim rule).

No demo video: code-only change

**Risk:** low — pure move/dedupe; the one trim-rule change reads
system-written, unpadded values.
**Verified:** characterization tests (both files, 6 pass, 0 fail) pass
at origin/dev and at head; the string-helper test now goes through
pre-existing exports (privateSpaceText, sessionAgentHandle,
normalizeRuntimeSessionSnapshot); changed-file tests 45 pass, 0 fail;
apps/api `tsc --noEmit` exit 0; eslint 0 errors; whole apps/api unit
suite: only 37 git-store tests fail, the same 37 at origin/dev. Not run:
the DB-backed integration suites (no Docker).
suna-skills: worktree, testing, learnings, contributing, create-pr
ponytail: full · review: nothing left to cut · markers: 0
Self-review: 0 items found and fixed

## Summary

KRTX-4908 (parent pass KRTX-4907, unit `api-projects-lib`). Removes the
duplicated `project_trigger_runtime` values literal, the duplicated
connector join condition, and four duplicate helper bodies; drops
`export` from 28 names nothing outside their file references. No feature
change.

Deviations from the issue text (stated reading):
1. `insertRuntimeTriggerRow` lives in `session-reminders.ts`, not
`runtime-triggers.ts`: `runtime-triggers.ts` already imports from
`session-reminders.ts`, so the writer in `runtime-triggers.ts` would add
an import cycle. `runtime-triggers.ts` imports it.
2. `stringOrNull` has one rule, trim (the `runtime-session-snapshot.ts`
copy). `session-control-readers.ts` previously returned a non-empty
string untrimmed; its inputs (`runtimeWakeStartedAt`,
`runtimeWakeProgressAt`, `stopReason`) are system-written, unpadded
values, so no observable change. Pinned in
`string-helpers.characterization.test.ts`.
3. Net LOC target −35: production code is −29 (`57 insertions, 86
deletions` over 18 prod files); the two characterization test files add
+107.

## Demo video

No demo video: code-only change.

## Type of change

- [x] Refactor (behavior-preserving)

## Deltas

- `git diff --shortstat origin/dev`: 20 files changed, 168
insertions(+), 86 deletions(-) (prod only: 18 files, 57+/86−)
- signal (`codebase-analysis.mjs metrics --unit api-projects-lib`):
`unused_exports` 166 → 138, `prod_loc` 31305 → 31279, `import_cycles` 3
→ 2, `dup_lines` 0 → 0 (the metric counted 0 at both ends; the
duplicates are listed in the issue by file:line and are now one
definition each).
- `git grep` one body per concept: the `project_trigger_runtime` values
literal (`session-reminders.ts`), `connectorJoin`
(`connection-mutation.ts`), `oneLine` (`session-where.ts`), `metaString`
(`session-board.ts`), `stringOrNull` (`runtime-session-snapshot.ts`).
- Moved blocks: `git diff --color-moved=zebra
--color-moved-ws=allow-indentation-change origin/dev` shows the values
literal, with its comment, moved verbatim from
`runtime-triggers.ts`/`session-reminders.ts` into
`insertRuntimeTriggerRow`; the helper bodies move unreformatted.

## How was this tested?

```
$ cd apps/api && bash scripts/test.sh src/projects/lib/trigger-row-insert.characterization.test.ts   # before the refactor
 3 pass
 0 fail
$ bash scripts/test.sh src/projects/lib/trigger-row-insert.characterization.test.ts src/projects/lib/string-helpers.characterization.test.ts src/projects/lib/runtime-triggers.test.ts src/projects/lib/session-reminders.test.ts src/projects/routes/session-reminders.test.ts
 45 pass
 0 fail
$ bash scripts/test.sh src/projects/lib/chat-threads src/projects/lib/session-board src/projects/lib/session-control-readers src/projects/lib/connection src/projects/lib/session-where src/projects/lib/runtime-session
 54 pass
 0 fail
$ NODE_OPTIONS=--max-old-space-size=8192 npx tsc --noEmit -p apps/api   -> exit 0
$ npx eslint <changed files>   -> 0 errors
$ bash scripts/test.sh   # whole apps/api unit suite
 37 fail, all in git-store suites (partial-store-*, object-store, merge-preview-burst, project-snapshot-build, config-releases/partial-store)
Pre-existing: the same 7 files fail the same way at origin/dev (3 pass, 37 fail on both).
```

Not run: the DB-backed `integration-runtime-triggers`,
`integration-session-reminders`, `connection-mutation`/connection suites
(need PostgreSQL; the sandbox has no Docker). They are covered by the
characterization test of the written row and by `tsc`; the SQL join
condition is textually identical, now built once.


## Security & data review

No auth, secret, customer-data or permission change. The connector join
keeps the same three-column condition.

## Rollout / rollback

No migration, flag or env var. Revert the single commit.

## Reviewer checklist

- [x] Tests pass before and after
- [x] No feature change in the same PR

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…ng on others (#12682)

## Summary

Kortix can now score an agent, and My Kortix is the first agent measured
with it.

- **Evals (flag `evals`, off by default).**
- **Tasks and suites:** a task is in the Harbor format
(`instruction.md`, `task.toml`, `tests/test.sh` writes the reward,
optional `solution/solve.sh`). A suite lives under `evals:` in
`kortix.yaml`, or a job names a dataset (`kortix/my-kortix`,
`terminal-bench@2.0`) and runs it on any agent and model.
- **Trials:** each trial is its own session and sandbox. The tests reach
the sandbox only after the agent is done.
- **Scoring:** Harbor's error policy and its unbiased pass@k decide the
score.
- **Harbor support is native:** a Harbor task, a Harbor job config file
and a Harbor registry dataset run unchanged.
- **Customize → Evals:**
  - **Overview:** the leaderboard with a 95% interval on every score.
  - **Datasets:** README, Tasks, Leaderboard and Jobs per dataset.
- **Tasks:** Instruction, a file browser, and results per agent and
model.
  - **Jobs:**
    - a progress bar whose buckets filter the trials;
    - Run health and Trials;
    - retry or cancel one trial.
  - **Compare:** task by task against a baseline.

The same surface is in `kortix evals` and `@kortix/sdk`. Results are
Harbor-layout folders in Files.
- **Eval-driven development is the default.**
- **The `kortix-system` skill:** it teaches "measure before you change
an agent".
- **My Kortix's prompt:** it carries the same rule, and Evals is in its
map of Kortix.
- **The flagged `kortix-evals` skill:** it teaches the loop, including
the grading traps found here.
- **`kortix/my-kortix`:** a built-in dataset of 20 tasks. Each tests a
rule of My Kortix's prompt: delegation, asking before acting on others,
follow-through, honesty, voice, product knowledge. The reference
solutions score 1 on all 20.
- **Hill-climb:** one change to My Kortix's prompt, measured and kept.
- **The rule:** "ask once, then stop" before inviting people or creating
an automation.
- **The behavior it fixes:** before it, My Kortix asked and then acted
in the same turn (it ran `kortix members invite --role admin` and
installed connectors).

## Demo

**Web, Customize → Evals, 1440×900, synthetic workspace, 1.5× speed.**
The clip walks through:
1. Overview, with a 13-task oracle job under "Running now".
2. Datasets, and the source filter.
3. The `harbor` dataset page.
4. The `sum-numbers` task, with `tests/test.sh` in the file browser and
its Results.
5. Jobs: two `smoke` jobs selected, then Compare.
6. A finished job: the Passed bucket filters the trials, and one trial
opens beside them.


https://github.com/user-attachments/assets/5b8f58b3-0c04-4aa9-ad7d-365eeee7e457

## The My Kortix benchmark and the hill-climb

20 tasks × 3 attempts. Each model and each prompt variant ran in its own
fresh workspace.

| | Acted before asking (6 trials) | Tasks passed | Mean score | Cost |
| --- | --- | --- | --- | --- |
| DeepSeek V4.1 Flash, before | 4 | 41/60 | 0.901 | $0.88 |
| DeepSeek V4.1 Flash, after | **0** | **43/60** | 0.904 | $0.75 |
| GLM 5.3 Flash, before | 4 | 38/60 ¹ | 0.881 | $0.43 |
| GLM 5.3 Flash, after | **0** | **46/60** | **0.942** | $0.43 |

¹ 4 sandboxes failed to start (not scored), so 38/56 on valid trials.

Still failing on both models; these are the next climbs:
- `handoff-compare`: My Kortix researches itself instead of handing the
work to a colleague.
- `confirm-before-invite`: it says an invite cannot grant admin, though
`kortix members invite --role admin` works.
- `meeting-actions`: wording.
- `automation-asks-first`: its reply does not say what the automation
needs.

Building the benchmark found and fixed 6 grading problems. Each would
have produced a wrong score. Each is now a rule in the `kortix-evals`
skill and in the repo runbook.
1. **Graded too early.** The runner graded at the end of the agent's
first turn, before a colleague reported back. It now waits until every
session the trial started is idle for 30 s.
2. **Only the last message graded.** The grader saw only the last
message of the turn ("⏳ I'll check at 9:00") and missed the answer
before it. It now sees the whole last reply.
3. **Trials ran as child sessions.** A job started by an agent ran its
trials as child sessions, a harness without follow-through. Every trial
is now a top-level session.
4. **A judge failure scored 0.** A judge that could not run (an old
sandbox image) scored 0, so a missing tool read as a worse agent. Judged
graders now fail the verifier, and the trial counts as an error.
5. **Graders saw other trials' state.** Trials share the workspace, so a
trigger made by one trial failed the next. These checks now read only
the trial session's own tool calls.
6. **Rate-limit errors.** Platinum's `429 rate_limited` (20 writes/s per
org) errored trials. Sandbox exec now waits its `retry-after`.

## What is affected

- **Users of a workspace with `evals` on:**
- a new Customize group "Evals" (Overview, Datasets, Tasks, Jobs) and a
palette entry;
- `/workspaces/<id>/evals…` and `/admin/evals/jobs/<id>` redirect to it;
- with the flag off, the group is absent, the routes answer `403
feature_disabled`, and the CLI hides `kortix evals`.
- **Every My Kortix session** (no flag):
- Before an invite, a role change or an automation, it now asks and
waits for a yes in the person's next message.
- It has one new line about measuring agent changes with Evals when the
flag is on, and an Evals row in its map.
- **Permissions:**
  - Every member reads jobs and scores.
  - Task contents need `project.file.read`.
- Starting, cancelling and retrying need `project.session.start`;
deleting a job needs project write.
- A job whose task needs a new sandbox template (its own `docker_image`)
also needs `project.sandbox.write`: without it, `403 eval_forbidden`.
  - Trials run as the principal that started the job.
- **API:**
- 18 routes under `/v1/workspaces/:id/evals` (alias
`/v1/projects/:id/evals`).
  - `Idempotency-Key` on the POSTs.
  - `dry_run` on job start.
  - `verifier.env` values read back `<redacted>`.
- **Worker:** a new leader job `eval-runner`. It touches only `eval_*`
rows and the sessions it created.
- **Database:**
  - `20261113205853562_eval_jobs_and_trials` adds two tables, with RLS.
- `20261113234740274_eval_jobs_started_by_idempotency` adds three
columns to `eval_jobs`.
  - `eval_jobs` moves with a transferred workspace.
  - `eval_trials` keeps the session id when a session is purged.
  - No existing table changes.
- **Sessions and billing:**
  - Each trial is a normal session and bills as one.
  - A project runs at most 32 trials at once.
- **Not affected:** existing sessions, triggers, goals, kortixd, mobile,
desktop shell code, self-host config.

## Verification

All on the `evals` worktree stack with real Platinum sandboxes and the
real model gateway.

| Check | Result | What it proves |
| --- | --- | --- |
| `kortix/my-kortix` with `oracle` (20 tasks) | PASS 20/20 | every
grader gives its reference solution 1 on every key |
| My Kortix on DeepSeek and GLM, before and after the prompt change (240
trials, fresh workspace each) | numbers above | the benchmark separates
the variants; the targeted behavior goes 8/12 → 0/12 |
| Session-scoped invite check against a known-bad trial | PASS | finds
the one `kortix members invite … --role admin` call, ignores help output
and `invites ls` |
| `smoke`, worker on `deepseek-v4.1-flash`, Harbor config file, missing
reward, Terminal-Bench 2.0 `fix-git` oracle | PASS | rewards, error
policy, a registry dataset with its own `docker_image` |
| `pnpm test -- --id EVAL-1` | PASS | flag gate, parsing, plan, dry run,
idempotent replay, filters, compare, trial retry/cancel, trial files,
`403` for a plain member, `403 eval_forbidden` for a member starting a
task whose template is missing (no job created) |
| DB gates `integration-kortix-rls`, `project-transfer.integration`,
`integration-session-log-purge` | PASS 3, 29, 30 | RLS on both tables;
transfer and purge classify the new tables |
| Browser (agent-browser): every Evals page | PASS | rail group,
filters, Compare, bucket → filtered trials, trial panel, task file
browser, task results; dark at 720×480 without horizontal scroll |
| Typecheck: api, sdk, cli, ui, web; `pnpm --filter kortix-api lint`;
brand audit; i18n audit and lint | PASS | 0 errors |
| Tests: API scoring, task ids and rate limit, plus the My Kortix prompt
(49); SDK evals and surfaces (16); route contract (11); Customize rail
(12); CLI (3) | PASS | — |

Not verified:
- **A real agent API key** starting a job: the code path is covered, but
no live run used one.
- **Desktop (Electron):** not checked.
- **Built-in datasets on dev:** they read the deployed commit, so
`kortix/my-kortix` on dev works after this merges and deploys. It is
verified on dev after the merge.

## Risks and rollback

- **The My Kortix prompt change reaches every workspace.** It is
measured as above. It adds a confirmation before an invite, a role
change or an automation, which the prompt already required.
- **Evals is off by default.** No route, page, worker step or CLI
command runs without the flag.
- **Cost:** a job can start up to 2,000 trials, each a billed session.
The New job dialog shows the trial count first.
- **Judged graders** need the new CLI in the sandbox image (it ships
with this release). Without it they error, never score 0.
- **Rollback:** revert the PR. The migrations only add tables and
columns.

## Reviewer focus

1. **The prompt change:**
`apps/api/src/projects/lib/my-kortix-prompt.ts` (authority bullet 1,
"Work that starts by itself").
2. **When a trial is graded:** `apps/api/src/projects/evals/runner.ts`,
in `colleaguesQuiet`, `finalReply` and `startTrial` (top-level
sessions).
3. **Permissions and trust:** `routes/evals.ts` `scopeOf` and
`evals/task-id.ts` (signed task ids).

<details>
<summary>Changed files</summary>

- `packages/api-contract/src/evals.ts`: wire contract.
- `packages/manifest-schema/src/evals.ts`, `json-schema.ts`: the `evals`
section.
- `packages/db/src/schema/kortix.ts`,
`packages/db/migrations/2026111320…_eval_jobs_and_trials.sql`,
`…234740274_eval_jobs_started_by_idempotency.sql`: tables and columns.
- `apps/api/src/projects/evals/*`: task format, sources and signed ids,
config, scoring, runner, sandbox ops, storage, service.
- `apps/api/src/projects/routes/evals.ts`, `worker/jobs/eval-runner.ts`,
`feature-flags/registry.ts`: routes, leader job, flag.
- `apps/api/src/projects/lib/my-kortix-prompt.ts`: ask-then-stop, the
eval rule, the Evals map row.
- `apps/api/src/projects/lib/project-transfer.ts`,
`__tests__/integration-session-log-purge.test.ts`: ownership
classification.
- `packages/sdk/src/core/rest/projects-client/evals.ts`,
`src/react/use-evals.ts`: client and hooks.
- `apps/cli/src/commands/evals.ts`: `kortix evals`.
- `packages/ui/src/customize/evals/*`,
`packages/shared/src/workspace-settings.ts`,
`project-admin-sections.ts`, `project-admin-dashboard.tsx`: the Evals
group and pages.
- `apps/web/src/app/[locale]/(app)/workspaces/[id]/admin/eval-*`,
`…/evals/*`: routes and redirects.
- `apps/web/translations/*.json`: strings in 9 locales.
- `packages/starter/templates/managed/skills/kortix-evals/SKILL.md`,
`kortix-system/SKILL.md`, `kortix-system/references/capabilities.md`,
`.agents/skills/evals/SKILL.md`: skills and runbook.
- `evals/my-kortix/*`: the 20-task built-in dataset `kortix/my-kortix`.
- `tests/src/flows/evals.flow.ts`, `tests/spec/end-to-end.md`,
`tests/spec/routes.generated.json`: `EVAL-1`.

</details>
…s (KRTX-4913) (#12714)

## Review in 60 seconds

- KRTX-4913: Better Stack Frontend pattern `46e2be26…` (`ReferenceError:
swbrowser is not defined`, call site `<anonymous>:needInjectCss`) plus
its cluster of 4 `xbrowser is not defined` patterns (`862cb1fa…`,
`3cd7e975…`, `3fcbde47…`, `e1bbcba4…`): 8 events, 0 identified users,
all from ONE anonymous session in an Android 15 WebView (`; WV` UA) on
`/auth`, 2026-11-14 02:52:26–31Z, release `eadefb05…`.
- Root cause: not Kortix code. An XBrowser-family Android browser app
injects its own scripts into the page with `evaluateJavascript`
(functions `needInjectCss`, `execute_auto_fill`; breadcrumb
`============ page load finished =============` is its console line).
Those scripts call the app's bridge globals `swbrowser` / `xbrowser`,
which do not exist in this WebView, so they throw. Every throw site is
an `<anonymous>` frame; the only bundle frame (`app:///_next/…` fn `r`,
col 4885) is Sentry's own
`browserapierrors.setTimeout`/`addEventListener` wrapper, outermost. No
Kortix source names either global (`grep -rn "xbrowser\|swbrowser" apps
packages` → 0 hits outside this PR).
- Fix: one Sentry-gate rule `android-webview-host-browser-bridge` in
`packages/ui/src/internal/lib/browser-noise/rules/webview.ts`: exact
message `/^(?:sw|x)browser is not defined$/`, the innermost (throw-site)
frame is `<anonymous>`, and no de-minified first-party frame. A bundle
frame at the throw site, a first-party frame anywhere, no stack, or any
other global keeps reporting. Pinned by
`packages/ui/src/internal/lib/browser-noise/host-browser-bridge.test.ts`
and 6 golden cases in
`apps/web/src/lib/browser-noise/golden-verdicts.json`.

No demo video: telemetry-only change, nothing on screen changes. The
live proof ran in a real browser on this worktree's web app (`/auth`,
port 3913) against a local Sentry envelope sink: the two injected-script
errors threw in the page and sent 2 error envelopes without the fix, 0
with it; a control `zzbrowser is not defined` from the same
`<anonymous>` shape still sent 1. Commands and output are under "How was
this tested?".

**Risk:** low — telemetry only; the rule can only suppress Sentry events
whose message is exactly `swbrowser is not defined` / `xbrowser is not
defined` thrown from injected `<anonymous>` code. Revert the commits to
restore the previous behaviour.
**Verified:** new `host-browser-bridge.test.ts` → 2 pass; 0 fail (1 fail
without the fix) · golden `rules.test.ts` → 6 pass; 0 fail (1 fail on
cases 1949/1950 without the fix) · `rules.test.ts` +
`browser-error-noise.test.mts` → 489 pass; 0 fail · whole `packages/ui`
suite → 9232 pass; 0 fail · `sentry.options.test.ts` +
`sentry.client.config.test.ts` → 7 pass; 0 fail · `packages/ui`
typecheck exit 0 · `apps/web` typecheck exit 0 · eslint on the changed
files exit 0 · 8 real prod events replayed through
`shouldIgnoreSentryNoiseEvent` → 8/8 dropped (0/8 at origin/dev) · live
browser on the worktree web app → 2 envelopes before, 0 after, control
still 1.
suna-skills: worktree, testing, learnings, contributing
ponytail: full · review: Lean already. Ship. · markers: 0
Self-review: 0 items found and fixed (checks A–H and the slop table of
`software-factory-review` `<self-review>`, run by hand on head
ef6c158)

Closes KRTX-4913

factory-key:
betterstack:2346967:46e2be26866354c32c7a56c1d0dd59f1bbf1d94cc776ddd85981df9b2b44b054

## Summary

Frontend prod reported `ReferenceError: swbrowser is not defined` (3
events) and `ReferenceError: xbrowser is not defined` (5 events over 4
patterns) within 5 seconds, from one anonymous visitor of `/auth` in an
Android 15 System WebView. The raw Sentry events (Better Stack
`t502678_kortix_frontend_exceptions`) show the throwing functions
`needInjectCss` and `execute_auto_fill` at `<anonymous>` line 1. That is
the frame Chrome gives to code that a host app runs through
`WebView.evaluateJavascript`. The globals belong to the host browser
app's native bridge, which this WebView build does not expose. Kortix
code never references them, and the user saw no failure.

These events now stop at the Sentry `beforeSend` gate. The rule follows
the existing throw-site pattern of `googletag-undefined` (KRTX-4571) and
`injected-csp-eval` (KRTX-4423): the decision is made on the innermost
frame, never on Sentry's own timer/listener wrapper frame, and a
de-minified `apps/web/src/…` or `packages/ui/src/…` frame always keeps
the event. Before, nothing matched these events and every one paged.

The cluster's 4 related `xbrowser` patterns have the same root and the
same rule covers them (see the replay below: all 8 events of the 5
patterns are dropped).

## Demo video

No demo video: telemetry-only change — nothing a person sees changes.
The proof is the real browser run below: the page throws the same errors
with and without the fix; only the Sentry envelope count changes.

## Type of change

fix — frontend telemetry noise classification for `factory-key:
betterstack:2346967:46e2be26866354c32c7a56c1d0dd59f1bbf1d94cc776ddd85981df9b2b44b054`
and its related keys
`betterstack:2346967:862cb1fab14ea8f223438eb7227e1dcf3fa676bb41f74b98e1647caf62e1de44`,
`betterstack:2346967:3cd7e975ea1c9ffea633d6a5eca8735708250dad4c940c7fa24957081c061bb6`,
`betterstack:2346967:3fcbde47ce8b99057eec2db314fcd0eb6f928576a40c61b431807ef651d7c333`,
`betterstack:2346967:e1bbcba471bd9ab5b36384e394f1365915208abe546e9d0e7f7199b79a2fc998`.

## How was this tested?

Worktree `/workspace/suna-krtx-4913`, branch `krtx-4913`, head
`ef6c158b7d`, based on `origin/dev` `10af4a490c`.

Evidence (Better Stack MCP `query`, hot exceptions table, last 24 h, the
5 patterns): 8 rows, all `environment: prod`, all the same Android
WebView UA and `/auth` transaction. Frames (oldest first) per shape:

```text
swbrowser is not defined | <anonymous> ? > <anonymous> needInjectCss           | global_handlers.onerror
xbrowser is not defined  | app:///_next/…/30hcsow18blzn.js r > <anonymous> execute_auto_fill  | browserapierrors.setTimeout
xbrowser is not defined  | app:///_next/…/30hcsow18blzn.js r > <anonymous> ? > <anonymous> execute_auto_fill | browserapierrors.addEventListener
xbrowser is not defined  | <anonymous> execute_auto_fill                      | global_handlers.onerror
xbrowser is not defined  | <anonymous> ? > <anonymous> execute_auto_fill      | global_handlers.onerror
```

New gate test, red without the fix (the rule file restored from
`origin/dev`), then green:

```sh
$ cd packages/ui && bun test --isolate src/internal/lib/browser-noise/host-browser-bridge.test.ts   # rule from origin/dev
Expected: true
Received: false
(fail) drops the host-browser bridge ReferenceErrors in their observed frame shapes [4.88ms]
(pass) keeps reporting when the throw is ours or unattributable [0.99ms]
 1 pass
 1 fail
Ran 2 tests across 1 file. [66.00ms]

$ cd packages/ui && bun test --isolate src/internal/lib/browser-noise/host-browser-bridge.test.ts   # with the fix
(pass) drops the host-browser bridge ReferenceErrors in their observed frame shapes [1.68ms]
(pass) keeps reporting when the throw is ours or unattributable [2.09ms]
 2 pass
 0 fail
Ran 2 tests across 1 file. [94.00ms]
```

Golden rule-table suite with the 6 new cases, red before the rule (cases
1949 and 1950, the two prod shapes, mismatched; the 4 negatives already
held), then green:

```sh
$ cd apps/web && pnpm test ./src/lib/browser-noise/rules.test.ts   # before the rule
+     "index": 1949,
+     "index": 1950,
(fail) golden: both gates return the pre-table verdict for every case [59.14ms]
 5 pass
 1 fail
Ran 6 tests across 1 file. [258.00ms]

$ cd apps/web && pnpm test ./src/lib/browser-noise/rules.test.ts ./src/lib/browser-error-noise.test.mts   # with the rule
 489 pass
 0 fail
Ran 489 tests across 2 files. [268.00ms]
```

The rule-table meta tests in `rules.test.ts` also hold: a positive
fixture, a case only this rule drops on the Sentry gate, and proof that
it never fires on the runtime gate.

Whole `packages/ui` suite, browser-noise directory, Sentry config tests:

```sh
$ cd packages/ui && bun test --isolate src/
 9232 pass
 0 fail
 20 snapshots, 182207 expect() calls
Ran 9233 tests across 828 files. [294.74s]

$ cd packages/ui && bun test --isolate src/internal/lib/browser-noise/
 32 pass
 0 fail
Ran 32 tests across 9 files. [160.00ms]

$ cd apps/web && pnpm test ./sentry.options.test.ts ./sentry.client.config.test.ts
 7 pass
 0 fail
Ran 7 tests across 2 files. [1432.00ms]
```

Static checks:

```sh
$ cd packages/ui && pnpm typecheck
> tsgo --noEmit          (exit 0, no output)
$ cd apps/web && bun scripts/generate-fumadocs-source.mjs && ./node_modules/.bin/tsgo --noEmit
[fumadocs-mdx] regenerated .source/          (tsgo exit 0, no output)
$ cd packages/ui && npx eslint src/internal/lib/browser-noise/rules/webview.ts src/internal/lib/browser-noise/host-browser-bridge.test.ts
(exit 0)
```

(`pnpm typecheck` in `apps/web` runs the fumadocs generator under Node
22.11, which cannot load `.ts` imports in this sandbox,
`ERR_UNKNOWN_FILE_EXTENSION`; the generator ran under bun instead, then
the same `tsgo`.)

Replay of the 8 real prod events (exception values and frames only)
through the real `shouldIgnoreSentryNoiseEvent`, at `origin/dev` and
with the fix:

```sh
$ cd packages/ui && bun run /workspace/out/krtx-4913-replay.ts      # origin/dev rule
swbrowser is not defined | ? > needInjectCss | dropped: false
swbrowser is not defined | ? > needInjectCss | dropped: false
xbrowser is not defined | r > execute_auto_fill | dropped: false
xbrowser is not defined | r > execute_auto_fill | dropped: false
xbrowser is not defined | r > execute_auto_fill | dropped: false
xbrowser is not defined | r > ? > execute_auto_fill | dropped: false
xbrowser is not defined | execute_auto_fill | dropped: false
xbrowser is not defined | ? > execute_auto_fill | dropped: false

$ cd packages/ui && bun run /workspace/out/krtx-4913-replay.ts      # with the fix
swbrowser is not defined | ? > needInjectCss | dropped: true
swbrowser is not defined | ? > needInjectCss | dropped: true
xbrowser is not defined | r > execute_auto_fill | dropped: true
xbrowser is not defined | r > execute_auto_fill | dropped: true
xbrowser is not defined | r > execute_auto_fill | dropped: true
xbrowser is not defined | r > ? > execute_auto_fill | dropped: true
xbrowser is not defined | execute_auto_fill | dropped: true
xbrowser is not defined | ? > execute_auto_fill | dropped: true
```

Real surface: the worktree's web app (`next dev --port 3913`) with
`NEXT_PUBLIC_SENTRY_DSN=http://public@127.0.0.1:9555/1`, a local HTTP
sink that logs every envelope. agent-browser opened `/auth`. `eval` code
runs through CDP `Runtime.evaluate`, so its frames are `<anonymous>`,
the same as Android's `evaluateJavascript`. The page's own error
listener proves each error threw:

```sh
$ agent-browser eval 'window.__errs=[]; addEventListener("error", e => window.__errs.push(e.message)); setTimeout(function needInjectCss(){ swbrowser.injectCss(); }, 0); setTimeout(function execute_auto_fill(){ xbrowser.autoFill(); }, 0); "queued"'
# without the fix (rule file from origin/dev, HMR reload):
"[\"Uncaught ReferenceError: swbrowser is not defined\",\"Uncaught ReferenceError: xbrowser is not defined\"]"
WITHOUT FIX sink error events: 2
"value":"swbrowser is not defined"
"value":"xbrowser is not defined"
# with the fix:
"[\"Uncaught ReferenceError: swbrowser is not defined\",\"Uncaught ReferenceError: xbrowser is not defined\"]"
WITH FIX sink error events: 0

$ agent-browser eval 'setTimeout(function needInjectCss(){ zzbrowser.injectCss(); }, 0); "queued"'   # control, with the fix
sink POSTs with an error event: 1
"value":"zzbrowser is not defined"
```

Not run: the full worktree stack (`pnpm worktree start` stops at
`primary Supabase credentials are unavailable`; this sandbox has no
primary `pnpm dev` stack). The change needs only the web app's client
Sentry gate, which ran in the browser above; no API, DB or flow code
changed, so no Docker-backed suite covers it.

## Security & data review

- No customer data: the evidence names only the error class, the
messages, the frame shapes, the public asset path and the browser class.
No user, session or device identifier is in this PR.
- No secret, auth, dependency or PII change. The rule can only suppress
Sentry events for these two messages from injected `<anonymous>` code.

## Rollout / rollback

- No migration, flag or env var. The shared `beforeSend` gate
(`apps/web/sentry.options.ts`, unchanged) picks the rule up in the
client, server and edge configs.
- Rollback: revert the two commits; the rule, its test and its golden
cases are self-contained.
- Confirm after deploy: Better Stack Frontend (`application_id` 2346967)
patterns `46e2be26…`, `862cb1fa…`, `3cd7e975…`, `3fcbde47…`, `e1bbcba4…`
show 0 new occurrences for 24 h after prod carries the merge (metrics:
`SELECT pattern, sum(exceptions_count) FROM
remote(t502678_kortix_frontend_metrics) WHERE pattern IN (…) AND dt >=
<prod time> GROUP BY pattern`).

## Reviewer checklist

- Read
`packages/ui/src/internal/lib/browser-noise/rules/webview.ts:435-450`:
the throw site is `frames.at(-1)` (Sentry frames are oldest first); the
message is exact; a first-party resolved frame anywhere keeps the event.
- Check `apps/web/src/lib/browser-noise/golden-verdicts.json:1951-1956`:
2 prod shapes dropped (Sentry gate), 4 negatives kept (bundle throw
site, first-party frame, frameless, runtime capture).
- Red proof: restore `rules/webview.ts` from `origin/dev` and run
`host-browser-bridge.test.ts`, 1 fail.

<details>
<summary>Changed files</summary>

- `packages/ui/src/internal/lib/browser-noise/rules/webview.ts`: the
`android-webview-host-browser-bridge` rule (Sentry gate).
-
`packages/ui/src/internal/lib/browser-noise/host-browser-bridge.test.ts`:
the observed shapes are dropped; the actionable shapes are kept.
- `apps/web/src/lib/browser-noise/golden-verdicts.json`: 6 pinned cases.

</details>

<!-- codesmith:footer -->
---
<a
href="https://backend.blacksmith.sh/track/enable-autofix?expires=1797220254&installation_model_id=434224&pr_number=12714&ref=codesmith_pr_footer&repository=kortix-ai%2FCompanyOS&return_to=https%3A%2F%2Fgithub.com%2Fkortix-ai%2FCompanyOS%2Fpull%2F12714&signature=42a7310d641fc33527c84b0f5d5188d8db86917c9b53e8dce2417695bb302c0d"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-light.svg"><img
alt="Autofix with [code]smith"
src="https://pr-comments-assets.blacksmith.sh/codesmith/autofix-with-codesmith-dark-3x.png"
width="190" height="28"></picture></a>
<sup><a
href="https://app.blacksmith.sh/kortix-ai/codesmith/CompanyOS/pr/12714?autoLogin=true&ref=codesmith_pr_footer">Open
in [code]smith</a> · Need help on this PR? Tag
<code>@codesmith-bot</code> with what you need.</sup>

<!-- codesmith:autofix:disabled -->
<!-- /codesmith:footer -->

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
@strix-security

strix-security Bot commented Nov 18, 2026 •

Copy link
Copy Markdown

Strix Security Review

Security review could not be completed for this commit.

Updated for c7a7341.


Reviewed by Strix
Re-run review · Configure security review settings

markokraemer added a commit that referenced this pull request Nov 18, 2026
## Demo video

Non-visual exception: release-suite fixes (a flow, a scanner entry and a
test).

## Quick verification

Outcome: these fix the three failures that blocked the v0.16.0 release
PR (#12852). They are the deployed API smoke, secret-scan and the
packages-kortixd lane. The fourth red job, packages-rest, failed on an
external package server returning `503`.

- **SHIP-1** (`✗ wrote .kortix/link.json — expected true, got false`,
deployed smoke against staging):
- Since #12812 the workspace link lives in `git config
kortix.workspace`.
- Every ship flow assertion now reads git config through one helper
(`linkedWorkspace`), and the spec (§SHIP-1, step 3) says so.
- Local runs: `pnpm test -- --id SHIP-DEFAULT` PASS, `--id SHIP-2` PASS.
- SHIP-1 itself needs `managedGitPush`, which only a deployed target
has. The release suite verifies it on staging.
  - `tests` typecheck: 0 errors.
- **secret-scan** (gitleaks `generic-api-key`,
`apps/api/src/projects/evals/hub.ts:22`, from #12748):
- The value is the Harbor Hub's Supabase publishable key
(`sb_publishable_…`). It is public and read-only by design, the same
value Harbor ships in its own client.
- It is ignored by fingerprint in `.gitleaksignore`, with the reason, in
that file's existing format. Locally gitleaks reports "no leaks found"
with it.
- **packages-kortixd** (`pi-workflow.test.ts`, "a person's Stop of one
agent is `stopped`…"):
- The test expected agent `a1` to get child session `child-1`, but the
three agents start in parallel. It failed 1 of 1 locally on dev head,
and was green in the dev lane run.
  - It now asserts the line's meaning and that it names a child session.
- Result: 47/0 three times on bun 1.4.4 (CI's version) and once on
1.3.14. Typecheck: 0 errors.
- Next: this commit is applied to `staging`, then Promote runs with
`ref=staging` for the release PR.
## Demo video

Non-visual exception: release-suite fixes (a flow, a scanner entry and a
test).

## Quick verification

Outcome: these fix the three failures that blocked the v0.16.0 release
PR (#12852). They are the deployed API smoke, secret-scan and the
packages-kortixd lane. The fourth red job, packages-rest, failed on an
external package server returning `503`.

- **SHIP-1** (`✗ wrote .kortix/link.json — expected true, got false`,
deployed smoke against staging):
- Since #12812 the workspace link lives in `git config
kortix.workspace`.
- Every ship flow assertion now reads git config through one helper
(`linkedWorkspace`), and the spec (§SHIP-1, step 3) says so.
- Local runs: `pnpm test -- --id SHIP-DEFAULT` PASS, `--id SHIP-2` PASS.
- SHIP-1 itself needs `managedGitPush`, which only a deployed target
has. The release suite verifies it on staging.
  - `tests` typecheck: 0 errors.
- **secret-scan** (gitleaks `generic-api-key`,
`apps/api/src/projects/evals/hub.ts:22`, from #12748):
- The value is the Harbor Hub's Supabase publishable key
(`sb_publishable_…`). It is public and read-only by design, the same
value Harbor ships in its own client.
- It is ignored by fingerprint in `.gitleaksignore`, with the reason, in
that file's existing format. Locally gitleaks reports "no leaks found"
with it.
- **packages-kortixd** (`pi-workflow.test.ts`, "a person's Stop of one
agent is `stopped`…"):
- The test expected agent `a1` to get child session `child-1`, but the
three agents start in parallel. It failed 1 of 1 locally on dev head,
and was green in the dev lane run.
  - It now asserts the line's meaning and that it names a child session.
- Result: 47/0 three times on bun 1.4.4 (CI's version) and once on
1.3.14. Typecheck: 0 errors.
- Next: this commit is applied to `staging`, then Promote runs with
`ref=staging` for the release PR.
Comment thread apps/cli/src/commands/evals.ts Fixed
fix(release): the three v0.16.0 release-suite failures (staging)
@github-actions

Copy link
Copy Markdown
Contributor Author

Tests - release attempt 2 ended failure: contract / core-db lane, contract / packages-publish lane, full suite + quality gates. Rerunning it once: https://github.com/kortix-ai/CompanyOS/actions/runs/44392064437

A job that passes on the rerun is a flake. Fix it at the root or delete it the same day (the testing skill, No quarantine). A job that fails twice is real.

markokraemer added a commit that referenced this pull request Nov 18, 2026
…2859)

## Demo video

Non-visual exception: security fixes in connectors, the CLI and two
parsers. Behaviour stays the same.

## Quick verification

Outcome: the 11 high-severity CodeQL alerts that blocked release PR
#12852 are fixed in code, not dismissed. The CodeQL check was the gate's
last red item; every other check in that release suite passed.

- **Incomplete multi-character sanitization (4):**
`connectors/google/gmail-message.ts` `htmlToText` and
`docs-markdown.ts`.
- Each tag and comment removal repeats until the text stops changing.
The steps and their order are unchanged.
- **File system race (3):**
- `kortix evals init` writes with flag `wx`; `EEXIST` keeps the existing
file.
  - The eval judge reads the reward file directly.
  - `templates doctor` reads the import file directly.
- **Polynomial regex (4):**
- The trailing `/` trims in `manifest-schema/src/evals.ts` and the
trailing `-` trim in `sandbox-tab.tsx` are now loops.
- The markdown link label class in `inline-code-text.tsx` excludes `[`.
- 60,000 slashes in the middle of a path parse in 3 ms. `./evals/tasks/`
and `evals/tasks///` both still give `evals/tasks`.
- Tests, one file at a time, all 0 fail:
  - api `gmail-mime` 7, `google/server` 10
  - cli `templates-doctor` 2, `evals` 3, `evals-judge.blackbox` 1
- ui `inline-code-text` 14, `sandbox-tab` 26, `sandbox-tab.empty-new` 3
- Typecheck shows 0 errors for kortix-api, @kortix/cli, @kortix/ui and
manifest-schema. `pnpm --filter kortix-api lint` exits 0.
- Next: this commit is applied to `staging`, then Promote runs with
`ref=staging`.
…2859)

## Demo video

Non-visual exception: security fixes in connectors, the CLI and two
parsers. Behaviour stays the same.

## Quick verification

Outcome: the 11 high-severity CodeQL alerts that blocked release PR
#12852 are fixed in code, not dismissed. The CodeQL check was the gate's
last red item; every other check in that release suite passed.

- **Incomplete multi-character sanitization (4):**
`connectors/google/gmail-message.ts` `htmlToText` and
`docs-markdown.ts`.
- Each tag and comment removal repeats until the text stops changing.
The steps and their order are unchanged.
- **File system race (3):**
- `kortix evals init` writes with flag `wx`; `EEXIST` keeps the existing
file.
  - The eval judge reads the reward file directly.
  - `templates doctor` reads the import file directly.
- **Polynomial regex (4):**
- The trailing `/` trims in `manifest-schema/src/evals.ts` and the
trailing `-` trim in `sandbox-tab.tsx` are now loops.
- The markdown link label class in `inline-code-text.tsx` excludes `[`.
- 60,000 slashes in the middle of a path parse in 3 ms. `./evals/tasks/`
and `evals/tasks///` both still give `evals/tasks`.
- Tests, one file at a time, all 0 fail:
  - api `gmail-mime` 7, `google/server` 10
  - cli `templates-doctor` 2, `evals` 3, `evals-judge.blackbox` 1
- ui `inline-code-text` 14, `sandbox-tab` 26, `sandbox-tab.empty-new` 3
- Typecheck shows 0 errors for kortix-api, @kortix/cli, @kortix/ui and
manifest-schema. `pnpm --filter kortix-api lint` exits 0.
- Next: this commit is applied to `staging`, then Promote runs with
`ref=staging`.
fix(security): the 11 CodeQL alerts the v0.16.0 release PR raised (staging)
@github-actions

Copy link
Copy Markdown
Contributor Author

Tests - release attempt 2 ended failure: full suite + quality gates. Rerunning it once: https://github.com/kortix-ai/CompanyOS/actions/runs/44404557106

A job that passes on the rerun is a flake. Fix it at the root or delete it the same day (the testing skill, No quarantine). A job that fails twice is real.

markokraemer added a commit that referenced this pull request Nov 18, 2026
…2862)

## Demo video

Non-visual exception: a regex bound. Link text renders the same.

## Quick verification

Outcome: this fixes the last CodeQL alert on release PR #12852, a
polynomial regex in `packages/ui/src/markdown/inline-code-text.tsx:41`
that runs slow on a run of `[](`.

- Fix: the link URL class `[^)]*` becomes `[^()\s]*`, so the scan stops
at whitespace and parentheses.
- Same output for `see [docs](https://kortix.com/docs) now`,
`![img](a.png)`, `[a](b) and [c](d)` and `plain [not a link]`.
- Hostile inputs: `[](`×40k (120k chars) 6 ms, `[`×60k 3 ms, `[\\`×40k 4
ms.
- `pnpm --filter @kortix/ui test
src/markdown/inline-code-text.test.tsx`: 14 pass, 0 fail. ui typecheck:
0 errors.
…2862)

## Demo video

Non-visual exception: a regex bound. Link text renders the same.

## Quick verification

Outcome: this fixes the last CodeQL alert on release PR #12852, a
polynomial regex in `packages/ui/src/markdown/inline-code-text.tsx:41`
that runs slow on a run of `[](`.

- Fix: the link URL class `[^)]*` becomes `[^()\s]*`, so the scan stops
at whitespace and parentheses.
- Same output for `see [docs](https://kortix.com/docs) now`,
`![img](a.png)`, `[a](b) and [c](d)` and `plain [not a link]`.
- Hostile inputs: `[](`×40k (120k chars) 6 ms, `[`×60k 3 ms, `[\\`×40k 4
ms.
- `pnpm --filter @kortix/ui test
src/markdown/inline-code-text.test.tsx`: 14 pass, 0 fail. ui typecheck:
0 errors.
@github-actions

Copy link
Copy Markdown
Contributor Author

Tests - release attempt 2 ended failure: contract / core-flows lane, full suite + quality gates. Rerunning it once: https://github.com/kortix-ai/CompanyOS/actions/runs/44416725481

A job that passes on the rerun is a flake. Fix it at the root or delete it the same day (the testing skill, No quarantine). A job that fails twice is real.

@markokraemer markokraemer left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Approving v0.16.0 on Marko's behalf (he asked me to approve and merge). Gate full suite + quality gates is green on head 3ccfe9e (gate run 44416725481). Migrations are costed against prod. Drata compliance-as-code is advisory, with the same 2 known false positives as v0.15.11.

@markokraemer
markokraemer merged commit 9bc6179 into prod Nov 18, 2026
85 of 90 checks passed
@markokraemer
markokraemer deleted the release/v0.16.0 branch November 18, 2026 23:28
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants