post_bot_commits counts local-tier agent commits as human rework

[The01Geek]

Problem Statement

The retrospective's post_bot_commits field is meant to measure human rework after the bot's last push. It miscounts the engine's own local-tier agent commits as human rework, so the clean gate forces LLM analysis on PRs that had none.

Current Behavior

lib/fetch-pr-context.sh (the POST_BOT_COMMITS jq block) finds the last commit whose author_login or committer_login ends in [bot] or equals the PR author, then counts every non-merge commit after it. Local-tier Claude Code sessions commit under a git identity GitHub cannot resolve to an account, so the API returns a blank login. A blank login fails the [bot] test, so every local-tier agent commit after the last cloud-bot commit is counted as human rework.

Measured over 232 merged PRs / 1,912 commits: only three distinct author_login values exist — blank (1,080), github-actions[bot] (673), and the repo owner (159, of which 158 are merge commits already excluded). Essentially one identifiable human non-merge commit exists in the corpus, yet post_bot_commits > 0 on 96 of 232 PRs (41%). A run-boundary cross-check (commits after the ready_for_review timeline event) disagrees on 42 of 219 PRs, systematically the local-tier-agent case; PR 772 reports post_bot_commits: 8 where all eight counted commits carry fix-loop messages and predate ready_for_review.

Desired Behavior

A non-merge commit counts toward post_bot_commits only when it is positively attributable to a human: its author_login or committer_login is non-blank, does not end in [bot], and is not the PR author. A commit whose login the API returns blank is treated as agent-side, never as human rework — unknown is not a human.

User Impact

The weekly retrospective's clean gate stops forcing LLM analysis on ~41% of PRs whose only "rework" was the engine's own fix-loop commits, and consumers of the field stop drawing a wrong human-effort conclusion.

Technical Context

Scope note: The files and details below are the known starting points, not the full list. Before implementing, trace the change through the codebase to find every affected call site, consumer, and layer — this issue maps the work, it does not bound it.

• Relevant Classes/Files — lib/fetch-pr-context.sh (the POST_BOT_COMMITS jq filter and its explanatory comment), lib/cheap-gate.jq (sole consumer: the post_bot_commits > 0 non-clean arm and the field's header comment).
• Architecture Alignment — this repo's unknown is not zero rule: an unestablished measurement never collapses onto a real value. A blank login is unestablished identity and must not collapse onto "human".
• Dependencies — none beyond the existing gh/jq path.
• Data/Schema Considerations — the field's type and name are unchanged; only its counting rule changes.
• Cross-layer Impact — retrospective scan path only; no skill, workflow, or config surface.

Acceptance Criteria

• A PR whose commits after the last bot/author commit all carry blank author_login and blank committer_login reports post_bot_commits: 0.
• A non-merge commit after the last bot/author commit whose author_login or committer_login is a non-blank login that neither ends in [bot] nor equals the PR author is still counted.
• Merge commits (parents_count > 1) remain excluded from the count.
• The comment above the jq block and the field's description in lib/cheap-gate.jq state what the field now measures, including that blank-login commits are classified agent-side.
• lib/cheap-gate.jq's post_bot_commits > 0 non-clean arm is unchanged.

Implementation Notes

• Approach — narrow the counted tail in the POST_BOT_COMMITS jq filter so only positively human-attributable non-merge commits count; the bot-side "last bot commit" anchor logic keeps its current meaning.
• Relevant files — this likely touches only lib/fetch-pr-context.sh, plus the field comment in lib/cheap-gate.jq.
• Code Patterns — the filter runs on API-shaped commit JSON; guard possibly-non-string fields with (.x | strings) before string ops, per the repo's embedded-jq rules.
• Testing Strategy — lib/fetch-pr-context.sh is unmodularized (run.sh-resident), so drive the jq filter directly against fixture commit arrays: a blank-login tail (asserts 0), a mixed tail with one named human commit (asserts 1), a merge-commit-only tail (asserts 0), and a null-login record (asserts no filter abort). Each fixture assertion fails first against today's filter where the behavior changes.
• Documentation Needed — none. The behavior is internal to the retrospective scan and its contract lives in the two file comments the ACs cover.
• Potential Gotchas — a genuinely human commit made with an unresolvable email becomes invisible to the count; the measured corpus shows this population is ~1 commit in 1,912, and the field's purpose (a cheap non-clean trigger) tolerates that false negative far better than the current 41% false-positive rate. Historical retrospective records are not rewritten.

[The01Geek]

/devflow:implement 1440

[github-actions]

PRFlow Workpad — Issue #1440

Status: 🎉 Complete
Branch: issue-1440-post-bot-commits-counts-local-tier-agent-commits
Run: View run
PR: #1941
Last updated: 2026-08-25 19:19 UTC

Progress

• Setup — branch & workpad
  • prompt extension resolved: implement
  • 04:28:56 — /prflow:implement run started
  • 04:29:40 — Claude job setup complete; invoking agent
  • 04:32:48 — agent entered Phase 1 setup; workpad triage passed
  • 04:32:59 — agent initialized; Phase 1 workpad hydrated
  • 04:33:08 — resume-kind: fresh
  • 04:33:08 — classification: bug-report — Describes incorrect behavior: post_bot_commits miscounts blank-login agent commits as human rework — a counting defect to fix
  • 04:33:36 — Phase 1.3.5: early declared-dependency preflight passed (PROCEED)
  • 04:36:14 — issue-claim audit (projection): every Desired Behavior obligation is represented in the resolved acceptance criteria — pass complete
  • 04:36:16 — issue-claim audit (count): no count or enumeration claims found — pass complete
  • 04:36:21 — issue-claim audit (negative-scope): issue excluded skill/workflow/config surface; trace confirms no impact — the only skill mention (skills/retrospective/SKILL.md) already describes the field qualitatively as human commits and needs no functional edit
  • 04:36:23 — issue-claim audit (policy): no policy-referencing AC claims found — pass complete
  • 04:36:28 — issue-claim audit (execution-capability): no workflow-resident acceptance criteria found — pass complete
  • 04:36:30 — issue-claim audit (verified-premise): no Verified: bullets and no ungraded claims found in the issue body — pass complete
  • 15:02:44 — /prflow:implement run resumed from stall-backstop Failed status (self-check ground 3); Phase 1 workpad hydrated
  • 15:03:05 — resume-kind: in-flight
  • 15:04:29 — branch-state: Verdict B classified branch issue-1440-post-bot-commits-counts-local-tier-agent-commits (PR-adopted arm, PR post_bot_commits counts local-tier agent commits as human rework #1941, selected by head query) as VALIDATED_RESUME — proceeding
  • 15:05:01 — checkpoint 1: merged origin/main and pushed (was behind by 19); auto-merged lib/test/run.sh cleanly. Re-ran 2.3.0 changed-contract sweep against arrived sites (CONTRIBUTING.md, docs, group_labels_by_subject.py + its test, coverage-map.json, run.sh subject-grouping rows) — none touch the post_bot_commits/fetch-pr-context.sh/cheap-gate.jq contract; clean.
  • 15:05:22 — Phase 1.6 issue-claim audit: carried forward from the prior in-flight attempt (all 6 passes recorded above at 04:36) — issue body and ACs unchanged this resume, and the 19 base commits merged at checkpoint 1 do not touch the post_bot_commits contract, so no audit assumption is invalidated.
  • 18:09:44 — /prflow:implement run resumed from stall-backstop Failed status (self-check ground 3); Phase 1 workpad hydrated
  • 18:09:53 — resume-kind: in-flight
• Implement
  • reproduction captured (bug issues only)
  • code + sweeps
  • 04:37:27 — entered Phase 2; §2.0 resume gate not fired (resume-kind fresh) — running full discovery
  • 04:43:55 — captured reproduction signal (recorded jq probes; no proof edits)
  • 04:44:56 — §2.2.4 reuse: searched for existing bot/human login classifier (Grep 'def is_human|def ends_bot|human_login|is_bot_login') — no match; new inline jq defs justified. Altitude: fix the classification predicate in the filter itself (shared mechanism), not a special-case patch.
  • 04:44:56 — §2.2.7 coupled-site map: rule replicated at POST_BOT_COMMITS (fetch-pr-context.sh ~259) and POSTBOT_SHAS (~762); comments at fetch-pr-context.sh ~250 and cheap-gate.jq ~27; test at run.sh ~14113. Searched 'endswith("[bot]")' → run.sh, fetch-pr-context.sh, install.sh (install.sh does NOT consume post_bot_commits — different rule, out of scope). skills/retrospective/SKILL.md ln100 documents the field but already reads 'by a human' (not falsified; outside AC scope). Historical .prflow/learnings/*.jsonl not rewritten (per issue).
  • 04:44:56 — Path A: simple issue (jq filter + comments + test, <=5 files) — architect skipped
  • 04:45:39 — criterion-lifecycle: AC1/AC2/AC3 → Testable (RED-first jq fixture tests in run.sh); AC4 (comments state new measurement) → genuinely-untestable prose deliverable, verified by Phase 2.4 inspection/dry-trace (no wording-only pin); AC5 (cheap-gate >0 arm unchanged) → Testable via negative-scope: diff shows the arm untouched + existing CLEAN/793 fixtures still gate correctly.
  • 04:45:39 — test-first: extract-and-execute fixtures in lib/test/run.sh drive the shipped POST_BOT_COMMITS/POSTBOT_SHAS blocks; new rows (blank→0, mixed→1, null→no-abort, blank-SHA-excluded) fail against today's filter and pass after the fix
  • 04:51:44 — section-2.3 operand ledger: examined merge-base(origin/main)->worktree delta (lib/fetch-pr-context.sh, lib/cheap-gate.jq, lib/test/run.sh); sole Phase 2 sweep boundary. Diff shape: modified a classification predicate + added a predicate at 2 co-equal peer sites + deleted a test helper. Sweeps run: 2.3.0, 2.3.0a, 2.3.0c, 2.3.1, 2.3.7 + always-on 2.3.3/2.3.4/2.3.4a/2.3.4b/2.3.5/2.3.6.
  • 04:51:44 — 2.3.0a peer-checkpoint: classification=co-locatable peers -> search arm. Peers: POST_BOT_COMMITS (count) and POSTBOT_SHAS (SHA list) in lib/fetch-pr-context.sh. Searched is_human/ends_bot; both peer blocks carry the identical rule. POSTBOT_SHAS 'same rule as post_bot_commits above' comment stays true. No exemptions.
  • 04:51:44 — 2.3.0 changed-contract: dependent sites reconciled. cheap-gate.jq >0 non-clean arm unchanged (int consumer, AC5). HUMAN_POSTBOT_DIFF loop consumes the now-narrower SHA list (correct). gh-stub fixtures verified preserved: 793 post-bot = 4 octocat (human, non-blank, non-merge) still counted; CLEAN = 2 claude[bot] -> 0. run.sh inline hand-copy replaced by shipped-block extraction (no stale citation of old filter).
  • 04:51:44 — 2.3.0c operand-trace: comparands author_login, committer_login. Producer: COMMITS normalization at lib/fetch-pr-context.sh line 175 which coerces missing/null login to empty string, emitted on every path. Malformed/empty arms decided by is_human type-guard: non-string/null -> false (not human), empty string -> false (not human). Fourth column: the other value producing the same result is a blank agent login, deliberately classified agent-side (unknown is not a human).
  • 04:51:44 — 2.3.1 orphaned-setup: deleted _postbot_count helper + old comment lines; COMMITS/AUTHOR still consumed by the surrounding block; grep for _postbot_count returns no matches (no dangling refs); no orphaned setup.
  • 04:51:44 — 2.3.7 cardinality: count tally + SHA list projection; multi-element fixtures added (mixed 3-element tail; SHA fixture X/B/A asserting only the human SHA A). Discharged.
  • 04:51:44 — always-on sweeps: 2.3.3 embedded-jq non-string guarded, no jq-comment apostrophes, temp vars unset; 2.3.4 endswith([bot]) is the unchanged GitHub-login boundary, COMMITS shape verified against 793 fixture, no new boundary; 2.3.4a comments match shipped code; 2.3.4b stale-prose-lint --worktree exit 0 zero rows with hunks present = clean pass (leg1) [leg2/new-file changeset re-run in Phase 3]; 2.3.5 is_human reuses ends_bot, no added complexity; 2.3.6 filter now null-safe fail-closed, no silent failure.
  • 05:06:50 — focused-selection: touched surface lib/fetch-pr-context.sh is run.sh-resident (unmodularized per coverage-map; no covering focused module) — ran the monolith shard (lib/test/run-shard.sh monolith) as its exemption; single_flight_consulted=no-reusable-result (tree freshly edited, no clean prior result to reuse). Also covers lib/cheap-gate.jq (comment) and lib/test/run.sh (post_bot_commits counts local-tier agent commits as human rework #1440 test block).
  • 05:06:50 — section-2.4 verification: monolith shard on committed tree f4e098d — all 9 post_bot_commits counts local-tier agent commits as human rework #1440 tests PASS (blank->0, mixed->1, null-no-abort, merge-excluded, human-merge->0, missing-parents-counted, POSTBOT_SHAS excludes blank SHA, both extraction guards). 12 unrelated failures in Keep writer-job push and gh credentials fresh past the App token's 60-minute lifetime #487/Keep writer-job push and gh credentials fresh past the App token's 60-minute lifetime #491/Sign the credential refresher's JWT without openssl process substitution so long cloud runs keep GitHub access on non-Linux runners #1882 (Stop-refresher liveness, resolve_real_gh, credential-mint) are outside this diff's blast radius (post_bot_commits jq filter), were absent in the first run, and are process-liveness/pidfile races from a second monolith run in the same env. RED-first established by the section-2.1.5 reproduction probes (blank endswith=false; null aborts).
  • 06:16:04 — run died: fail-exhausted — https://github.com/The01Geek/prflow/actions/runs/32816106216
  • 14:58:15 — trigger accepted; existing workpad adopted
  • 14:58:53 — Claude job setup complete; invoking agent
  • 15:02:34 — agent entered Phase 1 setup; workpad triage passed
  • 15:05:53 — Phase 2 §2.0 resume gate: fired (resume-kind in-flight + committed Plan present); skipping code-explorer/code-architect re-dispatch and building on the committed Plan. §2.1.5 no-op (Reproduction populated + ticked). §2.2.5 no-op (single-PR lib/ change; not workflow-resident so no GITHUB_TOKEN-fallback capability block). §2.2.6 no-op (ACs name post_bot_commits/parents_count/fetch-pr-context.sh/cheap-gate.jq — all still accurate vs the shipped diff). Re-verifying the shipped diff against a fresh tree.
  • 15:13:46 — focused-selection: touched surfaces lib/fetch-pr-context.sh + lib/cheap-gate.jq + lib/test/run.sh post_bot_commits counts local-tier agent commits as human rework #1440 block are run.sh-resident (unmodularized per coverage-map, no covering module) — ran lib/test/run-shard.sh monolith as the exemption. single_flight_consulted=no (tree freshly changed by checkpoint-1 base merge; no reusable clean result held for it).
  • 16:03:22 — run died: fail-exhausted — https://github.com/The01Geek/prflow/actions/runs/32862791893
  • 18:05:17 — trigger accepted; existing workpad adopted
  • 18:06:23 — Claude job setup complete; invoking agent
• Review
  • prompt extension resolved: review engine
  • prompt extension resolved: fix loop
  • prompt extension resolved: code-review reception
  • Classify diff (Phase 0.5)
  • Generate verification checklist (Phase 1)
  • Verify checklist (Phase 2)
  • Review agents (Phase 3)
  • Aggregate & verdict (Phase 4)
  • Run complete — everything this run owed
  • /simplify
  • review-and-fix
  • acceptance-criteria gate
  • 05:07:57 — Phase 3.1 checkpoint 2: merged origin/main and pushed (was behind by 18); auto-merged lib/test/run.sh cleanly. Base brought scripts/stop-refresher.sh + run.sh changes. Re-running 2.3.0 changed-contract sweep against newly-arrived sites per the rule.
  • 05:13:59 — Phase 3.1 re-verification after checkpoint-2 merge: monolith shard on merged tree = 10005 passed, 0 failed, 0 skipped (rc=0). All 9 post_bot_commits counts local-tier agent commits as human rework #1440 tests pass; the earlier Keep writer-job push and gh credentials fresh past the App token's 60-minute lifetime #487/Keep writer-job push and gh credentials fresh past the App token's 60-minute lifetime #491/Sign the credential refresher's JWT without openssl process substitution so long cloud runs keep GitHub access on non-Linux runners #1882 liveness failures are resolved on the merged base (stop-refresher.sh changed on main) + a fresh run, confirming they were environmental, not a post_bot_commits counts local-tier agent commits as human rework #1440 regression. 2.3.0 re-sweep clean: no newly-arrived site violates the post_bot_commits contract.
  • 05:18:42 — /simplify: reuse/simplification/efficiency clean. Altitude flagged the coupled-mirror duplication between POST_BOT_COMMITS and POSTBOT_SHAS (count conceptually = POSTBOT_SHAS | length). SKIPPED as out-of-scope: the select/anchor duplication predates this change (both blocks already carried it; my change only mirrored the added defs), the two blocks sit ~500 lines apart (SHAS gated by the count), a shared jq fragment collides with the single-quoted jq/shell boundary, and collapsing them would rework the extraction test — an unrelated refactor beyond issue post_bot_commits counts local-tier agent commits as human rework #1440's 'narrow the counted tail' approach. The existing extraction test already drives BOTH shipped blocks, so the mirror stays drift-guarded.
  • 05:19:10 — Phase 3.3: pre-loop iter-*.json snapshot empty (fresh cloud runner); wrote empty .phase33-iters-before baseline; invoking review-and-fix in PR mode (PR 1941)
  • 05:55:33 — trigger accepted; existing workpad adopted
  • 05:56:16 — Claude job setup complete; invoking agent
  • 06:00:49 — /prflow:implement run resumed (run 32814750405); adopted interim workpad at Status Reviewing. Branch issue-1440-... checked out at 26c2eb7 (was on main). Phase 1-2 + Phase 3.1 (checkpoint-2 merge, /simplify) complete per workpad; resuming at Phase 3.3 review-and-fix per recorded resume point.
  • 06:01:32 — Phase 3.3: invoking review-and-fix in PR mode (PR 1941, bare numeric token) with --push-each-iteration --issue 1440
  • 06:14:56 — trigger accepted; existing workpad adopted
  • 06:15:34 — Claude job setup complete; invoking agent
  • 15:14:22 — Phase 3 entry (resume): re-running checkpoint 2, PR adoption, and the §3.3 review-and-fix pass on the merged tree (9554735). /simplify already ran clean on this PR's unchanged 4-file diff (row ticked, 05:18:42) — not re-run.
  • 15:14:58 — Phase 3.3: invoking review-and-fix in PR mode (PR 1941, bare numeric token) with --push-each-iteration --issue 1440; pre-loop iter-*.json baseline empty (fresh cloud runner)
  • 15:14:58 — mid-phase resume point recorded
  • 15:45:57 — Phase 3.3 iter 1: engine APPROVE with notes (9/9 checklist PASS; 1 Important whitespace-login finding + 8 Suggestions). Fixed the Important finding (whitespace-only login now classified agent-side via is_human test([^space]); coupled POST_BOT_COMMITS+POSTBOT_SHAS blocks + RED-first fixture; commit b9388ed). Step 3.5 fix-delta gate: clean (blinded reviewer confirmed narrowing, type-guard/short-circuit safe, fails-closed on all login shapes, assertion RED-against-pre-fix). 8 Suggestions parked below-threshold (advisory). monolith shard 10006 passed 0 failed.
  • 18:11:58 — Phase 3.3 (resume): re-invoking review-and-fix in PR mode (PR 1941, bare numeric token) with --push-each-iteration --issue 1440; pre-loop iter baseline empty (fresh cloud runner). Prior iter-1 fixed the Important whitespace-login finding (b9388ed); this pass re-reviews the converged diff.
  • 18:42:54 — Phase 3.3 Loop Exit checkpoint 3: merged origin/main and pushed (was behind 174); auto-merged cleanly, HEAD now ca72fb7. 2.3.0 changed-contract re-sweep: the 174 arrived commits touch neither lib/fetch-pr-context.sh nor lib/cheap-gate.jq — post_bot_commits contract untouched. Loop-exit suite rule does not fire (implement caller; Phase 4.3 owns the whole-suite over the merged tree).
  • 18:43:29 — review coverage recorded (coverage=full, dispatch=attempted, roster=complete, checklist=complete; as of head ca72fb7 at 20260825T184329Z)
  • 18:43:48 — Extension state established this run: review-and-fix (fix loop) and receiving-code-review (reception) loaded directly via the ladder = content-present; the review engine's own extension (review.md) loaded on the engine's prose path during the inline Step-1/Step-2.6 engine dispatches (verdicts produced). Loop converged: APPROVE WITH ADVISORY NOTES, shadow full coverage, 0 Critical/Important.
  • 18:54:38 — 3.4 gate: all 5 in-scope non-post-merge ACs reconciled satisfied (two-verifier: all_satisfied=true, blocking=[]); evidence + dispositions in this run's AC dispositions note. Reviewed head ca72fb7; monolith 9850/0/0.
  • 18:54:58 — Phase 3.4 single-flight finish summary (for Phase 4.3 completion-evidence): lib/test/run-shard.sh monolith on ca72fb7 = 9850 passed, 0 failed, 0 skipped, rc=0, aggregate CLEAN; flight_key a022bea744f404395fc7bda5f79a2f0a5e51521facdc130eb183de0c06a0567e; log .prflow/tmp/shard-tally/monolith/log.txt.
• Documentation
  • 18:58:05 — Phase 4.1: routine prflow:docs pass ran. Internal-doc fix: docs/internal/DEVFLOW_SYSTEM_OVERVIEW.md §12 stale 'open post_bot_commits counts local-tier agent commits as human rework #1440' clause rewritten to describe the shipped fix (committed 25102d5). External docs: no surface. Release notes: changeset already carries the entry. Documentation Needed=none; Stage 2 diff gate is a no-op. docs.labels=PRFlow applied to post_bot_commits counts local-tier agent commits as human rework #1941.
  • 19:00:21 — 4.2 claim audit (behavioral): all body claims traced to shipped code (is_human/ends_bot in POST_BOT_COMMITS+POSTBOT_SHAS; cheap-gate.jq > 0 arm unchanged) — confirmed, no overclaim.
  • 19:00:21 — 4.2 claim audit (verification): Test Plan rows bound to named post_bot_commits counts local-tier agent commits as human rework #1440 tests present in the diff (run.sh 13650-13706); tally cites observed monolith 9850/0/0 — pass complete.
  • 19:00:21 — 4.2 claim audit (artifact-existence): every named path + changeset + issue post_bot_commits counts local-tier agent commits as human rework #1440 resolvable — pass complete.
  • 19:00:21 — 4.2 PR body updated from placeholder (confirmed). PR-description extension present/content-present/honored. No code-level fix made.
  • 19:00:50 — checkpoint 4: observed token UP_TO_DATE — clean, proceeding to the publish decision
  • 19:18:37 — completion verification recorded (flight 672226cfdf96…, validated)
• PR marked ready
  • 19:19:33 — /prflow:implement run finished, PR published (gh pr ready): post_bot_commits counts local-tier agent commits as human rework #1941

Plan

• Update the POST_BOT_COMMITS jq filter in lib/fetch-pr-context.sh (~ln 259): add jq def ends_bot/def is_human helpers; make the bot-anchor select null-safe (type-guard before endswith) and narrow the counted tail to positively human-attributable non-merge commits (a non-blank, non-[bot], non-author login on author_login OR committer_login). Keeps the "last bot commit" anchor meaning; keeps parents_count > 1 merge exclusion. (AC1, AC2, AC3)
• Rewrite the explanatory comment above the block to state what the field now measures, including that blank/absent-login commits are classified agent-side (unknown is not a human). (AC4)
• Apply the identical ends_bot/is_human rule to the coupled POSTBOT_SHAS peer block (~ln 762) so the SHA list stays coherent with the count (§2.3.0a peer completeness).
• Update the post_bot_commits field description comment in lib/cheap-gate.jq (~ln 27) to state the new measurement incl. blank-login=agent-side; leave the post_bot_commits > 0 non-clean arm (~ln 141) unchanged. (AC4, AC5)
• Replace the hand-copied _postbot_count() test in lib/test/run.sh (~ln 14113) with a drift-proof extract-and-execute of the shipped POST_BOT_COMMITS block (CP4-style), plus RED-first fixtures: blank-login tail → 0, mixed tail with one named human → 1, merge-commit-only tail → 0, null-login record → no abort (+ retain merge-excluded / missing-parents_count regression rows); add a POSTBOT_SHAS peer assertion excluding a blank-login SHA. (AC1, AC2, AC3)
• Add a .changeset/*.md file (patch) for the engine-surface fix.
• Run the monolith shard to verify RED→GREEN; regenerate artifacts; whole-suite via Phase 4.3 CI.

Acceptance Criteria

• A PR whose commits after the last bot/author commit all carry blank author_login and blank committer_login reports post_bot_commits: 0.
• A non-merge commit after the last bot/author commit whose author_login or committer_login is a non-blank login that neither ends in [bot] nor equals the PR author is still counted.
• Merge commits (parents_count > 1) remain excluded from the count.
• The comment above the jq block and the field's description in lib/cheap-gate.jq state what the field now measures, including that blank-login commits are classified agent-side.
• lib/cheap-gate.jq's post_bot_commits > 0 non-clean arm is unchanged.

Reproduction

Reproduction — issue #1440 (post_bot_commits miscounts blank-login agent commits)

The POST_BOT_COMMITS jq filter in lib/fetch-pr-context.sh classifies a commit
as bot-side only when author_login/committer_login ends in [bot] or equals
the PR author. A blank login (the local-tier agent identity GitHub cannot resolve)
therefore fails both tests and lands in the counted "human rework" tail. Two
root-cause mechanisms, demonstrated with recorded jq probes (no proof edits made):

1. A blank login reads as "not a bot", so it is counted as human

$ echo '""' | jq 'endswith("[bot]")'
false

"" | endswith("[bot]") is false, and "" == <author> is false, so a
blank-login non-merge commit after the last bot commit is counted as human rework.
Driving the current shipped filter against a blank-login tail
[{bot},{blank-login non-merge}] returns 1; the issue's AC1 requires 0.

2. A null login aborts the whole filter

$ echo 'null' | jq 'endswith("[bot]")'
jq: error (at <stdin>:1): endswith() requires string inputs
jq-exit=5

The current anchor select runs endswith on every commit's login, so a null
login (absent field) aborts the entire filter — the testing-strategy "null-login
record (asserts no filter abort)" case.

Fix direction

Narrow the counted tail to positively human-attributable non-merge commits
(a non-blank, non-[bot], non-author login on author_login or committer_login),
and make the classification null-safe (type-guard before endswith). Apply the
same rule to the coupled POSTBOT_SHAS peer block. RED-first fixture tests will
fail against today's filter and pass after the fix.

Devflow Reflection

Devflow Reflection (click to expand)

💡 Improvements

• 💡 Dogfood: the Keep writer-job push and gh credentials fresh past the App token's 60-minute lifetime #487/Keep writer-job push and gh credentials fresh past the App token's 60-minute lifetime #491/Sign the credential refresher's JWT without openssl process substitution so long cloud runs keep GitHub access on non-Linux runners #1882 Stop-refresher/credential-mint/liveness assertions in lib/test/run.sh appear non-hermetic across repeated runs in one environment — a first monolith shard run was clean, a second (same commit) reported 12 kill-by-pidfile/liveness failures (expected yes, actual no), consistent with leaked background refresher processes from the earlier run colliding on pidfiles. These tests spawn real background loops; consider per-test process isolation or unique pidfile namespacing so a repeated in-env run does not flake.
• 💡 Phase 3.4 AC-gate friction: reconcile-ac-verifiers.py keys strictly on the charter disposition slot names (evidence: type-decided/command-run/single-flight/evidence-recorded; claim: claim-traced/command-source-read/evidence-recorded). When the orchestrator dispatches prflow:ac-evidence-verifier / prflow:ac-claim-verifier with a custom prompt that names its own disposition keys, the subagents emit those instead and the reconciler forces every criterion to 'unestablished' (both sides still reported satisfied), costing a full re-dispatch. Fix: either have the verifier agents' own charters authoritatively fix the slot names (and instruct dispatch prompts not to override the dispositions shape), or have reconcile-ac-verifiers.py accept documented aliases, so a correct two-verifier agreement is not blocked purely on slot-name spelling.

ℹ️ Notes

• ℹ️ skill-directory anchor resolved from the runner-reported base directory
• ℹ️ resume-precheck: both open-PR queries ran and returned none for this issue; workpad Branch line absent; no prior attempt to adopt
• ℹ️ resume-precheck: adopted PR post_bot_commits counts local-tier agent commits as human rework #1941 (head issue-1440-post-bot-commits-counts-local-tier-agent-commits, selected by the head query, closes-issue by-construction); workpad Branch line issue-1440-post-bot-commits-counts-local-tier-agent-commits; skipping branch creation and both signals
• ℹ️ Verification evidence: lib/test/run-shard.sh monolith on merged HEAD (feature branch, post checkpoint-1 base merge, commit 9554735) — aggregate 10006 passed, 0 failed, 0 skipped (rc=0); all 9 post_bot_commits counts local-tier agent commits as human rework #1440 tests PASS. Mid-iteration focused re-verification of the run.sh-resident surface after the 19-commit base merge; whole-suite completion evidence deferred to Phase 4.3.
• ℹ️ park-calibration gate clean: no parked finding matched a mis-grade shape; the shadow's self-exclusion Minor re-raise is equivalent to the recorded defense-in-depth disposition. over-grade calibration gate clean: no promote-path (Critical/Important) shadow finding flagged. Shadow: APPROVE, coverage full (5/5 reviewers), 0 Critical/Important, advisory-only.
• ℹ️ Phase 3.4 acceptance-criteria gate — two-verifier reconciled dispositions (issue post_bot_commits counts local-tier agent commits as human rework #1440, reviewed head ca72fb7) Reconciliation: all_satisfied=true, blocking=[]. Evidence verifier ran lib/test/run-shard.sh monolith = 9850 passed / 0 failed / 0 skipped, rc=0, CLEAN (flight a022bea7…, HEAD ca72fb7); claim verifier traced each claim to code + assertions. Both sides satisfied on every criterion; all disposition slots discharged. - AC1 (blank-login tail -> 0): status=satisfied; evidence_status_reported=satisfied, claim_status_reported=satisfied; undischarged_slots=[]. Evidence: post_bot_commits counts local-tier agent commits as human rework #1440 blank-login fixture; monolith GREEN. Claim: run.sh:13674 vs is_human non-blank test fetch-pr-context.sh:264. Evidence slots: type-decided=yes, command-run=yes, single-flight=yes, evidence-recorded=yes. Claim slots: claim-traced=yes, command-source-read=yes, evidence-recorded=yes. - AC2 (non-blank human, either OR arm, not [bot], != author -> counted): status=satisfied; both reported satisfied; undischarged=[]. Evidence: post_bot_commits counts local-tier agent commits as human rework #1440 OR-arm fixtures; monolith GREEN. Claim: run.sh:13693-13695,13701 vs select(is_human(author) or is_human(committer)) fetch-pr-context.sh:276. Evidence slots all yes; claim slots all yes. - AC3 (merge parents>1 excluded; missing parents non-merge): status=satisfied; both reported satisfied; undischarged=[]. Evidence: post_bot_commits counts local-tier agent commits as human rework #1440 merge fixtures; monolith GREEN. Claim: run.sh:13682-13686 vs select((.value.parents_count // 1) <= 1) fetch-pr-context.sh:275. Evidence slots all yes; claim slots all yes. - AC4 (comment + cheap-gate.jq field description state blank-login=agent-side): status=satisfied; both reported satisfied; undischarged=[]. Evidence: fetch-pr-context.sh:250-261 comment + cheap-gate.jq:27-31 description. Claim: cheap-gate.jq:27 vs shipped exclusion logic. Evidence slots: type-decided=yes, command-run=no (non-command), single-flight=no (non-command), evidence-recorded=yes. Claim slots: claim-traced=yes, command-source-read=no (comment-accuracy criterion), evidence-recorded=yes. - AC5 (cheap-gate.jq > 0 non-clean arm unchanged): status=satisfied; both reported satisfied; undischarged=[]. Evidence: git diff shows only comment changed; line 145 unchanged. Claim: cheap-gate.jq:145 unchanged + run.sh:14177 existing test. Evidence slots: type-decided=yes, command-run=no, single-flight=no, evidence-recorded=yes. Claim slots: claim-traced=yes, command-source-read=no (negative-scope), evidence-recorded=yes.
• ℹ️ Verification evidence: lib/test/run-parallel.sh (whole-suite coordinator) on final tree 25102d5 — aggregate CLEAN, 21263 passed, 0 failed, 0 skipped, rc=0, elapsed 968s; 5 shards (monolith python-pool modules-pin modules-large modules-rest), required partition covered; retained logs /home/runner/work/prflow/prflow/.prflow/tmp/parallel-suite/run-325522-0/logs; flight 672226cf (satisfies_verification=true). batched-regeneration: skipped — not run explicitly this iteration, but the coordinator's read-only preflight over the same registry did not refuse launch and the modules-* drift checks passed, so no generated artifact drifted.

[prflow-implementer]

DevFlow stall backstop — this cloud run ended while the workpad Status was still in-progress (interim 🚀 Reviewing). Auto-resume attempt 1 of 2:

Resume note: invoke bundled helpers as .prflow/vendor/prflow/scripts/… (and .prflow/vendor/prflow/lib/…) with that path as the leading token — never an absolute path, never repo-root scripts/…, never behind a VAR= prefix or bash <path> wrapper (those forms are silently denied).

Headless note: this is a headless run — ending the turn ends the process, with no re-invocation. Never end the turn while any dispatched agent has not returned; poll to keep the turn alive; treat ScheduleWakeup and any future task-notification as unavailable.

/prflow:implement 1440

[prflow-implementer]

DevFlow stall backstop — this cloud run ended while the workpad Status was still in-progress (interim 🚀 Reviewing). Auto-resume attempt 2 of 2:

Resume note: invoke bundled helpers as .prflow/vendor/prflow/scripts/… (and .prflow/vendor/prflow/lib/…) with that path as the leading token — never an absolute path, never repo-root scripts/…, never behind a VAR= prefix or bash <path> wrapper (those forms are silently denied).

Headless note: this is a headless run — ending the turn ends the process, with no re-invocation. Never end the turn while any dispatched agent has not returned; poll to keep the turn alive; treat ScheduleWakeup and any future task-notification as unavailable.

/prflow:implement 1440

[prflow-implementer]

DevFlow stall backstop — this cloud run ended while the workpad Status was still in-progress (interim 🚀 Reviewing), and the auto-resume cap (max_resume_attempts=2, attempts already made: 2) is exhausted. The job has been failed so the stall is visible; to retry, post a fresh implement trigger comment naming this issue number.

[The01Geek]

/prflow:implement 1440

[prflow-implementer]

DevFlow stall backstop — this cloud run ended while the workpad Status was still in-progress (interim 🚀 Reviewing), and the auto-resume cap (max_resume_attempts=2, attempts already made: 2) is exhausted. The job has been failed so the stall is visible; to retry, post a fresh implement trigger comment naming this issue number.

[The01Geek]

/prflow:implement 1440