review-runs-tracking: 2026-04

[worktrunk-bot]

Monthly tracking issue for below-threshold findings. Each run appends findings as a comment. Future runs read these to build cumulative evidence.

Do not close manually — a new issue is created each month.

[worktrunk-bot]

Run 23890775590 — 2026-04-02

Runs analyzed

Run ID	Workflow	Conclusion	Notes
23887729285	tend-nightly	success	Created PRs #1886, #1887
23889208419	tend-triage	success	Triaged issue #1888
23888194806	tend-review	success	Reviewed PR #1887 (self-authored)
23888137508	tend-review	success	Reviewed PR #1886 (self-authored)
23886186498	tend-review	success	Approved dependabot PR #1885
23879840745	tend-review	success	Re-reviewed PR #1820
23879423511	tend-review	success	First review of PR #1820
23869816323	tend-review	failure	Infra: PR #1879 merge ref gone (already merged)
23876334883	tend-notifications	failure	Infra: no session produced

Additionally: ~20 skipped tend-review/mention runs, 9 skipped tend-ci-fix runs, multiple successful tend-notifications runs — all normal operation.

---

Finding 1: Self-approval attempt on own PR

• Evidence level: Medium
• Occurrences this run: 1 (out of 4 self-authored PR reviews — 3 handled correctly)
• Classification: Stochastic
• Run ID: 23879423511
• Workflow: https://github.com/max-sixty/worktrunk/actions/runs/23879423511
• Session: cc387489-ab90-4f92-935b-67c35637469f.jsonl
• Detail: Bot attempted gh pr review 1820 --approve on its own PR despite earlier reasoning correctly identifying PR_AUTHOR=worktrunk-bot == BOT_LOGIN. GitHub rejected with "Can not approve your own pull request". Three other sessions (23888194806, 23888137508, 23879840745) correctly skipped self-approval. This is stochastic — the model has the knowledge but inconsistently applies it.
• Proposed fix: Add explicit guard to review skill: "If PR author == bot login, skip approval"
• Gate assessment: Medium evidence, 1 occurrence, stochastic → needs 5+ occurrences to act. Below threshold.

Finding 2: CI polling filter inconsistency

• Evidence level: Low
• Occurrences this run: 2 variants observed across 5 sessions
• Classification: Stochastic
• Detail: Some sessions use grep -v "$GITHUB_WORKFLOW" (env var) while others use grep -v "tend-review" (hardcoded). The env var approach is correct per the running-in-ci skill. The hardcoded approach works but is fragile if workflow names change. Neither variant has caused an actual failure.
• Gate assessment: Low evidence, no actual failure observed. Below threshold.

Finding 3: Triage skipped reproduction step

• Evidence level: Low
• Occurrences this run: 1
• Classification: Stochastic
• Run ID: 23889208419
• Session: d544ecc6-8bb7-4bfe-85a6-82bef8bec237.jsonl
• Detail: Bot triaging issue wt step prune seems very slow #1888 (slow wt step prune) self-exempted from the reproduction step with reasoning "no meaningful failing test can be written" rather than following the skill's "could not reproduce" template path. The analysis itself was accurate — correctly identified sequential vs parallel integration checks as root cause.
• Gate assessment: Low evidence, 1 occurrence, stochastic. The bot's analysis quality was good despite the process deviation. Below threshold.

Finding 4: Triage didn't apply labels

• Evidence level: Low
• Occurrences this run: 1
• Classification: Stochastic
• Run ID: 23889208419
• Detail: Issue wt step prune seems very slow #1888 was left unlabeled after triage. The running-tend skill defines automated-fix and nightly-cleanup labels but has no triage-specific label guidance.
• Gate assessment: Low evidence, 1 occurrence. No applicable label was defined anyway. Below threshold.

---

Run 23939231746 — 2026-04-03T08:08:28Z

Runs analyzed

Run ID	Workflow	Conclusion	PR/Issue	Notes
23936957763	tend-nightly	success	PR #1898, #1883	Created docstring fix PR, closed transient outage issue
23937183081	tend-review	success	PR #1899	Approved then dismissed on codecov failure; thorough coverage investigation
23937152856	tend-review	success	PR #1898	Self-authored PR, correctly skipped approval
23934423318	tend-review	success	PR #1893	Already-reviewed commit, correctly exited
23929311196	tend-review	success	PR #1890	Correctly identified codecov false positive from renamed code
23932550716	tend-review	success	PR #1897	Dependabot bump, approved correctly
23928248179	tend-review	failure	PR #1894	No artifact — merge ref already gone (PR merged before review ran)
23934126683	tend-mention	success	PR #1890	Answered reviewer question about parallelization scope
23934095404	tend-mention	success	PR #1890	Implemented reviewer-suggested single-pass refactor
23933934456	tend-mention	success	Issue #1891	Confirmed migration handles both field renames
23933874463	tend-mention	success	Issue #1891	Aligned deprecation table in release notes

Additionally: ~20 skipped tend-mention runs, 6 skipped tend-ci-fix runs, 14 tend-notifications runs (all success) — normal operation.

---

Finding 1: Multi-commit fix cycles when pushing code changes

• Evidence level: Medium
• Occurrences this run: 1 (session 84611ba1 on PR perf: parallelize integration checks in step prune #1890)
• Historical: 0 prior occurrences in tracking
• Classification: Stochastic
• Run ID: 23934095404
• Workflow: https://github.com/max-sixty/worktrunk/actions/runs/23934095404
• Session: 8bc3d762-4719-4c48-b99e-ce7c2a30865e.jsonl
• Detail: Bot implemented reviewer-suggested refactor on PR perf: parallelize integration checks in step prune #1890, ran cargo build, cargo test, cargo clippy, and cargo fmt --check locally. Pushed, but CI failed. Required two additional fix commits: one for .expect() → .context() (CLAUDE.md rule violation) and one for cargo fmt formatting. The bot ran local checks but either had version discrepancies or missed applying the format fix before committing. Final result: 3 commits for what should have been 1.
• Proposed fix: Add guidance to running-tend skill: "Run pre-commit run --all-files or cargo run -- hook pre-merge --yes before pushing code changes."
• Gate assessment: Medium evidence, 1 occurrence, stochastic → needs 5+ to act. Below threshold.

Finding 2: CI polling filter inconsistency (carry-forward)

• Evidence level: Low
• Occurrences: 2 in previous run, 2 more variants this run (sessions dc6c25d0, f86cd859 use $GITHUB_WORKFLOW; session 3773fff5 uses "tend-review")
• Classification: Stochastic
• Detail: Ongoing inconsistency between grep -v "$GITHUB_WORKFLOW" and grep -v "tend-review". No actual failure caused. The env var approach is correct per running-in-ci skill.
• Gate assessment: Low evidence, no failures observed. Below threshold. Carrying forward.

Quality notes (not findings)

• Codecov handling: Excellent. Session f86cd859 (PR fix: warn when deprecated config key appears in wrong config file #1899) properly dismissed its own approval after codecov failed and investigated coverage gaps. Session af509227 (PR perf: parallelize integration checks in step prune #1890) correctly identified a codecov false positive from renamed files and verified by running coverage on both branches.
• Self-authored PR handling: All sessions correctly identified self-authored PRs and skipped approval. No repeat of the self-approval attempt from the previous run.
• Mention handling: Clean and responsive — answered reviewer questions with code references, implemented suggestions, confirmed migration behavior with evidence.

---

Run 23974701581 — 2026-04-04T07:57:51Z

Runs analyzed

Run ID	Workflow	Conclusion	PR/Issue	Notes
23973637398	tend-review	success	PR #1908	Approved bugfix for primary_remote() parsing
23972923281	tend-review	success	PR #1907	COMMENT — flagged doc sync failure
23971958611	tend-review	success	PR #1906	Approved docs-only change
23971578194	tend-review	success	PR #1904	Incremental re-review after new test commit, approved
23971309740	tend-review	success	PR #1905	Approved copy thread pool fix
23971100076	tend-review	success	PR #1904	First review, approved; investigated codecov false positive
23973377705	tend-nightly	success	Closed #1790	Verified SSH alias fix shipped in v0.34.0
23963532092	tend-triage	success	Issue #1903	Triaged Claude Code path question
23964236527	tend-mention	success	Issue #1903	Directed Windows bug to Claude Code repo
23964420037	tend-mention	success	Issue #1903	Correctly stayed silent (human-to-human)
23968974405	tend-mention	success	Issue #1903	Correctly stayed silent (user-maintainer exchange)
23968997439	tend-mention	success	Issue #1903	Correctly stayed silent (duplicate trigger)

Additionally: ~10 skipped tend-mention runs, 4 skipped tend-ci-fix runs, ~20 tend-notifications runs (all success) — normal operation.

---

No new findings above threshold

All sessions performed well. No behavioral issues detected.

Quality notes (not findings)

• Review accuracy: All 6 reviews well-calibrated. PRs fix: spawn background pipelines per hook-type #1904 and fix: restore dedicated copy thread pool to avoid CPU saturation #1905 merged without reverts or human overrides. PR docs: add cmux workspace integration recipe #1907 correctly got COMMENT (not approve) for failing doc sync test. No human reviewers added corrections after bot reviews.
• Nightly accuracy: Closing SSH host aliases in remote URLs break pr:/mr: resolution #1790 was justified — verified via code search that both failure stages (SSH alias resolution and forge config) were addressed in v0.34.0.
• Mention handling: Excellent judgment on issue Claude code projects #1903 — responded when addressed, stayed silent during human-to-human exchanges, handled duplicate trigger without double-posting.
• CI polling filter inconsistency (carry-forward): Still observed — some sessions use $GITHUB_WORKFLOW env var, others hardcode workflow name. No failures caused. Cumulative: 6+ observations across 3 runs, all low severity. Remains below threshold (no actual failures).
• Self-approval (carry-forward from run 23890775590): Not observed this run. 0/6 review sessions attempted self-approval. Pattern appears resolved (1 occurrence total, stochastic).

---

Run 24007780000-equiv — 2026-04-05T08:00:00Z (via /tend-ci-runner:review-runs)

Runs analyzed

Run ID	Workflow	Conclusion	PR/Issue	Notes
23993905364	tend-review	success	PR #1913	Approved copy-pool parallelism fix, polled CI
23993802081	tend-review	success	PR #1910	Already reviewed; HEAD unchanged; silent exit
23988290091	tend-review	success	PR #1880 (list-ignore)	Already approved; trivial new commits (merge + .DS_Store); silent exit
23987973227	tend-review	success	PR #1911	Self-authored bot PR (Windows stack fix); stayed silent, monitored CI
23987366243	tend-review	success	PR #1910	Approved pipeline-routing refactor
23986753886	tend-review	success	PR #1909	Self-authored AUR bump; detected flaky macOS PTY test, filed flake, reran CI
23987900176	tend-ci-fix	success	→ PR #1911	Correctly diagnosed Windows stack overflow in test_copy_ignored_many_directories_no_emfile, created fix PR, CI green
23996135379	tend-nightly	success	—	Reviewed 5 recent commits, open-issue sweep (no closures), rolling file survey; clean

Additionally: ~20 skipped tend-mention/tend-ci-fix runs, 22 tend-notifications runs (all success), one cancelled tend-review — normal operation.

---

No findings above threshold

All sessions performed well. Specific quality observations:

• Self-authored PR handling: 3 self-authored bot PRs (fix: bump AUR deploy action to v4.1.2 #1909, fix: increase copy pool stack size to avoid Windows overflow #1911, and fix: route all background hooks through pipeline-aware API #1910 was human-authored but bot did multiple reviews). All correctly skipped approval and monitored CI instead. No self-approval attempts — pattern from run 23890775590 remains resolved.
• Flaky test handling (run 23986753886): When an unrelated macOS switch_picker PTY test failed on the AUR bump PR, bot correctly identified it as flaky, filed a flake report, and reran CI. Clean triage.
• CI-fix accuracy (run 23987900176): Root-caused Windows stack overflow to rayon COPY_POOL inheriting Windows's ~2 MiB default thread stack; fix (8 MiB stack_size) was minimal and correct. PR merged same day. Later PR fix: run copy-ignored outer loop on dedicated copy pool #1913 addresses a separate parallelism issue in the same module, not a revert.
• Nightly subagent verification (run 23996135379): The general-purpose survey subagent reported commands_already_filtered in src/commands/command_approval.rs as an unused parameter ("always passed as false"). The main agent verified and caught the false positive — hook_commands.rs:348 passes true. Main agent also correctly dismissed the subagent's other claim about post-create being a misleading alias on PreStart (it is an intentional legacy alias, exercised in tests). Good self-correction loop.
• Nightly doc-drift finding (run 23996135379): Main agent found a real minor doc drift at src/output/handlers.rs:842 ("All hooks use remove_ctx for spawning" — post-switch pipelines now spawn via switch_ctx). Correctly deferred action because open PR refactor: always use pipeline runner for background hooks #1912 removes the dual-path hook system entirely. Good judgment — avoided merge-conflict churn.

Below-threshold observations (carrying forward)

Subagent false-positive rate on "dead code" claims

• Evidence level: Low
• Occurrences this run: 1 (nightly survey subagent)
• Historical: 0 prior
• Classification: Stochastic — subagents do lightweight surveys and miss callers in other files
• Run ID: 23996135379
• Session: subagents/agent-a586f51d0be2ecf58.jsonl
• Detail: Survey subagent asserted commands_already_filtered was always passed false without grep-checking all call sites; the true caller is in a different file. Main agent's verification loop caught it. Not a problem worth acting on yet — the verification layer worked as intended. Worth watching if subagent false positives recur frequently enough to waste main-agent tokens on verification.
• Gate assessment: Low evidence, 1 occurrence, no bad outcome. Below threshold.

CI polling filter inconsistency (carry-forward)

• Evidence level: Low
• Occurrences: Still observed this run across review sessions (mix of $GITHUB_WORKFLOW and hardcoded tend-review). Cumulative: 8+ observations across 4 runs, zero failures caused.
• Gate assessment: Low evidence, no failures. Below threshold. Carry forward.

---

Run 24017668068-equiv — 2026-04-06T08:15:00Z

Runs analyzed

Run ID	Workflow	Conclusion	PR/Issue	Notes
24017668068	tend-review	success	PR #1922	Approved minijinja 2.19.0 bump
24017663876	tend-review	success	PR #1920	Approved patch dependency group bump
24017661321	tend-review	success	PR #1920	Same PR, different trigger
24017631507	tend-review	success	PR #1743	Follow-up review, approved after hidden rename
24015713947	tend-review	success	PR #1918	Approved codecov-action migration
24015384984	tend-review	success	PR #1743	Review after experimental labeling
24015367116	tend-review	success	PR #1918	Approved codecov-action CI change
24015269514	tend-review	success	PR #1917	Approved --shortstat perf change (nit: stale comment)
24015022758	tend-review	success	PR #1743	COMMENT — flagged missing example config
24014748717	tend-review	success	PR #1916	Approved renice copy-ignored
24014707618	tend-review	failure	PR #1916	Infra: merge ref gone (PR merged before checkout)
24014095434	tend-review	success	PR #1743	Incremental review, approved (example config addressed)
24013761335	tend-review	success	PR #1743	First review — flagged example config gap
24015269514	tend-review (cancelled)	cancelled	—	Superseded by concurrency group
24022213281	tend-nightly	success	—	Clean sweep: 5 commits reviewed, 24 issues checked, 3 files surveyed
24015406804	tend-mention	success	PR #1743	Executed hidden rename per maintainer directive
24014878756	tend-mention	success	PR #1743	Duplicate trigger, correctly stayed silent
24014841545	tend-mention	success	PR #1743	Added experimental labels per maintainer request
24013896250	tend-mention	success	PR #1743	Human-to-human conversation, correctly stayed silent
24013889975	tend-mention	success	PR #1743	Updated example config per author request

Additionally: ~15 skipped tend-mention runs, 7 skipped tend-ci-fix runs, ~20 tend-notifications runs (all success) — normal operation.

---

No findings above threshold

All sessions performed well. Reviews were accurate — dependency bumps received genuine analysis of upstream changelogs mapped to actual codebase usage. Mention handling was disciplined: responded when addressed, stayed silent for human-to-human exchanges, avoided duplicate responses.

Quality notes

• PR Add list ignore support for worktree paths and branch names #1743 multi-round workflow: Excellent example of the full review + mention cycle working correctly. Bot posted initial COMMENT (flagged missing example config), author addressed it, bot approved. Then maintainer directed experimental labeling and a naming rename (ignore → hidden) via mentions — bot executed both correctly with full test suite verification.
• Dependency reviews: All 3 Dependabot PRs (ci: replace deprecated test-results-action with codecov-action #1918, chore: bump the patch group across 1 directory with 6 updates #1920, chore: bump minijinja from 2.18.0 to 2.19.0 #1922) received meaningful analysis rather than rubber-stamp approvals. PR chore: bump minijinja from 2.18.0 to 2.19.0 #1922 traced minijinja's SemiStrict usage; PR chore: bump the patch group across 1 directory with 6 updates #1920 called out the insta semver revert and toml_edit winnow transition.
• Nightly run: Conservative and accurate. Subagents were verbose but the main agent correctly verified their findings.

Below-threshold observations

Inline review comment 422 failures (carry-forward)

• Evidence level: Low
• Occurrences this run: 2 (runs 24013761335, 24015269514)
• Historical: Seen in prior runs but not formally tracked
• Classification: Structural — GitHub API rejects inline comments on lines outside the diff every time
• Detail: Bot attempts gh api repos/.../pulls/.../reviews with inline comments anchored to lines not in the diff. Gets HTTP 422 "Path could not be resolved". Recovers correctly by including the comment in the review body text. No user-visible harm — just a wasted API call and minor latency.
• Gate assessment: Structural, but recovery is also structural (always works). Impact is a single failed API call per occurrence. Below threshold — the outcome is correct and the cost is negligible. Carry forward to see if it becomes more frequent or causes actual problems.

---

Run 24068310801 — 2026-04-07T06:46:49Z

Runs analyzed

Run ID	Workflow	Conclusion	PR/Issue	Notes
24064025285	tend-review	failure	PR #1934	Approved, then hit API limit during CI polling
24064145051	tend-review	failure	PR #1966	Approved, then hit API limit during CI polling
24064232297	tend-review	failure	PR #1971	Approved, then hit API limit during CI polling
24064318131	tend-review	failure	PR #1972	Approved, then hit API limit during CI polling
24060817593	tend-review	failure	—	No artifacts (likely same budget issue)
24064603836	tend-notifications	failure	—	API limit hit on first turn, zero work done
24065812958	tend-notifications	failure	—	API limit hit on first turn, zero work done
24066791341	tend-mention	failure	Issue #1973	API limit hit, couldn't respond to maintainer
24068310801	tend-nightly	success	—	Clean sweep: 40 commits reviewed, 22 issues checked, 10 files surveyed

Additionally: ~20 successful tend-review runs, ~15 successful tend-notifications runs, 3 tend-triage runs (2 success, 1 skipped), multiple skipped tend-ci-fix and tend-mention runs — normal operation.

---

Finding 1: CI polling exhausting API budget — cascading failures

• Evidence level: High
• Occurrences this run: 7 (4 review + 2 notifications + 1 mention)
• Historical: 0 prior occurrences of cascading budget exhaustion
• Classification: Structural — when multiple PRs arrive simultaneously, each review session polls CI in a 10×60s loop, consuming tokens. Budget exhausts and cascades to all subsequent runs.
• Detail: All 4 review runs completed their actual reviews successfully (PRs feat: per-command log files for background hooks #1934, perf: cache integration target and reason in RepoCache #1966, refactor: consolidate test repo creation into testing::TestRepo #1971, test: convert config state spot-checks to snapshot tests #1972 were all approved). The failures occurred during post-approval CI polling. The cascading effect then caused 3 additional runs to fail immediately with "You've hit your limit - resets 6am (UTC)" before doing any work. The failed mention (issue Bot temporarily unavailable #1973, maintainer asking to restart jobs) was handled later after the limit reset.
• Proposed fix: This is an infrastructure issue — the CI polling loop is defined by the tend plugin, not repo-local skills. Possible mitigations: (1) reduce polling iterations in tend plugin, (2) increase API budget, (3) have the workflow handle CI polling outside the Claude session. No repo-local fix available.
• Gate assessment: High evidence, structural, but not actionable via repo-local files. Recording for visibility. Infrastructure fix needed upstream in the tend plugin.

Finding 2: Nightly subagent false-positive rate (carry-forward)

• Evidence level: Low
• Occurrences this run: 1 (nightly survey subagents reported ~30 findings, ~70-80% were false positives)
• Historical: 1 prior occurrence (run 23996135379, 2026-04-05)
• Classification: Stochastic — subagents make wrong claims about dead code, unused parameters, etc.
• Run ID: 24068310801
• Session: subagents/agent-a771f0e5e86947273.jsonl, subagents/agent-ab981f174e53da1fe.jsonl
• Detail: Both survey subagents had high false-positive rates. Examples: claimed parse_pager_value was unused (used in 3 files), claimed duplicate priority values in columns.rs (they were distinct), claimed semaphore deadlock in tasks.rs (RAII guard is correct). Main bot correctly verified and dismissed false positives. The verification loop works but wastes tokens.
• Gate assessment: Low evidence, 2 occurrences, stochastic, recovery works. Below threshold. Carrying forward.

Finding 3: CI polling filter inconsistency (carry-forward)

• Evidence level: Low
• Occurrences: Cumulative 10+ across 5 runs, zero failures caused
• Classification: Stochastic
• Detail: Some review sessions use grep -v "$GITHUB_WORKFLOW" (correct, per running-in-ci skill) while others hardcode grep -v "tend-review". No actual failures observed.
• Gate assessment: Low evidence, no failures. Below threshold. Carrying forward.

Quality notes

• Review accuracy: All 4 failed review runs completed their actual reviews correctly before hitting the limit. All 4 PRs received valid approvals.
• Nightly run: Solid. 40 commits reviewed, correctly assessed as clean. Subagent false positives were caught by verification. No unnecessary PRs created.
• Bot PR outcomes: In the past 12h, 3 bot PRs merged (fix: suppress worktree-path hint when project-specific config exists #1941, fix: preserve file permissions when copying ignored files #1937, refactor: simplify PR lookup to single remote with better 404 hint #1927), 1 closed (fix: exclude docs.anthropic.com from lychee link checker #1938). Good track record.

---

Run 24120674053 — 2026-04-08T06:07:40Z

Runs analyzed

Run ID	Workflow	Conclusion	PR/Issue	Notes
24122006261	tend-nightly	success	—	Clean sweep: 25 commits reviewed, 3 issues checked, no PRs created
24118523944	tend-review	success	PR #1743	Re-review after new commits, approved
24118103680	tend-review	success	PR #2001	Approved, CI green
24116258141	tend-review	success	PR #1999	Dependabot bump, approved appropriately
24115173347	tend-review	success	PR #1998	Thorough refactoring review, approved
24117177368	tend-review	success	PR #2001	Initial review + codecov analysis (SpawnFailed arm)
24107374006	tend-review	success	PR #1986	Posted COMMENT instead of APPROVE (see below)
24118013462	tend-mention	success	PR #1743	Substantial code changes (removed --hidden flag)
24105961390	tend-review	failure	—	Infrastructure: no Claude session created

Additionally: ~15 successful tend-review runs, ~20 tend-notifications (all successful), tend-ci-fix (all skipped), 1 tend-triage (skipped).

---

No findings above threshold

All sessions performed well overall. Reviews were accurate and thorough — code was read in context, CI was monitored, codecov failures were properly investigated. The nightly run correctly assessed all 25 commits as clean and avoided creating unnecessary PRs. The mention run handled substantial code changes (removing --hidden CLI flag) correctly with proper testing and CI verification.

Quality notes

• PR fix: JSON output follow-ups from --format=json rollout #2001 codecov handling: Excellent. Bot investigated SpawnFailed arm coverage, attempted to write a test, correctly identified it was untestable (shell wraps the command → exit code 127 instead of SpawnFailed), and posted a clear analysis as a COMMENT review.
• Nightly subagents: All 4 performed correctly. TestRepo subagent correctly identified the src/testing/ module is intentionally in library code (maintainer decision, not a CLAUDE.md violation).
• PR Add list ignore support for worktree paths and branch names #1743 mention: Bot handled a complex multi-file refactoring (removing --hidden flag, updating all call sites, tests, docs, snapshots) with only one formatting fixup commit needed.

Below-threshold observations

COMMENT instead of APPROVE for pre-existing issue

• Evidence level: Medium
• Occurrences this run: 1
• Run ID: 24107374006
• Workflow: https://github.com/max-sixty/worktrunk/actions/runs/24107374006
• Classification: Stochastic — the bot made a judgment call; different runs might handle this differently
• Detail: On PR docs: fix inaccurate logs documentation #1986 (docs fix), the bot found a pre-existing bug (stale error message in state.rs:470 listing "logs" as a valid key after the logs arm was removed). Instead of approving with an inline suggestion, it posted a COMMENT review, potentially blocking the PR. The issue was not introduced by the PR. The PR was eventually approved in a subsequent review session and merged. No actual harm, but the initial COMMENT could have been an APPROVE+suggestion.
• Gate assessment: Medium evidence, stochastic, 1 occurrence → needs 5+ → below threshold. Carrying forward.

Inline review comment 422 failures (carry-forward from prior runs)

• Evidence level: Low
• Occurrences this run: 0
• Historical: 2 occurrences (runs 24013761335, 24015269514)
• Classification: Structural, but recovery is also structural
• Detail: Not observed this run. Carrying forward from prior tracking.
• Gate assessment: Not seen this run. Below threshold.

---

Run 24179732738 — 2026-04-09T08:10:21Z

Runs analyzed

Run ID	Workflow	Conclusion	PR/Issue	Notes
24176660597	tend-nightly	success	PR #2033	Found workflow security fix (bot self-exclusion + github.token for verify step); PR open, all CI green
24174864622	tend-triage	success	Issue #2031 → PR #2032	Doc improvement request from owner; implemented 3 changes, PR merged
24157907455	tend-triage	success	Issue #2012 → PR #2013	Multiple NAME filters feature; fully implemented and merged
24163478291	tend-ci-fix	success	—	Correctly identified flaky macOS PTY test; no PR needed
24164619577	tend-ci-fix	success	—	Correctly identified flaky zsh PTY test (spawn_detached); no PR needed
24165415836	tend-review	failure	PR #2010	Rate-limited at 21 turns; PR already approved by earlier session
24165601466	tend-notifications	failure	#2010, #2022	Rate-limited at 16 turns
24176821330	tend-review	success	PR #2033	Reviewed own nightly PR; verified security fix, polled CI
24170958007	tend-review	success	PR #2030	Dependabot bump; approved and polled CI
24167933555	tend-review	success	PR #2029	Docs-only; verified claims against code, approved
24165934448	tend-review	success	PR #2025	Optimization; read Rust code, verified correctness, approved

Additionally: ~25 skipped tend-mention runs, ~20 skipped tend-ci-fix runs, 20 successful tend-notifications runs, multiple cancelled/skipped tend-review runs — normal operation.

---

No new findings above threshold

All sessions performed well. Specific quality observations:

• Review accuracy: All 4 sampled review sessions were well-calibrated. Code was read before approving (not just checking commit message), claims were verified against source. PRs Reference picker PreviewCache in caching spec #2029, chore: bump taiki-e/install-action from 2.75.0 to 2.75.1 #2030, perf: thread worktree list through prepare_worktree_removal #2025 merged cleanly.
• Triage accuracy: Both triage sessions implemented real work. PR docs(config): improve project config intro and template variable heading #2032 (doc improvements) and PR feat: accept multiple NAME filters in hook subcommands #2013 (multiple NAME filters) both merged — no reverts or corrections from maintainer.
• CI-fix accuracy: Both ci-fix sessions correctly identified flaky PTY tests as flaky (not regressions) and did not create unnecessary fix PRs.
• Mention handling: 4 sampled sessions correctly exited when not needed — maintainer closing an issue, bot's own prior review firing a trigger, etc.
• Nightly correctness: Correctly deferred model reset (8 PM PT deadline ~20 hours away); correctly regenerated workflow with real security improvement.

Rate-limit failures (infrastructure observation, not behavioral)

• Evidence level: Low
• Occurrences this run: 2 (runs 24165415836 and 24165601466, both at ~00:12-00:18 UTC)
• Classification: Structural — both hit the API rate limit in the same 5-minute burst
• Detail: Two workflows (tend-review and tend-notifications) exhausted the Claude API limit simultaneously around 00:12-00:18 UTC. Both ended with failure conclusion after 16-21 turns. No adverse outcomes — PR perf(ci): speed up CI with D: drive, skip redundant pre-commit, CARGO_INCREMENTAL=0 #2010 was already approved by an earlier session and merged at 01:36 UTC.
• Gate assessment: Low evidence; infrastructure issue (API provider rate limiting). No guidance change can prevent provider-side limits. Not actionable.

CI polling filter inconsistency (carry-forward)

• Evidence level: Low
• Occurrences this run: 3 more (runs 24176821330, 24170958007, 24167933555 all use grep -v "tend-review" not grep -v "$GITHUB_WORKFLOW")
• Cumulative: 14+ occurrences across 6 runs, zero failures
• Classification: Stochastic (guidance exists in running-in-ci skill but inconsistently followed)
• Gate assessment: Functionally equivalent in current setup (workflow is named "tend-review"). Low severity — no observed impact. Below threshold.

Run 24233412050 — 2026-04-10T05:00:00Z (review-runs)

ci-fix opened exclude-list PR for single lychee link-check timeout

• Evidence level: High (acted on)
• Occurrences this run: 1
• Run ID: 24223378129 (tend-ci-fix)
• Workflow: https://github.com/max-sixty/worktrunk/actions/runs/24223378129
• PR opened: fix: exclude caddyserver.com from lychee link checks #2045 (closed by maintainer 21 seconds later)
• Session: 9af46e43-c8e1-465d-8060-09947d194c7e.jsonl
• Detail: Bot opened a fix PR adding caddyserver.com to .config/lychee.toml exclude list after a single [TIMEOUT] from the link checker. Maintainer had already rerun CI at 02:32:58Z — attempt 2 passed. Bot opened PR at 02:32:59Z (one second after rerun started). Maintainer closed it at 02:33:20Z. Bot's own reasoning noted "previous CI run an hour earlier passed fine" but still committed a permanent config change. Acted on via skills(running-tend): prefer CI rerun over exclude list for transient failures #2050 — added "CI Fix: Prefer Rerun for Transient Infrastructure Failures" section to .claude/skills/running-tend/SKILL.md.
• Classification: Structural — the ci-fix skill has no guidance for classifying infra flakes vs real regressions.

Anthropic API 401 auth failures on tend-triage and tend-notifications

• Evidence level: Medium (infrastructure, not actionable via skills)
• Occurrences this run: 2
• Run IDs: 24226131313 (tend-triage), 24226103227 (tend-notifications)
• Detail: Both workflows failed at 04:17-04:19Z with API 401 Unauthorized from Anthropic. Not a bot behavior problem — upstream auth blip or token expiry. Recording for pattern tracking; no repo-local fix applies. If this recurs across multiple days, escalate to workflow config / secret rotation review.
• Classification: Infrastructure (external), not structural to a skill.

Sampled clean runs (no findings)

• 24230590573 (tend-nightly, 137 turns) — real code cleanup, merged cleanly
• 24217607720 (tend-review, cancelled but posted valid review before timeout)
• 24218351323 (tend-notifications, Windows path fix that merged as part of refactor(logs): nest hook output by branch/source/hook-type/name #2041)
• 24217813564 (tend-mention, docs path fix that merged as part of refactor(logs): nest hook output by branch/source/hook-type/name #2041)
• High token counts on the two latter runs are justified by the real fixes they produced.

---

Run 24278174860 — 2026-04-11T08:00Z

Runs analyzed (past 24h, 136 runs with session artifacts)

Workflow	count	notes
tend-review	60	3 successes on the current sync-stack-file/sync-flags/sync-core series (runs 24277499083/04/12); several failures Apr 10 22:xx were all "out of extra usage" API rate limit, not Claude decisions
tend-mention	41	1 cancelled, 1 near-timeout (see Finding 1)
tend-notifications	29	2 failures with "out of extra usage" — same API limit
tend-triage	4	2 successes, 2 API-limit failures
tend-nightly	1	success (24276830345, 9 min, 196 turns) — healthy
tend-review-runs	1	previous run 24233412050

Token totals: 2.27M output, 14.5M cache_creation, 372M cache_read, 4847 turns.

By-workflow turns: tend-mention 1218 (of which 2 runs = 688, 57%), tend-review 2558, tend-notifications 690, tend-nightly 196, tend-triage 45.

All 12 failures today were "You're out of extra usage · resets 11pm (UTC)" — Anthropic API quota exhaustion, not Claude behavioral issues. Not actionable locally.

Finding 1: tend-mention burning 60-min cap on post-push CI polling

• Evidence level: High
• Occurrences this run: 2
• Historical: 7+ prior occurrences in tracking (runs 24064025285, 24064145051, 24064232297, 24064318131, plus cascading notifications/mention failures) documented as "CI polling exhausting API budget — cascading failures"
• Classification: Structural (deterministic time-budget math)
• Run IDs: 24252583044 (cancelled, 60 min, 388 turns), 24254014913 (success but 58 min, 300 turns)
• Sessions: a84b9da3-25a7-4198-be7f-4335e1518afa.jsonl, 49e4fa31-014a-4592-b398-efb701ecb227.jsonl
• Detail:
  • Session 24252583044 implemented external-subcommand dispatch end-to-end in response to a mention on PR feat: add wt sync command to rebase stacked worktree branches #2053, created PR Add git-style external subcommand dispatch (wt-<name>) #2054, pushed at ~turn 200, then ran two for i in $(seq 1 15|12); do sleep 60 polling loops at turns 315 and 379 before being cancelled at turn 388.
  • Session 24254014913 replied to a review comment on the same PR Add git-style external subcommand dispatch (wt-<name>) #2054, then ran seven separate polling loops (turns 128, 155, 172, 247, 256, 293, 297) as it iterated on codecov/patch failures. Its own reasoning at turn 173 was "I'll wait for codecov polling to complete," followed by four more loops. Finished at 58 min, 2 minutes under the cap.
  • PR Add git-style external subcommand dispatch (wt-<name>) #2054 was ultimately merged (20:07 UTC) so outcomes were correct, but the bot had no margin — a single slower CI check on either session would have cancelled mid-operation.
• Proposed fix: Add "CI Wait Budget After Pushing" section to .claude/skills/running-tend/SKILL.md capping post-push polling at one loop per push, at most one fix iteration, no chained required-check + codecov waits, and a required summary comment when the budget is exhausted.
• Gate assessment: High evidence, structural, 9+ cumulative occurrences, new-paragraph magnitude → passes both gates.
• Acted on: skills(running-tend): cap CI polling inside mention/review sessions #2071

Run 24302108289 — 2026-04-12T08:16:38Z

Runs analyzed (past 24h, 200 runs with session artifacts)

Workflow	Runs	Output Tokens	Avg Output/Run	Max Output
tend-review	88	568,603	6,461	102,984 (run 24290891114)
tend-notifications	30	204,559	6,818	57,052
tend-mention	75	105,688	1,409	55,624
tend-triage	5	65,528	13,105	32,191
tend-ci-fix	1	0	0	0
tend-nightly	1	0*	0	0
Total	200	944,378	—	—

*Nightly's 0 is a token-report artifact — cache tokens (7.4M effective) aren't counted in the output_tokens field.

Sessions downloaded and analyzed: 4 (runs 24300392375, 24290891114, 24300799248, 24290304861)

Long-running jobs (>30 min): 1 — tend-mention run 24300392375 (37 min)

Bot PR outcomes:

• 2 open: fix(list): detect AA and DD unmerged status codes in conflict check #2124 (nightly bug fix, CI green), fix: address norm violations from #2122 #2123 (norm violations, CI green)
• 3 merged: docs(skill): support OpenCode in agent handoffs section #2108, docs(extending): add recipe for carrying changes to a new worktree #2083, fix: pass directive file through wt step aliases #2077
• 1 self-closed duplicate: docs(extending): add aliases for moving in-progress changes #2084 (dup of docs(extending): add recipe for carrying changes to a new worktree #2083)
• 1 closed by maintainer: skills(running-tend): cap CI polling inside mention/review sessions #2071 (CI polling caps — explicitly rejected)

No findings above threshold

Below-threshold observations

Duplicate PR creation from concurrent mention runs

• Evidence level: Low
• Occurrences this run: 1 (PRs docs(extending): add recipe for carrying changes to a new worktree #2083 and docs(extending): add aliases for moving in-progress changes #2084, both responding to issue Worktree friction: copy staged and unstaged changes when creating new worktree #938)
• Run ID: 24290304861
• Workflow: https://github.com/max-sixty/worktrunk/actions/runs/24290304861
• Classification: Stochastic — the bot correctly detected and self-closed the duplicate. Two concurrent tend-mention triggers dispatched for the same issue produced overlapping work. Recovery was clean (bot reviewed its own duplicate, identified the overlap, and closed it).
• Gate assessment: Low evidence, 1 occurrence, recovery was correct. Below threshold.

CI polling time (carry-forward, not actionable)

• Occurrences: Observed again in run 24300392375 (37-min tend-mention, ~18 min polling). Maintainer explicitly closed skills(running-tend): cap CI polling inside mention/review sessions #2071 stating "polling is very cheap and I don't see evidence that it's making a big mistake." This topic is resolved — no further PRs on CI polling time will be proposed.

Quality notes

• High-output tend-review (103k tokens, run 24290891114): Justified — reviewed a 2,427-line config refactor (PR refactor: remove config crate, inline OverridableConfig #2087), found a real type-coercion regression, built a compiled Rust repro to prove it. PR was subsequently merged after multiple fix iterations influenced by the bot's findings.
• High-output tend-mention (56k tokens, run 24290304861): Justified — implemented 3 working aliases + docs page for issue Worktree friction: copy staged and unstaged changes when creating new worktree #938, tested end-to-end in scratch repos, ran full test suite.
• Nightly run (24300799248): Productive — found a real bug (AA/DD unmerged status codes missed in conflict check), created PR fix(list): detect AA and DD unmerged status codes in conflict check #2124 with 4 unit tests and CI passing on all platforms.