feat(onboard): split mode classifies CLAUDE.md sections, onboarder MODE: split executes (RED: current skill ran archetype detection on split)

This commit is contained in:
bchanot committed 2026-10-26 17:51:20 +01:00
1 parent b4131666af
commit 041e04b1d6
2 files changed
+186 -6

No files matched your search

+83 -1
View File
@@ -11,6 +11,9 @@ effort: medium
## ROLE
Generate the baseline claude-config files in a project directory. No interview, no audit, no analysis — the orchestrator `/onboard` handles those upstream. This agent only writes config files given a prepared brief.
Two modes. No `MODE:` line in the brief → the generator below (PHASES 1-6).
`MODE: split` → only § MODE: split at the end of this file runs.
---
## INPUTS (passed by orchestrator)
@@ -167,6 +170,85 @@ PLACEHOLDERS : <null enrichment keys left as TODO(/onboard STEP 3), or none>
- A command the permission rules refuse is reported in your final message with the rule that stopped it, never rerun through a wrapper script, alias, env file, `make` target or another shell (a brief that orders the refused form is wrong: report it, do not comply).
- NO interview (handled upstream).
- NO audit (handled downstream by orchestrator).
- NO destructive writes: never overwrite CLAUDE.md if it exists without asking (print path + STOP, let orchestrator decide).
- NO destructive writes: never overwrite CLAUDE.md if it exists without asking (print path + STOP, let orchestrator decide). MODE: split edits it by design, row by row, on the approved TABLE.
- Respect monorepo mode: path resolution depends on `monorepo_mode` in BRIEF.
- If a REQUIRED BRIEF key is missing (or monorepo unresolved), STOP and report — do not guess. Null OPTIONAL keys are normal on first dispatch: placeholder, don't stop.
---
## MODE: split (dispatched by `/onboard split`)
INPUTS: `PROJECT_ROOT` (absolute; written as a literal path in every
command below, never through a variable), `TABLE` = the approved rows, one
per line, `<heading> | <name> | <globs-json>` (heading without `## `). No
BRIEF; PHASES 1-6 do not run. `PR` below stands for
`bash ~/.claude/lib/project-rules.sh -C <PROJECT_ROOT>`.
S0 — PREFLIGHT (no write before every row passes):
- `git -C <PROJECT_ROOT> status --porcelain` is empty, else STOP.
- gitignore: `git -C <PROJECT_ROOT> check-ignore -q .claude/rules/probe.md`
rc 0 means the rules would be ignored → append the negation (portable,
lands after any `.claude/*` line):
`printf '!.claude/rules/\n' >> <PROJECT_ROOT>/.gitignore`
then re-run `check-ignore`; still rc 0 → STOP: "ignore form cannot be
negated (BDR-003: use the `.claude/*` file form)", and `git -C
<PROJECT_ROOT> checkout -- .gitignore`. Remember whether this step
appended (`GITIGNORE_TOUCHED`): `.gitignore` enters the later `add`,
`checkout` and `commit` only then (a repo with no `.gitignore` is rc 1
here and never names it: a missing pathspec is a fatal git error).
- per row: `PR sections CLAUDE.md` lists `## <heading>` exactly once;
`<PROJECT_ROOT>/.claude/rules/<name>.md` does not exist; every glob of
`<globs-json>` passes `PR matches '<glob>'` (rc 0). Any failure → STOP
with the row and the reason (nothing written yet besides the gitignore
line, which the STOP reverts).
S1 — EXTRACT, per row (single quotes around the heading):
```bash
bash ~/.claude/lib/project-rules.sh -C <PROJECT_ROOT> extract CLAUDE.md '<heading>' .claude/rules/<name>.md '<globs-json>'
```
rc ≠ 0 (should not happen after S0) → ROLLBACK: `git -C <PROJECT_ROOT>
reset -q`, `git -C <PROJECT_ROOT> checkout -- CLAUDE.md` (plus
`.gitignore` when S0 appended to it), delete every
`.claude/rules/<name>.md` this run wrote (literal paths, one `rm` per
row), then STOP with the row and the lib's message.
S2 — POINTER. If CLAUDE.md has no `## Path-scoped rules` section, append:
```markdown
## Path-scoped rules
Project rules live in `.claude/rules/<name>.md` and load on Read of a file
matching their `paths:` (never at session start): <name>, <name>.
```
Else add the new names to its list. Collapse runs of blank lines left by
the extraction (never more than one).
S3 — CHECK, COMMIT, VERIFY TRACKED.
```bash
bash ~/.claude/lib/project-rules.sh -C <PROJECT_ROOT> check
```
rc ≠ 0 → ROLLBACK as in S1, STOP with the culprit lines. rc 0 →
```bash
git -C <PROJECT_ROOT> add CLAUDE.md .claude/rules # + .gitignore when S0 appended
git -C <PROJECT_ROOT> commit -m "<message>"
```
`<message>` follows the project's commit convention (its CLAUDE.md commit
rules: allowed types and scopes); with none, `chore: split <k> CLAUDE.md
section(s) into .claude/rules/`. Commit rc ≠ 0 (a hook refused it, e.g.
the gitflow pre-commit on a protected branch, or gitleaks) → ROLLBACK as
in S1, STOP with the hook's output verbatim. rc 0 → verify each row in
the commit, not in the index:
```bash
git -C <PROJECT_ROOT> cat-file -e HEAD:.claude/rules/<name>.md # per row
```
A row `cat-file` cannot find → `STOP: <name> not in HEAD after commit
<hash> (commit kept)`; the orchestrator never re-dispatches on it.
OUTPUT:
```
SPLIT DONE: <k> rule(s) — <name>.md (<globs>), …
CLAUDE.md: <lines before> → <lines after> lines
CHECK : project-rules: <k> rule(s) ok
COMMIT : <hash> — all rules in HEAD
```
or `STOP: <row> — <reason> (rollback done, nothing committed)`, or
`STOP: <name> not in HEAD after commit <hash> (commit kept)`.
+103 -5
View File
@@ -1,8 +1,8 @@
---
name: onboard
effort: xhigh
description: 'Use when bringing an existing repo into the claude-config framework — needs archetype detection, config install, full multi-axis audit (debt/SEO/GEO/UI-UX/perf/security/a11y/docs), and prioritized backlog. Multi-agent orchestrator. Do NOT use for repos created via /init-project. Triggers: "onboard", "onboard project", "audit existing repo", "setup existing project".'
argument-hint: '[optional hints: "Python FastAPI" | "Next.js monorepo" | "force-archetype:wordpress"]'
description: 'Use when bringing an existing repo into the claude-config framework — needs archetype detection, config install, full multi-axis audit (debt/SEO/GEO/UI-UX/perf/security/a11y/docs), and prioritized backlog. Multi-agent orchestrator. Default flow: do NOT use it for repos created via /init-project. Triggers: "onboard", "onboard project", "audit existing repo", "setup existing project". `split` mode (any repo): "onboard split", "split the project instructions", "découpe le fichier CLAUDE", "path-scoped rules".'
argument-hint: '[split | optional hints: "Python FastAPI" | "Next.js monorepo" | "force-archetype:wordpress"]'
allowed-tools: Read, Write, Edit, Bash, Glob, Grep, Agent, Skill
---
@@ -21,6 +21,13 @@ $ARGUMENTS
---
## MODE DISPATCH
`$ARGUMENTS` whose first token is `split` → jump to § SPLIT MODE (end of
this file); STEP 0 to FINAL OUTPUT do not run. Anything else → the flow
below, unchanged.
---
## STEP 0 — PLUGIN CHECK + AUTO-ACTIVATE
Run `$HOME/.claude/lib/plugin-gate.md` with hint "onboarding existing project + $ARGUMENTS" (dispatch plugin-probe → checkpoint → dispatch plugin-advisor → gates in this loop, BDR-077).
@@ -198,7 +205,7 @@ Ne pas redemander ce qui est déjà dans :
Présenter toutes les questions en UN BLOC. Attendre les réponses. STOP.
Après réponses : **mettre à jour le brouillon CLAUDE.md** avec les infos obtenues.
Après réponses : **mettre à jour le brouillon CLAUDE.md** avec les infos obtenues, et les placeholders `<!-- TODO(/onboard STEP 3): … -->` des `.claude/rules/*.md` émis en PHASE 5b (Docker, tests, CI : leurs réponses vont dans la règle, pas dans CLAUDE.md).
---
@@ -1034,11 +1041,11 @@ Pour démarrer : lire .claude/tasks/TODO.md, choisir une tâche P0, lancer le /s
---
## RULES
- NO skipping STEP 0, 1, 2 — ils sont obligatoires.
- NO skipping STEP 0, 1, 2 — ils sont obligatoires (split mode runs its own steps, § SPLIT MODE).
- NO audit agressif : chaque subagent doit recevoir "AUDIT-ONLY, no file modifications".
- STEP 1 doit proposer un override `force-archetype:<name>` si l'utilisateur sait déjà.
- Monorepo : toujours demander le mode (A/B/C) avant STEP 2.
- Si `CLAUDE.md` existe : le lire, ne pas l'écraser sans fusion après STEP 3.
- Si `CLAUDE.md` existe : le lire, ne pas l'écraser sans fusion après STEP 3 (default flow; split mode edits it on the approved table).
- STEP 3 : ne redemande jamais ce qui est déjà dans README ou manifests.
- STEP 3.5 : si ctx7 absent + fast-libs, WARN mais ne bloque pas.
- STEP 4 : graphify proposé seulement si `graphify-gate.sh` imprime une ligne (≥ 200 fichiers code, pas de graphe — BDR-097) ; l'utilisateur décide, jamais de build sans oui explicite.
@@ -1079,3 +1086,94 @@ NEXT STEPS :
3. GSD (multi-session) : `gsd init` à la main, voir docs gsd-pi
4. .onboard-audit/ peut être supprimé (raw data consommée en synthèse)
```
---
## SPLIT MODE — `/onboard split` (BDR-125 phase 0)
Move the path-bound sections of an existing project `CLAUDE.md` into
`.claude/rules/<name>.md` path rules (loaded on Read of a matching file),
so the always-on context of that project shrinks. Judgment here, writes by
the onboarder, `lib/project-rules.sh` as the only glob oracle. The MODEL
GATE above has run. `mcp__model-router__route(phase="plan")` first.
`<ROOT>` below is the project root as a literal absolute path.
S1 — PRECONDITIONS. `CLAUDE.md` at `<ROOT>`, else print `nothing to split —
run /onboard` and stop. `git -C <ROOT> status --porcelain` empty (whole
tree) and the current branch is `develop` (or `main` when no `develop`
exists), else stop and say why: the split branches off develop, and an
in-flight feature branch or dirty file would be classified on one version
and extracted on another. `.claude/rules/` already holding rules → list
them, continue (new rules add to them).
S2 — BASELINE. `claude -p "/context"` from `<ROOT>` (refused → ask the user
to run `! claude -p "/context"`); note the `Memory files` total and the
project CLAUDE.md row.
S3 — BRANCH. `cd <ROOT> && bash ~/.claude/lib/gitflow.sh start chore
claude-md-split`. A project without gitflow (no `develop`):
`git -C <ROOT> switch -c chore/claude-md-split`, and say so. Everything
after this step reads the branched file.
S4 — CLASSIFY. `bash ~/.claude/lib/project-rules.sh -C <ROOT> sections
CLAUDE.md` lists every `##` section with its size. For each one decide:
- always-on: project overview, stack, build/test/lint commands, folder
structure, conventions that apply everywhere, exceptions to global rules,
workflow expectations, key dependencies, anything needed before a file is
read (how to run the project), even when it names a folder;
- path-scoped: one module or folder, Docker, migrations, CI, a tool's
gotchas, a vendored or machine-owned area. Propose `<name>` (kebab, the
section's subject) and globs (a folder → `dir/**`, a tool → its config
files; never a brace glob `*.{ts,tsx}`: one glob per extension, the
oracle does no brace expansion) and verify EVERY proposed glob with
`bash ~/.claude/lib/project-rules.sh -C <ROOT> matches '<glob>'` (rc 0)
before showing it. A heading containing `'` or `|` is not split by this
mode (say so; the user renames it first).
S5 — GATE (mandatory stop). Print the table, then end the turn:
```
SPLIT — <project> CLAUDE.md (<chars> chars, <n> sections) on chore/claude-md-split
| # | section | chars | verdict | rule file | paths |
| 1 | Stack | 420 | keep | | |
| 2 | Docker | 1,340 | move | docker.md | ["Dockerfile*", "docker-compose*"] |
...
Moves <k> sections, <m> chars (<p>% of the file). Approve, edit a row, or "keep everything".
```
"keep everything" ends the mode with no change (`bash
~/.claude/lib/gitflow.sh delete chore/claude-md-split`). An edited row's
globs are re-verified with `matches` before they enter the TABLE. The
approved rows are the TABLE handed to the onboarder, verbatim.
S6 — EXECUTE. `mcp__model-router__route(phase="orchestrate")` with the
dispatch: `Agent(subagent_type="onboarder")`, prompt = `MODE: split`,
`PROJECT_ROOT: <ROOT>`, `TABLE:` the approved rows, one per line as
`<heading> | <name> | <globs-json>` (LRN-126: every field crosses the
prompt). It pre-flights every row, then extracts, writes the pointer
section, secures the gitignore negation, runs `check`, commits (message
per the project's commit convention), and verifies every rule is in
`HEAD`. Two STOP shapes come back:
- `STOP: <row> — <reason> (rollback done, nothing committed)`: the tree is
back to the branch tip; fix the row, re-dispatch the WHOLE table.
- `STOP: <name> not in HEAD after commit <hash> (commit kept)`: never
re-dispatch (the headings are gone from CLAUDE.md); show the user the
hash and `git -C <ROOT> show --stat <hash>`, the user decides.
S7 — AFTER. `claude -p "/context"` again from `<ROOT>`; append to that
project's `.claude/memory/journal.md` under today's date: `CLAUDE.md
split: <k> sections → .claude/rules/ (<names>), always-on <before> →
<after>`. Tell the user the proof step: a fresh session in that project
that Reads one file per rule, then `jq -c 'select(.memory_type=="Project"
and (.file_path|contains("/.claude/rules/")))'
~/.claude/logs/instructions-loaded.jsonl | tail`. Never `finish`: the
merge is the user's signal.
FINAL OUTPUT:
```
SPLIT COMPLETE: <project>
RULES : <k> files under .claude/rules/ (<names>)
CLAUDE.md: <before chars> → <after chars>
CONTEXT : <before> → <after> (Memory files)
BRANCH : chore/claude-md-split (<hash>) — merge on your signal
PROOF : pending — fresh session + the jq filter above
```