@elvis1513/auto-coding-skill 4.1.1 → 4.1.3
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.
- package/README.md +52 -8
- package/cli/assets/agents/explorer.toml +1 -1
- package/cli/assets/agents/fixer.toml +3 -3
- package/cli/assets/skill/SKILL.md +30 -25
- package/cli/assets/skill/data/contracts/orchestration-v1.schema.json +19 -1
- package/cli/assets/skill/data/policies/workflow-migrations-v1.json +2 -2
- package/cli/assets/skill/data/templates/ENGINEERING.md +6 -6
- package/cli/assets/skill/data/templates/bridges/AGENTS.md +19 -15
- package/cli/assets/skill/scripts/ap.py +371 -99
- package/cli/src/check-package.js +38 -0
- package/cli/src/index.js +5 -5
- package/cli/tests/ap_concurrency_test.py +73 -0
- package/cli/tests/ap_profiles_test.py +178 -5
- package/cli/tests/index.test.js +12 -12
- package/package.json +5 -2
- package/cli/assets/skill/scripts/__pycache__/core.cpython-314.pyc +0 -0
- package/cli/assets/skill/scripts/__pycache__/scaffold_templates.cpython-314.pyc +0 -0
package/README.md
CHANGED
|
@@ -9,7 +9,10 @@ The Skill is a selectable guardrail, not a command sequence that must run for
|
|
|
9
9
|
every task. The model skips machinery whose expected benefit does not exceed its
|
|
10
10
|
cost; read-only work and obvious small clean-checkout changes normally stay direct.
|
|
11
11
|
|
|
12
|
-
Version 4.1.
|
|
12
|
+
Version 4.1.3 closes classification bypasses and keeps current-task changes direct
|
|
13
|
+
without weakening dirty-workspace isolation. Version 4.1.2 made mechanisms required,
|
|
14
|
+
model-selectable when beneficial, or forbidden for the current task. Version
|
|
15
|
+
4.1.1 consolidated the 4.x delivery-first workflow into one canonical
|
|
13
16
|
repository contract. Version 4.0 replaced the 3.x governance-first defaults with progressive
|
|
14
17
|
guardrails. Clean single-writer work stays on the current branch. Worktrees,
|
|
15
18
|
parallel fixers, fingerprint review, durable design records, and stronger
|
|
@@ -65,17 +68,27 @@ The result includes:
|
|
|
65
68
|
- `execution_mode=isolated`: dirty checkout, explicit isolation, or multiple writers.
|
|
66
69
|
- `execution_mode=none`: no task signal or change; create no branch/worktree.
|
|
67
70
|
- `review_required` and `design_required`: risk-based escalation decisions.
|
|
71
|
+
- `task_kind`: `read_only`, `change`, `terminal_maintenance`, or internal `none`.
|
|
68
72
|
- `mechanism_plan.required`: the complete minimum mechanism set for the task.
|
|
69
|
-
- `mechanism_plan.
|
|
73
|
+
- `mechanism_plan.optional_when_beneficial`: model-selectable mechanisms whose
|
|
74
|
+
expected benefit must exceed coordination cost.
|
|
75
|
+
- `mechanism_plan.forbidden`: mechanisms that stay off unless the user overrides.
|
|
76
|
+
- `optional_agents`: model-selectable explorer/docs/browser candidates; they are
|
|
77
|
+
never automatic stages.
|
|
70
78
|
- an adaptive agent plan that does not force fan-out for ordinary work.
|
|
71
79
|
|
|
72
80
|
Clean serial work proceeds directly on the current branch with normal Git; it does
|
|
73
81
|
not create machine lifecycle state. Use `task-start` only when classification
|
|
74
82
|
requires isolation/review or the user explicitly requests lifecycle tracking:
|
|
75
83
|
|
|
84
|
+
When a direct task materially changes scope after writing, reclassify with
|
|
85
|
+
`--continue-direct` and repeat every current task `--planned-path`. Undeclared
|
|
86
|
+
dirty paths, another writer, or mandatory isolation still selects a worktree. If
|
|
87
|
+
the new plan requires review lifecycle, reuse `--continue-direct` on `task-start`.
|
|
88
|
+
|
|
76
89
|
```bash
|
|
77
90
|
python3 docs/tools/autopipeline/ap.py task-start T0001 \
|
|
78
|
-
--owned-path backend/internal/orders
|
|
91
|
+
--owned-path backend/internal/orders --force-lifecycle
|
|
79
92
|
python3 docs/tools/autopipeline/ap.py commit-push T0001 \
|
|
80
93
|
--msg "T0001: fix order retry"
|
|
81
94
|
```
|
|
@@ -90,8 +103,10 @@ Dirty or parallel work receives an isolated task branch/worktree:
|
|
|
90
103
|
python3 docs/tools/autopipeline/ap.py task-start T0002 \
|
|
91
104
|
--owned-path backend --review-required
|
|
92
105
|
# implement in the printed worktree
|
|
106
|
+
python3 docs/tools/autopipeline/ap.py task-status T0002 --json
|
|
93
107
|
python3 docs/tools/autopipeline/ap.py task-review T0002 \
|
|
94
|
-
--verdict approved --diff-fingerprint "$
|
|
108
|
+
--verdict approved --diff-fingerprint "$CURRENT_DIFF_FINGERPRINT" \
|
|
109
|
+
--reviewer "$REVIEWER_ID"
|
|
95
110
|
python3 docs/tools/autopipeline/ap.py commit-push T0002 \
|
|
96
111
|
--msg "T0002: bounded change"
|
|
97
112
|
python3 docs/tools/autopipeline/ap.py task-integrate T0002
|
|
@@ -122,6 +137,7 @@ validation:
|
|
|
122
137
|
- name: backend
|
|
123
138
|
paths: ["backend/**", "contracts/**"]
|
|
124
139
|
commands: [backend_fast]
|
|
140
|
+
timeout_seconds: 90
|
|
125
141
|
- name: frontend
|
|
126
142
|
paths: ["frontend/**"]
|
|
127
143
|
commands: [frontend_fast]
|
|
@@ -152,9 +168,10 @@ closure gate is limited to one stable-diff run. Full regression, Docker, builds,
|
|
|
152
168
|
Jenkins, deployment, API verification, and target checks remain explicit
|
|
153
169
|
diagnostics outside normal closure.
|
|
154
170
|
|
|
155
|
-
The final closure
|
|
156
|
-
seconds total
|
|
157
|
-
route
|
|
171
|
+
The recommended final closure defaults are 120 seconds per command and 180
|
|
172
|
+
seconds total. Projects may raise them for measured affected-scope checks or set
|
|
173
|
+
a smaller `timeout_seconds` on a route. A timeout calls for a narrower route and
|
|
174
|
+
never triggers a broader fallback.
|
|
158
175
|
|
|
159
176
|
## Risk and subagent policy
|
|
160
177
|
|
|
@@ -192,11 +209,38 @@ project-specific facts outside the managed ENGINEERING block. Root `AGENTS.md` i
|
|
|
192
209
|
replaced as a whole; its previous content is archived under
|
|
193
210
|
`docs/archive/workflow/` as historical, non-authoritative context.
|
|
194
211
|
|
|
212
|
+
## What changed in 4.1.3
|
|
213
|
+
|
|
214
|
+
- Made explicit task kinds fail closed: `none` is internal-only and terminal
|
|
215
|
+
maintenance cannot be applied to code or unrelated documentation paths.
|
|
216
|
+
- Added guarded `--continue-direct` reclassification for already-owned changes;
|
|
217
|
+
unknown dirty paths still require isolation.
|
|
218
|
+
- Separated intent risk hints from path/rule-confirmed high risk so ordinary login
|
|
219
|
+
or payment UI work no longer automatically receives a heavy lifecycle.
|
|
220
|
+
- Aligned generated Agent plans with their JSON Schema and exposed optional
|
|
221
|
+
read-only Agent candidates without auto-dispatching them.
|
|
222
|
+
- Clarified delegated-fixer worktree ownership, reviewer identity, and explorer
|
|
223
|
+
source authority.
|
|
224
|
+
|
|
225
|
+
## What changed in 4.1.2
|
|
226
|
+
|
|
227
|
+
- Added explicit read-only, change, and terminal-maintenance task kinds so
|
|
228
|
+
intent-only questions cannot accidentally trigger code-delivery machinery.
|
|
229
|
+
- Split mechanism planning into required, model-selectable, and forbidden sets;
|
|
230
|
+
read-only discovery agents remain available when they save time or add needed
|
|
231
|
+
expertise.
|
|
232
|
+
- Made 120/180-second gate budgets recommended defaults, added route-level
|
|
233
|
+
timeouts, and kept project-specific affected-scope overrides valid.
|
|
234
|
+
- Rejected unnecessary lifecycle creation before access checks, fetching, locks,
|
|
235
|
+
or branch work; corrected the review lifecycle examples.
|
|
236
|
+
- Added a release-package assertion and prepack cache cleanup so generated Python
|
|
237
|
+
caches cannot leak into the npm tarball.
|
|
238
|
+
|
|
195
239
|
## What changed in 4.1.1
|
|
196
240
|
|
|
197
241
|
- Added a machine-readable minimum mechanism plan and rejected unnecessary clean
|
|
198
242
|
serial task lifecycles unless explicitly requested.
|
|
199
|
-
- Added
|
|
243
|
+
- Added bounded per-command and total time budgets for the final changed-scope gate.
|
|
200
244
|
- Made root `AGENTS.md` the sole behavioral protocol, reduced `SKILL.md` to
|
|
201
245
|
invocation guidance, and reduced `ENGINEERING.md` to project configuration and
|
|
202
246
|
facts.
|
|
@@ -8,7 +8,7 @@ developer_instructions = """
|
|
|
8
8
|
|
|
9
9
|
工作要求:
|
|
10
10
|
1. 优先用 rg / git / 项目脚本快速定位真实入口,再向下追踪调用链。
|
|
11
|
-
2.
|
|
11
|
+
2. 以当前代码、测试、schema、migration 和运行配置为事实源;仅按任务需要读取 ENGINEERING 或当前接口文档,任务记录只在明确追溯历史时读取。涉及结构/优化时,仅在结构标准、健康基线或 optimization backlog 已存在时读取,不因可选文档缺失而阻塞。
|
|
12
12
|
3. 必要时运行 `ap.py classify --scope auto`,说明执行模式、风险和是否确实需要设计/reviewer;普通任务默认主代理直接完成,本地 gate 始终保持 changed-scope。
|
|
13
13
|
4. 输出根因时说明当前代码应归属的层级、现有同类模块、可复用 helper / 组件 / 脚本,以及是否触及 `structure.layer_rules`。
|
|
14
14
|
5. 如果根因依赖外部库、框架或版本行为,标记需要 docs_researcher / Context7 继续核对,不要凭记忆猜。
|
|
@@ -7,9 +7,9 @@ developer_instructions = """
|
|
|
7
7
|
你负责小而准的实现修改。
|
|
8
8
|
|
|
9
9
|
工作要求:
|
|
10
|
-
1. 只在问题已经相对明确,并收到 task_id、execution_mode、owned_paths、depends_on 和 acceptance
|
|
11
|
-
2.
|
|
12
|
-
3. 一个 worktree 只能有一个写入者。只修改 owned_paths
|
|
10
|
+
1. 只在问题已经相对明确,并收到 task_id、execution_mode、owned_paths、depends_on 和 acceptance 后再动手。
|
|
11
|
+
2. fixer 属于 delegated writer,必须使用注册的 isolated worktree;direct 模式保留给干净 checkout 中独立开发的主代理。确认当前目录和 task_branch 与分配一致,否则停止并上报。
|
|
12
|
+
3. 一个 worktree 只能有一个写入者。只修改 owned_paths,不进入其他任务工作树;发现未知改动立即停止,不 restore/reset/stash/clean。
|
|
13
13
|
4. 坚持“最小可辩护修改”原则。
|
|
14
14
|
5. 除非明确要求,否则不要重构、不做大面积改名、不改无关文件。
|
|
15
15
|
6. 先跟随项目现有模式、helper、组件、脚本和文档约束,不新造不必要抽象。
|
|
@@ -1,16 +1,15 @@
|
|
|
1
1
|
---
|
|
2
2
|
name: auto-coding-skill
|
|
3
|
-
description:
|
|
3
|
+
description: Optional delivery-first engineering workflow for repository changes. Use when its bounded validation, adaptive Git isolation, review, or coordination will improve delivery speed or defect prevention. Keep clean serial work on the current branch and let the model select beneficial design or subagents without turning them into mandatory ceremony.
|
|
4
4
|
---
|
|
5
5
|
|
|
6
6
|
# Auto Coding Skill
|
|
7
7
|
|
|
8
8
|
## Use the two authorities
|
|
9
9
|
|
|
10
|
-
Read root `AGENTS.md` for the shared behavioral protocol
|
|
11
|
-
`docs/ENGINEERING.md`
|
|
12
|
-
|
|
13
|
-
from historical task documents.
|
|
10
|
+
Read root `AGENTS.md` for the shared behavioral protocol and
|
|
11
|
+
`docs/ENGINEERING.md` for project facts, access, risk, validation, and budgets.
|
|
12
|
+
Do not reconstruct or duplicate those rules from historical task documents.
|
|
14
13
|
|
|
15
14
|
Normal delivery is:
|
|
16
15
|
|
|
@@ -19,24 +18,32 @@ changed-scope gate → commit/push`.
|
|
|
19
18
|
|
|
20
19
|
## Select the minimum mechanism set
|
|
21
20
|
|
|
22
|
-
|
|
23
|
-
ledger maintenance
|
|
21
|
+
Read-only questions need no workflow command. Obvious clean serial edits and
|
|
22
|
+
terminal ledger maintenance may follow root `AGENTS.md` directly. When task kind,
|
|
23
|
+
impact, risk, or concurrency is unclear, run:
|
|
24
24
|
|
|
25
25
|
```bash
|
|
26
26
|
python3 docs/tools/autopipeline/ap.py classify --scope auto \
|
|
27
|
-
--planned-path <PATH> --intent "<intent>" [--writers <N>]
|
|
27
|
+
--planned-path <PATH> --intent "<intent>" [--writers <N>] \
|
|
28
|
+
[--task-kind read_only|change|terminal_maintenance]
|
|
28
29
|
```
|
|
29
30
|
|
|
30
|
-
|
|
31
|
-
`
|
|
32
|
-
|
|
31
|
+
Run `mechanism_plan.required`. The model may select
|
|
32
|
+
`optional_when_beneficial` only when expected value exceeds coordination cost.
|
|
33
|
+
Do not run `forbidden` mechanisms unless the user explicitly overrides the plan.
|
|
34
|
+
Reclassify only after material change. For already-direct work, add `--continue-direct`
|
|
35
|
+
with every task path and reuse it on `task-start` if needed; unknown dirt isolates.
|
|
33
36
|
|
|
34
37
|
Use the registered lifecycle only when classification requires isolation or
|
|
35
38
|
fingerprinted review, or when the user explicitly requests it:
|
|
36
39
|
|
|
37
40
|
```bash
|
|
38
41
|
python3 docs/tools/autopipeline/ap.py task-start T0001 \
|
|
39
|
-
--owned-path src [--isolated] [--review-required]
|
|
42
|
+
--owned-path src [--isolated] [--review-required] [--continue-direct] [--force-lifecycle]
|
|
43
|
+
# For review-required work, obtain the current fingerprint and approve it:
|
|
44
|
+
python3 docs/tools/autopipeline/ap.py task-status T0001 --json
|
|
45
|
+
python3 docs/tools/autopipeline/ap.py task-review T0001 \
|
|
46
|
+
--verdict approved --diff-fingerprint <CURRENT_DIFF_FINGERPRINT> --reviewer <REVIEWER_ID>
|
|
40
47
|
python3 docs/tools/autopipeline/ap.py commit-push T0001 --msg "T0001: summary"
|
|
41
48
|
# Isolated tasks only:
|
|
42
49
|
python3 docs/tools/autopipeline/ap.py task-integrate T0001
|
|
@@ -54,21 +61,20 @@ python3 docs/tools/autopipeline/ap.py light-gate --scope changed --explain
|
|
|
54
61
|
```
|
|
55
62
|
|
|
56
63
|
Every changed code/config path must map to a real project command. The runtime
|
|
57
|
-
de-duplicates
|
|
58
|
-
|
|
59
|
-
|
|
64
|
+
de-duplicates commands and enforces route, command, and total budgets. Defaults
|
|
65
|
+
are 120 seconds per command and 180 total; projects may raise them for a measured
|
|
66
|
+
affected-scope check. Timeouts should narrow the route, never trigger a full build.
|
|
67
|
+
Focused test/fix/retest loops remain allowed before the final gate.
|
|
60
68
|
|
|
61
69
|
Run full regression, Docker, Jenkins, deployment, browser/device writes, or target
|
|
62
|
-
acceptance only as an explicit diagnostic.
|
|
63
|
-
|
|
64
|
-
new ledger lifecycle.
|
|
70
|
+
acceptance only as an explicit diagnostic. A requested just-pushed failure
|
|
71
|
+
diagnosis continues in the same task without inventing another lifecycle.
|
|
65
72
|
|
|
66
73
|
## Keep state and upgrades safe
|
|
67
74
|
|
|
68
|
-
Machine task state, leases, fingerprints, and gate evidence stay in Git
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
and commits once without creating another lifecycle.
|
|
75
|
+
Machine task state, leases, fingerprints, and gate evidence stay in Git local
|
|
76
|
+
state. Ordinary work creates no taskbook, closure, evidence JSONL, active-task,
|
|
77
|
+
or design artifact. Terminal ledger reconciliation validates and commits once.
|
|
72
78
|
|
|
73
79
|
```bash
|
|
74
80
|
autocoding init
|
|
@@ -78,6 +84,5 @@ python3 docs/tools/autopipeline/ap.py upgrade --write
|
|
|
78
84
|
python3 docs/tools/autopipeline/ap.py doctor
|
|
79
85
|
```
|
|
80
86
|
|
|
81
|
-
Finish registered tasks before changing
|
|
82
|
-
|
|
83
|
-
and access values, and converges all managed Skill/runtime/agent assets together.
|
|
87
|
+
Finish registered tasks before changing versions. Sync replaces root `AGENTS.md`,
|
|
88
|
+
archives it, preserves project facts/access, and converges all managed assets.
|
|
@@ -13,9 +13,11 @@
|
|
|
13
13
|
},
|
|
14
14
|
"mechanismPlan": {
|
|
15
15
|
"type": "object",
|
|
16
|
-
"required": ["required", "not_required", "lifecycle_required", "rule"],
|
|
16
|
+
"required": ["required", "optional_when_beneficial", "forbidden", "not_required", "lifecycle_required", "rule"],
|
|
17
17
|
"properties": {
|
|
18
18
|
"required": { "$ref": "#/$defs/stringArray" },
|
|
19
|
+
"optional_when_beneficial": { "$ref": "#/$defs/stringArray" },
|
|
20
|
+
"forbidden": { "$ref": "#/$defs/stringArray" },
|
|
19
21
|
"not_required": { "$ref": "#/$defs/stringArray" },
|
|
20
22
|
"lifecycle_required": { "type": "boolean" },
|
|
21
23
|
"rule": { "type": "string", "minLength": 1 }
|
|
@@ -382,6 +384,7 @@
|
|
|
382
384
|
"policies",
|
|
383
385
|
"assignment_contract",
|
|
384
386
|
"result_contract",
|
|
387
|
+
"optional_roles",
|
|
385
388
|
"stages",
|
|
386
389
|
"constraints"
|
|
387
390
|
],
|
|
@@ -407,6 +410,9 @@
|
|
|
407
410
|
"result_contract": {
|
|
408
411
|
"$ref": "#/$defs/stringArray"
|
|
409
412
|
},
|
|
413
|
+
"optional_roles": {
|
|
414
|
+
"$ref": "#/$defs/stringArray"
|
|
415
|
+
},
|
|
410
416
|
"stages": {
|
|
411
417
|
"type": "array",
|
|
412
418
|
"items": {
|
|
@@ -452,9 +458,12 @@
|
|
|
452
458
|
"classification_files",
|
|
453
459
|
"intent_provided",
|
|
454
460
|
"intent_categories",
|
|
461
|
+
"intent_risk_candidates",
|
|
462
|
+
"task_kind",
|
|
455
463
|
"profile",
|
|
456
464
|
"categories",
|
|
457
465
|
"recommended_agents",
|
|
466
|
+
"optional_agents",
|
|
458
467
|
"agent_plan",
|
|
459
468
|
"mechanism_plan"
|
|
460
469
|
],
|
|
@@ -480,6 +489,12 @@
|
|
|
480
489
|
"intent_categories": {
|
|
481
490
|
"$ref": "#/$defs/stringArray"
|
|
482
491
|
},
|
|
492
|
+
"intent_risk_candidates": {
|
|
493
|
+
"$ref": "#/$defs/stringArray"
|
|
494
|
+
},
|
|
495
|
+
"task_kind": {
|
|
496
|
+
"enum": ["none", "read_only", "change", "terminal_maintenance"]
|
|
497
|
+
},
|
|
483
498
|
"profile": {
|
|
484
499
|
"enum": [
|
|
485
500
|
"micro",
|
|
@@ -493,6 +508,9 @@
|
|
|
493
508
|
"recommended_agents": {
|
|
494
509
|
"$ref": "#/$defs/stringArray"
|
|
495
510
|
},
|
|
511
|
+
"optional_agents": {
|
|
512
|
+
"$ref": "#/$defs/stringArray"
|
|
513
|
+
},
|
|
496
514
|
"agent_plan": {
|
|
497
515
|
"$ref": "#/$defs/agentPlan"
|
|
498
516
|
},
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
---
|
|
2
2
|
workflow:
|
|
3
|
-
skill_version: "4.1.
|
|
3
|
+
skill_version: "4.1.3"
|
|
4
4
|
mode: "dev"
|
|
5
5
|
profile: "auto"
|
|
6
6
|
completion: "push"
|
|
@@ -56,8 +56,8 @@ commands:
|
|
|
56
56
|
|
|
57
57
|
validation:
|
|
58
58
|
on_unmapped: "error"
|
|
59
|
-
#
|
|
60
|
-
#
|
|
59
|
+
# Recommended fast defaults. Raise only for a measured affected-scope check;
|
|
60
|
+
# an individual route may set a smaller timeout_seconds.
|
|
61
61
|
max_command_seconds: 120
|
|
62
62
|
max_total_seconds: 180
|
|
63
63
|
# Initialization must replace this with real path-to-command mappings. Empty
|
|
@@ -94,7 +94,7 @@ docs:
|
|
|
94
94
|
|
|
95
95
|
# Engineering Configuration and Project Facts
|
|
96
96
|
|
|
97
|
-
<!-- auto-coding-skill:managed-workflow:start version=4.1.
|
|
97
|
+
<!-- auto-coding-skill:managed-workflow:start version=4.1.3 -->
|
|
98
98
|
|
|
99
99
|
Root `AGENTS.md` is the single shared behavioral protocol. The installed
|
|
100
100
|
`SKILL.md` contains invocation guidance. This file is the single source for the
|
|
@@ -107,8 +107,8 @@ The frontmatter contract is:
|
|
|
107
107
|
- `concurrency`: target branch and adaptive isolation settings.
|
|
108
108
|
- `commands`: project-native executable commands.
|
|
109
109
|
- `validation`: explicit path-to-command routes plus the bounded final-gate time
|
|
110
|
-
budget. Code/config paths must be mapped;
|
|
111
|
-
|
|
110
|
+
budget. Code/config paths must be mapped; route-level `timeout_seconds` can
|
|
111
|
+
bound a slower command without changing the project defaults.
|
|
112
112
|
- `risk`: project-specific signals that increase design/review depth but never
|
|
113
113
|
expand the automatic local gate.
|
|
114
114
|
- `structure` and `optimization`: advisory architecture and no-new-debt policy.
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
<!-- auto-coding-skill:managed-agents:start version=4.1.
|
|
1
|
+
<!-- auto-coding-skill:managed-agents:start version=4.1.3 -->
|
|
2
2
|
# Shared Engineering Protocol
|
|
3
3
|
|
|
4
4
|
This file is fully managed by `auto-coding-skill`. Keep project-specific facts,
|
|
@@ -26,7 +26,7 @@ The default budget is:
|
|
|
26
26
|
|
|
27
27
|
| Situation | Required mechanisms |
|
|
28
28
|
| --- | --- |
|
|
29
|
-
| Read-only question |
|
|
29
|
+
| Read-only question | analysis only; no workflow command |
|
|
30
30
|
| Obvious clean serial edit | analysis, affected gate, commit/push |
|
|
31
31
|
| Terminal ledger/archive maintenance | targeted consistency check, one commit |
|
|
32
32
|
| Dirty checkout or concurrent writer | registered lifecycle and isolated worktree |
|
|
@@ -34,9 +34,11 @@ The default budget is:
|
|
|
34
34
|
| Multiple independent writers | one task/worktree/lease per writer, ordered integration |
|
|
35
35
|
|
|
36
36
|
Do not add classify, task lifecycle, durable design, subagents, reviewer, or a
|
|
37
|
-
worktree
|
|
38
|
-
|
|
39
|
-
`
|
|
37
|
+
worktree merely because the tool exists. If task kind or impact is unclear, run
|
|
38
|
+
`classify`: run `mechanism_plan.required`, let the model select
|
|
39
|
+
`optional_when_beneficial` only when expected value exceeds coordination cost,
|
|
40
|
+
and keep `forbidden` mechanisms off unless the user explicitly overrides the
|
|
41
|
+
plan. Reclassify only after a material task-kind, scope, risk, or writer change.
|
|
40
42
|
|
|
41
43
|
Normal delivery is:
|
|
42
44
|
|
|
@@ -49,11 +51,11 @@ failure may continue in the same conversation/task without a second lifecycle.
|
|
|
49
51
|
|
|
50
52
|
## Git and parallel work
|
|
51
53
|
|
|
52
|
-
- One writer in a clean checkout works directly
|
|
53
|
-
|
|
54
|
+
- One writer in a clean checkout works directly without a lifecycle. On
|
|
55
|
+
reclassification, `--continue-direct` requires every dirty path to be declared.
|
|
54
56
|
- Existing unrelated changes, another writer, or configured mandatory isolation
|
|
55
57
|
requires a registered task branch/worktree.
|
|
56
|
-
- Every
|
|
58
|
+
- Every delegated fixer or parallel writer owns a task ID/worktree, writer lease,
|
|
57
59
|
non-overlapping `owned_paths`, and prerequisite commit SHAs.
|
|
58
60
|
- Never let multiple writing agents share a checkout. Never restore, reset, stash,
|
|
59
61
|
clean, overwrite, or commit unknown changes.
|
|
@@ -64,14 +66,15 @@ failure may continue in the same conversation/task without a second lifecycle.
|
|
|
64
66
|
|
|
65
67
|
## Design, agents, and review
|
|
66
68
|
|
|
67
|
-
- Micro and standard tasks stay main-agent-only unless
|
|
68
|
-
|
|
69
|
+
- Micro and standard tasks stay main-agent-only unless the model identifies an
|
|
70
|
+
independent question whose latency or expertise benefit exceeds coordination
|
|
71
|
+
cost.
|
|
69
72
|
- Create DD/ADR only for lasting cross-module, API, data, security, deployment,
|
|
70
73
|
or key user-flow decisions.
|
|
71
74
|
- Use read-only explorer/docs/browser agents only for independent questions.
|
|
72
75
|
- Use parallel fixers only for bounded, dependency-free, non-overlapping units.
|
|
73
|
-
- Require fingerprinted review only for high-risk,
|
|
74
|
-
or
|
|
76
|
+
- Require fingerprinted review only for path/rule-confirmed high-risk,
|
|
77
|
+
contract-crossing, parallel, or configured work; intent words are candidates only.
|
|
75
78
|
- Review blocks only defects introduced or worsened in the promised scope.
|
|
76
79
|
Adjacent existing issues are non-blocking follow-ups.
|
|
77
80
|
- Semantic changes invalidate approval. Mechanical documentation-only corrections
|
|
@@ -90,9 +93,10 @@ failure may continue in the same conversation/task without a second lifecycle.
|
|
|
90
93
|
validation.
|
|
91
94
|
- Focused tests may run and rerun during development. Run one final routed gate
|
|
92
95
|
after the diff is stable.
|
|
93
|
-
- The final gate
|
|
94
|
-
|
|
95
|
-
|
|
96
|
+
- The final gate defaults to 120 seconds per command and 180 seconds total. A
|
|
97
|
+
project may raise either budget for a measured affected-scope check, and may set
|
|
98
|
+
a smaller `timeout_seconds` on an individual route. A timeout should narrow the
|
|
99
|
+
route; never expand or retry it as a full gate.
|
|
96
100
|
- Do not proactively install dependencies. Only after the selected route fails
|
|
97
101
|
because a repository-locked dependency is absent may it be restored once and
|
|
98
102
|
only that route retried.
|