@elvis1513/auto-coding-skill 3.0.1 → 3.0.3

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.
@@ -39,7 +39,7 @@ access:
39
39
  password: ""
40
40
 
41
41
  concurrency:
42
- isolation: "worktree"
42
+ isolation: "adaptive"
43
43
  base_ref: ""
44
44
  target_branch: ""
45
45
  branch_prefix: "codex/"
@@ -121,7 +121,7 @@ docs:
121
121
 
122
122
  # Engineering Workflow
123
123
 
124
- <!-- auto-coding-skill:managed-workflow:start version=3.0.1 -->
124
+ <!-- auto-coding-skill:managed-workflow:start version=3.0.3 -->
125
125
 
126
126
  This file is the single project workflow configuration. Keep it in Git. Its YAML
127
127
  frontmatter and content outside these versioned markers are manually maintained;
@@ -162,8 +162,8 @@ Required for normal development:
162
162
  initialization incomplete and block `doctor`; keep each value as an inline
163
163
  YAML string, quoting values that resemble numbers, dates, booleans, or YAML
164
164
  collections
165
- - `concurrency.isolation`: the only supported value is `worktree`; shared-checkout
166
- and legacy write modes are rejected
165
+ - `concurrency.isolation`: use `adaptive`; a clean serial task works directly on
166
+ the current branch, while dirty or parallel write work uses isolated worktrees
167
167
  - `concurrency.base_ref` and `target_branch`: leave blank to derive the current
168
168
  upstream, or set them explicitly for a fixed integration branch
169
169
  - `concurrency.branch_prefix` and `worktree_root`
@@ -179,10 +179,21 @@ callable explicitly with `ap.py run <name>` but never replace the automatic
179
179
  changed gate. `quick_test` remains the only fallback when `gate_changed` is
180
180
  absent. Run `doctor` for exact missing-field diagnostics.
181
181
 
182
+ If the actual changed gate fails only because a dependency already declared in
183
+ the repository lockfile is not installed in the working checkout, restore that
184
+ locked dependency locally (for example with `npm ci`) and rerun the same changed
185
+ gate once. Do not install dependencies or restart a standard/full diagnostic
186
+ that was invoked outside the normal development gate, and never promote this
187
+ recovery into a full-gate run.
188
+
182
189
  ## Concurrent write tasks
183
190
 
184
- Run every write task in a registered Git worktree and task branch. Do not share
185
- the primary worktree between writers.
191
+ For one serial write task, inspect the current checkout first. If it is clean,
192
+ develop directly on the current branch, run the changed gate once, commit, and
193
+ push without creating a temporary branch. If it already contains changes, or if
194
+ multiple writers will run concurrently, preserve the checkout and use one
195
+ registered task branch/worktree per writer. Never share one worktree between
196
+ writers. Clean temporary worktrees and merged task branches after integration.
186
197
 
187
198
  ```bash
188
199
  python3 docs/tools/autopipeline/ap.py task-start <TASK_ID> \
@@ -261,9 +272,10 @@ work concurrently.
261
272
  it is not post-push acceptance.
262
273
  2. The main agent merges discovery evidence and decides the design, dependency
263
274
  graph, and non-overlapping path ownership.
264
- 3. Assign each independent write unit to exactly one `fixer`, one registered task
265
- branch, and one worktree. Never run two writers in one worktree, and do not let
266
- the main agent edit a fixer-owned worktree concurrently.
275
+ 3. Assign each independent write unit to exactly one `fixer`. A single serial
276
+ fixer may use a clean current branch; dirty checkouts and parallel fixers each
277
+ require a registered task branch/worktree. Never run two writers in one worktree.
278
+ Do not let the main agent edit a fixer-owned worktree concurrently.
267
279
  4. Parallelize fixer units only when they are dependency-free and their owned
268
280
  paths do not overlap. Complete implementation → review → gate/integration for
269
281
  the current dependency wave, then start the next wave only after its
@@ -285,7 +297,8 @@ work concurrently.
285
297
  `classify` emits a machine-readable `agent_plan` with this stage and ownership
286
298
  contract. Micro tasks remain main-agent-only unless useful independent work
287
299
  clearly exceeds delegation overhead. If the client cannot run subagents, execute
288
- the same stages sequentially without weakening worktree isolation.
300
+ the same stages sequentially and check for a clean current branch before creating
301
+ any temporary worktree.
289
302
 
290
303
  ## Default workflow
291
304
 
@@ -295,7 +308,7 @@ Development:
295
308
  2. Analyze and decompose the request; run or reason through `classify --scope auto`.
296
309
  3. Fan out independent discovery roles, then fan their results back into one main-agent design.
297
310
  4. Record the smallest useful design note; create DD/ADR only when indicated.
298
- 5. Start one registered worktree per independent write unit with explicit owned paths, dependency SHAs, and writer lease.
311
+ 5. Use the clean current branch for one serial writer; only dirty or parallel work starts registered worktrees with explicit owned paths, dependency SHAs, and writer leases.
299
312
  6. Review stable diffs, record their fingerprints with `task-review`, and return findings to their owning fixers until approved.
300
313
  7. Hand the writer lease to the main agent. The main agent runs `commit-push` for each write task; it enforces ownership,
301
314
  dependency, lease, and current-review gates, executes the one changed-scope fast gate, records
@@ -362,8 +375,9 @@ python3 docs/tools/autopipeline/ap.py docs-ledger-check
362
375
  python3 docs/tools/autopipeline/ap.py gate-profile
363
376
  ```
364
377
 
365
- `light-gate` is available for an explicit gate-diagnostic task. Do not run it
366
- before normal `commit-push`; `commit-push` owns the single automatic gate.
378
+ `light-gate` is available for an explicit gate-diagnostic task. Do not add it to
379
+ normal closure. Run exactly one configured changed gate directly on a clean
380
+ serial branch or through `commit-push` for isolated work.
367
381
 
368
382
  The launcher under `docs/tools/autopipeline` delegates to the single runtime in
369
383
  `.agents/skills/auto-coding-skill/scripts`; do not duplicate that runtime.
@@ -1,12 +1,16 @@
1
+ <!-- auto-coding-skill:managed-agents:start version=3.0.3 -->
2
+
1
3
  Follow `docs/ENGINEERING.md` strictly. Source of truth: `docs/ENGINEERING.md`.
2
4
  At task start, use `ap.py classify --scope auto` when changed-file impact is not obvious.
3
5
  Before code changes, read the configured structure standard and place new code in the correct layer; do not add new responsibilities to already-large files.
4
6
  Obey `structure.layer_rules` import boundaries unless a DD/ADR explicitly records the exception.
5
7
  Before optimization reviews, read the project health baseline and optimization backlog, and report only new, worsened, unrecorded P0/P1, or upgraded-priority items.
6
8
  Route work through already available MCP servers, installed skills, plugins, and app connectors when they provide current docs, design context, browser evidence, GitHub/CI state, security review, or artifact rendering.
7
- Use the `.agents/agents` role model when available. The main agent decomposes and fans out independent explorer/docs_researcher/browser_debugger work, fans evidence into one design, assigns each independent fixer exactly one registered branch/worktree with explicit non-overlapping paths, routes stable diffs through reviewer and back to the owning fixer, then alone runs the fast gate, commits, integrates in dependency order, pushes, and cleans. Never run two writers in one worktree; start dependent write work only after its prerequisite is integrated. If the client cannot run subagents, execute the same stages sequentially without weakening isolation.
9
+ Use the `.agents/agents` role model when available. The main agent decomposes and fans out independent explorer/docs_researcher/browser_debugger work, then fans evidence into one design. For one serial writer, use the current branch when its checkout is clean; create registered task branches/worktrees only when that checkout is already dirty or multiple writers run concurrently. Give every writer explicit non-overlapping paths, route stable diffs through reviewer, then let the main agent run the one fast gate, commit, integrate, push, and clean temporary branches. Never run two writers in one worktree; start dependent write work only after its prerequisite is integrated.
8
10
  Use exactly one fast changed-scope local gate before pushing. Keep advisory structure, docs-ledger, full regression, Docker, API, Jenkins, and target checks outside the default development gate.
9
11
  Keep active docs ledgers small: `taskbook.md`, `closure-log.md`, and top-level `docs/design/T*.md` must be physically archived when over budget. Use `docs-ledger-archive --plan` before `--write`; `archive-index.md` is only navigation.
10
12
  Keep `docs/tasks/evidence.jsonl` and closure Markdown aligned with actual executed checks.
11
13
  Treat target-branch push as coding completion. Do not wait for, poll, diagnose, or fix Jenkins/build/deploy results unless the user starts a separate diagnostic task; the project owner performs acceptance.
12
14
  Use the direct URL, username, and password values configured under `access.*` in `docs/ENGINEERING.md`; do not invent credentials.
15
+
16
+ <!-- auto-coding-skill:managed-agents:end -->