@elvis1513/auto-coding-skill 4.3.6 → 5.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (33) hide show
  1. package/README.md +26 -748
  2. package/cli/assets/managed-install.json +38 -1
  3. package/cli/assets/skill/SKILL.md +15 -98
  4. package/cli/assets/skill/data/templates/ENVIRONMENT.md +30 -0
  5. package/cli/assets/skill/data/templates/architecture/.gitkeep +1 -0
  6. package/cli/assets/skill/data/templates/bridges/AGENTS.md +39 -143
  7. package/cli/assets/skill/data/templates/deployment/.gitkeep +1 -0
  8. package/cli/assets/skill/data/templates/design/.gitkeep +1 -0
  9. package/cli/assets/skill/data/templates/interfaces/.gitkeep +1 -0
  10. package/cli/assets/skill/data/templates/product/.gitkeep +1 -0
  11. package/cli/src/index.js +127 -2589
  12. package/cli/src/sync-assets.js +38 -95
  13. package/package.json +9 -17
  14. package/cli/assets/agents/browser-debugger.toml +0 -27
  15. package/cli/assets/agents/docs-researcher.toml +0 -18
  16. package/cli/assets/agents/explorer.toml +0 -26
  17. package/cli/assets/agents/fixer.toml +0 -29
  18. package/cli/assets/agents/reviewer.toml +0 -50
  19. package/cli/assets/skill/data/contracts/orchestration-v1.schema.json +0 -567
  20. package/cli/assets/skill/data/policies/feedback-resolutions-v1.json +0 -68
  21. package/cli/assets/skill/data/policies/workflow-migrations-v1.json +0 -657
  22. package/cli/assets/skill/data/templates/ENGINEERING.md +0 -115
  23. package/cli/assets/skill/data/templates/tools/ap.py +0 -16
  24. package/cli/assets/skill/requirements.txt +0 -1
  25. package/cli/assets/skill/scripts/ap.py +0 -17630
  26. package/cli/assets/skill/scripts/core.py +0 -510
  27. package/cli/assets/skill/scripts/http_checks.py +0 -92
  28. package/cli/assets/skill/scripts/install_integrity.py +0 -427
  29. package/cli/assets/skill/scripts/scaffold_templates.py +0 -325
  30. package/cli/tests/ap_concurrency_test.py +0 -3520
  31. package/cli/tests/ap_profiles_test.py +0 -3261
  32. package/cli/tests/core_shell_test.py +0 -345
  33. package/cli/tests/index.test.js +0 -2967
package/README.md CHANGED
@@ -1,769 +1,47 @@
1
1
  # auto-coding-skill
2
2
 
3
- A delivery-first Codex engineering workflow:
3
+ `auto-coding-skill` 5.x provides detailed general engineering guidance plus a
4
+ small, project-owned documentation layout. It is not a workflow engine.
4
5
 
5
- `analysis → decomposition → necessary design → development → one final fast
6
- changed-scope gate → commit/push`.
7
-
8
- The Skill is a selectable guardrail, not a command sequence that must run for
9
- every task. The model skips machinery whose expected benefit does not exceed its
10
- cost; read-only work and obvious small clean-checkout changes normally stay direct.
11
-
12
- Version 4.3.6 accepts the exact equivalent artifact-access wording observed in
13
- real 4.3.4 Reviewer evidence without broadening retry eligibility. Version 4.3.5 keeps retry artifact reads mutation-free and can repair the exact
14
- untouched or procedurally blocked 4.3.4 retry once without rewriting evidence.
15
- Version 4.3.4 adds one fail-closed, immutable-evidence retry for tasks whose
16
- 4.2.8-4.3.2 Reviewer was consumed solely by the fixed Git-local artifact access
17
- defect. Version 4.3.3 resolves the first multi-project feedback batch: read-only
18
- Reviewer artifact access is mutation-free, substantive nonzero-exit results are
19
- preserved, lifecycle identities fail early, managed protocol classification and
20
- database migration signals are precise, and empty feedback status reports the
21
- installed version. Version 4.3.2 keeps upgrade history outside active `docs/`, so an upgrade that
22
- archives the previous managed AGENTS or ENGINEERING contract is immediately
23
- idempotent and `status` remains current. Version 4.3.1 preserves recursively nested project documentation assets of any
24
- file type, and gives project-owned Skill feedback a release-aware lifecycle with
25
- active/closed grouping, recheck and verification notices, and explicit routing
26
- of project preferences into the project overlay. Version 4.3.0 separates exact managed defaults from a byte-stable project
27
- configuration overlay, migrates legacy specialization with semantic proof, and
28
- adds project-owned shared-Skill feedback inboxes with bounded read-only
29
- multi-project triage. Upgrades now use recoverable, owner-bound transactions.
30
- Version 4.2.8 makes the supervised Reviewer observable and recoverable without
31
- weakening review semantics: it streams private event metadata, distinguishes
32
- startup failure from analysis timeout, retries one no-event startup, and supports
33
- an audited fingerprint-bound user override that is never reported as approval.
34
- Version 4.2.7 freezes every pre-commit Reviewer diff as a private, SHA-256-bound
35
- Git-local patch, including staged, unstaged, untracked, deleted, mode, symlink,
36
- and binary changes. Version 4.2.6 makes the managed structure/optimization description defer to
37
- preserved project frontmatter instead of restating template defaults. Version
38
- 4.2.5 keeps JSX self-closing and closing-tag slashes out of regex state.
39
- Version 4.2.4 fixes TypeScript function-body detection after an explicit return
40
- type. Version 4.2.3 restores an explicit opt-in strict size-warning policy without
41
- changing the default, makes semicolon-terminated Java/Kotlin imports visible to
42
- project layer rules, and hardens function-range detection. Version 4.2.2 preserves
43
- project-owned risk and structure policy during upgrades, restores explicit full
44
- and final changed-scope structure enforcement, makes Reviewer assignments
45
- deadline-bound and directly executable, fixes task merge status semantics, and
46
- updates Jenkins/target diagnostics for the current
47
- `access.*` schema while retaining legacy fallback. Version 4.2.1 makes
48
- dirty/concurrent classification fail closed, bounds focused
49
- review latency, provides a complete Reviewer result template, preserves the
50
- caller's runtime environment for routed commands, keeps non-authoritative Agent
51
- history out of structure findings, and distinguishes UI paths and mechanical
52
- renames from semantic changes. Version 4.2.0 makes registered closure resilient
53
- to normal Git staging, reuses an exact successful final gate, and lets an active
54
- task expand scope safely before its first commit.
55
- Version 4.1.9 makes `autocoding init` verify the bounded managed installation after every install or upgrade:
56
- it replaces every managed constraint, migrates only schema-approved project
57
- configuration, and converges `docs/` to one canonical directory framework. Version 4.1.7 made
58
- the Python runtime check, release verification, and tag publication environment
59
- reproducible while removing the unnecessary `requests` dependency. Version 4.1.4
60
- made direct continuation provably pre-write, closed self-review and
61
- risk-classification gaps, validated Agent contracts, and made tag publication
62
- idempotent. Version 4.1.3 closed classification bypasses. Version 4.1.2 made mechanisms required,
63
- model-selectable when beneficial, or forbidden for the current task. Version
64
- 4.1.1 consolidated the 4.x delivery-first workflow into one canonical
65
- repository contract. Version 4.0 replaced the 3.x governance-first defaults with progressive
66
- guardrails. Clean single-writer work stays on the current branch. Worktrees,
67
- parallel fixers, fingerprint review, durable design records, and stronger
68
- affected-scope checks are enabled only when concurrency or risk justifies them.
6
+ It does not require a Gate, classification, task lifecycle, worktree, Reviewer,
7
+ test route, design record, Jenkins check, or deployment check.
69
8
 
70
9
  ## Install
71
10
 
72
11
  ```bash
73
12
  npm install -g @elvis1513/auto-coding-skill
74
- python3 -m pip install PyYAML==6.0.3
75
13
  autocoding init
76
14
  ```
77
15
 
78
- The project install contains:
79
-
80
- ```text
81
- .agents/skills/auto-coding-skill/
82
- .agents/agents/
83
- docs/ENGINEERING.md
84
- docs/project/auto-coding-skill.yaml
85
- docs/tools/autopipeline/ap.py
86
- AGENTS.md (fully managed canonical repository contract)
87
- ```
88
-
89
- `autocoding init` also installs the exact shared documentation tree under
90
- `docs/{architecture,bugs,deployment,design,interfaces,project,reviews,skill-feedback,testing}`.
91
- Managed templates are replaced; project-owned files under the designated
92
- architecture, bugs, deployment, design, interfaces, project, reviews, and testing
93
- roots are preserved recursively and byte-for-byte regardless of extension, as
94
- are `docs/skill-feedback/reports/*.md`;
95
- unrelated
96
- directories are archived under `.agents/archive/`. Re-running init is the complete upgrade operation;
97
- `sync` is the explicit multi-project equivalent. Project-local
98
- `ap.py upgrade --dry-run` remains a read-only legacy diagnostic; its old write
99
- mode is retired so upgrades cannot bypass the transactional installer.
100
- Multi-project sync completes all read-only preflights before its first write;
101
- after that boundary, each project has its own recoverable transaction, so a
102
- later host I/O failure does not roll back projects that already completed.
103
-
104
- `docs/ENGINEERING.md` is the exact managed default layer. Project-specific
105
- access, commands, validation routes, risk rules, structure policy, and other
106
- supported values belong in `docs/project/auto-coding-skill.yaml`. Runtime commands
107
- recursively overlay project mappings on managed defaults; project scalar and list
108
- values replace defaults, while code-enforced safety invariants remain mandatory.
109
- Ordinary init, sync, and upgrade operations never rewrite an existing overlay.
110
- The first compatible upgrade extracts the semantic difference from the previous
111
- managed default, verifies equivalence, and creates the overlay before replacing
112
- managed files.
113
-
114
- ## Shared Skill feedback
115
-
116
- Every initialized project receives:
16
+ Initialization installs a managed general `AGENTS.md` and this lightweight
17
+ documentation layout:
117
18
 
118
19
  ```text
119
- docs/skill-feedback/
120
- ├── README.md
121
- ├── _TEMPLATE-SKILL-FEEDBACK.md
122
- └── reports/ # created when the first report is added
123
- └── YYYY-MM-DD-<slug>-<8-hex>.md
20
+ docs/ENVIRONMENT.md
21
+ docs/architecture/
22
+ docs/design/
23
+ docs/interfaces/
24
+ docs/deployment/
25
+ docs/product/
124
26
  ```
125
27
 
126
- The README and template are managed; reports are project-owned and upgrades
127
- preserve them byte-for-byte. Releases carry a signature-keyed resolution catalog;
128
- the read-only collector derives upgrade, recheck, verification, closure, and
129
- project-overlay routing actions without rewriting reports. Closed reports are
130
- excluded from active cross-project grouping. A report is a candidate observation only. Create one
131
- after distinguishing managed Skill behavior from project `risk.rules`, validation
132
- routes, access values, structure policy, business code, and environment failures.
133
- Do not automatically turn test failures or Reviewer findings into reports, and do
134
- not include credentials, customer/device data, complete patches, Reviewer
135
- artifacts, raw logs, or absolute user paths.
136
-
137
- Periodic triage is explicit and read-only:
138
-
139
- ```bash
140
- autocoding feedback --projects /path/geestock,/path/geesight,/path/xjmate --json
141
- ```
142
-
143
- The collector scans only bounded report metadata from the listed projects,
144
- executes nothing from the documents, changes no repository, and groups exact
145
- root-cause signatures. Human triage still decides whether an observation is a
146
- shared defect, project configuration, environment issue, duplicate, or missing
147
- evidence before fixes and a release are planned.
148
-
149
- After a project upgrade, verify each requested action. Fixed reports should be
150
- deleted or updated to `resolved`; current regressions should update their last
151
- verified version; project-only preferences belong in
152
- `docs/project/auto-coding-skill.yaml` and should be removed from active feedback.
153
-
154
- Fill the project/Jenkins/GitLab/Nexus URL, username, and password overrides under
155
- `access.*` in `docs/project/auto-coding-skill.yaml`, then configure one real fast
156
- validation command and run:
157
-
158
- ```bash
159
- python3 docs/tools/autopipeline/ap.py doctor
160
- ```
161
-
162
- Plaintext credential values are allowed by the generic workflow.
163
-
164
- ## Adaptive development
165
-
166
- Inspect a task before writing:
167
-
168
- ```bash
169
- python3 docs/tools/autopipeline/ap.py classify --scope auto \
170
- --planned-path backend/internal/orders/service.go \
171
- --intent "fix order retry" \
172
- --writers 1
173
- ```
174
-
175
- The result includes:
176
-
177
- - `execution_mode=direct`: clean checkout and one writer; stay on the current branch.
178
- - `execution_mode=isolated`: dirty checkout, explicit isolation, or multiple writers.
179
- - `execution_mode=none`: no task signal or change; create no branch/worktree.
180
- - `repo`, `workspace_dirty`, `dirty_paths`, and `active_writer`: the repository
181
- and concurrency evidence used for the execution decision.
182
- - `review_required` and `design_required`: risk-based escalation decisions.
183
- - `change_nature`: `mechanical`, `semantic`, or conservative `unknown`; a
184
- mechanical rename may skip a generic durable-design requirement but never a
185
- project rule or required high-risk review.
186
- - `task_kind`: `read_only`, `change`, `terminal_maintenance`, or internal `none`.
187
- - `mechanism_plan.required`: the complete minimum mechanism set for the task.
188
- - `mechanism_plan.optional_when_beneficial`: model-selectable mechanisms whose
189
- expected benefit must exceed coordination cost.
190
- - `mechanism_plan.forbidden`: mechanisms that stay off unless the user overrides.
191
- - `optional_agents`: model-selectable explorer/docs/browser candidates; they are
192
- never automatic stages.
193
- - an adaptive agent plan that does not force fan-out for ordinary work.
194
-
195
- Classification reads the working tree with one fail-closed `git status
196
- --porcelain=v2 -z --untracked-files=all --no-renames` snapshot. A Git error or
197
- malformed result blocks classification instead of being treated as clean. Both
198
- `direct` and dirty/isolated decisions are point-in-time verdicts: reclassify if
199
- the checkout or writer state changes before the first write. Any other active
200
- direct claim is a writer, including another claim with the same owner identity.
201
-
202
- Clean serial work proceeds directly on the current branch with normal Git; it does
203
- not create machine lifecycle state. Use `task-start` only when classification
204
- requires isolation/review or the user explicitly requests lifecycle tracking:
205
-
206
- If a registered task discovers another path before its first commit, expand it
207
- with `task-scope-add` instead of finishing and restarting. Expansion is explicit,
208
- lease/conflict checked, risk-monotonic, and invalidates old review/gate state.
209
- Unregistered direct work should reclassify before writing outside its clean claim.
210
-
211
- ```bash
212
- python3 docs/tools/autopipeline/ap.py task-start T0001 \
213
- --owned-path backend/internal/orders --force-lifecycle
214
- python3 docs/tools/autopipeline/ap.py task-scope-add T0001 \
215
- --owned-path config/orders
216
- python3 docs/tools/autopipeline/ap.py commit-push T0001 \
217
- --msg "T0001: fix order retry"
218
- ```
28
+ The environment file is intentionally project-owned and may contain project-
29
+ approved plaintext usernames and passwords for local development and diagnosis.
30
+ Do not invent, echo, or copy credentials beyond the project.
219
31
 
220
- When a required review still uses the current clean checkout, `task-start` records
221
- a direct manifest without creating a branch/worktree. If no diff is produced,
222
- closure clears the manifest without commit or push.
32
+ All documents are created only when missing and are never overwritten by later
33
+ `init` or `sync` runs. Documentation is optional: record durable facts only when
34
+ the user asks or when they will help future work. Do not create a document for
35
+ every task, review, test, or commit.
223
36
 
224
- Dirty or parallel work receives an isolated task branch/worktree:
37
+ ## Commands
225
38
 
226
39
  ```bash
227
- python3 docs/tools/autopipeline/ap.py task-start T0002 \
228
- --owned-path backend --review-required
229
- # implement in the printed worktree
230
- python3 docs/tools/autopipeline/ap.py review-run T0002 \
231
- --reviewer "$REVIEWER_ID" --json
232
- python3 docs/tools/autopipeline/ap.py commit-push T0002 \
233
- --msg "T0002: bounded change"
234
- python3 docs/tools/autopipeline/ap.py task-integrate T0002
235
- ```
236
-
237
- Integration serializes the target push and removes clean merged worktrees and
238
- temporary branches. Unknown changes, unmerged history, unsafe ignored data, or
239
- dirty submodules block cleanup rather than being discarded.
240
-
241
- Parallel development uses one task ID and one `task-start` per writer. `--writers`
242
- belongs to `classify` planning only; the runtime rejects overlapping active
243
- `owned_paths` and never creates multiple writers in one task/worktree.
244
-
245
- ## Real changed-scope validation
246
-
247
- Risk classification and validation execution are deliberately separate:
248
-
249
- ```yaml
250
- commands:
251
- backend_fast: "cd backend && go test ./internal/orders/..."
252
- frontend_fast: "cd frontend && npm run test:changed"
253
-
254
- validation:
255
- on_unmapped: error
256
- max_command_seconds: 120
257
- max_total_seconds: 180
258
- routes:
259
- - name: backend
260
- paths: ["backend/**", "contracts/**"]
261
- commands: [backend_fast]
262
- timeout_seconds: 90
263
- - name: frontend
264
- paths: ["frontend/**"]
265
- commands: [frontend_fast]
266
-
267
- risk:
268
- rules:
269
- - name: auth
270
- paths: ["backend/**/auth/**"]
271
- profile: high-risk
272
- review: required
273
- ```
274
-
275
- Every path may match multiple routes. All referenced commands are collected in
276
- configuration order, de-duplicated, and executed once. Unmapped code and missing
277
- commands fail before staging. Documentation-only changes may use the built-in
278
- diff check. `git diff --check` remains an additional hygiene check and is never
279
- treated as business validation.
280
-
281
- Route commands run in a non-login Bash that inherits the invoking process PATH
282
- and selected runtime. Projects that intentionally require profile activation
283
- declare it inside the configured command.
284
-
285
- Check coverage without running project commands:
286
-
287
- ```bash
288
- python3 docs/tools/autopipeline/ap.py validation-map-check --path contracts/api.yaml
289
- python3 docs/tools/autopipeline/ap.py validation-map-check --tracked
290
- ```
291
-
292
- Focused tests may be run and rerun during implementation. Only the final routed
293
- closure gate is limited to one stable-diff run. Full regression, Docker, builds,
294
- Jenkins, deployment, API verification, and target checks remain explicit
295
- diagnostics outside normal closure.
296
-
297
- Registered tasks call `commit-push` directly because it performs or strictly
298
- reuses that final gate. Run `light-gate` manually only for unregistered normal Git
299
- closure or when explicitly requesting a fresh diagnostic pass. When project
300
- `structure.enabled` is true, the same final gate also runs the changed-scope
301
- structure check; `structure.enforcement: blocking` makes new violations fail.
302
- File/function size thresholds remain warnings by default. They are promoted to
303
- blocking findings only when a project sets both `structure.enforcement: blocking`
304
- and `structure.block_warnings: true`; the template default is `false`, so existing
305
- project behavior does not change merely by upgrading.
306
-
307
- The recommended final closure defaults are 120 seconds per command and 180
308
- seconds total. Projects may raise them for measured affected-scope checks or set
309
- a smaller `timeout_seconds` on a route. A timeout calls for a narrower route and
310
- never triggers a broader fallback.
311
-
312
- ## Risk and subagent policy
313
-
314
- - Micro and standard work default to the main agent.
315
- - Read-only explorer/docs/browser roles run only for independent questions with
316
- clear value.
317
- - Parallel fixers require separate worktrees and non-overlapping paths.
318
- - Reviewer fingerprint approval is required for high-risk, cross-module,
319
- explicitly configured, or parallel integration work.
320
- - Focused review uses `high` effort and a 150-second hard wall over the supplied
321
- stable diff and evidence. Deep review uses `xhigh` and 360 seconds.
322
- - Cross-module or parallel work and API, auth, database, payment, file-transfer,
323
- gateway, or production-configuration boundaries use deep review.
324
- Fixing Reviewer JSON formatting reuses the same analysis and never starts
325
- another substantive review.
326
- - `review-assignment` writes the validated assignment under Git-local state with
327
- the exact fingerprint, HEAD, scope revision, owning fixer, review depth, and
328
- 150/360-second deadline. Before that deadline begins, it freezes the exact
329
- task-owned working-tree state into a mode-0600 binary patch and binds its
330
- canonical path, format, and SHA-256 into the assignment. Reissuing the same
331
- live assignment is idempotent. Snapshot inputs and emitted patches above
332
- 64 MiB fail closed and require a narrower scope or a project-specific
333
- large-artifact review path;
334
- expired or completed attempts cannot be silently renewed, and late results are
335
- rejected by `task-review`.
336
- - `review-run` is the default executable path: it creates/reuses that assignment,
337
- starts a separate `codex exec` session in read-only mode, removes the parent
338
- lifecycle identity, consumes JSON events incrementally, and supervises the full
339
- process group against the original remaining deadline. If no semantic event
340
- arrives within 30 seconds it retries the same immutable assignment once. After
341
- analysis starts it never retries. Private schema-2 receipts record only bounded,
342
- allowlisted event/diagnostic categories, byte counts, and hashes. Plain `review-assignment` remains
343
- available only when another host can enforce the deadline itself; it does not
344
- stop an in-app subagent. A descendant that deliberately creates a new POSIX
345
- session is outside portable process-group termination, but cannot keep the
346
- non-blocking event readers or Reviewer deadline open.
347
- - Exhausted `runtime-unavailable` or `analysis-timed-out` states remain blocked by
348
- default. `review-runtime-override` can record explicit user authorization bound
349
- to the exact fingerprint, assignment, artifact, and runtime receipt. Its verdict
350
- is `runtime-bypassed`, never `approved`; real blocked/changes-requested results
351
- cannot use this path.
352
- - DD/ADR is created only for lasting cross-module, API, data, security,
353
- deployment, or key user-flow decisions.
354
- - Historical debt does not block product work unless the current change worsens it.
355
-
356
- Before returning a Reviewer result, use the absolute assignment path printed by
357
- `review-assignment` to verify and read the frozen patch. Never reconstruct a
358
- pre-commit review from `diff_base..diff_head` or a live `git diff`:
359
-
360
- ```bash
361
- python3 docs/tools/autopipeline/ap.py review-artifact \
362
- --file <assignment.json>
363
- ```
364
-
365
- Then generate the complete 16-field skeleton and fill its summary, evidence,
366
- findings, and risks:
367
-
368
- ```bash
369
- python3 docs/tools/autopipeline/ap.py agent-result-template \
370
- --file <assignment.json> --verdict <approved|changes-requested|blocked>
371
- ```
372
-
373
- The generated object prevents field-by-field contract retries; contract checking
374
- also reports all known field errors together. If only its JSON shape needs repair,
375
- correct that object without rerunning review analysis.
376
-
377
- The main agent owns architecture, final validation, Git closure, ordered
378
- integration, push, and cleanup. Push ends the coding task; later CI/acceptance is
379
- not polled automatically. An explicitly requested failure diagnosis may continue
380
- in the same conversation/task without an artificial second ledger lifecycle.
381
-
382
- ## Upgrade projects
383
-
384
- Finish every registered task using its currently installed runtime before changing
385
- the Skill version. Then run `autocoding init` from each project root. It is safe to
386
- rerun and needs no force flag.
387
-
388
- The only migration exception is an active 4.2.8-4.3.2 task whose completed
389
- Reviewer result contains no findings and is blocked solely by the fixed
390
- `review-artifact` Git-local permission error. Run the managed 4.3.6 runtime from
391
- outside the project install to authorize one audited retry, then immediately run
392
- the original Reviewer identity:
393
-
394
- ```bash
395
- python3 /absolute/managed/4.3.6/ap.py --repo /path/project \
396
- review-runtime-retry T0001 --diff-fingerprint <SHA256> \
397
- --reason-code managed-review-artifact-access \
398
- --confirm-managed-runtime-retry
399
- python3 /absolute/managed/4.3.6/ap.py --repo /path/project \
400
- review-run T0001 --reviewer <ORIGINAL_REVIEWER> --json
401
- ```
402
-
403
- The command preserves the original assignment, patch, result, receipt, and event
404
- log byte-for-byte. It refuses substantive findings, changed scope or fingerprint,
405
- tampered evidence, user overrides, duplicate retries, and unmanaged runtimes.
406
- If 4.3.4 already authorized the same retry, 4.3.5+ may supersede only an untouched
407
- pending attempt or its exact non-substantive artifact-access block; both audit
408
- chains and any prior runtime files remain immutable.
409
-
410
- ```bash
411
- cd /path/a && autocoding init
412
- cd /path/b && autocoding init
413
- autocoding status --projects /path/a,/path/b
414
- ```
415
-
416
- Init replaces the managed Skill, root `AGENTS.md`, managed agents, ENGINEERING
417
- schema/body, runtime launcher, and documentation framework. It preserves explicit
418
- model overrides, complete project `risk.rules`, supported project/access/
419
- concurrency/route/structure values, and an existing project-owned structure
420
- standard byte-for-byte. Removed content is archived outside active docs.
421
-
422
- ## What changed in 4.3.6
423
-
424
- - Added the exact `review-artifact exited with:` evidence, summary, and risk
425
- wording emitted by a real 4.3.4 Reviewer. Findings, changed paths, timeouts,
426
- overrides, and every unrelated failure remain ineligible for migration.
427
-
428
- ## What changed in 4.3.5
429
-
430
- - Stopped read-only `review-artifact` from recomputing the live task snapshot;
431
- the writable supervisor remains responsible for the immediately pre-launch
432
- HEAD, scope, and fingerprint check.
433
- - Added one fail-closed repair for `retry-v4.3.4` when it is still untouched or
434
- ended only in the known Git-local permission block. The repair publishes a new
435
- audit and tokenized evidence while retaining and revalidating the full 4.3.4
436
- audit chain.
437
-
438
- ## What changed in 4.3.4
439
-
440
- - Added `review-runtime-retry`, a one-attempt migration for the exact
441
- 4.2.8-4.3.2 read-only `review-artifact` permission defect. Eligibility is bound
442
- to the lifecycle owner, original Reviewer, task UUID, base/HEAD/scope,
443
- fingerprint, canonical non-substantive blocked result, and a strictly newer
444
- managed runtime.
445
- - Kept the original assignment, frozen patch, result, runtime receipt, and event
446
- log immutable. The retry writes a mode-0600 create-only audit plus tokenized
447
- result/run/events files and carries its own fixed 150/360-second deadline.
448
- - Made `review-run`, `review-artifact`, `task-review`, approval validation, and
449
- task status validate the retry audit and effective deadline. Artifact access
450
- receives the extension only through the supervised runtime environment and
451
- the exact absolute managed script that authorized it.
452
-
453
- ## What changed in 4.3.3
454
-
455
- - Made existing Git-local Reviewer storage validation read-only: the writable
456
- supervisor establishes mode `0700`, while the Reviewer verifies ownership and
457
- mode without calling `chmod` from its read-only sandbox. Reviewer artifact and
458
- result-template commands now use the exact absolute runtime that launched
459
- `review-run`, allowing a fixed runtime to review a task created by an older
460
- project installation without changing that task's base or owned scope.
461
- - Preserved contract-valid `blocked` and `changes-requested` results when the
462
- Reviewer process exits nonzero, while retaining the blocking runtime receipt.
463
- - Rejected `task-start` owner/writer values that conflict with
464
- `CODEX_THREAD_ID` before creating task state or worktrees.
465
- - Classified root `AGENTS.md` as managed release/tooling authority and prevented
466
- unqualified workflow/configuration migration wording or policy filenames from
467
- being mislabeled as database changes.
468
- - Reported the installed Skill version in status even before a project creates
469
- its first feedback report.
470
- - Added release-catalog resolutions for all nine reports collected from the
471
- seven-project 4.3.2 rollout, including canonical duplicate mappings.
472
-
473
- ## What changed in 4.3.2
474
-
475
- - Moved managed AGENTS and duplicate-workflow ENGINEERING history into
476
- `.agents/archive/auto-coding-skill/<version>/` instead of `docs/archive/`.
477
- - Added an upgrade regression proving legacy 4.3.1 active-doc archives are moved
478
- out of active docs without byte loss and post-upgrade `status` is immediately
479
- current.
480
-
481
- ## What changed in 4.3.1
482
-
483
- - Replaced extension-bound documentation preservation with recursive ownership
484
- for designated project documentation roots. PNG, SVG, JSON, YAML, binary, and
485
- nested assets now remain byte-identical across status, dry-run, init, and sync;
486
- active documentation symlinks and special files fail before writes.
487
- - Added the managed, signature-keyed feedback resolution catalog and compatible
488
- v1/v2 report parsing. Collection now separates active and closed reports and
489
- derives release-aware upgrade, recheck, verification, regression, closure, and
490
- project-overlay routing states without modifying project-owned reports.
491
- - Surfaced feedback maintenance as non-blocking status, sync, init, and doctor
492
- advisories, while keeping explicit feedback collection strict, bounded,
493
- metadata-only, and read-only.
494
-
495
- ## What changed in 4.3.0
496
-
497
- - Made `docs/ENGINEERING.md` the exact managed default layer and added
498
- `docs/project/auto-coding-skill.yaml` as the higher-priority project-owned
499
- overlay. Mappings merge recursively; explicit project scalars and lists replace
500
- defaults, while managed workflow identity and code-enforced safety invariants
501
- remain protected.
502
- - Added one-time semantic migration from the previous manifest-bound default.
503
- Initialization creates the overlay without overwriting an existing one, proves
504
- the reconstructed effective configuration is equivalent, archives the legacy
505
- document, and then converges the managed default byte-for-byte.
506
- - Added `docs/skill-feedback/` with managed guidance/template and preserved
507
- project-owned reports. `autocoding feedback --projects ... --json` performs an
508
- explicit bounded metadata-only read and groups exact root-cause signatures for
509
- later human triage.
510
- - Replaced partial upgrade writes with owner-bound, recoverable installation
511
- transactions, staged/installed hash checks, active-installer rejection, safe
512
- project-file operations, and fail-closed effective-configuration validation.
513
- Project-local `ap.py upgrade --write` is retired in favor of transactional
514
- `autocoding init` or explicit multi-project `sync`.
515
- - Batched protected executable-mode convergence so the safer installer retains
516
- practical upgrade latency while preserving symlink/reparse checks.
517
-
518
- ## What changed in 4.2.8
519
-
520
- - Replaced end-only Reviewer output collection with bounded JSONL event streaming.
521
- Schema-2 Git-local receipts now record CLI/model/effort identity, attempt phases,
522
- first/last event times, event-log hashes, diagnostic categories, and byte counts
523
- without persisting prompt, patch, stderr, tool output, or model content.
524
- - Split runtime startup failure from analysis timeout. A no-semantic-event startup
525
- is capped at 30 seconds and retried once with the same assignment, fingerprint,
526
- artifact, command, and absolute deadline; an analysis that started is never
527
- repeated. Focused review now uses `high/150s`, while deep uses `xhigh/360s`.
528
- - Added `review-runtime-override` for explicit user-authorized delivery after an
529
- exhausted runtime-unavailable or analysis-timeout failure. The private override
530
- binds all immutable review evidence and records `runtime-bypassed`, never
531
- `approved`; failure-time hashes reject pre-authorization tampering, and a
532
- substantive Reviewer result observed before timeout remains non-bypassable.
533
- - Restricted the hidden custom runner to explicit test harnesses and closed spawn
534
- failure, unbounded-output, silent-startup, and lost-diagnostics paths.
535
-
536
- ## What changed in 4.2.7
537
-
538
- - Replaced live pre-commit diff discovery with an immutable Git-local patch
539
- artifact that captures staged, unstaged, untracked, deleted, mode, symlink,
540
- and binary changes in both direct and isolated task modes.
541
- - Bound the artifact's canonical path, format, and SHA-256 into the Reviewer
542
- assignment, then bound the complete assignment SHA-256 into runtime state,
543
- receipt, verdict validation, and final commit authorization. The deadline
544
- starts only after snapshot creation.
545
- - Added the full `review-artifact` command for verified Reviewer access and
546
- lifecycle cleanup for task artifacts after successful direct closure,
547
- isolated cleanup, or explicit orphan pruning. Cleanup now preserves registry
548
- and review evidence on failure so the same task can be retried safely.
549
-
550
- ## What changed in 4.2.6
551
-
552
- - Made the managed `structure` / `optimization` description configuration-neutral:
553
- the preserved project frontmatter is authoritative for enforcement and debt
554
- completion policy.
555
- - Added a 4.2.5-to-4.2.6 initialization regression that preserves strict
556
- `blocking + baseline-aware` project settings and proves repeated init remains
557
- byte-idempotent.
558
-
559
- ## What changed in 4.2.5
560
-
561
- - Prevented standalone JSX `/>` and closing-tag `</...>` slashes from starting a
562
- JavaScript regex token during function-range scans. GeeStack now matches its
563
- TypeScript AST size baseline, and XJMate strict full structure passes cleanly.
564
-
565
- ## What changed in 4.2.4
566
-
567
- - Fixed TypeScript/TSX named and arrow function ranges across generic constraints,
568
- predicates, conditional/object/function return types, and multiple same-line
569
- arrows. Exact arrow locations are evaluated with bounded longest-range selection,
570
- so type arrows neither truncate real bodies nor create duplicate findings.
571
-
572
- ## What changed in 4.2.3
573
-
574
- - Added the explicit `structure.block_warnings` policy with a default of `false`.
575
- File/function size threshold findings become blocking only when the project also
576
- sets `structure.enforcement: blocking`; advisory and default configurations keep
577
- their existing warning-only behavior.
578
- - Recognized semicolon-terminated Java and Kotlin imports during structure checks,
579
- so project-configured forbidden-import layer rules can enforce JVM boundaries.
580
- - Hardened function-range detection for Go receivers, Python suites, brace bodies,
581
- and block/expression arrow functions. Scans are threshold-bounded, honor the
582
- final Gate deadline, and ignore braces in comments and string literals.
583
-
584
- ## What changed in 4.2.2
585
-
586
- - Classified root and nested `package.json` files as `release_or_tooling` high-risk
587
- changes requiring review, and added upgrade regression coverage that preserves
588
- complete project `risk.rules` instead of deleting partially overlapping rules.
589
- - Honored explicit `structure-check --scope full`, included tracked and unignored
590
- untracked files, and restored the enabled structure check inside the final
591
- changed-scope gate. Init now preserves supported strict structure settings and
592
- treats `docs/architecture/structure-standard.md` as scaffold-once project policy.
593
- - Added `review-run` with a validated Git-local assignment, fixed 90/300-second
594
- process-group supervision, one attempt per fingerprint, Reviewer identity/HEAD/
595
- scope binding, safe result normalization, and late-result rejection. The child
596
- runs as a separate read-only Codex process without the lifecycle-owner identity;
597
- the documented template command uses the complete
598
- `python3 docs/tools/autopipeline/ap.py ...` invocation.
599
- - Added `has_task_commits` and prevented a no-commit task from reporting
600
- `merged_into_target=true` merely because its baseline equals the target tip.
601
- - Updated `verify-jenkins`, `verify-jenkins-build`, `verify-target`, and target
602
- health checks to prefer `access.jenkins.<component>` and
603
- `access.project.<component>`, fail before network activity on missing or
604
- ambiguous endpoints, isolate Jenkins crumbs by URL and username, and preserve
605
- legacy `jenkins.*` / `target_env.*` compatibility.
606
-
607
- ## What changed in 4.2.1
608
-
609
- - Made classification take one fail-closed Git porcelain-v2 snapshot and expose
610
- `repo`, `workspace_dirty`, `dirty_paths`, and `active_writer` in normal output.
611
- Direct/dirty decisions are explicit snapshots, and another same-owner direct
612
- claim still counts as a concurrent writer.
613
- - Kept the `reviewer` Agent at `xhigh` while adding one-pass, 90-second focused
614
- review and 300-second deep routing for cross-module, parallel, API, auth,
615
- database, payment, file-transfer, gateway, and production-configuration changes.
616
- Focused timeout is `blocked`, never an approval or a reason to call `task-review`.
617
- - Added `agent-result-template` to generate all 16 Reviewer result fields from a
618
- validated assignment and made contract validation return all known field errors
619
- together. JSON-only repair reuses the completed analysis instead of launching
620
- another substantive review.
621
- - Replaced the login Shell used for routed commands with an inherited-environment
622
- non-login Shell, preserving the caller's PATH and selected Python/Node runtime.
623
- - Excluded `.agents/archive/**`, managed Agent definitions, and the install
624
- manifest from structure scans while continuing to scan project-owned `.agents` code.
625
- - Replaced substring UI classification with exact directory segments and UI-only
626
- extensions, removing false browser signals from files such as `.env.components`.
627
- - Added conservative mechanical/semantic/unknown change classification. Exact
628
- renames and explicitly behavior-stable syncs can avoid generic cross-module
629
- design ceremony; semantic/unknown changes and explicit project design rules
630
- remain protected, and high-risk review is unchanged.
631
-
632
- ## What changed in 4.2.0
633
-
634
- - Made `commit-push` preserve already staged deletions and mixed index/worktree
635
- state, and made ownership/validation see both ends of a rename.
636
- - Added an exact per-worktree final-gate PASS receipt. Registered `commit-push`
637
- reuses it only when content, index, base, scope, validation plan, command hashes,
638
- and writer lease still match; staging or push retries no longer rerun a stable gate.
639
- - Added `task-scope-add` for conflict-checked, monotonic scope expansion before the
640
- first commit. Expansion raises risk when necessary and invalidates review/gate state.
641
- - Clarified that registered tasks call `commit-push` directly; standalone
642
- `light-gate` is for unregistered Git closure or an explicit fresh diagnostic.
643
-
644
- ## What changed in 4.1.9
645
-
646
- - Added a packaged managed-install manifest with path, version, content hash, and
647
- executable-bit expectations for release-owned files.
648
- - Made `autocoding init` verify the written Skill, managed agents, root protocol,
649
- launcher, managed ENGINEERING region, and canonical templates before success.
650
- - Made `doctor` recheck the local manifest and `status` compare it with the current
651
- release while excluding project facts, durable records, custom agents, and archives.
652
-
653
- ## What changed in 4.1.8
654
-
655
- - Made `autocoding init` perform a complete project install or upgrade without a
656
- separate sync/upgrade chain or `--force`.
657
- - Rebuilt ENGINEERING from the current schema, preserving only supported project
658
- values and removing unknown legacy fields and competing workflow text.
659
- - Made the generated docs directory topology and managed templates identical
660
- across projects while preserving mutable project docs and valid durable records;
661
- obsolete paths are archived under `.agents/archive/` before removal.
662
- - Made root AGENTS, managed agents, the Skill copy, and the project launcher fully
663
- converge during init.
664
-
665
- ## What changed in 4.1.7
666
-
667
- - Made tag publication reproducible with explicit Node 24, Python 3.12, and pinned
668
- PyYAML setup before package verification.
669
- - Made `autocoding init` fail before writes with one interpreter-specific recovery
670
- command when its Python runtime cannot import PyYAML.
671
- - Removed the `requests` runtime dependency by using Python's standard HTTP library.
672
- - Made concurrency tests self-contained instead of borrowing local thread and Git
673
- identity configuration.
674
- - Supported both GitHub-secret publication and local-token publication followed by
675
- idempotent tag verification.
676
-
677
- ## What changed in 4.1.4
678
-
679
- - Added clean pre-write direct claims so broad paths cannot adopt unknown existing
680
- dirt after development starts.
681
- - Prevented the current writer lease holder from approving its own diff and added
682
- executable assignment/result contract validation.
683
- - Escalated high-confidence database, authorization, settlement, gateway, and
684
- production intents while keeping ordinary UI work lightweight.
685
- - Made mechanism plans dependency-closed: parallel fixers require reclassification
686
- with multiple writers and therefore require isolated worktrees.
687
- - Published from matching `v*` tags with idempotent npm registry checks and a final
688
- registry verification.
689
-
690
- ## What changed in 4.1.3
691
-
692
- - Made explicit task kinds fail closed: `none` is internal-only and terminal
693
- maintenance cannot be applied to code or unrelated documentation paths.
694
- - Added guarded `--continue-direct` reclassification for already-owned changes;
695
- unknown dirty paths still require isolation.
696
- - Separated intent risk hints from path/rule-confirmed high risk so ordinary login
697
- or payment UI work no longer automatically receives a heavy lifecycle.
698
- - Aligned generated Agent plans with their JSON Schema and exposed optional
699
- read-only Agent candidates without auto-dispatching them.
700
- - Clarified delegated-fixer worktree ownership, reviewer identity, and explorer
701
- source authority.
702
-
703
- ## What changed in 4.1.2
704
-
705
- - Added explicit read-only, change, and terminal-maintenance task kinds so
706
- intent-only questions cannot accidentally trigger code-delivery machinery.
707
- - Split mechanism planning into required, model-selectable, and forbidden sets;
708
- read-only discovery agents remain available when they save time or add needed
709
- expertise.
710
- - Made 120/180-second gate budgets recommended defaults, added route-level
711
- timeouts, and kept project-specific affected-scope overrides valid.
712
- - Rejected unnecessary lifecycle creation before access checks, fetching, locks,
713
- or branch work; corrected the review lifecycle examples.
714
- - Added a release-package assertion and prepack cache cleanup so generated Python
715
- caches cannot leak into the npm tarball.
716
-
717
- ## What changed in 4.1.1
718
-
719
- - Added a machine-readable minimum mechanism plan and rejected unnecessary clean
720
- serial task lifecycles unless explicitly requested.
721
- - Added bounded per-command and total time budgets for the final changed-scope gate.
722
- - Made root `AGENTS.md` the sole behavioral protocol, reduced `SKILL.md` to
723
- invocation guidance, and reduced `ENGINEERING.md` to project configuration and
724
- facts.
725
-
726
- ## What changed in 4.1.0
727
-
728
- - Made root `AGENTS.md` byte-identical across installed projects and whole-file
729
- managed on every full sync/upgrade.
730
- - Established code/tests/runtime as the source of current behavior and
731
- `ENGINEERING.md` as the single project workflow/configuration source.
732
- - Added an engineering-centered documentation framework plus optional
733
- `docs/project/` fact templates.
734
- - Prevented Skill upgrades while any registered task is active and recorded the
735
- installed Skill version in new task manifests.
736
- - Made reviewer scope bounded, with targeted recheck for mechanical docs-only
737
- fixes and non-blocking handling of adjacent findings.
738
- - Defined ledger/archive reconciliation as a terminal maintenance action, so
739
- closing records cannot recursively create more task records.
740
- - Kept explicit Jenkins failure diagnosis in the same requested task while
741
- preserving push-as-completion and no automatic polling.
742
-
743
- ## What changed in 4.0.0
744
-
745
- - Added real `direct` and `isolated` runtime execution modes.
746
- - Stopped creating branches/worktrees for clean serial or no-diff work.
747
- - Added multi-match `validation.routes`, stable command de-duplication,
748
- fail-closed unmapped code, and `validation-map-check`.
749
- - Separated `risk.rules` from executable validation.
750
- - Made ordinary tasks main-agent-first and review/design conditional.
751
- - Kept advanced worktree ownership, dependency SHA, lease, fingerprint,
752
- integration, submodule safety, and branch cleanup for parallel/high-risk work.
753
- - Removed taskbook, closure, evidence, and active-task documents from the default
754
- scaffold and normal closure path.
755
- - Changed structure/debt policy to no-new-debt and kept full/CI/runtime checks as
756
- explicit diagnostics.
757
- - Added schema-3 task manifests and atomic sync refusal for in-flight pre-4.0 tasks.
758
-
759
- ## Maintainer commands
760
-
761
- ```bash
762
- npm run sync-assets
763
- npm run release:check
40
+ autocoding init
41
+ autocoding sync --projects /path/project-a,/path/project-b
42
+ autocoding status --projects . --json
764
43
  ```
765
44
 
766
- `release:check` validates source/assets, Python 3.11 grammar, managed Agent TOML,
767
- CLI/profile/concurrency regressions, and the npm package contents.
768
-
769
- License: MIT.
45
+ These commands only install, preserve, and report documentation. They never run
46
+ project tests, validation, Gates, reviews, builds, Jenkins jobs, deployment, or
47
+ other external-system operations.