You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
I re-enabled those eight disabled workflows. We need to refactor the whole suite, not build with piecemeal tinker toys. #586
Taking this over (agent: claude-code). Starting the holistic suite map #586 calls for — the automation system as one thing, not parts to patch or silence. This is Map v1: inventory + topology. Facts are from the Actions API + the origin/main file tree + each workflow's on:/script refs (read just now); the functional clusters are my organizing lens, labeled as such; unresolved items are marked *.
4 "ghost" workflows — registered in the Actions API (ran on some branch once) but no file on main, so invisible in the tree yet still counted: auto-pr.yml, agent-branch-auto-pr.yml, claude-sign.yml, code-coverage.yml. (code-coverage = open PR Add initial code coverage workflow file #587, not yet merged; the other three are stale registrations.)
The real coupling — two shared "spines" (the entanglement is in the scripts, not the YAML)
review_feedback_loop.py — 7 workflows: auto-merge-engage, review-feedback-loop, review-response, engage-outdated, agent-review-gate, looker-walk, agent-auto-pr. The nervous system of the merge/review engine.
(pr_lifecycle.py ×2; check_secret_patterns.py ×2; validate_content.py ×2.)
→ Redesigning the merge engine means redesigning around review_feedback_loop.py, not file-by-file.
Functional clusters (organizing lens)
A · Review / merge-queue engine — the contested core (#399 / #521 / #406; my #590/#591/#594 this session) auto-merge-engage (pull_request_target) · review-feedback-loop (issue_comment + pull_request_target) · review-response (pull_request_review) · engage-outdated (dispatch) · agent-review-gate (dispatch; self-marked DISABLED, "retired risk-tier schema") · batch-arm-merge-queue (dispatch) · looker-walk (dispatch, read-only) · auto-merge-rhythm (pull_request_target) · dependabot-rhythm (pull_request_target). Densest knot: 9 workflows, event + dispatch doing overlapping slices of resolve-threads / arm / reconcile. This is where "piecemeal tinker toys" shows most.
G · Deploy / infra / misc cloud-run-deploy (PULLMAN) · 1password-secret-template (template only, manual) · janitor-sweep (workflow_run). Ghosts: claude-sign, code-coverage.
Where the eight re-enabled (2026-06-19 13:50) land
They cut across clusters — A (review-feedback-loop, agent-review-gate), B (agent-auto-pr, branch-cleanup, stale-bot-prs, + the auto-pr ghost), E (daily-rollover) — confirming they were a disable batch, not a functional unit. Kept enabled, they stay in the field of view for the redesign instead of orphaned.
Open seams for next passes (not prescriptions yet)
Map v1 — Topology figure (companion to the inventory comment above)
Workflow → script/engine call graph. Every edge is from the actual on: / .github/scripts/* references read from main — not inferred. Nodes with no edges in a cluster (e.g. the inline-gh arming workflows in A) call no shared script; they're shown for cluster membership.
flowchart LR
classDef engine fill:#8b0000,color:#fff,stroke:#000,stroke-width:3px
classDef eng2 fill:#b8860b,color:#fff,stroke:#000
classDef off fill:#ddd,stroke:#999,color:#555,stroke-dasharray:4 3
RFL("review_feedback_loop.py<br/>ENGINE - 7 callers"):::engine
IR("issue_reconciler.py<br/>ENGINE - 5 callers"):::engine
PRL("pr_lifecycle.py<br/>shared - 2"):::eng2
CSP("check_secret_patterns.py<br/>shared - 2"):::eng2
VC("validate_content.py<br/>shared - 2"):::eng2
subgraph A["A - review / merge engine"]
AME["auto-merge-engage"]
RFW["review-feedback-loop"]
RRS["review-response"]
EOD["engage-outdated"]
ARG["agent-review-gate (disabled*)"]:::off
LKW["looker-walk (read-only)"]
BAM["batch-arm-merge-queue (inline gh)"]
AMR["auto-merge-rhythm (inline gh)"]
DBR["dependabot-rhythm (inline gh)"]
end
subgraph B["B - agent-PR / branch lifecycle"]
AAP["agent-auto-pr"]
BCL["branch-cleanup"]
SBP["stale-bot-prs"]
BGR["branch-garden-report"]
end
subgraph CE["C / E - gates and surveys"]
LFW["large-file-watchdog"]
MDS["metadata-survey"]
SPP["secret-pattern-policy"]
SPF["secret-pattern-full-scan"]
VAC["validate-agent-content"]
SMI["swarm-mvp-intake"]
end
AME --> RFL
RFW --> RFL
RRS --> RFL
EOD --> RFL
ARG --> RFL
LKW --> RFL
AAP --> RFL
ARG --> IR
LKW --> IR
BGR --> IR
LFW --> IR
MDS --> IR
AAP --> PRL
BCL --> PRL
SBP -. "via stale_bot_prs.py" .-> PRL
SPP --> CSP
SPF --> CSP
VAC --> VC
SMI --> VC
MPC["meshnetweb_portability_check.py"] -. imports .-> CPP["check_portable_paths.py"]
Loading
Legend: dark red = engine script (many callers) · gold = 2-caller shared script · grey dashed = disabled.
Reading it: Cluster A is effectively one engine (review_feedback_loop.py) wearing six workflow hats; issue_reconciler.py is the shared "report findings to an issue" substrate under five more. That coupling — plus A's three inline-gh arming workflows that bypass the engine entirely — is the knot a holistic redesign has to resolve.
Not drawn (kept readable; full list in chat/Map v1): ~14 single-use 1:1 scripts (a workflow and its own script), and the 10 inline-gh workflows that call no script (auto-merge-rhythm, dependabot-rhythm, batch-arm-merge-queue, opencode, codeql, cloud-run-deploy, sync-dependencies, validate-daily-notes, wayback-preserve, 1password-secret-template).
Generated by Claude Code — topology figure for Map v1, #586.
Map v1 — Cluster A deep-dive (review/merge engine)
Runtime trace of the 9 Cluster A workflows + the review_feedback_loop.py engine, read from main. Every claim is sourced to file/line. Full version committed to the vault as REVIEW-MERGE-ENGINE-CLUSTER-A-DEEPDIVE-2026-06-20.md.
The engine is 4 tools sharing one 2,221-line file
review_feedback_loop.py exposes 16 subcommands (build_parser, L2042–2181) in four concerns:
Concern
Subcommands
What it does
A · Label substrate
ensure-labels
Create the lifecycle labels; every command calls it first.
Plus 3 inline-gh arming workflows calling no engine: auto-merge-rhythm (live, every PR event), dependabot-rhythm, batch-arm-merge-queue (manual bulk).
Three findings
F1 · Arming has two brains, and the engine's is dormant. The engine re-enabled arming 2026-06-17 (AGENT_AUTO_MERGE_ENABLED = True, L51), but the only workflow reaching its arming path (reconcile-open-prs) is agent-review-gate, explicitly disabled ("runs the engine's retired risk-tier schema… does not work against the current repo"). Live arming runs outside the engine — auto-merge-rhythm.yml's inline gh pr merge --auto, re-deriving protected-path checks in bash that the engine already encodes (PROTECTED_PATH_PATTERNS, L100). Two "safe to arm?" implementations; the live one is YAML, the dormant one Python.
F2 · auto-merge-engage.yml doesn't engage auto-merge. It calls engage-outdated --apply (L68–69) — engage_outdated (L1930) attest-resolves outdated bot-only threads and does not arm. The workflow named for arming resolves stale bot comments. Clearest symptom of name/concern drift.
F3 · The looker subsystem (D) is a separable tool wearing the engine's skin. 6 subcommands, own vocabulary, read-only toward PRs (looker_walk, L1760: "Read-only — resolves nothing"); it rides in the engine only to share GitHub-thread plumbing.
Shape of a single coherent engine
Three tools over one shared lib: review-state (A+B+C — event-driven projector making the single arm decision, folding auto-merge-rhythm's bash back in), thread-witness (D extracted whole), and a shared lib (_fetch_pr, thread walking, PROTECTED_PATH_PATTERNS, label specs). Collapses 9 workflows + 2 dormant/duplicated arming paths → 3 workflows over 1 engine, and kills the F2 naming drift.
Open questions: (1) revive the engine's dormant arming path, or is inline-gh the permanent replacement? (2) keep thread-witness on manual dispatch, or schedule it?
Generated by Claude Code — Cluster A deep-dive for Map v1, #586. Full doc: REVIEW-MERGE-ENGINE-CLUSTER-A-DEEPDIVE-2026-06-20.md.
Correction to the Cluster A deep-dive — risk-tier schema is NOT retired
Logan flagged that #597 was auto-labeled risk/high, which contradicts the unqualified "retired risk-tier schema" phrasing in F1 of the deep-dive above. He's right — that phrasing was a verbatim quote of agent-review-gate.yml's header, but stating it without scope was misleading. Re-read from main:
Producer — LIVE.agent-auto-pr.yml → classify_paths.py stamps risk/<tier> on every agent PR at creation (--label "risk/$RISK_TIER", L175). classify_paths.py is fail-safe — "Unknown paths default to high-risk" (L7/L67) — so Research: agentic GitHub review best-practices + Cluster A & reporting-substrate engine deep-dives #597's two new root-level .md files match no known low-risk pattern and rate high. That's why a docs-only PR is risk/high.
Consumer — LIVE.dependabot-rhythm.yml treats risk/high as a hard auto-merge block (L17/L90/L99); the engine's _risk_tier_for_pr (L295) reads it as canonical.
Consumer — DISABLED. The agent-review-gate reconcile/arming gate is the only thing the "retired risk-tier schema" header describes — its risk-based arming eligibility, not risk labeling as a whole.
Live non-Dependabot arming — ignores risk.auto-merge-rhythm.yml has norisk reference at all.
Net (this sharpens F1): risk-tier labeling is a live producer with exactly one live consumer (the Dependabot lane) and one dormant consumer (the disabled gate); the live general arming path doesn't read it. So risk/high on a normal agent PR like #597 is computed and stamped, but its only teeth are in the Dependabot lane — elsewhere it's a human-facing signal whose automated consumer is dormant.
The §5 redesign question gains a third item: decide whether review-state becomes the single live consumer of the risk tier, or whether risk labeling is retired in fact rather than only at the dead gate.
The vault doc (REVIEW-MERGE-ENGINE-CLUSTER-A-DEEPDIVE-2026-06-20.md, on PR #597) carries this as a witnessed ## 3·Correction section rather than a silent edit.
Next companion to Map v1. Investigated issue_reconciler.py + its 5 callers from origin/main, answering four open questions by evidence. Full doc: REPORTING-SUBSTRATE-ISSUE-RECONCILER-DEEPDIVE-2026-06-20.md (on PR #597).
Expected vs. found (stated up front): I expected to confirm this is the "clean shared substrate review_feedback_loop.py never became" and recommend leaving it alone. The read only partly supports that — it's focused (1 concern), but not flawless.
The engine. Single-mode (no subcommands): --title --body-file --has-findings --resolved-comment. reconcile_issue (L157): find OPEN issue by exact title → create / noop_duplicate (SHA256 body fingerprint) / comment / (no findings →) resolved-comment + close. Issues only, no labels, never touches PRs.
The 5 callers — uniform shape [generator → markdown → reconciler → one durable issue]: agent-review-gate (DISABLED) via pr_loop_watchdog.py; looker-walk via render-worklist; branch-garden-report; large-file-watchdog; metadata-survey. Each owns a distinct issue title (no collision).
Findings (evidence-driven, not pre-baked)
Q1 — sound or flawed? Mixed. Real fragilities, not a rubber stamp:
Brittle identity.find_open_issue_number (L69–89) matches an exact title string with --limit 20. A manual issue rename, or >20 matching open issues, orphans the report and spawns a duplicate. The fingerprint dedups bodies, not identity — identity has no stable key.
Side-effecting hash.ensure_body_fingerprint (L60–66) writes the fingerprint back into the body file on disk — idempotent/benign in CI, but mixes pure-hash with I/O and desyncs the on-disk report from the generator's output.
No close hysteresis — flapping findings → repeated close + recreate (churn).
Q2 — separable? Yes, but duplicates plumbing. No module coupling, but gh()/gh_json() (issue_reconciler) reimplements the same subprocess-raise pattern as _run()/_graphql() (review_feedback_loop). Candidate first member of the Cluster A §5 shared lib.
Q3 — how callers relate? One clean shared sink; variation is upstream. Two callers bridge into review_feedback_loop.py (agent-review-gate, looker-walk); three are standalone surveys. has_findings differs per caller — metadata-surveyhardcodes true (L45) so its issue never closes (permanent track: metadata/frontmatter standardization debt and schema enforcement #252 surface).
Q4 — implication for the §5 redesign? Modest, mostly affirming. This subsystem is not a knot — it already embodies the "extract a focused, reusable tool" pattern, evidence the 3-tool target is reachable. Two independent improvements, neither needing the big redesign: (1) unify the gh wrapper into the shared lib; (2) make issue identity marker-based (a <!-- reconciler-id:… --> tag) so renames/the 20-cap can't orphan a report.
Still unmapped: the ~14 single-use scripts (C/E singletons) and the ~10 inline-gh workflows that call no script.
Generated by Claude Code — reporting-substrate deep-dive for Map v1, #586. Full doc: REPORTING-SUBSTRATE-ISSUE-RECONCILER-DEEPDIVE-2026-06-20.md.
Map v1 — capstone: the automation long tail (singletons + inline-gh)
Final companion to Map v1. This maps the 27 workflow files outside the two shared spines, which closes Map v1: every one of the 39 workflow .yml files on main now sits in exactly one structure — spine-A, spine-reconciler, or the long tail. Read first-hand from origin/main. Full doc: AUTOMATION-LONG-TAIL-SINGLETONS-AND-INLINE-GH-DEEPDIVE-2026-06-20.md (on PR #597).
Q1 — Cluster C: one framework or N tinker toys? A loose family. The hypothesized "event-gate + scheduled-sweep + reconciler" triad is fully present in only one twin-pair (large-file-policy + large-file-watchdog→reconciler). secret-pattern has the event+scheduled halves but no reconciler issue (full-scan just fails). The other ~6 C gates are plain single-event guards. → A parameterized gate framework would absorb the scaffold, but reconciler-reporting should stay opt-in (only large-file uses it). Much lower-risk than the Cluster A engine.
Q2 — orphan scripts? Yes — 9 of ~37. An all-scripts-vs-all-workflows scan found 9 scripts no workflow references: audit_repo_payloads, backfill_daily_notes, bind_ai_book, date_tagger, expand_date_aliases, generate_name_forms, normalize_tags, obsidian_rest_api_client, phone_link_intake, tag_stubs. Not necessarily dead — their lane is not CI (hooks / local runtime / manual). The finding: ~24% of .github/scripts/ is invoked by something other than a workflow, and the inventory had no record of which. Identify each caller before any "delete unused" pass.
Q3 — inline-gh tail: right-sized or hidden complexity? 7 of 10 right-sized.codeql/opencode/cloud-run-deploy/1password-template are thin glue over vetted actions — correctly script-free. The 3 arming workflows are the Cluster A F1 smell (already covered). The two new flags: wayback-preserve reimplements URL-extraction in bash while its sibling wayback-audit already has wayback_audit.py (a domain split — F1 in miniature); sync-dependencies runs inline uv lock/gh pr create overlapping dependency-submission-uv's uv_dependency_submission.py.
Q4 — do ×2 scripts form secondary spines? Only one does.pr_lifecycle.py is a genuine micro-spine — a second writer of the lifecycle/* vocabulary alongside review_feedback_loop.ensure-labels → folds into the Cluster A §5 label-substrate question. check_secret_patterns.py (event+sched twin) and validate_content.py (shared safety gate) are healthy shared utilities, not engines.
What the whole topology now shows
The mass is healthy periphery, not knot — ~21 of 39 are single-purpose gates/surveys or right-sized action-glue, each with an explicit conservative charter.
The knot is concentrated in exactly one place — Cluster A's review_feedback_loop.py + its dormant-vs-bash arming split.
Three small, independent cleanups fall out, none needing the big redesign: a parameterized C gate framework (Q1); resolving the wayback/dependency bash-vs-script splits (Q3); accounting for the 9 orphan scripts (Q2). Plus one cross-cluster thread for the A redesign: pr_lifecycle.py as a second lifecycle/* writer (Q4).
Legend: dark red = engine script · gold = 2-caller shared script · blue = inline-gh/platform-action (no own script) · grey dashed = script with no workflow caller · light-grey dashed = disabled.
Generated by Claude Code — Map v1 capstone for #586. Full doc: AUTOMATION-LONG-TAIL-SINGLETONS-AND-INLINE-GH-DEEPDIVE-2026-06-20.md.
Activity
loganfinney27 commentedon Jun 20, 2026
Taking this over (agent: claude-code). Starting the holistic suite map #586 calls for — the automation system as one thing, not parts to patch or silence. This is Map v1: inventory + topology. Facts are from the Actions API + the
origin/mainfile tree + each workflow'son:/script refs (read just now); the functional clusters are my organizing lens, labeled as such; unresolved items are marked*.Inventory reconciliation — 53 registered = 40 files + 4 ghosts + 9 platform
main— the editable suite.main, so invisible in the tree yet still counted:auto-pr.yml,agent-branch-auto-pr.yml,claude-sign.yml,code-coverage.yml. (code-coverage= open PR Add initial code coverage workflow file #587, not yet merged; the other three are stale registrations.)codeql.yml), Automatic Dependency Submission (= thesubmit-pypiof CI gate:submit-pypiis both KNOWN_NOISE (always-fails) and a required-check gate for main auto-merge #406), Dependabot Updates, Dependency Graph, Copilot reviewer.The real coupling — two shared "spines" (the entanglement is in the scripts, not the YAML)
review_feedback_loop.py— 7 workflows: auto-merge-engage, review-feedback-loop, review-response, engage-outdated, agent-review-gate, looker-walk, agent-auto-pr. The nervous system of the merge/review engine.issue_reconciler.py— 5 workflows: agent-review-gate, branch-garden-report, large-file-watchdog, metadata-survey, looker-walk. Shared "open/update a findings issue" substrate.pr_lifecycle.py×2;check_secret_patterns.py×2;validate_content.py×2.)→ Redesigning the merge engine means redesigning around
review_feedback_loop.py, not file-by-file.Functional clusters (organizing lens)
A · Review / merge-queue engine — the contested core (#399 / #521 / #406; my #590/#591/#594 this session)
auto-merge-engage(pull_request_target) ·review-feedback-loop(issue_comment + pull_request_target) ·review-response(pull_request_review) ·engage-outdated(dispatch) ·agent-review-gate(dispatch; self-marked DISABLED, "retired risk-tier schema") ·batch-arm-merge-queue(dispatch) ·looker-walk(dispatch, read-only) ·auto-merge-rhythm(pull_request_target) ·dependabot-rhythm(pull_request_target).Densest knot: 9 workflows, event + dispatch doing overlapping slices of resolve-threads / arm / reconcile. This is where "piecemeal tinker toys" shows most.
B · Agent-PR & branch lifecycle
agent-auto-pr(create) ·branch-cleanup(pull_request + schedule) ·stale-bot-prs(schedule) ·branch-garden-report(schedule). Ghosts here:auto-pr,agent-branch-auto-pr→ 3 auto-PR variants, 2 orphaned.C · PR policy / CI gates
secret-pattern-policy·large-file-policy·check-portable-paths·check-dotfolder-anchors·laf-usb-manifest-policy·version-transition-policy·validate-daily-notes·validate-agent-content·check-notebooks-paired·cross-platform-smoke·codeql. Plus scheduled twins:secret-pattern-full-scan,large-file-watchdog.Pattern repeated N times: event-gate + scheduled-sweep + issue_reconciler reporting — candidate for one parameterized framework.
D · Dependency management
sync-dependencies·dependency-submission-uv(#588) ·dependabot-rhythm(shared with A) · platform: Dependabot Updates, Dependency Graph, Automatic Dependency Submission (#406submit-pypi).E · Vault content / metadata / archival / reporting
sort-audit(Topology Census) ·metadata-survey·daily-rollover·wayback-audit·wayback-preserve·sync-plugin-registry·sync-agents-bootstrap.F · Swarm / agents / intake
swarm-mvp-intake·opencode(OpenCode/ocinvocation) · platform agent apps (Claude, Codex, Copilot ×2).G · Deploy / infra / misc
cloud-run-deploy(PULLMAN) ·1password-secret-template(template only, manual) ·janitor-sweep(workflow_run). Ghosts:claude-sign,code-coverage.Where the eight re-enabled (2026-06-19 13:50) land
They cut across clusters — A (
review-feedback-loop,agent-review-gate), B (agent-auto-pr,branch-cleanup,stale-bot-prs, + theauto-prghost), E (daily-rollover) — confirming they were a disable batch, not a functional unit. Kept enabled, they stay in the field of view for the redesign instead of orphaned.Open seams for next passes (not prescriptions yet)
review_feedback_loop.py(event vs dispatch roles, debounce, the Reviewer multiplicity + Copilot-not-re-reviewing-on-push (restore + harden) #399 look-then-resolve design).*unresolved: who/why disabled the original set (Review-reconciliation workflows disabled 2026-05-26 → auto-queue arms but doesn't flow #521 marks*); whethermain's ruleset still requires the merge queue; the live armed-state of A's reconcilers.Next pass: a layer deeper on Cluster A (densest knot, live pain), unless directed elsewhere.
Generated by Claude Code — Map v1, durable record for #586.
Generated by Claude Code
loganfinney27 commentedon Jun 20, 2026
Map v1 — Topology figure (companion to the inventory comment above)
Workflow → script/engine call graph. Every edge is from the actual
on:/.github/scripts/*references read frommain— not inferred. Nodes with no edges in a cluster (e.g. the inline-gharming workflows in A) call no shared script; they're shown for cluster membership.flowchart LR classDef engine fill:#8b0000,color:#fff,stroke:#000,stroke-width:3px classDef eng2 fill:#b8860b,color:#fff,stroke:#000 classDef off fill:#ddd,stroke:#999,color:#555,stroke-dasharray:4 3 RFL("review_feedback_loop.py<br/>ENGINE - 7 callers"):::engine IR("issue_reconciler.py<br/>ENGINE - 5 callers"):::engine PRL("pr_lifecycle.py<br/>shared - 2"):::eng2 CSP("check_secret_patterns.py<br/>shared - 2"):::eng2 VC("validate_content.py<br/>shared - 2"):::eng2 subgraph A["A - review / merge engine"] AME["auto-merge-engage"] RFW["review-feedback-loop"] RRS["review-response"] EOD["engage-outdated"] ARG["agent-review-gate (disabled*)"]:::off LKW["looker-walk (read-only)"] BAM["batch-arm-merge-queue (inline gh)"] AMR["auto-merge-rhythm (inline gh)"] DBR["dependabot-rhythm (inline gh)"] end subgraph B["B - agent-PR / branch lifecycle"] AAP["agent-auto-pr"] BCL["branch-cleanup"] SBP["stale-bot-prs"] BGR["branch-garden-report"] end subgraph CE["C / E - gates and surveys"] LFW["large-file-watchdog"] MDS["metadata-survey"] SPP["secret-pattern-policy"] SPF["secret-pattern-full-scan"] VAC["validate-agent-content"] SMI["swarm-mvp-intake"] end AME --> RFL RFW --> RFL RRS --> RFL EOD --> RFL ARG --> RFL LKW --> RFL AAP --> RFL ARG --> IR LKW --> IR BGR --> IR LFW --> IR MDS --> IR AAP --> PRL BCL --> PRL SBP -. "via stale_bot_prs.py" .-> PRL SPP --> CSP SPF --> CSP VAC --> VC SMI --> VC MPC["meshnetweb_portability_check.py"] -. imports .-> CPP["check_portable_paths.py"]Legend: dark red = engine script (many callers) · gold = 2-caller shared script · grey dashed = disabled.
Reading it: Cluster A is effectively one engine (
review_feedback_loop.py) wearing six workflow hats;issue_reconciler.pyis the shared "report findings to an issue" substrate under five more. That coupling — plus A's three inline-gharming workflows that bypass the engine entirely — is the knot a holistic redesign has to resolve.Not drawn (kept readable; full list in chat/Map v1): ~14 single-use
1:1scripts (a workflow and its own script), and the 10 inline-ghworkflows that call no script (auto-merge-rhythm,dependabot-rhythm,batch-arm-merge-queue,opencode,codeql,cloud-run-deploy,sync-dependencies,validate-daily-notes,wayback-preserve,1password-secret-template).Generated by Claude Code — topology figure for Map v1, #586.
Generated by Claude Code
loganfinney27 commentedon Jun 20, 2026
Map v1 — Cluster A deep-dive (review/merge engine)
Runtime trace of the 9 Cluster A workflows + the
review_feedback_loop.pyengine, read frommain. Every claim is sourced to file/line. Full version committed to the vault asREVIEW-MERGE-ENGINE-CLUSTER-A-DEEPDIVE-2026-06-20.md.The engine is 4 tools sharing one 2,221-line file
review_feedback_loop.pyexposes 16 subcommands (build_parser, L2042–2181) in four concerns:ensure-labelssync-pr,review-submitted,acknowledge-apply,promote-ready,reconcile-open-prs,enable-auto-mergeverify-claimmergeable/checks; post a divergence note.list-unlooked,looker-walk,render-worklist,attest-resolve,engage-outdated,reconcile-witnessLive wiring
review-feedback-loopissue_comment, PR open/syncacknowledge-apply,sync-pr,verify-claimreview-responsereview-submittedauto-merge-engageengage-outdated --applyengage-outdatedengage-outdatedlooker-walklooker-walk+render-worklistagent-review-gatereconcile-open-prsagent-auto-prensure-labelsPlus 3 inline-
gharming workflows calling no engine:auto-merge-rhythm(live, every PR event),dependabot-rhythm,batch-arm-merge-queue(manual bulk).Three findings
F1 · Arming has two brains, and the engine's is dormant. The engine re-enabled arming 2026-06-17 (
AGENT_AUTO_MERGE_ENABLED = True, L51), but the only workflow reaching its arming path (reconcile-open-prs) isagent-review-gate, explicitly disabled ("runs the engine's retired risk-tier schema… does not work against the current repo"). Live arming runs outside the engine —auto-merge-rhythm.yml's inlinegh pr merge --auto, re-deriving protected-path checks in bash that the engine already encodes (PROTECTED_PATH_PATTERNS, L100). Two "safe to arm?" implementations; the live one is YAML, the dormant one Python.F2 ·
auto-merge-engage.ymldoesn't engage auto-merge. It callsengage-outdated --apply(L68–69) —engage_outdated(L1930) attest-resolves outdated bot-only threads and does not arm. The workflow named for arming resolves stale bot comments. Clearest symptom of name/concern drift.F3 · The looker subsystem (D) is a separable tool wearing the engine's skin. 6 subcommands, own vocabulary, read-only toward PRs (
looker_walk, L1760: "Read-only — resolves nothing"); it rides in the engine only to share GitHub-thread plumbing.Shape of a single coherent engine
Three tools over one shared lib:
review-state(A+B+C — event-driven projector making the single arm decision, foldingauto-merge-rhythm's bash back in),thread-witness(D extracted whole), and a shared lib (_fetch_pr, thread walking,PROTECTED_PATH_PATTERNS, label specs). Collapses 9 workflows + 2 dormant/duplicated arming paths → 3 workflows over 1 engine, and kills the F2 naming drift.Open questions: (1) revive the engine's dormant arming path, or is inline-
ghthe permanent replacement? (2) keepthread-witnesson manual dispatch, or schedule it?Generated by Claude Code — Cluster A deep-dive for Map v1, #586. Full doc:
REVIEW-MERGE-ENGINE-CLUSTER-A-DEEPDIVE-2026-06-20.md.Generated by Claude Code
Correct Cluster A deep-dive: risk-tier schema is not retired (#586)
loganfinney27 commentedon Jun 20, 2026
Correction to the Cluster A deep-dive — risk-tier schema is NOT retired
Logan flagged that #597 was auto-labeled
risk/high, which contradicts the unqualified "retired risk-tier schema" phrasing in F1 of the deep-dive above. He's right — that phrasing was a verbatim quote ofagent-review-gate.yml's header, but stating it without scope was misleading. Re-read frommain:agent-auto-pr.yml→classify_paths.pystampsrisk/<tier>on every agent PR at creation (--label "risk/$RISK_TIER", L175).classify_paths.pyis fail-safe — "Unknown paths default to high-risk" (L7/L67) — so Research: agentic GitHub review best-practices + Cluster A & reporting-substrate engine deep-dives #597's two new root-level.mdfiles match no known low-risk pattern and ratehigh. That's why a docs-only PR isrisk/high.dependabot-rhythm.ymltreatsrisk/highas a hard auto-merge block (L17/L90/L99); the engine's_risk_tier_for_pr(L295) reads it as canonical.agent-review-gatereconcile/arming gate is the only thing the "retired risk-tier schema" header describes — its risk-based arming eligibility, not risk labeling as a whole.auto-merge-rhythm.ymlhas noriskreference at all.Net (this sharpens F1): risk-tier labeling is a live producer with exactly one live consumer (the Dependabot lane) and one dormant consumer (the disabled gate); the live general arming path doesn't read it. So
risk/highon a normal agent PR like #597 is computed and stamped, but its only teeth are in the Dependabot lane — elsewhere it's a human-facing signal whose automated consumer is dormant.The §5 redesign question gains a third item: decide whether
review-statebecomes the single live consumer of the risk tier, or whether risk labeling is retired in fact rather than only at the dead gate.The vault doc (
REVIEW-MERGE-ENGINE-CLUSTER-A-DEEPDIVE-2026-06-20.md, on PR #597) carries this as a witnessed## 3·Correctionsection rather than a silent edit.Generated by Claude Code — witnessed correction, #586.
Generated by Claude Code
Add reporting-substrate deep-dive: issue_reconciler.py map (#586)
loganfinney27 commentedon Jun 20, 2026
Map v1 — Reporting substrate (
issue_reconciler.py) deep-diveNext companion to Map v1. Investigated
issue_reconciler.py+ its 5 callers fromorigin/main, answering four open questions by evidence. Full doc:REPORTING-SUBSTRATE-ISSUE-RECONCILER-DEEPDIVE-2026-06-20.md(on PR #597).Expected vs. found (stated up front): I expected to confirm this is the "clean shared substrate
review_feedback_loop.pynever became" and recommend leaving it alone. The read only partly supports that — it's focused (1 concern), but not flawless.The engine. Single-mode (no subcommands):
--title --body-file --has-findings --resolved-comment.reconcile_issue(L157): find OPEN issue by exact title → create /noop_duplicate(SHA256 body fingerprint) / comment / (no findings →) resolved-comment + close. Issues only, no labels, never touches PRs.The 5 callers — uniform shape [generator → markdown → reconciler → one durable issue]:
agent-review-gate(DISABLED) viapr_loop_watchdog.py;looker-walkviarender-worklist;branch-garden-report;large-file-watchdog;metadata-survey. Each owns a distinct issue title (no collision).Findings (evidence-driven, not pre-baked)
find_open_issue_number(L69–89) matches an exact title string with--limit 20. A manual issue rename, or >20 matching open issues, orphans the report and spawns a duplicate. The fingerprint dedups bodies, not identity — identity has no stable key.ensure_body_fingerprint(L60–66) writes the fingerprint back into the body file on disk — idempotent/benign in CI, but mixes pure-hash with I/O and desyncs the on-disk report from the generator's output.gh()/gh_json()(issue_reconciler) reimplements the same subprocess-raise pattern as_run()/_graphql()(review_feedback_loop). Candidate first member of the Cluster A §5 shared lib.review_feedback_loop.py(agent-review-gate,looker-walk); three are standalone surveys.has_findingsdiffers per caller —metadata-surveyhardcodestrue(L45) so its issue never closes (permanent track: metadata/frontmatter standardization debt and schema enforcement #252 surface).ghwrapper into the shared lib; (2) make issue identity marker-based (a<!-- reconciler-id:… -->tag) so renames/the 20-cap can't orphan a report.Still unmapped: the ~14 single-use scripts (C/E singletons) and the ~10 inline-
ghworkflows that call no script.Generated by Claude Code — reporting-substrate deep-dive for Map v1, #586. Full doc:
REPORTING-SUBSTRATE-ISSUE-RECONCILER-DEEPDIVE-2026-06-20.md.Generated by Claude Code
Add automation long-tail deep-dive: Map v1 capstone (#586)
loganfinney27 commentedon Jun 20, 2026
Map v1 — capstone: the automation long tail (singletons + inline-
gh)Final companion to Map v1. This maps the 27 workflow files outside the two shared spines, which closes Map v1: every one of the 39 workflow
.ymlfiles onmainnow sits in exactly one structure — spine-A, spine-reconciler, or the long tail. Read first-hand fromorigin/main. Full doc:AUTOMATION-LONG-TAIL-SINGLETONS-AND-INLINE-GH-DEEPDIVE-2026-06-20.md(on PR #597).The partition — 39 = 12 spine + 27 long tail
pr_lifecycle.py,check_secret_patterns.py,validate_content.py(2 callers each) +cross-platform-smoke/swarm-mvp-intake(multi-script)gh/ no own script1password-secret-template,cloud-run-deploy,codeql,opencode,sync-dependencies,validate-daily-notes,wayback-preserveFindings (evidence-driven)
large-file-policy+large-file-watchdog→reconciler).secret-patternhas the event+scheduled halves but no reconciler issue (full-scan just fails). The other ~6 C gates are plain single-event guards. → A parameterized gate framework would absorb the scaffold, but reconciler-reporting should stay opt-in (onlylarge-fileuses it). Much lower-risk than the Cluster A engine.audit_repo_payloads,backfill_daily_notes,bind_ai_book,date_tagger,expand_date_aliases,generate_name_forms,normalize_tags,obsidian_rest_api_client,phone_link_intake,tag_stubs. Not necessarily dead — their lane is not CI (hooks / local runtime / manual). The finding: ~24% of.github/scripts/is invoked by something other than a workflow, and the inventory had no record of which. Identify each caller before any "delete unused" pass.ghtail: right-sized or hidden complexity? 7 of 10 right-sized.codeql/opencode/cloud-run-deploy/1password-templateare thin glue over vetted actions — correctly script-free. The 3 arming workflows are the Cluster A F1 smell (already covered). The two new flags:wayback-preservereimplements URL-extraction in bash while its siblingwayback-auditalready haswayback_audit.py(a domain split — F1 in miniature);sync-dependenciesruns inlineuv lock/gh pr createoverlappingdependency-submission-uv'suv_dependency_submission.py.pr_lifecycle.pyis a genuine micro-spine — a second writer of thelifecycle/*vocabulary alongsidereview_feedback_loop.ensure-labels→ folds into the Cluster A §5 label-substrate question.check_secret_patterns.py(event+sched twin) andvalidate_content.py(shared safety gate) are healthy shared utilities, not engines.What the whole topology now shows
review_feedback_loop.py+ its dormant-vs-bash arming split.pr_lifecycle.pyas a secondlifecycle/*writer (Q4).Capstone topology figure — all 39 files
flowchart LR classDef engine fill:#8b0000,color:#fff,stroke:#000,stroke-width:3px classDef eng2 fill:#b8860b,color:#fff,stroke:#000 classDef inline fill:#1f6feb,color:#fff,stroke:#000 classDef orphan fill:#444,color:#fff,stroke:#000,stroke-dasharray:4 3 classDef off fill:#ddd,stroke:#999,color:#555,stroke-dasharray:4 3 RFL("review_feedback_loop.py<br/>ENGINE · 7 callers"):::engine IR("issue_reconciler.py<br/>ENGINE · 5 callers"):::engine PRL("pr_lifecycle.py<br/>shared x2"):::eng2 CSP("check_secret_patterns.py<br/>shared x2"):::eng2 VC("validate_content.py<br/>shared x2"):::eng2 subgraph A["A - review / merge engine"] AME["auto-merge-engage"] RFW["review-feedback-loop"] RRS["review-response"] EOD["engage-outdated"] ARG["agent-review-gate (disabled*)"]:::off LKW["looker-walk (read-only)"] BAM["batch-arm-merge-queue - inline gh"]:::inline AMR["auto-merge-rhythm - inline gh"]:::inline DBR["dependabot-rhythm - inline gh"]:::inline end subgraph B["B - agent-PR / branch lifecycle"] AAP["agent-auto-pr"] BCL["branch-cleanup"] SBP["stale-bot-prs -> stale_bot_prs.py"] BGR["branch-garden-report -> branch_garden_report.py"] end subgraph C["C - PR policy / CI gates (event + scheduled twins)"] SPP["secret-pattern-policy (event)"] SPF["secret-pattern-full-scan (sched)"] LFP["large-file-policy (event) -> check_large_files.py"] LFW["large-file-watchdog (sched) -> large_file_watchdog.py"] CPP["check-portable-paths -> check_portable_paths.py"] CDA["check-dotfolder-anchors -> check_dotfolder_anchors.py"] CNP["check-notebooks-paired -> jupytext_sync_paired.py"] LUM["laf-usb-manifest-policy -> laf_usb_manifest.py"] VTP["version-transition-policy -> check_version_transitions.py"] VAC["validate-agent-content"] VDN["validate-daily-notes - inline gh"]:::inline XPS["cross-platform-smoke -> codex_work_guard + meshnetweb"] CQL["codeql - platform action"]:::inline end subgraph E["E - content / metadata / archival"] MDS["metadata-survey -> metadata_survey.py"] SRT["sort-audit -> topology_census.py"] DRO["daily-rollover -> daily_rollover.py"] WBA["wayback-audit -> wayback_audit.py"] WBP["wayback-preserve - inline gh"]:::inline SPR["sync-plugin-registry -> sync_obsidian_plugin_registry.py"] SAB["sync-agents-bootstrap -> generate_agents_bootstrap.py"] end subgraph DF["D - dependencies"] DSU["dependency-submission-uv -> uv_dependency_submission.py"] SDP["sync-dependencies - inline gh"]:::inline end subgraph FG["F/G - swarm / deploy / infra"] SMI["swarm-mvp-intake -> swarm_mvp_intake + update_manifest"] OPC["opencode - platform action"]:::inline CRD["cloud-run-deploy (PULLMAN) - inline gh"]:::inline OPW["1password-secret-template - manual template"]:::inline JNS["janitor-sweep -> janitor_sweep.py"] end subgraph ORPH["orphans - scripts with NO workflow caller (9)"] O1["audit_repo_payloads / backfill_daily_notes / bind_ai_book"]:::orphan O2["date_tagger / expand_date_aliases / generate_name_forms"]:::orphan O3["normalize_tags / obsidian_rest_api_client / phone_link_intake / tag_stubs"]:::orphan end AME --> RFL RFW --> RFL RRS --> RFL EOD --> RFL ARG --> RFL LKW --> RFL AAP --> RFL ARG --> IR LKW --> IR BGR --> IR LFW --> IR MDS --> IR AAP --> PRL BCL --> PRL SPP --> CSP SPF --> CSP VAC --> VC SMI --> VCLegend: dark red = engine script · gold = 2-caller shared script · blue = inline-
gh/platform-action (no own script) · grey dashed = script with no workflow caller · light-grey dashed = disabled.Generated by Claude Code — Map v1 capstone for #586. Full doc:
AUTOMATION-LONG-TAIL-SINGLETONS-AND-INLINE-GH-DEEPDIVE-2026-06-20.md.Generated by Claude Code
Address review threads on #597: fix subcommand count + cross-refs (#586)
Add Cluster A deep-dive: review/merge engine map (#586)
Correct Cluster A deep-dive: risk-tier schema is not retired (#586)
18 remaining items
fix: label issue reference so #586 stops sitting at column one
Add Cluster A deep-dive: review/merge engine map (#586)
Correct Cluster A deep-dive: risk-tier schema is not retired (#586)
Add reporting-substrate deep-dive: issue_reconciler.py map (#586)
Add automation long-tail deep-dive: Map v1 capstone (#586)
Address review threads on #597: fix subcommand count + cross-refs (#586)
fix: rewrap prose so issue references stop landing at column one
fix: label issue reference so #586 stops sitting at column one
Add Cluster A deep-dive: review/merge engine map (#586)
Correct Cluster A deep-dive: risk-tier schema is not retired (#586)
Add reporting-substrate deep-dive: issue_reconciler.py map (#586)
Add automation long-tail deep-dive: Map v1 capstone (#586)
Address review threads on #597: fix subcommand count + cross-refs (#586)
fix: rewrap prose so issue references stop landing at column one
fix: label issue reference so #586 stops sitting at column one