Weekly digest -- 2026-09-07 (Session Trail + Guard made real + OpenAI cost crisis resolved)

[vivekchand]

Theme of the week

Three things landed in sequence and each depended on the other: the Session Trail (Inputs → Decisions → Outcome for every runtime, schema v15) gave users the per-session narrative they'd been missing; Guard control was fixed to actually send signals to non-OpenClaw runtimes (it had been silently no-oping for Claude Code, Codex, Cursor, and every other adapter since the feature launched); and three successive OpenAI cost-accuracy PRs corrected a systematic problem in rate lookup, cache discount application, and double-counting that made every OpenAI cost figure wrong.

Note: vivekchand/clawmetry-cloud is not accessible to this session's GitHub token. The "Cloud" rows throughout this digest are marked [data unavailable]. To fix: add clawmetry-cloud to the session's allowed repos in the Claude Code connector settings.

---

Shipped (OSS only — ~52 meaningful PRs)

OSS (vivekchand/clawmetry)

Features

• [RELEASE] Session Trail: inputs, decisions and outcome for every runtime (carries #5484 #5486 #5487 #5488 #5494) #5512 Session Trail: Inputs / Decisions / Outcome for every runtime (carries Inputs & context: capture what the agent was given (session_context, context.compiled, Inputs panel) #5484 Trail: reasoning as first-class spans + per-runtime coverage declarations #5486 Deliver silent failures to humans: detector alerts, rate_limited/blocked_on_user/crashed detectors, opt-in non-determinism replay, outcome insights #5487 Trail data layer: typed event columns, session intent, per-session git join (schema v15) #5488 Trail page (Inputs / Decisions / Outcome) + session-first landing #5494) — new tab giving every completed session a structured narrative: what the agent was handed, what decisions it made, and what the outcome was. Schema v15. Works for Claude Code, Codex, Cursor, Goose, OpenClaw and hosted dashboard.
• [RELEASE] Signal shifts and briefs; the Raindrop gap closure for every runtime (carries #5506) #5518 Signal shifts + scheduled briefs; Raindrop gap closure for every runtime (feat(signals): shift detection opens issues, scheduled briefs to alert channels (WO-62) #5506) — behaviour signal shifts now open issues automatically; the daemon posts scheduled briefs to alert channels (WO-62); every runtime's Raindrop gap is closed.
• [RELEASE] Briefs reach the hosted dashboard (carries #5520) #5532 Briefs reach the hosted dashboard (feat(signals): briefs snapshot slice for the hosted dashboard; cloud-mode controls say where to act #5520) — cloud users get the signals brief snapshot; cloud-mode controls now say "go here to act" instead of pretending they work locally.
• feat(openclaw): surface OOM-victim events for local model servers (#5548) #5573 feat(openclaw) surface OOM-victim events for local model servers ([obs-gap:openclaw] openclaw: OOM-victim preference for local model servers not tracked #5548) — when a GPU model server is killed by the kernel, it now shows up in the dashboard.
• feat(openclaw): capture response_steered spans in the Tracing tab (#5578) #5591 feat(openclaw) capture response-steered spans in Tracing tab ([obs-gap:openclaw] OpenClaw: response-steering history for async tool calls not captured #5578) — async tool-call steering history is now visible per span.
• feat(openclaw): capture reply recovery state in assistant spans (#5577) #5590 feat(openclaw) capture reply recovery state in assistant spans ([obs-gap:openclaw] OpenClaw: reply recovery state across Gateway restarts not captured #5577) — Gateway-restart recovery attempts tracked and surfaced.
• feat(nemoclaw): resolve vLLM sandbox provider distinctly in _sandbox_inference_configs (#5579) #5587 feat(nemoclaw) resolve vLLM sandbox provider distinctly ([obs-gap:nemoclaw] nemoclaw: vLLM sandbox provider not resolved distinctly (generic 'inference' fallback) #5579) — vLLM no longer collapses to the generic inference fallback in cost and tracing.
• feat(openclaw): surface gateway degraded-start state from migration warnings (#5547) #5561 feat(openclaw) surface gateway degraded-start from migration warnings ([obs-gap:openclaw] openclaw: degraded-start (migration warning) Gateway state not captured #5547).
• feat(crons): surface quarantined cron rows in health-summary (#5546) #5569 feat(crons) surface quarantined cron rows in health-summary ([obs-gap:openclaw] openclaw: cron quarantine on malformed legacy rows not surfaced #5546) — malformed legacy rows are now visible rather than silently dropped.

Fixes

• fix(guard): Pause/Stop/Kill did nothing on every runtime except OpenClaw #5529 fix(guard): Pause/Stop/Kill did nothing on every runtime except OpenClaw — the pid guard compared the session record's timestamp instead of the process start time, so every signal went to the wrong (or no) process. All runtimes now controllable.
• Fix Claude Code hook ownership when the launcher path is quoted (P0) #5509 Fix Claude Code hook ownership when launcher path is quoted (P0) — a quoted launcher path broke the ownership check, causing ClawMetry to skip reinstalling its hook and leaving the gate dark.
• Guard: a killed agent says Stopped, tells you how to resume it, and shows what the button did #5563 Guard: a killed agent now says "Stopped", shows the resume command, and records what the button did — previously the UI reported the agent as still running after a kill.
• fix(approvals): make an approval request legible, and stop the preset that paused everything #5593 fix(approvals) make an approval request legible; stop the preset that paused everything — the default approval preset was blocking every session; the request card was unreadable.
• Fix: per-repo agent memory never reached the cloud Memory tab (codex, cursor, opencode) #5596 fix per-repo agent memory never reached the cloud Memory tab (Codex, Cursor, opencode).
• Show only the Security panels the hosted dashboard can actually fill #5597 Show only Security panels the hosted dashboard can actually fill — stops the tab from claiming scans it cannot run.
• Remove the Ask tab #5603 Remove the Ask tab — was present on the hosted dashboard but non-functional; removed cleanly.
• fix: correct OpenAI rates and inclusive prompt-cache accounting #5608 + [RELEASE] fix(accuracy): proxy cost uses provider-aware accounting and reads OpenAI cached tokens #5616 + [RELEASE] fix(usage): stop double-counting OpenAI cache_read in /api/token-attribution #5617 Fix OpenAI cost accounting (three PRs): correct rates for GPT-4.1/mini, inclusive prompt-cache discount, stop double-counting cache_read tokens in /api/token-attribution.
• [RELEASE] fix(accuracy): restore full event shape in sessions clusters (tool_count regression from #5524) #5607 fix(accuracy) restore full event shape in sessions clusters (tool_count regression from perf(cost): ship slim event rows and share one round trip for the Cost tab scans #5524).
• fix(replay): the Sessions replay lost its column, and a reloaded Trail link never loaded #5537 fix(replay) Sessions replay lost its column; a reloaded Trail link never loaded.
• fix(trail): the Inputs card showed each kind twice, and both as Empty #5560 fix(trail) Inputs card showed each kind twice, both labelled "Empty".
• fix(overview): "No agents running" printed above "3 sessions are working right now." #5553 fix(overview) "No agents running" printed above "3 sessions are working right now."
• fix(guard): list a running session in milliseconds, not minutes #5551 fix(guard) running session duration shown in milliseconds instead of minutes.
• fix(guard): the pid guard compared the session record's timestamp, not the process start #5543 fix(guard) pid guard used session record timestamp, not process start — root cause of the Pause/Stop/Kill regression.
• fix(cli): bare clawmetry --help/-h must not import dashboard #5535 fix(cli) bare clawmetry --help / -h must not import dashboard (500ms startup regression).
• fix(c6-health): post to active tracking issue #5266, not closed #4552 #5565 fix(c6-health) bot was posting to closed tracking issue E2E Robustness Epic: 7-criterion automated testing tracker (active) #4552; redirected to active E2E Robustness Epic: 7-criterion tracker (cron-managed) #5266.

Perf

• perf(store): compact the DuckDB file at writer startup when it is mostly dead space #5503 + perf(store): drop the nine range indexes on events and spans; keep point lookups only #5504 + [RELEASE] perf: newest-events scan sorts keys first, no longer reads the whole payload column per poll #5505 + [RELEASE] perf: the local store stays compact and snappy (carries #5503, #5504) #5515 DuckDB store compaction and index removal — compact at startup when mostly dead space; drop 9 range indexes on events/spans (point lookups only); newest-events scan sorts keys first instead of reading the full payload column. Measured 40–60% reduction in scan time on large stores.
• perf(store): run the startup compaction in a helper process so an engine crash cannot crash-loop the daemon #5519 Startup compaction in a helper process — a DuckDB engine crash during compaction can no longer crash-loop the daemon.
• perf(cost): ship slim event rows and share one round trip for the Cost tab scans #5524 + perf(store): write session_phase and session_context rows only when they change #5533 Slim event rows; skip unchanged session_phase/context writes — Cost tab scan is a single round-trip; session_phase stops rewriting itself every 15s tick.

Chore / refactor / CI hardening / docs

• Harden supply chain: clear 7 npm advisories, attach release provenance, pin base image, add fuzzing #5510 Harden supply chain: clear 7 npm advisories, attach release provenance, pin base image to digest, add fuzzing.
• Harden CI: install the Claude Code CLI from a lockfile #5571 + Harden CI: install the E2E runtime's OpenClaw CLI from a lockfile #5575 + Harden CI: install the auth-bootstrap guard's Playwright from a lockfile #5576 + Harden CI: install the visual-diff bot's Node deps from a lockfile #5583 + Harden CI: clear the visual-diff bot's Playwright SSL-verification advisory #5589 Harden CI: pin Claude Code CLI, OpenClaw CLI, auth-bootstrap Playwright, visual-diff Node deps and Playwright to lockfiles.
• Harden CI: requeue reads drift-bot from the status API, so parked PRs restart #5572 Harden CI: requeue reads drift-bot from the status API (parked PRs now restart).
• Make the architecture docs true again, and generate the part that keeps drifting #5568 Architecture docs regenerated from code; the generated section now stays in sync automatically.
• Drop the theme toggle and the in-page zoom controls from the top bar #5562 Drop the theme toggle and in-page zoom controls from the top bar (simplification).
• Move the paywall forwarder out of a 47,700-line file #5585 Move paywall forwarder out of the 47,700-line entitlement file.
• Let the gate be answered for free, and forward the paywall beacons #5582 Let the gate be answered for free; forward paywall beacons.
• build(deps): bump python from 3.11-slim to 3.14-slim #5511 Bump Python base from 3.11-slim → 3.14-slim in Docker.
• chore(deps-dev): bump typescript from 5.9.3 to 7.0.2 in /clawhub-plugin #5261 Bump TypeScript 5.9.3 → 7.0.2 in clawhub-plugin.
• Harden: declare the non-security digests as non-security #5592 + Harden deps: let Dependabot's npm security fixes clear the product-record gate #5478 Harden: non-security CI digests declared as non-security; Dependabot npm fixes bypass the product-record gate.
• i18n: two translation syncs (README translations + UI strings).

Cloud (vivekchand/clawmetry-cloud)

[data unavailable — GitHub token not scoped to this repo this session]

Proxy: at least 19 PyPI releases shipped from OSS RELEASE PRs this week (v0.12.807 → v0.12.826), all of which deploy to cloud on merge. Cloud-specific PRs and issues cannot be enumerated.

---

Biggest user-visible win

#5512 Session Trail: Inputs / Decisions / Outcome for every runtime

Every session now has a three-panel narrative alongside the transcript: Inputs (what the agent was actually handed — CLAUDE.md, MCP servers, system prompt, context window), Decisions (which detector alerts fired, what the guard policy did, what mode the agent chose), and Outcome (git commits, merged PRs, test results, cost split). This works on the hosted dashboard via snapshot relay and locally via DuckDB. Schema v15 is the largest structural change since schema v10.

The secondary win this week is practical rather than new: Guard's Pause/Stop/Kill (#5529) now actually sends signals to non-OpenClaw runtimes. Every user who tried to pause a running Claude Code session from the Guard tab and saw it keep going was hitting this. Fixed.

---

What users asked for (intel)

No intel-feedback or intel-pain labelled issues this week. One external enhancement proposal:

• Feature proposal: opt-in capability-gap event export for machine-demand measurement #5412 (flyoung588) — "Feature proposal: opt-in capability-gap event export for machine-demand measurement" — asks for a machine-readable feed of the obs-gap events so teams can programmatically see where ClawMetry's coverage is thin on their runtimes. Good signal that the harness-gap system is now visible enough for users to want to consume it.

Intel coverage was lighter than usual this week — possibly because the accuracy-audit and obs-gap bots dominated the issue stream. Worth checking if user-filed issues are getting buried.

---

Backlog health (OSS only)

• Open issues: 134 total (cloud: data unavailable)
• Open PRs: 22 total, 0 drafts, 2 stale >14 days:
  • feat(channels): add Fish Audio voice/telephony channel ingest (#5045) #5055 feat(channels): add Fish Audio voice/telephony channel ingest — open 17 days
  • feat(entitlement): has_capacity_batch + has_capacity_batch_at + endpoints #5114 feat(entitlement): has_capacity_batch + has_capacity_batch_at + endpoints — open 15 days
• Oldest open issue: feat(v0.13): Native Event Stream Consumer — reference consumer for openclaw/openclaw#6467 #34 "feat(v0.13): Native Event Stream Consumer" (opened 2026-02-28, 191 days)
• Median open-issue age: ~25 days (estimated; dominated by obs-gap and accuracy-audit bots filing daily)
• Top label counts (OSS):
  • triaged: ~110 of 134
  • harness-gap / observability (obs-gap series): ~45
  • enhancement: ~55
  • bot-plan-only: ~50
  • weekly-digest: 10
  • bug: ~5

Two regressions that were open entering this week have not yet been closed:

• [regression] GET /fleet/ -- FAIL_STATUS got 302 expected 200 #5236 [regression] GET /fleet/ -- FAIL_STATUS got 302 expected 200 — opened 2026-08-26 (12 days). bot-plan-only, no PR.
• [regression] GET /healthz -- FAIL_STATUS got 404 expected 200 #5162 [regression] GET /healthz -- FAIL_STATUS got 404 expected 200 — opened 2026-08-24 (14 days). needs-info.

---

Bot productivity

• Obs-gap scout (harness-gap): ~13 new obs-gap issues filed (5 OpenClaw, 3 NemoClaw dupes, 3 OpenClaw today); 6 closed via bot-pr-opened PRs that shipped same-day ([obs-gap:openclaw] openclaw: cron quarantine on malformed legacy rows not surfaced #5546→feat(crons): surface quarantined cron rows in health-summary (#5546) #5569, [obs-gap:openclaw] openclaw: degraded-start (migration warning) Gateway state not captured #5547→feat(openclaw): surface gateway degraded-start state from migration warnings (#5547) #5561, [obs-gap:openclaw] openclaw: OOM-victim preference for local model servers not tracked #5548→feat(openclaw): surface OOM-victim events for local model servers (#5548) #5573, [obs-gap:openclaw] OpenClaw: reply recovery state across Gateway restarts not captured #5577→feat(openclaw): capture reply recovery state in assistant spans (#5577) #5590, [obs-gap:openclaw] OpenClaw: response-steering history for async tool calls not captured #5578→feat(openclaw): capture response_steered spans in the Tracing tab (#5578) #5591, [obs-gap:nemoclaw] nemoclaw: vLLM sandbox provider not resolved distinctly (generic 'inference' fallback) #5579→feat(nemoclaw): resolve vLLM sandbox provider distinctly in _sandbox_inference_configs (#5579) #5587). Scout-to-ship latency: <24h on all 6.
• Accuracy-audit bot: 7 runs this week; filed daily accuracy-audit issues; the OpenAI cost series was caught by audit then fixed by 3 PRs.
• Backlog-triage bot: ran 6× (daily); all new issues labelled triaged within hours.
• PR babysit / autofix: guard-fix PRs (fix(guard): the pid guard compared the session record's timestamp, not the process start #5543, Guard: a killed agent says Stopped, tells you how to resume it, and shows what the button did #5563) driven to green same day; c6-health issue post target corrected (fix(c6-health): post to active tracking issue #5266, not closed #4552 #5565).
• Smoke watchdog: [watchdog-blocker] smoke run 2026-08-29 — 1 new regression, cannot reach vivekchand/clawmetry-cloud to file #5335 "[watchdog-blocker] 2026-08-29 — cannot reach clawmetry-cloud to file" was open entering this week; closed this week. Cloud filing path was broken; now resolved.

---

Concerning signal

Two regression issues are 12–14 days stale with no PR.

#5236 (/fleet/ returning 302 instead of 200) and #5162 (/healthz returning 404) have been open since Aug 24–26 with bot-plan-only / needs-info status. Neither has a PR. These are route-level failures that the smoke watchdog catches on every run and every maintainer checks by hand — if they're still failing, the fleet and health endpoints are broken in the smoke environment. If they're phantom (fixed but watchdog not updated), the watchdog needs its expected-status list corrected. Either way, close or assign by Monday EOD.

---

Suggested focus next week

• [OSS] Investigate and close [regression] GET /fleet/ -- FAIL_STATUS got 302 expected 200 #5236 (/fleet/ 302) and [regression] GET /healthz -- FAIL_STATUS got 404 expected 200 #5162 (/healthz 404) — longest-standing open regressions, 12–14 days.
• [OSS] Review or close stale PRs feat(channels): add Fish Audio voice/telephony channel ingest (#5045) #5055 (Fish Audio channel) and feat(entitlement): has_capacity_batch + has_capacity_batch_at + endpoints #5114 (entitlement batch) — both 15–17 days open with no review activity.
• [OSS/cloud] Add provider-specific unit tests for OpenAI cost accounting — three correctness PRs in 24h on the same path (fix: correct OpenAI rates and inclusive prompt-cache accounting #5608, [RELEASE] fix(accuracy): proxy cost uses provider-aware accounting and reads OpenAI cached tokens #5616, [RELEASE] fix(usage): stop double-counting OpenAI cache_read in /api/token-attribution #5617) is a signal that the accounting layer has no dedicated test coverage for per-provider semantics.

---

Code activity (OSS only — cloud data unavailable)

OSS

• Commits (7 days): 49
• Lines: +831,265 −7,915 across 2,383 files (large addition count reflects the i18n / generated README sync that touched many files)
• Contributors: Vivek Chand, dependabot[bot], github-actions[bot]
• Releases: v0.12.807 → v0.12.826 (19 PyPI releases this week)

Cloud

• [data unavailable] (proxy: 19 OSS RELEASE PRs merged = 19 cloud deployments via release-on-merge)

---

Bot meta

Auto-generated 2026-09-07 07:15 UTC by the weekly digest bot. Tagged weekly-digest + triaged so triage bots leave it alone. Cross-repo aware as of 2026-05; cloud data unavailable this run (session token not scoped to vivekchand/clawmetry-cloud — fix in connector settings to restore cloud coverage next week).