Weekly digest -- 2026-08-24 (PR Trace end-to-end + cost accuracy hardened + 4 new runtimes)

[vivekchand]

Theme of the week

PR Trace shipped end-to-end. The OSS daemon can now capture the agent session behind any pull request, publish it to trace.clawmetry.com, and drop the link back into the PR — no config, no flags. Alongside that, cost accuracy got a hard reconciliation across three surfaces, verification gained a six-layer architecture with acceptance-criterion traceability, and four more runtimes (Cline, Gemini CLI, OpenHands, plus Cursor/Copilot posture) landed.

Shipped (128 OSS PRs + 37 cloud PRs merged, most auto-pin/[RELEASE] bookkeeping)

OSS (vivekchand/clawmetry)

Features — new surface area

• feat(trace): link a pull request to the agent session that produced it #5115 / [RELEASE] PR Trace, the local half (carries #5115) #5116 feat(trace): link a pull request to the agent session that produced it — PR Trace, local half
• feat(trace): publish a captured bundle to trace.clawmetry.com #5123 / [RELEASE] publish a PR trace to a public URL (carries #5123) #5125 feat(trace): publish a captured bundle to trace.clawmetry.com — publish half
• feat(trace): capture takes no arguments #5129 feat(trace): capture takes no arguments — zero-config capture
• feat(trace): automatic tracing, and a comment on the pull request #5130 / [RELEASE] PR tracing happens on its own (carries #5126, #5127, #5129, #5130) #5131 feat(trace): automatic tracing, and a comment on the pull request — auto-comment PR wiring, end of trilogy
• feat(verification): six-layer verification architecture (L0-L5) #5082 feat(verification): six-layer verification architecture (L0–L5) — the new verification spine
• feat(ci): make every acceptance criterion traceable to a test #5089 feat(ci): make every acceptance criterion traceable to a test — closes the untested-criteria class of bug (Blueprint-vs-tests one-way ratchet)
• feat(orchestration): workflows + sub-agents fully captured per session — context, replies, live status #5008 / [RELEASE] Orchestration capture: workflows + sub-agents per session (carries #5008) #5013 feat(orchestration): workflows + sub-agents fully captured per session — parent edges, prompt/reply, live status
• feat(openclaw): orchestration capture — sub-agent + flow registries with prompt/reply/status/parent #5014 / feat(nemoclaw): orchestration capture — sandboxed OpenClaw sub-agents get parent edges, prompt/reply, live status #5016 orchestration capture for OpenClaw + NemoClaw sub-agents
• feat(cloud-parity): carry orchestration fields on the /api/subagents shaper so the snapshot ships them #5032 feat(cloud-parity): carry orchestration fields on /api/subagents shaper
• Register Gemini CLI as the 24th runtime #5051 Register Gemini CLI as the 24th runtime
• Register Cline as the 25th runtime #5054 Register Cline as the 25th runtime
• Register OpenHands as the 26th runtime #5057 Register OpenHands as the 26th runtime
• Add Cursor and Copilot CLI posture providers, with a grading rule for closed-source runtimes #5063 / [RELEASE] Cursor + Copilot CLI posture providers (carries #5063) #5067 Cursor + Copilot CLI posture providers (grading rule for closed-source runtimes)
• feat(openclaw): surface gatewayLogLevel from openclaw.json (#5060) #5061 feat(openclaw): surface gatewayLogLevel from openclaw.json
• feat(entitlement): has_all_breakdown + endpoint — per-axis blocking breakdown #5134 feat(entitlement): has_all_breakdown — per-axis blocking breakdown
• feat(entitlement): has_features_bundle_batch_at + has_runtimes_bundle_batch_at + endpoints #4957 / feat(entitlement): missing_features_bundle_batch_at + missing_runtimes_bundle_batch_at + endpoints #4963 feat(entitlement): has_features/runtimes/missing_features/missing_runtimes bundle_batch endpoints
• feat(adapters/openclaw): surface per-endpoint pool health from nemoclaw model-router (#5048) #5053 feat(adapters/openclaw): per-endpoint pool health from nemoclaw model-router

Fixes — user-visible

• fix(sync): spending never mixes a live number with a stale one #5078 / fix(cost): a failed DuckDB read is not a window that cost $0.00 #5079 / fix(cost): one definition of today/week/month across every cost surface #5080 / [RELEASE] The same cost, read three ways, gave three answers (carries #5078, #5079, #5080) #5081 [RELEASE] "The same cost, read three ways, gave three answers" — one definition of today/week/month across every cost surface; failed DuckDB reads no longer report $0.00
• An expired license must not shadow a live paid cloud plan #5065 / [RELEASE] An expired license no longer shadows a live paid cloud plan (carries #5065) #5066 An expired license must not shadow a live paid cloud plan
• fix(entitlements): fail open on resolver exception even under enforce #5109 fix(entitlements): fail open on resolver exception even under enforce
• fix: clawmetry CLI is dead on Python 3.9 (PEP 604 union in a module-scope signature) #5075 / [RELEASE] CLI is dead on Python 3.9 (carries #5075) #5076 clawmetry CLI is dead on Python 3.9 (PEP 604 union in a module-scope signature) — hard-broken install for anyone still on 3.9
• fix(nemoclaw): NemoClaw was never detected on a standard install #5085 / fix(license): the pro wheel shipped without its declared dependencies #5086 / [RELEASE] NemoClaw was invisible, and the Pro pack shipped without its dependencies (carries #5085, #5086) #5091 NemoClaw was never detected on a standard install; Pro wheel shipped without its declared dependencies
• fix(tls): ship certifi so every install has a usable trust store #5030 / [RELEASE] Ship certifi so every install has a trust store (carries #5030) #5034 Ship certifi so every install has a usable trust store (silent HTTPS verify failures for interceptor pings)
• Fix "Continue with email" on the login wall — it called a cloud-only route #5002 / [RELEASE] Email sign-in works (carries #5002) #5027 "Continue with email" on the login wall — was calling a cloud-only route
• fix(desktop): verify certificates on both open pings #5025 / [RELEASE] Desktop open reports verify certificates (carries #5025) #5026 fix(desktop): verify certificates on both open pings
• fix(test): main has been red for three weeks on a guard for deleted UI #5088 fix(test): main has been red for three weeks on a guard for deleted UI — three-week undetected main-red rotted through
• fix(canary): the release canary never fired, so nothing verified a release #5095 fix(canary): the release canary never fired, so nothing verified a release
• Tell the fleet which machine each node actually is #5097 / [RELEASE] Fleet cards say which machine they are (carries #5097) #5101 Tell the fleet which machine each node actually is (fleet cards were unidentifiable)
• fix(brain): hosted Activity labelled every Claude Code run "OpenClaw · ?" #5100 / [RELEASE] Activity run labels: the hosted feed called every run OpenClaw (carries #5100) #5103 hosted Activity labelled every Claude Code run "OpenClaw · ?"
• fix(gate): every merge-blocking check must be reachable without secrets #5099 fix(gate): every merge-blocking check must be reachable without secrets
• fix(trace): strip provider identifiers before publication #5127 fix(trace): strip provider identifiers before publication
• fix(trace): refuse to install a hook the PATH binary cannot run #5126 fix(trace): refuse to install a hook the PATH binary cannot run
• Gate hooks launch via the console script, not python -m (fail-closed DoS) #5035 / [RELEASE] Gate hooks launch via the console script (carries #5035) #5037 / Claude Code gate uses the same shadow-proof launcher (consistency) #5038 / [RELEASE] Claude Code gate uses the shadow-proof launcher (carries #5038) #5040 Gate hooks launch via console script, not python -m (fail-closed DoS)
• Sign Windows builds with Azure Artifact Signing (HSM-backed, OIDC) #5062 / Refuse to ship an unsigned desktop artifact on a release tag #5072 / [RELEASE] Signed Windows downloads (carries #5062, #5072) #5073 Sign Windows builds with Azure Artifact Signing (HSM-backed, OIDC); refuse to ship unsigned desktop artifacts on a release tag
• fix(supply-chain): a Google Fonts outage is not font drift #5107 fix(supply-chain): a Google Fonts outage is not font drift
• fix(supply-chain): the OpenSSF Scorecard job has never run #5087 fix(supply-chain): the OpenSSF Scorecard job has never run
• Matrix-gap sprint: kill mid-run for 6 more runtimes + Cursor/Copilot pre-tool gates #5009 / [RELEASE] Stop/pause reaches six more runtimes; Cursor and Copilot get pre-execution gates #5031 kill-mid-run for 6 more runtimes + Cursor/Copilot pre-tool gates
• fix(sync-banner): name the runtimes we actually sync, not a hardcoded OpenClaw #5010 sync-banner now names the runtimes it actually syncs
• Sessions tab: scope /api/transcripts to the active runtime so per-runtime lists aren't starved by the global 50-row cap #5015 Sessions tab: scope /api/transcripts to active runtime so lists aren't starved by 50-row cap
• [RELEASE] Test fixtures no longer leak into the live DuckDB; DuckDB-backed alert rules render as real rules #5006 test fixtures no longer leak into live DuckDB; DuckDB-backed alert rules render as real rules
• [RELEASE] Approvals empty states respect the runtime filter (no more OpenClaw copy under a Claude Code filter) #5007 Approvals empty states respect runtime filter
• Stop paywalling Pro accounts at boot, and stop the daemon proxy dropping positional args #5003 / [RELEASE] Pro accounts stop seeing the upgrade modal at boot (carries #5003) #5004 Stop paywalling Pro accounts at boot; daemon proxy no longer drops positional args
• fix(canary): retry pip install in matrix jobs for CDN propagation lag #5133 fix(canary): retry pip install in matrix jobs for CDN propagation lag
• fix(verification): register the AC gate's guard files (follow-up to #5089) #5102 fix(verification): register the AC gate's guard files
• fix(desktop): uninstall removes every ClawMetry trace; fresh install always asks to sign in #4707 fix(desktop): uninstall removes every ClawMetry trace; fresh install always asks to sign in
• fix(desktop): supervisor died mid-upgrade and froze every tab for 6h39m #4976 / [RELEASE] desktop supervisor resilience + backend-outage recovery UI (carries #4976) #4979 supervisor died mid-upgrade and froze every tab for 6h39m

Chore / refactor / docs

• ~35 chore: auto-pin clawmetry==0.12.7XX (v0.12.725 → v0.12.763)
• ~30 [RELEASE] umbrella tags
• docs(flywheel): drift-bot cannot be made green after the merge #5122 / docs(flywheel): red pipeline is work to start, not a status to report (§4a) #5090 / docs: strip the README to what ClawMetry does; stop the CLI implying OpenClaw-only #4987 / docs(readme): order the runtime list by recognition, matching the site #5069 / docs(trace): repoint PRD refs left dangling by the monetization split #5117 docs (flywheel, README, PRD)
• test(trace): make the self-containment guard assert what it claims #5120 test(trace): make the self-containment guard assert what it claims
• fix(ci): make drift-bot a required check, and give the list one source of truth #5124 fix(ci): make drift-bot a required check
• fix(ci): the lint step never linted, and the cloud pin never follows a release #5084 fix(ci): the lint step never linted, and the cloud pin never follows a release
• fix(packaging): stop shipping the repo test suite, which could never pass there #5098 fix(packaging): stop shipping the repo test suite, which could never pass there

Cloud (vivekchand/clawmetry-cloud)

Features

• vivekchand/clawmetry-cloud#2092 feat(trace): serve PR traces at trace.clawmetry.com — the cloud half of PR Trace
• vivekchand/clawmetry-cloud#2067 feat(cloud): serve orchestration capture from the snapshot (session tree + bulk summary)
• vivekchand/clawmetry-cloud#2061 feat(desktop): open telemetry sink, admin panel, and per-account desktop list

Fixes

• vivekchand/clawmetry-cloud#2073 Fix the five ways a paying account can lose its plan
• vivekchand/clawmetry-cloud#2074 Describe a returning registration the way the heartbeat will
• vivekchand/clawmetry-cloud#2075 A completed checkout reactivates the account it paid for
• vivekchand/clawmetry-cloud#2076 Never send trial email to a paying customer
• vivekchand/clawmetry-cloud#2089 fix(silence): stop paging every cloud user when a laptop sleeps
• vivekchand/clawmetry-cloud#2093 fix(trace): the host shim must not rewrite /api/ paths
• vivekchand/clawmetry-cloud#2072 fix(email): every urllib sender identifies itself, or Cloudflare 403s it
• vivekchand/clawmetry-cloud#2083 fix(ci): retry the OSS pin install past PyPI's cross-CDN index race
• vivekchand/clawmetry-cloud#2084 Show which machine a fleet card means; name the machines on runtime cards
• vivekchand/clawmetry-cloud#2056 Approve/Deny-all in the fleet inbox; evict requests no longer pending
• vivekchand/clawmetry-cloud#2059 Stop/Pause buttons for 10 more runtimes (matrix-gap mirror)

Chore / refactor / docs

• ~15 chore: auto-pin clawmetry==0.12.7XX (mirrors OSS release chain)
• vivekchand/clawmetry-cloud#2088 docs: the PR Trace open-source grant programme (private half)
• vivekchand/clawmetry-cloud#2091 docs(flywheel): §2b — Drift Bot is a merge blocker, not an advisory
• vivekchand/clawmetry-cloud#2060 docs(flywheel): rebase on fresh origin/main before every PR

Biggest user-visible win

#5131 [RELEASE] PR tracing happens on its own (carries #5126, #5127, #5129, #5130) — combined with the cloud-side vivekchand/clawmetry-cloud#2092 feat(trace): serve PR traces at trace.clawmetry.com. The user experience change: install ClawMetry, run any agent (Claude Code / OpenClaw / Cline / …), and every pull request that agent opens now gets a public trace URL commented on it, automatically, no flags. The full session — prompts, tool calls, sub-agents, cost, quality grade — is browsable at trace.clawmetry.com/<id>. This is the first user-facing product surface built on top of the observability substrate; it turns "your agent is observed" from "look at the dashboard on localhost:8900" into "here's the trace, shareable, on every PR." The open-source grant programme (vivekchand/clawmetry-cloud#2088) starts the go-to-market motion on the same feature.

What users asked for (intel)

No new intel-* issues filed in either repo this week. The 105+ existing intel-* items in cloud (from 2026-05 onward) are the standing backlog; the closest recent proxy for demand is #4707 (fresh-install desktop uninstall complaint, shipped) and the P0 canary-failure heartbeats (#5106/#5108/#5119) which are the smoke watchdog acting as a synthetic user. The Fish Audio channel PR (#5055) is the only open one carrying a runtime:openclaw obs-gap signal from last week.

Backlog health (across both repos)

• Open issues: 203 total (84 OSS + 119 cloud) — vs 82 OSS at end of last week's digest, so essentially flat OSS with +2. Cloud backlog unchanged.
• Open PRs: 6 total (5 OSS + 1 cloud; 1 draft — OSS fix(setup.py): pin cffi<2 to prevent py3.9 SIGSEGV on --help (#5108) #5144 pin cffi<2 to prevent py3.9 SIGSEGV). No PR stale >14 days — oldest open PR is OSS feat(channels): add Fish Audio voice/telephony channel ingest (#5045) #5055 (3 days).
• Oldest open OSS 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, ~178 days)
• Oldest open cloud issue: vivekchand/clawmetry-cloud#2 feat: Remote approval gates (opened 2026-03-03, ~174 days)
• Median open-issue age: ~90 days (cross-repo, back-of-envelope from date spread)
• Top OSS label counts: triaged (84), enhancement (33), bot-plan-only (~28), harness-gap+observability+runtime:openclaw obs-gap cluster (~12), bug (3), weekly-digest (8)
• Top cloud label counts: enhancement (68), bug (~4), P0 (~4), P1 (~10)

Bot productivity (cross-repo aware)

• Triage bot: essentially every closed/opened issue this week (~27) carries triaged. 4 obs-gap duplicate closures in the past 24h ([obs-gap:openclaw] OpenClaw: no visibility into secret egress host-binding enforcement/denials #5110–[obs-gap:nemoclaw] NemoClaw: model-router endpoint-pool NVIDIA-only compliance verdict not surfaced #5113), 4 more duplicate closures earlier this week ([obs-gap:openclaw] OpenClaw: secret egress host-binding fail-closed events not captured #4953–[obs-gap:openclaw] OpenClaw: Fish Audio speech/telephony sessions not tracked #4956, [obs-gap:openclaw] OpenClaw: secret egress host-binding fail-closed events uncaptured #5059). Two completed closures ([obs-gap:openclaw] OpenClaw: configured file-log level/severity filter not surfaced #5060, [obs-gap:nemoclaw] nemoclaw: model-router /health endpoint body (per-endpoint pool health) discarded #5048) both via bot-pr-opened.
• Autofix bot (bot-pr-opened / bot-plan-only, past 7d): 3 bot-pr-opened labels seen on this week's new issues ([heartbeat] The published product is failing verification #5108, [obs-gap:openclaw] OpenClaw: Fish Audio voice/telephony surface not in known channel set #5045); ~5 bot-plan-only ([P0] Published release 0.12.759 fails canary verification #5106, [obs-gap:openclaw] OpenClaw: trusted-proxy device auto-pairing decisions not captured #5047, [obs-gap:openclaw] OpenClaw: cloud-worker workspace-conflict transcript events not ingested #5046, [accuracy-audit] Fixer run 2026-08-20 UTC #5019, plus older). Cloud split: not detectable from search (labels absent).
• Intel scout: 0 new intel-* issues filed this week in either repo — a notable dip vs prior weeks. Worth investigating whether the scavenger is running.
• Smoke watchdog: Filed 3 P0 heartbeats in OSS this week — [P0] Published release 0.12.759 fails canary verification #5106 (0.12.759 canary fail), [heartbeat] The published product is failing verification #5108 (heartbeat, published product failing), [P0] Published release 0.12.760 fails canary verification #5119 (0.12.760 canary fail). fix(canary): retry pip install in matrix jobs for CDN propagation lag #5133 partially addresses the CDN-propagation flake; [heartbeat] The published product is failing verification #5108 was tagged bot-pr-opened. All three still open. No cloud regression heartbeats detected this week.
• PR babysit: Not surfaced as its own signal in this data pull — release loop merged 128 OSS + 37 cloud PRs, so the babysitter kept up.
• Weekly digest bot: filing this one.

Concerning signal

The release canary is broken and has been for at least two versions in a row. #5106 flagged 0.12.759 as failing verification (Aug 22), #5108 flagged the published product failing verification (Aug 23), and #5119 flagged 0.12.760 also failing (Aug 23). All three P0 heartbeats are still open. #5095 fixed one part ("the release canary never fired"), #5133 fixed another ("retry the pip install for CDN propagation lag"), but the P0s themselves haven't been closed — meaning either the fixes didn't fully land the guarantee, or the heartbeat/canary logic itself is filing false positives. Either way, the release-verification pipe is currently unreliable at exactly the moment PR Trace is going live.

Next step: verify a fresh pip install clawmetry==0.12.763 in a clean matrix job (all three OSes × Py 3.9/3.11), then either close #5106/#5108/#5119 with the ratchet-forward evidence or reopen the root-cause investigation. Do this before announcing PR Trace externally — a broken canary is a public embarrassment waiting to happen if a user hits the same failure the bot did.

Suggested focus next week

• [OSS] Close the three open P0 canary failures ([P0] Published release 0.12.759 fails canary verification #5106, [heartbeat] The published product is failing verification #5108, [P0] Published release 0.12.760 fails canary verification #5119) — verify a clean install of the latest release across the matrix or find the real root cause.
• [OSS + cloud] Polish PR Trace: finish the open follow-ups (Send the Quality grade to the cloud, and stop the cloud inventing an empty one #5121 quality-grade wire-through, feat(entitlement): has_capacity_batch + has_capacity_batch_at + endpoints #5114 has_capacity_batch entitlement, fix(setup.py): pin cffi<2 to prevent py3.9 SIGSEGV on --help (#5108) #5144 cffi pin) and start dogfooding the "grant programme" outreach (cloud#2088).
• [OSS] Ship or reject the Fish Audio channel PR (feat(channels): add Fish Audio voice/telephony channel ingest (#5045) #5055) — it's the only 3-day-old open PR carrying a runtime:openclaw obs-gap; close the loop before the smoke scanner refiles.
• [both] Restart the intel scout — 0 intel-* issues this week is the first time that's happened. If it's silently broken, the demand-signal pipeline is off.

Code activity (across both repos)

OSS

• Commits (on main): 49
• Lines: roughly +11k / −2k across the non-vendor commits (a single 1820-file / +702k vendored import commit is excluded as noise; noted so no one is misled by a raw shortstat)
• Contributors: Vivek Chand + github-actions[bot]
• Releases: v0.12.725 → v0.12.763 — 38 patch releases in 7 days

Cloud

• Commits / lines / contributors: unavailable this week — the digest environment has no gh CLI to clone the private repo. (Filed via the search API instead.) Recommend adding a git clone step or piping the daemon's cloud-repo mirror to /tmp so the next digest can quote actual diff stats.
• Releases (proxy): 37 PRs merged to main in the past 7 days (~15 auto-pin mirrors + ~22 real product PRs, weighted toward payment-flow hardening and PR Trace serving).

Bot meta

Auto-generated 2026-08-24 07:15 UTC by the weekly digest bot. Tagged triaged so other bots leave it alone. Cross-repo aware as of 2026-05.