feat: autocoder history logging + /retro command#12
Merged
Conversation
Workflow commands had GitHub-specific gh calls hardcoded, so file-backend users hit gh-auth errors and saw irrelevant label-setup output. Gate the GitHub-only blocks on ISSUE_SOURCE=github and route issue ops through the backend-neutral issue_update/issue_comment functions (or /update-issue slash command in user-facing prose). Mirrored to both .agent/workflows/ and plugins/autocoder/commands/. Also adds .gitignore entries for __pycache__, .playwright-mcp/, and the machine-local .autocoder.json config, plus the critical review doc for the pluggable-issue-source implementation plan. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
Six CDR cycles + thorough-brainstorming verification produced docs/specs/2026-05-21-fix-loop-token-efficiency-design.md. Anthropic pricing verified against platform docs; Opus 4.7 corrected from $15→$5/MTok. Spec includes Appendix A with empirical verification trail (19 assumptions). 🤖 Generated with Claude Code
Phase 1 of fix-loop token-efficiency plan. issues-file.py 'update --add-label LBL --if-unset' exits 9 if label already present, race-free under the existing flock(LOCK_EX). Enables fix-loop-gate.sh to do CAS-style atomic issue claiming. Tests: 4/4 pass — claim-succeeds, claim-fails-when-set, add-label-idempotent (regression), 10-process concurrent race (exactly one winner, no duplicate labels). 🤖 Generated with Claude Code
Phase 2.0 of fix-loop token-efficiency plan (deterministic subset; PostSessionEnd probe and TTL calibration deferred). - gate-log.py: append-only JSONL writer per spec §13.2 - analyze-gate-log.py: defensive parser, CSV + markdown output, fails on schema drift, warns on unknown fields - sample-session.jsonl + sample-gate.jsonl: golden-output fixtures (1024/8192/0/37 tokens, claude-sonnet-4-6) - test_analyze_gate_log.py: 3/3 pass (golden / schema-fail / unknown-warn) - README-measurement-tooling.md: usage docs - fix-loop-gate.sh: pre-LLM gate from earlier in plan Real Anthropic jsonl schema confirmed: usage at message.usage.* (not top-level), model at message.model. KNOWN_USAGE_EXTRAS whitelists undocumented fields (server_tool_use, service_tier, cache_creation, inference_geo, iterations, speed). 🤖 Generated with Claude Code
Phase 0 of fix-loop token-efficiency plan. Default IDLE_SLEEP_MINUTES 15→4 keeps consecutive ticks inside the 5-min prompt cache TTL. See docs/specs/2026-05-21-fix-loop-token-efficiency-design.md §6 for the math: ~9× cheaper per idle tick on Sonnet at the cost of 4× tick frequency. - plugins/autocoder/commands/fix-loop.md + .agent/workflows/fix-loop.md: IDLE_SLEEP_MINUTES default 15→4, help text updated, comment cites spec - plugins/autocoder/scripts/run-phase-experiment.sh: experiment runner skeleton; dry-run default, --execute stubbed with PostSessionEnd TODO until Phase 2.0 verifies hook availability - plugins/autocoder/scripts/phase0-1h-ttl-probe.md: runbook for the 1-h TTL vs 5-min TTL comparison mandated by §13.4 Phase 0 acceptance (both configs measured; cheaper wins) Queue-state fixtures the runner expects under tests/fixtures/ queue-state-*.sh are pending (separate task). 🤖 Generated with Claude Code
Four fixtures under tests/fixtures/ for the experiment runner:
queue-state-{empty, p2-bug, triage, active}.sh. Each script
takes setup|teardown, honors ISSUE_DIR_PATH, backs up existing
.issues/ to a sidecar before installing fixture state, restores
on teardown. Marker file (not $$) so teardown works from a
different process.
Fixture states:
- empty: no open issues
- p2-bug: 1 P2 bug
- triage: 5 issues without P0-P3 labels
- active: 3 issues — P0 mid-fix (working), P1 ready, P2 enhancement
🤖 Generated with Claude Code
Empirical test 2026-05-21: spawned Task subagents with model=haiku and model=opus from an Opus 4.7 parent. Both reported their own distinct model identity (claude-haiku-4-5-20251001 vs claude-opus-4-7[1m]). A2 (Task model= override) ✅ confirmed. A1 (Skill inheritance) not directly tested but moot for Phase 2 (same-model only); flagged for re-test only if Phase 3 design changes to need cross-model Skill. Phase 2 + Phase 3 + Phase 4 dispatch architecture all unblocked. 🤖 Generated with Claude Code
The /loop cron now invokes /autocoder:gate instead of /autocoder:fix. Gate body is 1973 bytes (≪ 75 KB fix.md), so idle ticks load the slim command via prompt cache instead of the full triage workflow. - plugins/autocoder/commands/gate.md (new, 1973 B): reads /tmp/autocoder-work.json from fix-loop-gate.sh, branches on phase (triage handled inline with condensed decision matrix; fix/enhance delegates to /autocoder:fix N via Skill). - plugins/autocoder/commands/fix-loop.md + .agent/workflows/ fix-loop.md: invocation target switched to /autocoder:gate. /autocoder:fix N (direct invocation with issue number) unchanged for backward compat. Phase 4 (slim fix.md itself) is separate. Phase 2 acceptance per §13.4 requires measured per-tick input ≤15% of pre-Phase-2 baseline; that measurement awaits the §10 Phase 2.0 PostSessionEnd hook probe (runtime experiment, separate ticket). 🤖 Generated with Claude Code
Opt-in model-split routing for fix-loop. Set LOOP_MODEL_SPLIT=1 to use /autocoder:dispatch (Haiku) instead of /autocoder:gate (same-model). Default is unchanged — dispatch is opt-in until Phase 3 acceptance per §13.4 is measured. - plugins/autocoder/commands/dispatch.md (new, 2700 B): frontmatter model=claude-haiku-4-5 (verified id from §7.1.1). Triage path runs inline on Haiku; fix/enhance path spawns Task(model="sonnet"|"opus") subagent to load /autocoder:fix N. - plugins/autocoder/commands/fix-loop.md + .agent/workflows/ fix-loop.md: LOOP_TARGET dispatch switch; default unchanged. - tests/triage_corpus/README.md (placeholder for the Haiku-vs-Sonnet correctness corpus mandated by §13.4 Phase 3). §7.1.1 verified A2 (Task model= override) so the Haiku→Sonnet handoff is viable. Acceptance criteria (≥80% triage agreement with Sonnet baseline, ≤30% per-tick cost) await the corpus and runtime measurement. 🤖 Generated with Claude Code
Per .issues/001.md item 1: 10 synthetic GitHub-style issue bodies under tests/triage_corpus/issues/ with expected.csv ground truth covering P0×2, P1×3, P2×3, P3×2. Two deliberately ambiguous cases (SSO down for 65% of users; "feels slow sometimes") flagged so Haiku/Sonnet disagreement on borderline calls doesn't penalize either model. README rewritten with purpose (§13.4 Phase 3 gate), runner pseudocode, ≥80% Haiku/Sonnet agreement threshold, extension instructions. Actual benchmark runner is a runtime experiment (separate, runtime-only). 🤖 Generated with Claude Code
… items 3,5) Per .issues/001.md items 3, 5: - run-phase-experiment.sh: --inject-at <N> + --inject-fixture <name> now produce a documented dry-run preview line and a commented-out --execute block showing the background-sleeper (sleep N && fixture.sh setup) & pattern. Timing model documented: fires from runner PID, not a separate cron tick. - docs/reports/TEMPLATE-cumulative.md (new, 101 lines): fill-in template for the post-Phase-4 cumulative report per §13.5. Six sections + test env disclosure; v3 trigger checklist references the §11/§13.5 thresholds. Item 2 (pytestmark=unit on test_issues_file_if_unset.py) skipped: agents repo doesn't register a 'unit' marker anywhere (pyproject.toml/pytest.ini/conftest.py all absent, no other pytestmark usage in tests/). Adding the marker would just emit PytestUnknownMarkWarning. Documented as skip in commit. 🤖 Generated with Claude Code
Per .issues/001.md item 4. gate.md's phase routing is LLM-prose, not bash, so functional per-phase assertions aren't possible without an integration test driving a real LLM. What this test covers (and the honest scope): - bash -n over all fenced blocks in gate.md (syntactic validity) - Structural assertions on the main block: sources issue-fns.sh, honors AUTOCODER_WORK_JSON, invokes fix-loop-gate.sh, captures rc - Parameterized over triage/fix/enhance: shimmed gate.sh exits 1/2/0, GATE_RC propagates, work.json path reachable - Guard: asserts NO bash 'case $PHASE' dispatcher exists. Fails if someone migrates routing into shell, forcing real per-phase tests to be added at that point - Two prose-presence tests so phase docs / rc docs don't silently disappear 14 tests, 0.6s, all pass. 🤖 Generated with Claude Code
Verified against Claude Code upstream docs (code.claude.com/docs/ en/slash-commands frontmatter reference) and grep of bundled plugins: the model: frontmatter field expects short aliases (haiku/sonnet/opus/inherit), NOT dated IDs. Zero real-world usage of claude-haiku-4-5-20251001 found across feature-dev, plugin-dev, or superpowers plugins. - plugins/autocoder/commands/dispatch.md: model: claude-haiku-4-5 → model: haiku. Inline comment expanded with convention, verification date, §7.1.1 cross-reference. - docs/notes/2026-05-21-frontmatter-model-convention.md (new): research trail for future maintainers — what was checked, the short-alias convention, where to update if Claude Code adds full-id support. 🤖 Generated with Claude Code
Records the closed issue that drove the 5 follow-up commits (triage corpus, inject handler, cumulative template, gate.md tests, dispatch.md model fix) on this branch. Useful as a paper trail for the autonomous-fix workflow that closed it. 🤖 Generated with Claude Code
User-confirmed: those PNGs are local testing artifacts, not for commit. 🤖 Generated with Claude Code
Runs python3 -m pytest tests/ -v on ubuntu-latest with Python 3.12 on push and pull_request. Installs pytest in-job (no requirements file added). 19 lines. Closes #3. 🤖 Generated with Claude Code
Adds a Token-Efficient Loop Commands section to the autocoder plugin README covering: - /autocoder:gate — slim default for the loop cron path (1973 B vs ~75 KB fix.md) - /autocoder:dispatch — opt-in Haiku-pinned variant via LOOP_MODEL_SPLIT=1, with §7.1.1-verified Task model= dispatch - Configuration table: LOOP_MODEL_SPLIT, AUTOCODER_WORK_JSON, IDLE_SLEEP_MINUTES - Cross-references: spec §5.3, §7.1.1, §13.4 Also adds two rows to the existing Commands table for discoverability. Closes #2. 🤖 Generated with Claude Code
Measured 2026-05-21 with claude -p in fresh empty-queue sessions (opus, bypassPermissions). Headline: /autocoder:gate idle tick is $0.154 vs /autocoder:fix idle tick $0.381 — 60% cost reduction. Cache_creation ratio came in at 36%, not the original 15% target. Revised the §13.4 Phase 2 threshold to ≤40% cache_creation + ≥50% total-cost reduction, with total cost as the load-bearing metric. The 15% target was a naive byte-count prediction (18KB → 2KB → 11%) that ignored the slash-command rendering envelope's ~17K fixed overhead. New §13.4.1 documents: - The exact measurement table (cost, cw, cr, out, turns, duration) - Why 15% wasn't hittable (overhead floor) - $30K/year savings projection at observed cadence - Installation gotcha: gate.md must land in ~/.claude/plugins/cache/plugin-marketplace/autocoder/3.10.0/ not ~/.claude/plugins/autocoder/, because Claude Code v2.1.146 discovers commands from the versioned cache install path - Implication for §11 v3 trigger: Phase 2 alone doesn't cross the $0.20/day v3 threshold; Phase 3 (Haiku) still needed Also relaxed output-token budget from <50 to <1000 (measured 772) and gate-duration p95 from 2s to 60s (Claude Code session startup dominates first-tick latency, not the gate body). 🤖 Generated with Claude Code
Replaces flat .issues/*.md layout with open/working/closed buckets, defines atomic rename-based claim, extracts gh logic into issues-gh.sh behind a uniform 8-verb backend CLI, and adds any-claimable preflight for zero-cost idle fix-loop ticks. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Addresses 4 literal-wrongness findings from docs/criticalreviews/2026-05-22-issue-storage-and-backend-abstraction-design-critical-review-1.md: - 2.1 any-claimable decoupled from "claimable": add blocked/ bucket (Option A); blocking-label adds/removes trigger renames; any-claimable stays O(1) even when most open work is gated on human input. - 2.2 exit-code ambiguity: split codes (0 work, 1 no-work, 2 usage, 3 backend-error); preflight blocks in fix.md / fix-loop.md use case. - 2.3 comment-during-rename duplicate-file race: add normative fd-based I/O requirement under flock; tighten race-2 analysis to reference it. - 2.4 github any-claimable never exits 1: switch shown command to a count-and-bracket form that emits the right exit code. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Addresses 5 round-2 findings from docs/criticalreviews/2026-05-22-issue-storage-and-backend-abstraction-design-critical-review-2.md, plus a new explicit cross-worktree coordination requirement: - 2.1 §2/§4 ordering contradiction for claim/release: rewrite §4 file row to match §2's rename-first order for claim; document the asymmetry (claim rename-first, release flock-then-rename). - 2.2 --state blocked missing from CLI contract: add blocked to enum. - 2.3 release Python snippet hardcoded open/ destination: replace with label-aware target selection that flock-reads labels before rename. - 2.4 file-backend any-claimable can't emit exit 3: guard with directory-existence check; pipe find into grep -q . so exit code reflects file presence, not find's own quirks. - 2.5 Section E claim test unsatisfiable by gh: split into strict (file/jira) and best-effort (gh) variants. - Cross-worktree: add Goal #5, normative subsection in §1 requiring ISSUE_DIR_PATH resolution to the main worktree, and a cross-worktree end-to-end test in the testing plan. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Addresses finding 3.1 from docs/criticalreviews/2026-05-22-issue-storage-and-backend-abstraction-design-critical-review-3.md: expand migration plan step 6 to enumerate list-needs-design, list-needs-feedback, list-proposals, and the additional monitor-workers queries that all need --state open → --state blocked (or --state working) updates so the round-1 blocked-bucket fix doesn't silently break those slash commands. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Empirically verified 21 load-bearing claims about the codebase at spec-write time (file paths, function locations, existing query shapes, library/platform behavior). One discrepancy surfaced and addressed: monitor-workers.md:130 omits needs-feedback from its blocking-label filter (oversight); spec now notes the omission and the new layout corrects it. This makes the spec accepted as input by thorough-writing-plans. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… plan Translates docs/superpowers/specs/2026-05-22-issue-storage-and-backend-abstraction-design.md (spec SHA c0b3df6) into 8 tasks with 19 verified plan-level assumptions. Task ordering: file backend overhaul → gh extraction → fix-loop preflight → README → fix-loop-gate migration → state-query caller migration → new slash commands → .agent/ mirror. Tests fold into each task. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…plan Addresses 1 partial §1 + 5 §2 findings from docs/criticalreviews/2026-05-22-issue-storage-and-backend-abstraction-implementation-plan-critical-review-1.md: - §1.13 commit-convention assumption text broadened (codebase mixes scoped and unscoped commits; flag has no behavioral impact). - §2.1 cmd_update added to Step 9's resolve_path list; Step 7 code block made self-contained with proper variable bindings. - §2.2 update --add-label working stays supported (legacy/idempotent); claim is the atomic path. fix.md and fix-loop-gate.sh re-asserts unchanged. - §2.3 monitor-workers Step 2 expanded to show three distinct replacements (display / sort / count) instead of one. - §2.4 added cross-worktree and fix-loop-idle test steps to Task 1 Step 10, closing spec Testing-plan coverage gaps. - §2.5 Task 3 Step 1 line references corrected (276/282 instead of 350); preflight no longer duplicates the source line. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…d-based I/O Implements the file-backend overhaul per docs/superpowers/specs/2026-05-22-issue-storage-and-backend-abstraction-design.md: - open/working/blocked/closed bucket directories under .issues/ - atomic rename-based claim/release; --if-unset flag retired - any-claimable verb with exit-3 backend-error contract - migrate-layout subcommand for one-shot flat→bucket migration - .seq as authoritative monotonic counter (no more glob scan in create) - fd-based I/O under flock (no path re-open) preserves directory authority across concurrent renames - blocking labels (needs-design, needs-clarification, needs-feedback, needs-approval, too-complex, future, proposal) trigger bucket transitions automatically on update - cmd_close renames to closed/ regardless of source bucket - legacy update --add-label working still supported (idempotent label-set) 42 tests pass (list, get, update, comment, close, create, claim, release, any-claimable, migrate-layout, blocking transitions, parallel-claim stress). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Per spec §4 backend abstraction: - New plugins/autocoder/scripts/issues-gh.sh implements the uniform 9-verb CLI (list/get/update/comment/close/create/claim/release/ any-claimable) for the GitHub backend. Includes the blocking-label-exclusion search for --state open and the count-and-bracket exit-code form for any-claimable. - issue-fns.sh shrinks to pure dispatch — each issue_* function shells out to the configured backend script with the verb name and args. The _ifns_gh_* helper functions are gone. 11 dispatcher tests pass: list/get/update/claim/release/close/create/ any-claimable + bucket-transition routing through the file backend. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Per spec §5, inserts a case-statement preflight immediately after the issue-fns.sh source. When no claimable work exists, /fix prints one line and exits without invoking any sub-agent — zero LLM cost on idle. (fix-loop.md keeps its existing stop-hook + fix-loop-gate.sh dispatch; that gate is already the work-presence preflight for the outer loop. Task 5 migrates the gate's atomic-claim mechanism to issue_claim.) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The --if-unset exit-9 contract retires in this design; atomic claim moves to the new issue_claim verb (rename open/N.md → working/N.md on the file backend; best-effort label-add on github). The gate's race handling is unchanged at the boolean level — lost race still exits 1 and re-runs on next tick. 14 gate tests pass. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Per spec §3 migration plan step 6:
- list-needs-design.md, list-needs-feedback.md, list-proposals.md,
approve-proposal.md, brainstorm-issue.md: --state open --label X →
--state blocked --label X (blocked-labeled issues now live in
blocked/, not open/).
- monitor-workers.md:
- Line 79 display query: drop jq blocking-label filter (open/
excludes blocked by directory).
- Line 130 sort query: same simplification.
- Line 262 count query: WORKING from --state working (bucket
directly), UNBLOCKED from --state open count.
- Recovery step: issue_update --remove-label working → issue_release.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Covers spec Testing-plan goals #2 (zero-cost idle preflight) and #5 (cross-worktree coordination): - tests/test_issues_file_cross_worktree.sh (8 tests): bootstraps a temp git repo + secondary worktree; verifies ISSUE_DIR_PATH resolves identically from both, and that issue_list/claim/release/close from the secondary worktree all target the main worktree's .issues/. - tests/test_fix_loop_idle.sh (6 tests): verifies any-claimable exits 1 on empty open/, fix-loop-gate.sh exits 1 with no work plan when idle, and exits 0 + writes plan + claims the issue when work exists. Also fixes issue-config.sh and issue-fns.sh to tolerate `set -u` (unbound-variable strict mode) — both scripts are sourced by fix-loop-gate.sh which runs with set -euo pipefail. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Major bump for autocoder reflects the breaking changes from the issue-storage-and-backend-abstraction work on this branch: - --if-unset flag retired from issues-file.py - --state open semantics changed (now walks only open/ bucket, excludes working and blocked-labeled issues) - New 4-bucket file layout (open/working/blocked/closed) - New verbs (claim, release, any-claimable, migrate-layout) - New slash commands (/show-issue, /migrate-issues-layout) Also brings .claude-plugin/plugins/autocoder/plugin.json up to date (was lagging at 3.10.0) and registers the two new slash commands in both .claude-plugin and .factory-plugin plugin.json manifests. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Resolves three CDR findings: - GitHub backend: create history-log label before issue creation - Modernize: keep inline heredoc, add plugin copy as maintainer sync source - Worktrees: per-worker HISTORY.md merged to main worktree at cleanup Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
10-task plan covering shared append-to-history.sh extension, plugin distribution, fix.md/full-regression-test.md logging, and new /retro command. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Task 5 Step 2: unique old_string via simple-path PR body content - Task 5 Step 4: explicit old_string via "auto-resolved by autonomous fix workflow" anchor - Task 8: embed actual optional-skills-prelude block; remove placeholder Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
One finding: cut -d' ' -f2 truncates worktree paths with spaces. Fix: use sed 's/^worktree //' instead. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Preserves paths with spaces; addresses CIR round 2 finding 1. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…lags Adds GitHub backend (history-log issue), worktree-aware path resolution, and backward-compatible default behavior for modernize. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Points maintainers to the canonical script in plugins/modernize/scripts/. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…sion-test Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Overall: Very Good. Two caveats before heavy production use: 1. Sanitize $ISSUE_BODY before heredoc interpolation in append-to-history.sh 2. Wire test count variables in full-regression-test.md logging calls Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
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
1 participant
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
scripts/append-to-history.shwith--backend [file|github|auto]and--history-fileflags; GitHub backend creates/posts to ahistory-loglabeled issue; bare filenames auto-resolve to the main git worktree so parallel fix-loop workers write to one fileplugins/autocoder/scripts/,plugins/modernize/scripts/, and.agent/scripts/fix.md(simple PR, simple merge, complex PR, complex merge, blocking) and mirrored to.agent/workflows/fix.mdfull-regression-test.mdand its.agentmirror/retrocommand (plugins/autocoder/commands/retro.md+.agentmirror): 4-phase analysis (history, git log, issue tracker, test failures) →IMPROVEMENTS.mdretrocommandDesign process
Spec → CDR round 1 → spec updated → implementation plan → 2 CIR rounds (caught non-unique Edit anchors and
cutvssedworktree path bug) → subagent execution (10 tasks) → completion reviewCaveats (from completion review)
${ISSUE_BODY:0:150}infix.mdlogging calls flows into an unquoted heredoc — sanitize before heavy GitHub-backend useUNIT_PASS/UNIT_TOTALetc. not yet wired infull-regression-test.mdlogging calls; entries show?/?until connected.agent/workflows/retro.mdreplaces a pre-existing 146-line modernize retro stub (naming collision in Antigravity's flat namespace)Test plan
append-to-history.shverified: file backend, backward compat, auto-detect, error paths, worktree detection🤖 Generated with Claude Code