Skip to content

Release v0.16.3 — More reliable triggers and clearer usage tracking - #12969

Merged
markokraemer merged 11 commits into
prodfrom
release/v0.16.3
Nov 21, 2026
Merged

markokraemer merged 11 commits into
prodfrom
release/v0.16.3

Conversation

@github-actions

@github-actions github-actions Bot commented Nov 21, 2026 •

Copy link
Copy Markdown
Contributor

More reliable triggers and clearer usage tracking

New

Improved

Fixed

Security

  • No security-specific changes in this release.

Internal

agent-kortix and others added 11 commits November 20, 2026 18:03
…X-5031) (#12950)

## Demo video

No demo video: code-only change. The diff adds one unit test. No
renderer or runtime code changes, so there is no new screen to record.

## Quick verification

- **Outcome:** a `/command` message now has a test for its replay
privacy. The test checks that the command chip and the args the person
typed sit inside the `[data-replay-content]` region. That region is what
`REPLAY_CONTENT_SELECTOR` (existing,
`apps/web/src/lib/analytics/posthog-config.ts`) hides in a session
replay. Until now no test on `dev` checked the marker on user-message
text, so a renderer refactor could drop it and leak typed command args
into replays without failing anything.
- **Why this shape (the issue's acceptance, as I read it):** KRTX-5031
asked to add `data-replay-content` to the four byte-identical
command-markup fixtures. #12923 (KRTX-5013) has since deleted those
fixtures from `dev` under AGENTS.md testing rule 6 ("delete on red"),
and two earlier attempts stopped there as obsolete. I did not restore
the deleted fixtures. Instead I kept the issue's intent: the command
markup must carry the replay redaction marker. I assert it as a privacy
behavior (rule 6 keeps tests for secrets and privacy decisions). The
production renderer is unchanged (existing
`packages/ui/src/session/turn/user-message.tsx:1169`), and no assertion
is weakened.
- **Caller boundary:** the command comes either from `commandInfo`
passed by `transcript.tsx:1730` or from `detectCommandFromText`. Both
render through the single `UserMessageBubble` call site (existing,
`user-message.tsx:2110`), so this test covers the one region both paths
share.
- **Red before green:** I removed `data-replay-content=""` from the
renderer and ran the new test: `0 pass, 1 fail` with
`expect(received).not.toBeNull() Received: null`. I then restored the
renderer and it passed.
- **Risk:** low. The diff is test-only: one test and one `happy-dom`
import, which is already a devDependency of `@kortix/ui` and is used by
sibling tests.
- **Verified:** `bun test src/session/turn/user-message.test.tsx` gives
97 pass, 0 fail. The whole `@kortix/ui` suite (`pnpm test`) gives 9599
pass, 0 fail. `pnpm typecheck` exited 0, and eslint on the changed file
exited 0. Not run: the local stack and the browser, because the diff
changes no runtime code and nothing new is visible.
- suna-skills: contributing, create-pr (template), testing, worktree,
learnings (searched replay/posthog: no entry for this surface)
- ponytail: full · review: Lean already. Ship. · markers: 0
- Self-review: 1 items found and fixed (the first draft's
`hidden?.querySelector(...)` was `undefined` with no marker and still
passed `.not.toBeNull()`; it is now a descendant selector, which fails
with `null`)
- Closes KRTX-5031

```
$ cd packages/ui && bun test src/session/turn/user-message.test.tsx -t "replay"   # renderer marker removed
error: expect(received).not.toBeNull()
Received: null
 0 pass
 1 fail

$ cd packages/ui && bun test src/session/turn/user-message.test.tsx
 97 pass
 0 fail
 315 expect() calls
Ran 97 tests across 1 file. [1353.00ms]

$ cd packages/ui && pnpm test
 9599 pass
 0 fail
Ran 9599 tests across 872 files. [538.46s]

$ cd packages/ui && pnpm typecheck; echo exit=$?
exit=0
$ cd packages/ui && pnpm exec eslint src/session/turn/user-message.test.tsx; echo exit=$?
exit=0
```

<!-- codesmith:footer -->
---
<a
href="https://dashboardbackend.blacksmith.sh/track/enable-autofix?expires=1797779707&installation_model_id=434224&pr_number=12950&ref=codesmith_pr_footer&repository=kortix-ai%2FCompanyOS&return_to=https%3A%2F%2Fgithub.com%2Fkortix-ai%2FCompanyOS%2Fpull%2F12950&signature=7f218d6b347215fbb4b0137bfc7825c6cdc0ecf4fbd1bf712d7f5335532c2c9a"><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/12950?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>
…-5042) (#12956)

## Demo video

No demo video: code-only change. Gateway valuation and subscription
accounting only.

## Quick verification

- Outcome: price Sol, Terra, Morph's documented upstream proxy, and
Claude 5.5 Opus/Sonnet. Record Claude input/output/cache usage without
charging subscription calls or logging payloads. Closes KRTX-5042.
- Feedback: require observed input/output counters; malformed, oversized
or invalid frames stay incomplete and unpriced. Optional cache counters
default to zero. Relay bytes unchanged. Claude settlement cannot release
a paid request's block-budget reservation.
- Risk: unknown models stay unpriced; historical missing Claude usage
cannot be recovered. Morph retail proxy is not a bill. No migration.
Roll back by reverting this PR.
- Verified on head 92d8743 after merging origin/dev: the commands
below passed. No whole-package suites: repository AGENTS.md requires
named targeted files.

```sh
. scripts/hermetic-test-env.sh
bun --env-file=apps/api/scripts/test.env test --isolate packages/llm-gateway/src/claude-code/usage.test.ts packages/llm-gateway/src/claude-code/passthrough.test.ts packages/llm-gateway/src/create-gateway.test.ts apps/api/src/shared/llm-value.test.ts apps/api/src/llm-gateway/hooks.test.ts apps/api/src/__tests__/unit-gateway-budgets.test.ts apps/cli/src/__tests__/sessions-usage.test.ts apps/cli/src/__tests__/gateway.test.ts
131 pass
0 fail
394 expect() calls
Ran 131 tests across 8 files. [20.61s]

bun test packages/llm-gateway/src/claude-code/usage.test.ts packages/llm-gateway/src/create-gateway.test.ts
27 pass
0 fail
83 expect() calls
Ran 27 tests across 2 files. [259.00ms]

pnpm --filter @kortix/llm-gateway typecheck
# exit 0
pnpm --filter kortix-api typecheck
# exit 0
cd apps/api && ../../node_modules/.bin/eslint src/llm-gateway/hooks.ts src/llm-gateway/hooks.test.ts src/shared/llm-value.ts src/shared/llm-value.test.ts src/shared/integration-llm-value.test.ts
# exit 0; 0 errors, 3 ignored-test warnings
```

- Red before fix: usage parser 6 pass / 4 fail; real HTTP caller with
old parser 15 pass / 2 fail; reservation regression 29 pass / 1 fail.
Real local upstream tests assert 11/23/37/53 plain/read/write/output
becomes inclusive input 71/output 53, plus damaged JSON/SSE produce
estimated usage and unsuccessful traces without changing response bytes.
- Unrun this session: local API stack, GW-11, and DB suites. `pnpm
worktree start krtx-5042-rework --no-web` failed: primary Supabase REST
503 because schema `kortix` does not exist; primary Supabase credentials
unavailable. No deployed CLI/provider call or paid subscription call.
The prior worker's DB evidence remains in the earlier PR history; it is
not a run from this session.
- DB assertions inspected: integration-llm-value still prices exact
provider/model rows and daily series; integration-session-usage still
sums recorded counts; gateway-series and covering-index suites keep
unchanged columns/indexes. No schema change. The repair only invalidates
damaged usage and avoids releasing an unreserved Claude request.
- Self-review: 1 items found and fixed (block-budget release). A–H
checked against all acceptance criteria; caller/adapter traced from
gateway HTTP routes through createGateway, passthrough, observer and
settlement. No alternate production observer caller exists. Independent
G9 remains required.
- suna-skills: worktree, testing, learnings, contributing (attachments
and preview references), create-pr, claude-cli
- ponytail: full · review: Lean already. Ship. · markers: 0
- Issue: https://linear.app/kortix/issue/KRTX-5042

---------

Co-authored-by: Kortix Agent <292857086+agent-kortix@users.noreply.github.com>
…abels (#12962)

The reviewer path on prod is email, "Use a password", Allow, then Finish on the welcome tab; the prompt goes in the workspace home. The reference also records how the reviewer account is created.
## Review in 60 seconds

- Config commits now include `Kortix-Actor` and `Kortix-Session`
trailers from server authentication.
- Manifest, prompt, agent governance/default, and skill multi-file
writes share one formatter. Both Git CLI and GitHub Contents paths carry
attribution.
- A dedicated authentication-session field prevents route-target
sessions or request-body identity fields from changing attribution.
Missing identity/session is recorded as `none`.

No demo video: code-only change

**Risk:** Adds authenticated principal identifiers to workspace Git
history, as requested. No schema, permissions, author identity, or
public API change. Revert this PR to stop adding trailers; existing
commits remain unchanged.
**Verified:** Focused tests: 30 pass, 0 fail. Related tests: 130 pass, 0
fail. API typecheck exit 0. Changed-file eslint exit 0 (test files
ignored by existing config). Real local bare-Git commit and persisted
message verified. Full-stack flow could not run: fresh Supabase
PostgREST requires the absent `kortix` schema before stack
startup/migrations complete.

suna-skills: worktree, testing, learnings, contributing, create-pr
ponytail: full · review: Lean already. Ship. · markers: 0
Self-review: 1 items found and fixed

## Demo video

No demo video: code-only change. Real Git commit assertions and Hono
request/principal boundary tests provide the relevant proof.

## Quick verification

Outcome: platform config commits identify their authenticated caller
without accepting body-supplied identity.

`bun --env-file=scripts/test.env test --isolate
src/lib/request-context.test.ts
src/projects/git/manifest-cas.integration.test.ts
src/projects/lib/commit-manifest-cas.test.ts
src/middleware/auth-config-attribution.test.ts
src/projects/lib/config-commit-message.test.ts` → 30 pass; 0 fail; 74
expect() calls.

Closes KRTX-5049 — https://linear.app/kortix/issue/KRTX-5049

Root cause: config writers forwarded operation-only messages;
token-bound session identity was not propagated to asynchronous request
context.

Caller boundary: actual `patPrincipal`, `serviceAccountPrincipal`, and
`jwtPrincipal` run behind a Hono request adapter with forged body IDs
and a different route-target session. The commit formatter retains the
principal identity. An actual local bare Git repository proves the
stored manifest commit contains the trailers. GitHub Contents and Git
CLI adapters are separately asserted. Alternate direct agent-config and
project-skills callers explicitly use the same formatter; OAuth already
propagates the authenticated user and has no bound platform session.

Self-review A–H: all acceptance behaviors covered; no permission, schema
or rollout change. One test fixture lacked required project-row fields;
fixed without casts. Ponytail review: Lean already. Ship. No new
dependencies. Independent G9 remains the merge gate's responsibility.

## How was this tested?

`bun --env-file=scripts/test.env test --isolate
src/lib/request-context.test.ts
src/projects/git/manifest-cas.integration.test.ts
src/projects/lib/commit-manifest-cas.test.ts
src/middleware/auth-config-attribution.test.ts
src/projects/lib/config-commit-message.test.ts` → 30 pass; 0 fail; 74
expect() calls.

Commands ran from `apps/api` unless stated otherwise.

```text
bun test apps/api/src/projects/lib/config-commit-message.test.ts
# Before implementation (identity formatter returned the unmodified operation):
0 pass
4 fail
4 expect() calls
Ran 4 tests across 1 file. [54.00ms]

bun --env-file=scripts/test.env test --isolate src/middleware/auth-config-attribution.test.ts src/projects/lib/commit-manifest-cas.test.ts src/projects/git/manifest-cas.integration.test.ts
# Removed auth-session propagation and manifest-message attribution, retaining tests:
17 pass
5 fail
61 expect() calls
Ran 22 tests across 3 files. [4.77s]

bun --env-file=scripts/test.env test --isolate src/lib/request-context.test.ts src/projects/git/manifest-cas.integration.test.ts src/projects/lib/commit-manifest-cas.test.ts src/middleware/auth-config-attribution.test.ts src/projects/lib/config-commit-message.test.ts
30 pass
0 fail
74 expect() calls
Ran 30 tests across 5 files. [4.74s]

bun --env-file=scripts/test.env test --isolate src/middleware/auth.test.ts src/middleware/auth-characterization.test.ts src/middleware/auth-sso-sync.test.ts src/projects/lib/triggers.test.ts src/projects/git/manifest-imports.integration.test.ts src/connectors/manifest-crud-create-only.test.ts src/__tests__/unit-auth-logout-revocation.test.ts src/__tests__/unit-secret-grant-route.test.ts
130 pass
0 fail
413 expect() calls
Ran 130 tests across 8 files. [8.71s]

# Repository root:
pnpm --filter kortix-api typecheck
> tsgo --noEmit --incremental --tsBuildInfoFile node_modules/.cache/tsgo.tsbuildinfo
exit 0

# apps/api:
../../node_modules/.bin/eslint src/lib/request-context.ts src/middleware/auth-principal.ts src/middleware/auth-config-attribution.test.ts src/projects/lib/config-commit-message.ts src/projects/lib/config-commit-message.test.ts src/projects/lib/trigger-manifest.ts src/projects/routes/agent-config.ts src/projects/routes/project-skills.ts
0 errors, 2 warnings (test files ignored by existing configuration)
exit 0

# Repository root:
pnpm worktree start krtx-5049
supabase_rest_kortix-local unexpected status 503
Failed to load the schema cache: schema "kortix" does not exist
primary Supabase credentials are unavailable
exit 1

pnpm test -- --id PROJ-19
local database migration exited with code 1
code: "57P01" (stack startup stopped the database after failed health checks)
FAIL api-cli-flows
exit 1

git merge origin/dev --no-edit
Already up to date.
git diff --check
exit 0
```

Full API package suite was not run: repo AGENTS.md and testing skill
explicitly prohibit whole-package runs during development. Focused and
related files were run instead.

Unverified DB-backed coverage: `integration-iam-build-actor.test.ts`
uses principal bindings unchanged by the added context field;
`integration-session-list-poll-memo-attribution.test.ts` and
`integration-session-participants-roster.test.ts` retain
session/participant state unchanged. `PROJ-19` config bodies/responses,
CAS and permissions are unchanged; only Git messages gain trailers.
Local HTTP stack and deployed-dev verification remain unverified. The
Git-only `manifest-cas.integration.test.ts` and
`manifest-imports.integration.test.ts` did run successfully.

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

No demo video: code-only change.

## Quick verification

Outcome: KRTX-5050 rejects undeclared agents in `on`/`run` triggers.
POST and PATCH refuse new manifest errors before committing.

- The validator reports `triggers[i].run.agent`; declared agents,
`default`, `mykortix`, its legacy alias, and omitted agents remain
valid. Flat references remain checked.
- Root cause: cross-validation read only flat `agent`; trigger CRUD also
did not call the existing added-error validator. Reused
`addedManifestErrors` on both write paths.
- **Risk:** invalid trigger edits now return 400. Existing unrelated
manifest errors remain editable. No migration. Rollback: revert this PR.
- **Verified:** manifest-schema and API typechecks exit 0; changed API
file ESLint exits 0; 607 targeted parser tests pass; TRG-2 passes
through the local HTTP stack, including POST/PATCH refusal and unchanged
manifest read-back.

```sh
bun test packages/manifest-schema/src/__tests__/validator.test.ts packages/manifest-schema/src/__tests__/validator.v3.test.ts packages/manifest-schema/src/__tests__/validator.rules.test.ts packages/manifest-schema/src/__tests__/validator.coverage.test.ts packages/manifest-schema/src/__tests__/json-schema.conformance.test.ts
607 pass
0 fail
1601 expect() calls
Ran 607 tests across 5 files. [1336.00ms]

pnpm test -- --id TRG-2
results: 1/1 passed · 0 failed · 0 skipped · 0 todo · 1.5s
[test] PASS api-cli-flows 46.5s

pnpm --filter @kortix/manifest-schema typecheck
exit 0
pnpm --filter kortix-api typecheck
exit 0
cd apps/api && bunx eslint src/projects/routes/triggers.ts
exit 0
```

Before fix: validator regression run had `84 pass; 5 fail`. Before the
route guard, TRG-2 failed: `expected [400], got 201`. The final flow
verifies both POST and PATCH reject `ghost`, and reads the manifest back
unchanged. The existing create still returns 201.

Adapter check: `bun --no-env-file /workspace/out/run-adapter.ts`
launches the existing `trigger-on-run.test.ts` with the deterministic
local API environment: `8 pass; 0 fail; 15 expect() calls`.

Unrun: full package/release suites, prohibited during development by
repository AGENTS.md/testing. Dev deployment verification awaits merge.
No external managed Git repository or cloud sandbox is used; the local
flow uses its real bare-Git fixture and database.

Self-review: 2 items found and fixed. Final A–H review: no outstanding
defects. Caller boundary: POST/PATCH → draft → on/run serializer →
addedManifestErrors → canonical validator; the CLI shares the validator.
No helper duplication or dependency added.

suna-skills: worktree, testing, learnings, contributing (attachments and
preview references), create-pr.
ponytail: full · review: Lean already. Ship. · markers: 0

Closes https://linear.app/kortix/issue/KRTX-5050

<!-- codesmith:footer -->
---
<a
href="https://dashboardbackend.blacksmith.sh/track/enable-autofix?expires=1797812988&installation_model_id=434224&pr_number=12967&ref=codesmith_pr_footer&repository=kortix-ai%2FCompanyOS&return_to=https%3A%2F%2Fgithub.com%2Fkortix-ai%2FCompanyOS%2Fpull%2F12967&signature=1de79ac7a44ca6cc2772007901bcedf5260842cd408e94a12308196767d16d33"><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/12967?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 21, 2026 •

Copy link
Copy Markdown

Strix Security Review

✅ No open security findings.

PR summary: Release v0.16.3 bundles config-commit attribution (the authenticated actor and session are appended to configuration commit messages), recording and valuation of Claude subscription usage, a schema check that trigger `run.agent` references a declared agent, and CI changes that preload Supabase images into the runner template.

Review summary

This release PR changes no runtime behavior that opens a security hole. The parts worth scrutiny — the new commit-attribution trailer and its authSessionId plumbing, the provider-gated budget-reservation release, the Claude subscription usage observer and its ledger writes, and the widened trigger agent-reference check — were traced end to end and each suspected issue was rejected on the code. The attribution trailer reaches git as argv data and the GitHub Contents API as a JSON field, and nothing parses it for a decision; the budget-release guard only removes a release for a path that never takes a reservation, which closes a fail-open where an unconditional release could strip another in-flight request's project-scoped reservation; the response observer relays bytes, status and headers unchanged with a bounded 1 MiB parse buffer and records no prompt or completion content; the valuation SQL is built from a static price table through bind parameters and only feeds display reports. The remaining two suspicions — attribution text being spoofable through a caller-supplied name, and agent grants written through the agent-config route — are respectively unread by anything and a documented, pre-existing contract of an unchanged handler.

Ruled out (4)
  • Can a trigger name an undeclared agent by carrying a flat `agent` beside a `run:` table, and have the runtime honor it?
    The mixed shape does escape the roster check, but an undeclared name resolves to a default-deny grant that is strictly narrower than `default` or any declared agent, and governed projects refuse it when the session is created — so it only affects the manifest author's own configuration.
  • Can a usage settlement use the provider field to suppress or take another caller's block-budget reservation?
    Every user-facing emitter builds the event's provider, project and user from the authenticated principal or the resolved descriptor, the internal settlement route is token-gated, and the change only removes a release that is fail-closed and self-expires after five minutes.
  • Can a caller forge the Kortix-Actor/Kortix-Session trailer through a newline in the skill folder name?
    A newline in the folder does land extra trailer lines in the commit body, but no handler, job or authorization path parses those trailers, and the folder must already exist as a manifest value and repository path the caller can commit to.
  • Can a caller holding only project.agent.write escalate by writing kortix_permissions into an agent block?
    That write is the documented contract of the route and is pre-existing code the changed lines do not reach, human-only actions are enforced above the grant, and the same writer can already hand the agent the project's secrets, connectors and a sandbox shell.

Review 1 on this PR · db30285 · requested Nov 21 02:58 UTC when the PR was opened


Reviewed by Strix
Re-run review · Configure security review settings · Try a deeper review with @strix-security /deep

@markokraemer
markokraemer merged commit 983365d into prod Nov 21, 2026
31 checks passed
@markokraemer
markokraemer deleted the release/v0.16.3 branch November 21, 2026 11:57
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.

3 participants