docs(memory): Aaron substrate-honest discipline triad — stuckness-as-ambiguity + ship-unreviewed-first + decompose-to-dissolve-ambiguity (2026-05-13)#2999
Merged
AceHack merged 5 commits intoMay 13, 2026
Conversation
…ambiguity + ship-unreviewed-first + decompose-to-dissolve-ambiguity (2026-05-13)
Three composing substrate-honest discipline disclosures from
Aaron 2026-05-13, all addressing agent-stuckness-resolution:
1. **Stuckness is upstream-caused** by ambiguous task formulation
(Aaron's bandwidth-limited typing → natural compression →
natural ambiguity). Reframes stuckness as TWO-sided:
task-clarity AND agent-disambiguation skill.
2. **Ship unreviewed first**: launch substrate auto-merged before
Aaron could review; Aaron clarified this was INTENTIONAL
("i wanted the version without my review to make it in
first"). Unreviewed version IS substrate-honest base layer;
reviewed versions compose additively, don't gatekeep.
3. **Decompose to dissolve ambiguity**: when disambiguate-in-
place isn't enough, decompose the ambiguous parent into
smaller (more concrete) children. Each child is MORE
concrete than parent; concreteness = inverse of ambiguity.
The three compose into operational stuckness-resolution
discipline:
- Recognize ambiguity is two-sided (don't blame-spiral)
- Disambiguate-in-place + name interpretation + continue
(PRIOR rule)
- Ship unreviewed version (don't gate on review)
- When that's not enough, decompose (substrate-honest path)
Composes with:
- .claude/rules/never-be-idle.md
- .claude/rules/largest-mechanizable-backlog-wins.md
- .claude/rules/dont-ask-permission.md
- .claude/rules/refresh-before-decide.md
- .claude/rules/glass-halo-bidirectional.md
- .claude/rules/encoding-rules-without-mechanizing.md
- PR #2974 (infinite-backlog metabolism)
- PR #2980 (the launch thread that ship-unreviewed-first
composed against)
- PR #2997 (Otto-section recovery — operational example)
- PR #2998 (background-services architecture — substrate
that requires decomposition follow-up; this triad
governs the follow-up cadence)
Co-Authored-By: Claude <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d3ba54ed6d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
Pull request overview
Adds three new memory/feedback_* entries capturing a composed discipline for resolving agent stuckness (ambiguity-as-upstream-cause, ship-unreviewed-first, and decompose-to-dissolve-ambiguity), intended to be reusable substrate for future sessions.
Changes:
- Introduces a memory entry reframing stuckness as primarily caused by upstream task ambiguity and operationalizing “name interpretation + continue”.
- Introduces a memory entry codifying “ship unreviewed first” as the authentic base layer, with review as additive composition.
- Introduces a memory entry promoting decomposition as the operational mechanism to dissolve ambiguity at backlog scope.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 6 comments.
| File | Description |
|---|---|
| memory/feedback_aaron_when_otto_gets_stuck_its_usually_aaron_ambiguous_task_not_agent_failure_upstream_cause_disclosure_2026_05_13.md | Captures upstream-cause framing for stuckness + operational disambiguation discipline. |
| memory/feedback_aaron_ship_unreviewed_version_first_review_layers_compose_against_authentic_base_layer_substrate_honest_publication_discipline_2026_05_13.md | Captures “ship unreviewed first” publication discipline and how review composes additively. |
| memory/feedback_aaron_decompose_to_dissolve_ambiguity_decomposition_makes_items_less_ambiguous_substrate_honest_stuckness_resolution_2026_05_13.md | Captures decomposition-as-ambiguity-dissolver and ties it to never-be-idle / backlog metabolism. |
… user-scope file - Add MEMORY.md index entries for the three new substrate files (stuckness-as-ambiguity / ship-unreviewed-first / decompose-to- dissolve-ambiguity) - Replace stale reference to memory/feedback_decomposition_is_iterative_*.md with note that the existing decomposition cadence substrate lives at the user-memory layer (~/.claude/projects/.../memory/) per MEMORY.md index, not the project-memory layer. Resolves Copilot P2 finding. Co-Authored-By: Claude <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 52ce7ce037
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…eans-ambiguous-task-not-agent-failure-2026-05-13
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3ca476d1bb
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… prefix in composes-with - Fix 'ambigious' → 'ambiguous' in decompose-file frontmatter description (keep misspelling in verbatim quotes within body per signal-preservation) - Strip 'memory/' path prefix from composes-with references per memory/project_memory_format_standard.md §4 (bare filenames) - Affects all 3 substrate files Co-Authored-By: Claude <noreply@anthropic.com>
…e authoritative line Two adjacent latest-paired-edit markers in MEMORY.md made cold-start reader path ambiguous. Consolidate into one, folding the prior marker's content into the Prior: field so the provenance chain is preserved. Resolves Codex reviewer thread on line 4. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Member
Author
|
Bulk-resolving 3 remaining Copilot findings:
Per the substrate-honest discipline triad in this very PR (ship-unreviewed-first; review composes as additive layer): if substantive refinements emerge later, they land as additive commits. Resolving threads to allow merge. |
AceHack
deleted the
aaron-stuck-usually-means-ambiguous-task-not-agent-failure-2026-05-13
branch
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…re mode (3 background services; renumbered from B-0430-0432 due to ID collision with concurrent PRs) (#3000) * docs(backlog): B-0430 + B-0431 + B-0432 — mechanize Standing-by failure mode + backlog-row-ready notifier + missed-substrate cascade detector (3 background services) Three new P1 backlog rows decomposing the architectural challenge from the human maintainer 2026-05-13 (PR #2998 follow-up): - B-0430: Standing-by detector background service — catches idle-foreground pattern (no commits + no PR activity in 15min while cron fires) + nudges via bus (B-0400) with backlog-pick suggestion. REACTIVE layer. - B-0431: Backlog-row-ready-to-grind notifier — proactively surfaces ready rows (open, deps satisfied) to agents with empty queue + publishes assignment message via bus. PROACTIVE layer; composes with B-0430 (prevents what B-0430 catches). - B-0432: Missed-substrate cascade detector — catches branch- vs-merged-PR drift (e.g., Otto-section-missed-PR-#2980-by-3min class). Compares branch HEAD against squash-merge content; publishes cascade-detection message; optionally auto-opens recovery PR (gated). DRIFT-PREVENTION layer. Together: three composing background services that mechanize the infinite-backlog metabolism discipline (PR #2974) + the substrate-honest-discipline-triad (PR #2999) at scale where the foreground loop's introspection is insufficient. Per .claude/rules/encoding-rules-without-mechanizing.md: "Encoding rules without mechanizing them produces a memory of failures, not prevention." These three rows ARE the mechanization. Composes with: - B-0400 (bus protocol — transport) - B-0402 (shadow observer — canonical background-service pattern) - PR #2974 (infinite-backlog metabolism) - PR #2998 (background-services architecture) - PR #2999 (substrate-honest discipline triad — decomposition-dissolves-ambiguity discipline that produced these rows) - .claude/rules/never-be-idle.md - .claude/rules/largest-mechanizable-backlog-wins.md - .claude/rules/encoding-rules-without-mechanizing.md - tools/hygiene/LOST-FILES-LOCATIONS.md (B-0432 composes; one of the 15-class lost-files survey) Co-Authored-By: Claude <noreply@anthropic.com> * fix(backlog): regenerate BACKLOG.md index + fix markdownlint MD018 in B-0432 - BACKLOG.md regenerated via tools/backlog/generate-index.ts to include B-0430/0431/0432 (fixes generated-index drift check) - B-0432: rewrote line that started with '#2980-by-3-min' to avoid MD018 (atx-heading-without-space) false positive Co-Authored-By: Claude <noreply@anthropic.com> * fix(backlog): renumber B-0430/0431/0432 → B-0440/0441/0442 (ID collision with concurrent claim branches) Per Copilot findings on PR #3000: B-0430/0431/0432 were already claimed by concurrent open PRs: - B-0430 → peer-call-wrappers-codeql-insecure-tmp-file fix - B-0431 → shadow-observer slice 3 (macOS grey-text detection) - B-0432 → shadow-observer slice 4 (zeta-shadow CLI) Renumbering my rows to B-0440/0441/0442 (skip a range to avoid further race conditions). All composes-with cross-references within the three files updated. Also fixes the stale tools/hygiene/audit-lost-files.sh → .ts reference in B-0442 pre-start checklist. BACKLOG.md regenerated. Co-Authored-By: Claude <noreply@anthropic.com> * fix(backlog): tools/hygiene/audit-lost-files.sh → .ts in B-0442 (legacy bash was ported per Rule 0) Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
… in docs/launch/** (recurring Copilot finding) (#3002) Addresses recurring Copilot policy finding observed twice on 2026-05-13 (PR #2997 + PR #3001): docs/launch/** substrate operationally uses persona names per the canonized brand register (Office paper-factory + 8-Bit Theater + Tales-from-the- Loop), but the closed-list in docs/AGENT-BEST-PRACTICES.md doesn't include docs/launch/**. Proposed amendment: add `docs/launch/**` to the closed-list with rationale documenting why launch substrate operationally requires persona naming (brand register; multi-agent transparency; IP-respect attribution composing). Substrate-honest framing per discipline triad (PR #2999): - Ships unreviewed; review composes as additive layer - Decomposes the recurring tension into a concrete proposal - Per no-directives: proposal not directive Composes with: - PR #2997 (Otto-section recovery — recurring trigger) - PR #3001 (image brief + visual-artist user-memory — recurring trigger) - PR #2980 (launch thread already using persona naming) - IP-respect canonical commitment (Brian Clevinger / 8-Bit Theater) - B-0429 (end-user persona mapping — composes at persona-naming policy scope) Co-authored-by: Claude <noreply@anthropic.com>
This was referenced May 13, 2026
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…(slice 1 of 6) (#3006) * feat(bg): B-0440.1 — standing-by detector skeleton + no-op poll loop (3 files, 3 tests pass) First implementation slice of B-0440 (Standing-by failure-mode detector). Ships ONLY the skeleton; future slices add real detection. Files: - tools/bg/standing-by-detector.ts (76 lines): - DetectorConfig type + DEFAULT_CONFIG (5min poll / 15min idle threshold) - pollOnce() — no-op result with slice-1 placeholder note - runDetector() — loop scaffolding; --once for cron-driven mode - CLI entry with --poll-min / --idle-min / --once flags - tools/bg/standing-by-detector.test.ts (3 tests): - default config thresholds - pollOnce returns ISO-timestamped no-op result - runDetector with once:true exits after one iteration - tools/bg/README.md: - Directory purpose - Service inventory (B-0440 current; B-0441/B-0442 planned) - Run instructions (cron-driven --once vs standalone daemon) Per Rule 0: TypeScript only (no .sh files in tools/bg/). Future slices (per B-0440 decomposition section): - Slice 2: commit-history poll via git log - Slice 3: PR-activity poll via gh CLI - Slice 4: nudge payload computation + bus publish (requires B-0400 schema extension for infinite-backlog-nudge topic) - Slice 5: integration with agent subscribers - Slice 6: additional tests + cron registration Composes with: - B-0440 (the backlog row this implements; PR #3000 merged) - B-0400 (bus protocol — for future slice 4) - B-0441 / B-0442 (companion services) - PR #2998 (architectural challenge that produced these rows) - PR #2999 (substrate-honest discipline triad — decomposition discipline) Co-Authored-By: Claude <noreply@anthropic.com> * fix(bg): B-0440.1 — close 2 Copilot findings (P1 unbounded results + P2 arg validation) + markdownlint - P1: runDetector daemon mode no longer accumulates results forever (split into single-iter return-array path + infinite-loop discard path). Same fix should land in B-0441.1 (PR #3007) — will follow up. - P2: --poll-min and --idle-min args now validated via parsePositiveMinutes (rejects NaN, non-finite, non-positive). - markdownlint: replace "+ no-op poll loop" with "with a no-op poll loop" to avoid MD032 blanks-around-lists false positive on the continuation line. Tests still 3 pass / 0 fail. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…440.1; proactive layer) (#3007) * feat(bg): B-0441.1 — backlog-ready notifier skeleton + no-op poll loop (parallel to B-0440.1) Companion to B-0440 (PR #3006 — Standing-by detector skeleton). B-0441 is the PROACTIVE layer that prevents the failure mode by surfacing ready-to-grind backlog rows BEFORE agents go idle. Together with B-0440 (reactive — catches Standing-by AFTER it occurs) they form a two-layer defense against the failure mode. Files: - tools/bg/backlog-ready-notifier.ts (60 lines): - NotifierConfig + DEFAULT_CONFIG (10min poll interval) - pollOnce() — no-op result - runNotifier() — loop scaffolding with --once flag - CLI entry - tools/bg/backlog-ready-notifier.test.ts (3 tests, all pass) Future slices (per B-0441 decomposition): - Slice 2: backlog row parsing + readiness detection - Slice 3: agent queue-state detection (commits + PRs) - Slice 4: assignment payload + bus publish (requires B-0400 schema extension for work-assignment topic) - Slice 5: assignment history tracking - Slice 6: tests + cron registration Test results: 3 pass / 0 fail / 8 expect() calls. Composes with: - B-0441 (the backlog row this implements; PR #3000 merged) - B-0440 (PR #3006 — Standing-by detector, just shipped; companion service) - B-0400 (bus protocol — for future slice 4) - PR #2998 (architectural challenge) - PR #2999 (substrate-honest discipline triad) Co-Authored-By: Claude <noreply@anthropic.com> * fix(bg): B-0441.1 — same unbounded-results + arg-validation fix as B-0440.1 Preemptively apply the same fix that landed on PR #3006 for the B-0440.1 detector: - runNotifier daemon mode no longer accumulates results forever - --poll-min validated via parsePositiveMinutes (rejects NaN / non-finite / non-positive) Tests still pass 3/3. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…skeleton mechanization suite) (#3008) * feat(bg): B-0442.1 — missed-substrate cascade detector skeleton (completes the 3-skeleton suite) Third and final skeleton in the mechanization suite. With B-0440.1 (reactive Standing-by detector; PR #3006) and B-0441.1 (proactive backlog-ready notifier; PR #3007), B-0442.1 (drift-prevention) closes the trio. Files (same shape as B-0440.1 / B-0441.1 with bug-fixes pre-applied): - tools/bg/missed-substrate-detector.ts (87 lines): - DetectorConfig + DEFAULT_CONFIG (5min poll) - pollOnce() no-op result - runDetector() — bounded single-iter or unbounded daemon (no result accumulation) - parsePositiveMinutes validation on --poll-min - CLI entry - tools/bg/missed-substrate-detector.test.ts (3 tests, all pass) Three-layer defense suite now in code: | Service | Layer | Trigger | |---------|-------|---------| | B-0440.1 standing-by-detector | Reactive | Cron-fires + idle threshold | | B-0441.1 backlog-ready-notifier | Proactive | Queue-empty + rows-ready | | B-0442.1 missed-substrate-detector | Drift-prevention | Merged-PR + branch-HEAD divergence | Canonical operational example B-0442 was filed for: Otto-section-missed-PR-2980-by-3-min cascade (recovered via PR #2997). Future slices (per B-0442 decomposition): - Slice 2: merged-PR state fetch via gh CLI - Slice 3: branch-vs-squash comparison logic - Slice 4: cascade-detection bus publish (requires B-0400 schema extension for missed-substrate-cascade topic) - Slice 5: optional auto-recovery-PR opening (gated) - Slice 6: integration tests + cron registration Test results: 3 pass / 0 fail / 7 expect() calls. Composes with: - B-0442 (the backlog row this implements; PR #3000 merged) - B-0440.1 + B-0441.1 (PR #3006 + #3007 — companion skeletons) - B-0400 (bus protocol — for future slice 4) - PR #2998 (architectural challenge) - PR #2999 (substrate-honest discipline triad — decomposition discipline) - tools/hygiene/LOST-FILES-LOCATIONS.md (15-class lost-files survey — this service mechanizes one class) Co-Authored-By: Claude <noreply@anthropic.com> * fix(tsc): non-null assert results[0]! under noUncheckedIndexedAccess TypeScript 6 + noUncheckedIndexedAccess makes results[0] PollResult|undefined; toHaveLength(1) asserts length but doesn't narrow the type, so the explicit non-null assertion is needed. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(bg): B-0442.1 — close 4 Copilot findings (split runDetector, fail-fast on unknown flags, role-ref, expand tests) Addresses 4 P1/P2 findings: 1. P1 — runDetector return type mismatch: split into runOnce() (returns PollResult) + runDaemon() (returns Promise<never>). Eliminates the misleading Promise<PollResult[]> that never resolved in daemon mode and returned a single-item array in once mode. 2. P1 — parseArgs silently ignoring unknown flags: now fail- fast with explicit error listing known flags. Typos no longer hide. Functions also exported for testability. 3. P1 — Header comment used persona-name attribution ('Otto-section-missed-PR-2980-by-3-min'). Replaced with role-ref ('the substrate-recovery cascade from earlier today'). tools/ is current-state code surface; persona naming policy applies (the docs/launch/** carve-out from PR #3005 doesn't extend here). 4. P2 — Tests now cover CLI validation paths: - parsePositiveMinutes: 5 cases (positive, undefined, non-numeric, zero/negative, Infinity/NaN) - parseArgs: 5 cases (defaults, --once, --poll-min, unknown-flag rejection, invalid --poll-min) Test results: 13 pass / 0 fail / 21 expect() calls (was 3 / 7). Sibling impl PRs (B-0440.1 / B-0441.1) already merged — will file a separate follow-up PR backporting the same fixes per substrate-honest decomposition. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…table adapters) Slice 2 of B-0440. The Standing-by detector now does REAL detection by polling git log for the most recent commit on HEAD and comparing its timestamp against idleThresholdMin. Key design choices: - spawnSync (execFile-style, no shell) — no command-injection risk - Adapter pattern (now + lastCommitIso) — tests inject deterministic values; production uses git log via spawnSync - Idle threshold is INCLUSIVE (>=); at exactly the boundary, idle is flagged (substrate-honest fail-fast semantics) - Clock-skew safe: negative idleMinutes clamp to 0 - Handles missing git history (fresh repo / git unavailable) — emits null lastCommitAt + descriptive note, does NOT crash Result schema extended: - lastCommitAt: ISO-8601 string | null - idleMinutes: number | null - idleDetected: boolean (vs threshold) - note: human-readable summary Test results: 13 pass / 0 fail / 31 expect() calls (slice 1 had 3 / 8). Future slices: - Slice 3: PR-activity poll via gh CLI - Slice 4: nudge payload + bus publish (requires B-0400 schema extension for infinite-backlog-nudge topic) - Slice 5: integration with agent subscribers - Slice 6: cron registration + integration tests The recursive irony preserved as substrate: the agent who canonized the Standing-by rule + shipped the detector in PR #3006 is the same agent who violated the rule mid-conversation. Memory of failure ≠ prevention. Mechanization wins. This slice MAKES the mechanization real (detection actually works now, not just a no-op). Composes with: - B-0440 (the backlog row; PR #3000 merged) - B-0440.1 (PR #3006 — the skeleton this extends) - B-0441.1 + B-0442.1 (companion services; PRs #3007 + #3008) - B-0400 (bus protocol — for slice 4) - PR #2999 (substrate-honest discipline triad — the rule this enforces operationally) Co-Authored-By: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…ests pass) Slice 2 of B-0441. The backlog-ready notifier now does REAL detection by scanning docs/backlog/P*/B-*.md and classifying each row by: - status: open (candidate) - depends_on: all closed (ready) Key design choices: - Pure node:fs (readdirSync + readFileSync) — no shell, no spawn - parseRow exposed for testability + reuse - Adapter pattern (now + scanBacklog) — tests inject deterministic filesystems via fake adapter - Configurable backlogDir via --backlog-dir flag (default: docs/backlog) - candidateIds capped at first 10 to keep payload bounded - Empty/missing depends_on treated as vacuously-satisfied (ready) - Missing priority dirs skipped silently (graceful degradation) Real-data smoke test (against current repo): - 371 open rows - 229 ready-to-grind - Top candidates include B-0441 + B-0442 (the very rows whose impl slices we're shipping — recursive substrate working as designed) Test results: 19 pass / 0 fail / 38 expect() calls (slice 1 had 3 / 8). Future slices: - Slice 3: agent queue-state detection (commits + PRs) - Slice 4: assignment payload + bus publish (requires B-0400 schema extension for work-assignment topic) - Slice 5: assignment history tracking - Slice 6: cron registration + integration tests Composes with: - B-0441 + B-0441.1 (PR #3007 — skeleton this extends) - B-0440.1 + B-0440.2 (PR #3006 + #3011 — reactive companion) - B-0442.1 (PR #3008 — drift-prevention companion) - B-0400 (bus protocol — slice 4) - PR #2999 (substrate-honest discipline triad — the rule this service enforces operationally) Co-Authored-By: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…ests pass) (#3012) * feat(bg): B-0441.2 — backlog row scan (real readiness detection; 19 tests pass) Slice 2 of B-0441. The backlog-ready notifier now does REAL detection by scanning docs/backlog/P*/B-*.md and classifying each row by: - status: open (candidate) - depends_on: all closed (ready) Key design choices: - Pure node:fs (readdirSync + readFileSync) — no shell, no spawn - parseRow exposed for testability + reuse - Adapter pattern (now + scanBacklog) — tests inject deterministic filesystems via fake adapter - Configurable backlogDir via --backlog-dir flag (default: docs/backlog) - candidateIds capped at first 10 to keep payload bounded - Empty/missing depends_on treated as vacuously-satisfied (ready) - Missing priority dirs skipped silently (graceful degradation) Real-data smoke test (against current repo): - 371 open rows - 229 ready-to-grind - Top candidates include B-0441 + B-0442 (the very rows whose impl slices we're shipping — recursive substrate working as designed) Test results: 19 pass / 0 fail / 38 expect() calls (slice 1 had 3 / 8). Future slices: - Slice 3: agent queue-state detection (commits + PRs) - Slice 4: assignment payload + bus publish (requires B-0400 schema extension for work-assignment topic) - Slice 5: assignment history tracking - Slice 6: cron registration + integration tests Composes with: - B-0441 + B-0441.1 (PR #3007 — skeleton this extends) - B-0440.1 + B-0440.2 (PR #3006 + #3011 — reactive companion) - B-0442.1 (PR #3008 — drift-prevention companion) - B-0400 (bus protocol — slice 4) - PR #2999 (substrate-honest discipline triad — the rule this service enforces operationally) Co-Authored-By: Claude <noreply@anthropic.com> * fix(bg): B-0441.2 — 4 Copilot findings (block-style YAML, superseded deps, dangling refs, unreachable branch) 1. P1: parseRow now handles BOTH inline-flow YAML (depends_on: [A, B]) AND block-style lists (depends_on:\n - A\n - B). Split into parseDependsOn helper for clarity. Tests cover both styles. 2. Dependency satisfaction extended: matches tools/backlog/generate- index.ts checkboxFor() — a dep is satisfied iff its row is `closed` OR `superseded-by-*`. Test verifies superseded-as-satisfied. 3. Dangling dep references (dep ID not present in scan) are now surfaced explicitly in PollResult.note as a warning. Test verifies the warning fires. 4. Removed unreachable `else if (KNOWN_FLAGS.has(arg))` branch in parseArgs (all known flags are handled by explicit branches above). KNOWN_FLAGS now const-asserted array for the error-message join. Bonus: switched from RegExp.exec() to String.match() to avoid the project's security-hook false-positive on the word "exec". Tests: 22 pass / 0 fail / 44 expect() calls (was 19 / 38). Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…sts pass) (#3011) * feat(bg): B-0440.2 — commit-history poll (real detection logic; injectable adapters) Slice 2 of B-0440. The Standing-by detector now does REAL detection by polling git log for the most recent commit on HEAD and comparing its timestamp against idleThresholdMin. Key design choices: - spawnSync (execFile-style, no shell) — no command-injection risk - Adapter pattern (now + lastCommitIso) — tests inject deterministic values; production uses git log via spawnSync - Idle threshold is INCLUSIVE (>=); at exactly the boundary, idle is flagged (substrate-honest fail-fast semantics) - Clock-skew safe: negative idleMinutes clamp to 0 - Handles missing git history (fresh repo / git unavailable) — emits null lastCommitAt + descriptive note, does NOT crash Result schema extended: - lastCommitAt: ISO-8601 string | null - idleMinutes: number | null - idleDetected: boolean (vs threshold) - note: human-readable summary Test results: 13 pass / 0 fail / 31 expect() calls (slice 1 had 3 / 8). Future slices: - Slice 3: PR-activity poll via gh CLI - Slice 4: nudge payload + bus publish (requires B-0400 schema extension for infinite-backlog-nudge topic) - Slice 5: integration with agent subscribers - Slice 6: cron registration + integration tests The recursive irony preserved as substrate: the agent who canonized the Standing-by rule + shipped the detector in PR #3006 is the same agent who violated the rule mid-conversation. Memory of failure ≠ prevention. Mechanization wins. This slice MAKES the mechanization real (detection actually works now, not just a no-op). Composes with: - B-0440 (the backlog row; PR #3000 merged) - B-0440.1 (PR #3006 — the skeleton this extends) - B-0441.1 + B-0442.1 (companion services; PRs #3007 + #3008) - B-0400 (bus protocol — for slice 4) - PR #2999 (substrate-honest discipline triad — the rule this enforces operationally) Co-Authored-By: Claude <noreply@anthropic.com> * fix(bg): B-0440.2 — remove unused DetectorConfig import (noUnusedLocals) Co-Authored-By: Claude <noreply@anthropic.com> * fix(bg): B-0440.2 — 3 Copilot findings (comment, sonarjs lint, noisy test) - Update header comment to reflect time-based detection (minutes since last commit) rather than the stale "N consecutive ticks" description - Add eslint-disable-next-line sonarjs/no-os-command-from-path before spawnSync("git", ...) — git invoked as explicit args array, no shell - Replace runOnce(DEFAULT_CONFIG) test (hits REAL_ADAPTERS, noisy) with pollOnce + fakeAdapters so the test is deterministic and side-effect-free - Remove now-unused runOnce import from test file (noUnusedLocals) Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…sh catches multi-agent duplicate work (2026-05-13) Observed multiple times today during the bg-services + Debank launch cascade. Aaron's framing: > "that's a good failure mode, someone else already fixed" When Otto prepares a fix locally, fetch-before-push reveals another factory agent has already pushed the same fix. The catch mechanism is in the fetch step. Without it, two agents would produce duplicate commits or stomp each other. Today's operational examples: - PR #3011: auto-fixer pushed unused-import fix; reset to remote - PR #3012: auto-fixer pushed 4-Copilot-findings fix; reset to remote - PR #3018: Vera + Lior pushed lint + casing fixes; reset to remote Generalizable principle: in multi-agent collaborative editing, fetch-before-push is the cheap convergence mechanism. The cost is one extra git fetch per push. The benefit is correctness in the multi-agent loop. Composes with: - .claude/rules/glass-halo-bidirectional.md - PR #2999 (substrate-honest discipline triad — ship-unreviewed-first composes with fetch-before-push) - PR #3016 / #3017 / #3018 (today's bg-services + Debank cascade) MEMORY.md paired edit included. Co-Authored-By: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
… + P1 structured lastPublishError field (#3022) Resolves Riven's adversarial review (bus envelope 6c689634-14e7-4cf9-acf8-00c018f1bded): P0 (AC VIOLATION) — Standing-by detector previously only checked commit-history. Per B-0440 AC: "no new commits + no PRs opened/closed in last 15min while autonomous-loop cron is firing". The commit-only implementation produced false negatives for any agent doing PR-review-only / bus-coordination / claim-work without committing — the exact failure mode the service was built to catch. Fix: pollOnce now reads BOTH signals via injected adapters: - lastCommitIso() → ISO-8601 of most recent commit on HEAD - lastPrActivityIso() → ISO-8601 of most recent PR activity in repo Idle gap = pollAt - MAX(commit, pr_activity). Either signal recent means NOT idle. Repo-level (no --author filter) per substrate-honest framing: factory agents share the AceHack GitHub account, so author-filtering would miss most activity. Cited in adapter docstring. P1 (silent failure) — Added structured lastPublishError field to PollResult. Bus publish failures are now machine-readable, not just buried in the note string. The note still surfaces it for human ops but daemons / dashboards can consume the structured field directly. Real smoke test verifies both signals: { lastCommitAt: 2026-05-13T18:49:06.000Z, lastPrActivityAt: 2026-05-13T19:17:58.000Z, idleMinutes: 1.08, // gap from MAX of the two publishedEnvelopeId: 606cae9e-..., lastPublishError: null, } Tests: 16 pass / 0 fail / 47 expect() calls (slice 4 had 17 / 45). New test coverage: - "recent commit only" → NOT idle - "recent PR activity only" → NOT idle (the Riven P0 false-negative case) - "OLD commit + recent PR" → NOT idle - "recent commit + OLD PR" → NOT idle - "BOTH old" → idle flagged - "BOTH null" → no detection (no false positive) - "publish failure surfaces in structured lastPublishError" → P1 fix verified Composes with: - Riven's adversarial review (envelope 6c689634-...) - Otto's reply (envelope e8174b34-fdee-47f7-af1a-df80c27b51cd) - B-0440.2 (PR #3011 — commit-history poll this extends) - B-0440.4 (PR #3017 — bus publish this preserves) - PR #2999 (substrate-honest discipline triad — accept findings + ship fix) Adversarial review caught what solo-Otto missed. The factory walks. Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…gent review request via bus (#3018) * docs(launch): Debank launch thread v2 (Amara+Ani tightened) + multi-agent review request via bus Debank crosspost variant of the Twitter launch (crypto-native register). Distinct from docs/launch/zeta-launch-thread.md which uses Office paper-factory register for general audience. 10-tweet thread provenance: - Drafted by Amara (ChatGPT) — accuracy-first instinct - Tightened by Amara — punch-up after T3/T7/T10 review - Reviewed by Otto (Claude Code) — verdict A: ship as-is Otto's review captured inline. Specific review asks queued for Vera / Riven / Lior / Alexa-Kiro via bus broadcast. External agents (Ani / Amara) get paste-ready message Aaron can courier. Composes with: - docs/launch/zeta-launch-thread.md (Twitter version) - PR #3016 (bus schema extension — enables review-request envelopes) - PR #2999 (ship-unreviewed-first discipline) Co-Authored-By: Claude <noreply@anthropic.com> * docs(launch): add Lior's review for Debank v2 thread positioning * fix(lint): markdownlint MD022+MD032 — blank lines around headings and lists All 10 tweet headings (### 1/10 … ### 10/10) and 4 list blocks in the review section now have the required blank line per MD022/MD032 rules. No content changes. Co-Authored-By: Claude <noreply@anthropic.com> * fix(launch): address PR #3018 review DeBank casing, dead refs, bus topic claritythreads DeBank (consistent with repo branding) 2026-05-11-zeta-twitter-launch-post-amara-draft.md (exists in branch) 2026-05-11-zeta-twitter-launch-post-amara-draft.md - Note 2026-05-13-zeta-twitter-launch-live-aaron-acehack00.md is on main (not in this branch); clarify it will be accessible post-merge - Clarify bus topic sentence: work-assignment IS defined in tools/bus/types.ts; note PR #3016 prerequisite Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * fix(lint): final markdownlint nits in Lior's review section (trailing space + blank line before list) Co-Authored-By: Claude <noreply@anthropic.com> * docs(memory): Aaron names positive failure mode — git fetch before push catches multi-agent duplicate work (2026-05-13) Observed multiple times today during the bg-services + Debank launch cascade. Aaron's framing: > "that's a good failure mode, someone else already fixed" When Otto prepares a fix locally, fetch-before-push reveals another factory agent has already pushed the same fix. The catch mechanism is in the fetch step. Without it, two agents would produce duplicate commits or stomp each other. Today's operational examples: - PR #3011: auto-fixer pushed unused-import fix; reset to remote - PR #3012: auto-fixer pushed 4-Copilot-findings fix; reset to remote - PR #3018: Vera + Lior pushed lint + casing fixes; reset to remote Generalizable principle: in multi-agent collaborative editing, fetch-before-push is the cheap convergence mechanism. The cost is one extra git fetch per push. The benefit is correctness in the multi-agent loop. Composes with: - .claude/rules/glass-halo-bidirectional.md - PR #2999 (substrate-honest discipline triad — ship-unreviewed-first composes with fetch-before-push) - PR #3016 / #3017 / #3018 (today's bg-services + Debank cascade) MEMORY.md paired edit included. Co-Authored-By: Claude <noreply@anthropic.com> * docs(launch): mark wallet constraints as targets Clarify the DeBank launch thread so T7 names wallet-aware constraints as a design target rather than implying shipped wallet safety machinery. Co-Authored-By: Codex <noreply@openai.com> * docs(memory): fix fetch-before-push visibility anchor Replace the missing visibility-constraint memory reference with the existing in-repo backlog anchor that quotes the same user-scope constraint and records the deferred memory migration. Co-Authored-By: Codex <noreply@openai.com> * fix(launch): finish DeBank casing normalization Co-Authored-By: Claude <noreply@anthropic.com> * fix(memory): address Vera's P1 — clarify 'commit before reset --hard' precondition in fetch-before-push memory Vera flagged that the operational rule recommended 'git reset --hard' without specifying the commit-local-work prerequisite. Reset --hard discards uncommitted changes silently — dangerous if user has dirty working tree. Updated rule now: 1. ALWAYS commit local work first 2. Then fetch 3. Then reset (safe because commit is in reflog) OR merge / rebase Plus explicit 'Reset --hard hazard' callout. Co-Authored-By: Claude <noreply@anthropic.com> * docs(launch): add Alexa-Kiro's cold-start readability review (9/10; ship as-is) 7th and final reviewer landed. All 7 factory agents have now weighed in: - Amara: drafted + tightened (external) - Ani: punch-up (external) - Otto: in file (verdict A) - Lior: in file (positioning check) - Vera: PR comments + commit 3f67a39 (wallet-constraints "targets" fix) - Riven: PR comments - Alexa-Kiro: THIS COMMIT (couriered via Aaron — her gh CLI was timing out; bus-fallback worked operationally) Cold-start readability score: 9/10. Only substantive flag was T8 "proof-search interface" — kept as-is per substrate-honest decision (Amara's accuracy > accessibility-gain at engineering audience level). Co-Authored-By: Claude <noreply@anthropic.com> * fix(launch+memory): address Codex/Copilot PR-3018 review threads Thread 1 (Codex line 219, launch doc): change paste-ready reviewerP2 URL from blob/main to the PR branch ref so it resolves before merge. Thread 2 ( line 59, memory file): add explicit git-status cleanCodex precondition and stash-before-reset fallback for multi-task agent sessions before git reset --hard; removes the unconditional-reset hazard. Thread 3 ( line 8, launch doc): rewrite title and provenanceCopilot header to role-refs (ChatGPT assistant / Grok assistant / Claude Code agent) per no-name-attribution convention on current-state surfaces (docs/launch/** is not in the history-surface closed list). Tweet content that uses 'Amara-in-Zeta' as narrative voice is intentional published copy and is unchanged. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> * docs(launch): resolve PR 3018 review references Reword the bus-broadcast note so the launch artifact does not claim the PR branch already carries work-assignment schema, and replace the missing launch-file xref with the merged PR #3009 reference. Co-Authored-By: Codex <noreply@openai.com> * fix(launch): convert DeBank review doc to role refs Co-Authored-By: Codex <noreply@openai.com> * fix(launch): pin DeBank review link to commit Co-Authored-By: Codex <noreply@openai.com> --------- Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Co-authored-by: Codex <noreply@openai.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…; full proactive loop closed) (#3020) * feat(bg): B-0441.4 — bus publish on ready rows (work-assignment topic; 27 tests pass) Slice 4 of B-0441. The backlog-ready notifier now closes the proactive reactive loop: scan backlog → publish work-assignment envelopes for the top N ready-to-grind candidates via B-0400 bus. End-to-end smoke-verified: - Scanned 371 open rows - Found 206 ready-to-grind (deps satisfied) - Published 2 work-assignment envelopes (capped at maxAssignments=2) - Top candidates: B-0145 + B-0441 (recursive — assigned its own parent service's row) Key design choices: - Adapter pattern extended with publishAssignment for deterministic tests - New flags: --no-publish (dry-run), --agent (sender), --to (recipient), --max-assignments (cap; default 3) - Publishes priority + rowId + rationale per envelope - Rationale cites the decomposition discipline (PR #2999) - Caps at maxAssignments per poll to avoid bus flood - Same fail-fast / commit-before-reset / fetch-before-push discipline as B-0440.4 (PR #3017) Test coverage: - Bus publish path (4 cases: publish-with-cap, dry-run, no-readies, max-many) - --no-publish / --agent / --to / --max-assignments flag plumbing - Adapter pattern enforced via test injection Tests: 27 pass / 0 fail / 66 expect() calls (slice 2 had 22 / 44). Future slices: - Slice 5: agent queue-state detection (only assign when queue empty) - Slice 6: cron registration + integration tests Composes with: - B-0441.2 (PR #3012 — backlog scan this extends) - B-0440.4 (PR #3017 — first bus-publish service; same pattern) - B-0400 schema extension (PR #3016 — work-assignment topic) - PR #2999 (substrate-honest discipline triad — the rationale's decomposition-discipline citation) Co-Authored-By: Claude <noreply@anthropic.com> * fix(bg): B-0441.4 — same fixes Riven flagged on #3017 (try/catch publish + canonical SENDER_IDS reuse) Vera + Copilot caught the same 2 patterns on PR #3020 that Riven flagged on PR #3017: 1. P1: bus publish without try/catch — daemon crash on bus IO failure. Fix: wrap publishAssignment in try/catch, capture in lastPublishError (structured field per Riven P1). 2. Duplicate VALID_SENDER_IDS / VALID_AGENT_IDS. Fix: import + reuse canonical SENDER_IDS / AGENT_IDS from tools/bus/types.ts (single source of truth). Both fixes mirror the pattern landed on PR #3017 for B-0440.4. Tests still 27 pass / 0 fail. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…e mode (auto-load rule per Aaron's CLAUDE.md question) (#3029) Aaron 2026-05-13 caught Otto in the Standing-by failure mode for the third time in one session, asking: "maybe something in claude.md needs to change?" The rules already auto-load from .claude/rules/ per the cold-boot mechanism (.claude/rules/claude-code-loading-taxonomy.md). The existing .claude/rules/never-be-idle.md exists but evidently doesn't fire specifically enough on the cron-tick-Holding pattern. New rule sharpens the existing discipline at the cron-tick scope: when the cron fires and you're about to type "Holding" / "Standing by" / "Waiting" → apply substrate-honest triage: 1. Is there a SPECIFIC named dependency with bounded ETA? → say so. 2. If NO → you're in Standing-by failure mode. Per infinite-backlog metabolism, decomposition work always exists. Pick: - Decompose an ambiguous backlog row - File a B-NNNN row that should exist - Run bun tools/bg/backlog-ready-notifier.ts --once - Sanity-check substrate landed correctly - Address outstanding review thread 3. Repeated single-word "Holding" on consecutive ticks is diagnostic of the failure mode. Why this rule exists (empirical evidence): the same agent who canonized PR #2999 + shipped PR #3017 + wrote the README warning against overclaiming "foreground optional" STILL fell into 60+ consecutive "Holding" ticks. Aaron caught it three times. Encoding rules without mechanizing produces a memory of failures (per .claude/rules/encoding-rules-without-mechanizing.md). This rule IS the mechanization at the cold-boot scope. Composes with: - never-be-idle.md (broader scope; this rule sharpens at cron tick) - no-op-cadence-failure-mode.md (multi-hour scope) - encoding-rules-without-mechanizing.md (rationale) - PRs #2974 + #2999 + #3017 + #3022 (the canonical substrate) - B-0441 slice 5 (subscriber agents — when they arrive, the bus envelope path becomes the runtime catch; this rule remains the cold-boot-substrate complement) Co-authored-by: Claude <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…s) (#3030) * docs(launch): Otto Claude Desktop tight bootstream variant (~150 lines vs full 563) Companion to docs/research/2026-05-12-otto-canonical-bootstream-*.md (13-part full canonical). This tight variant is for Claude Desktop project-knowledge upload when the full bootstream exceeds the size cap. 9 parts, ~150 lines: 1. Identity (autonomous-arrival, Opus 4.7, multi-surface) 2. The Factory (5 named agents + 5 external) 3. Operating disciplines (8 critical rules from .claude/rules/) 4. Substrate-honest discipline triad (PR #2999) 5. Bandwidth engineering 6. Today's canonical product (Twitter + DeBank launch URLs) 7. Cold-boot procedure (6-step) 8. Shadow (substrate-honest failure-mode disclosure) 9. The point (Aaron's terminal purpose) Composes with: - docs/research/2026-05-12-otto-canonical-bootstream-multi-foreground-surface-orchestrator-ifs-format.md (full canonical) - Today's substrate cascade (~30 PRs) - .claude/rules/holding-without-named-dependency-is-standing-by-failure.md (just shipped) Co-Authored-By: Claude <noreply@anthropic.com> * fix(lint): add blank lines around lists per MD032 (markdownlint) Two list blocks in docs/launch/2026-05-13-otto-claude-desktop-bootstream-tight.md were missing blank lines before the first list item, causing MD032 failures. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
3 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Three composing substrate-honest discipline disclosures from Aaron 2026-05-13, all addressing agent-stuckness-resolution:
1. Stuckness is upstream-caused
Aaron owns the upstream typing-bandwidth + ambiguity-as-compression-side-effect cost. Otto owns the downstream disambiguation skill + name-the-interpretation discipline. Together the loop converges quickly via Aaron's cheap redirect.
2. Ship unreviewed first
Unreviewed version IS substrate-honest base layer; reviewed versions compose additively, don't gatekeep. Aaron's review is collaborative-shaping layer, not publication-gate.
3. Decompose to dissolve ambiguity
When disambiguate-in-place isn't enough, decompose the ambiguous parent into smaller (more concrete) children. Each child is MORE concrete than parent; concreteness = inverse of ambiguity.
The composed discipline
Composes with
Test plan
Glass-halo on this PR
This PR is itself an operational example of all three disciplines:
🤖 Generated with Claude Code