/we:setup

Every question has a skip path and a default; setup never blocks. Ask one question at a time, each
with one sentence on why the pipeline needs it. Idempotent: on a re-run, show the current value
and ask before replacing it. Create .weside/ only with the user's consent. Suggest a commit;
never commit yourself.

1. Detect

• Stack from marker files: pyproject.toml, package.json, Cargo.toml, go.mod; several →
  monorepo, one entry per component.
• Ticketing per ${CLAUDE_PLUGIN_ROOT}/references/ticketing.md.
• Tools, each by a check that actually runs, never by a path on disk:
  • gh auth status exits 0 (PRs, CI, review threads);
  • graphify: python3 -c "import graphify" (code graph for story and refine);
  • TurboVault: call mcp__turbovault__list_vaults; absent = not registered, error or hang =
    degraded (semantic doc search);
  • weside MCP: mcp__plugin_we_weside-mcp__get_companion_identity is listed (materialize,
    Companion-backed council).
    A missing tool is reported with its install command and never blocks.
• .pre-commit-config.yaml present → activate the hooks: pre-commit missing → offer
  pip install pre-commit; git config core.hooksPath set to a custom directory (Husky) → warn and
  skip; otherwise pre-commit install --hook-type <stage> once per stage the config declares
  (stages:, default_install_hook_types:, plus pre-commit), reporting a stage that errors and
  continuing with the rest.

2. Questions

1. Vision: a PRD at docs/plans/*/PRD.md is used as is; else a link, file or short text →
   .weside/vision.md, or /we:vision later writes the PRD. Skip → no vision check.
2. Ticketing tool confirmed; for Jira, the project key.
3. Stack confirmed or corrected.
4. Tests: when does a worker write them? tdd (failing test first) · tests-after (default,
   same change) · off. It decides when tests are written, not what a good test is.
5. Verification: must every PR declare how the change was observed, or why not? yes
   (recommended) · advisory. Green tests share the author's blind spots; with yes the plugin's
   PreToolUse hook refuses gh pr create without a ## Verification block whose **Oracle:** line
   says how the change was observed, or not-observed — <reason>. On yes, offer to scaffold
   .weside/verify.md from what step 1 found: the repo's CLI, its dev bring-up command, whether a
   browser driver exists.

Review gates are detected, not asked: a Claude or Codex review workflow in .github/workflows/, a
CodeRabbit or Greptile app. The ids go to review.available; /we:ci-review collects threads from
exactly these. Default ["claude"]. A Codex job named other than Codex Review goes to
review.codex_check: the auto-merge hook waits for that check's verdict.

Offer once: additive repo checks on top of the plugin's → .weside/dor.md and/or .weside/dod.md,
seeded with the user's concrete check. /we:story reads the DoR file, /we:orchestrate the DoD
rows. Additive, never a replacement.

3. .weside/config.json

Written always, also when step 4 is declined. Only keys a verb reads:

{
  "ticketing": { "tool": "jira|github-issues|none", "project_key": "<KEY or null>" },
  "stack": ["python", "node"],
  "tools": { "graphify": false, "turbovault": false },
  "test_discipline": "tests-after",
  "verification": { "required": true, "recipe": ".weside/verify.md", "staging_needs_ask": true },
  "review": { "available": ["claude"] }
}

A re-run re-detects tools and overwrites only that block. Print what was configured in a short
block (stack, ticketing, vision, test discipline, verification, review gates, missing tools).

4. Council (optional, ask once)

"Set up this repo's council now? /we:council and /we:meet convene it; without a weside account
every lens is generic and still works." On yes:

1. Extend config.json (never re-emit its keys) with vault, framework_version: 1, and the
   meeting rosters /we:meet and /we:council read:
   "council": {"default": ["product_owner","architect","scrum_master"], "meetings": {"vision": ["product_owner","architect","ux_researcher","marketing","orchestrator"], "saga": ["product_owner","architect","orchestrator"], "epic": ["product_owner","architect","orchestrator"], "story": ["product_owner","architect"]}}.
2. TurboVault: check whether this repo's vault is in list_vaults by name (another repo's vault
   passing a loose "any vault" check leaves every later search answering from the wrong tree).
   Missing → add_vault(name=<basename>, path=<root>) and set_active_vault; write vault only
   after both succeeded. A small repo may want no vault: vault: null is valid.
3. .weside/weside.md and .weside/council.json are written by no verb. /we:council reads
   council.json (members per role: members.<slug>.name, .role, optional lens), which the user
   edits by hand; without it every lens is generic.

5. Rule bridge and statusline

• Rule bridge, when the repo has .claude/rules/: Claude Code loads the rules itself, every other
  agent (Codex, Gemini) does not. .agents/skills/claude-rules/SKILL.md exists → skip. Otherwise copy
  ${CLAUDE_PLUGIN_ROOT}/templates/agents-skill/SKILL.md there unchanged and name the instruction-file
  line that points at it. The template resolves scripts/load-rules.py from the plugin cache
  (~/.claude/plugins/cache/weside-ai/we/*/) at run time, so the repo carries no copy of the loader.
• Statusline (model · branch · PR · context · cost; follows /we:orchestrate's
  ~/.claude/we-focus/<session>.json): run python3 ${CLAUDE_PLUGIN_ROOT}/scripts/install_statusline.py --status.
  offer to install → ask once, on yes --apply. already active with a NOTE that the copy differs → --apply
  refreshes ~/.claude/we-statusline.js. keep theirs → keep the user's statusline and
  mention --apply --force (backed up, --revert restores it). This and the council agents are the only
  user-scope files setup touches.

6. Instruction loop (optional, ask once)

"Keep this repo's rules, skills and AGENTS.md fit for the current model?" Store and keys:
${CLAUDE_PLUGIN_ROOT}/references/optimization-store.md. On yes, offer each item on its own:

1. Store skeleton: .weside/optimization/ with CHARTER.md (frontmatter last_optimize:
   empty; sections Goal, Decisions, Findings, Next steps), LEDGER.md (the ledger table header),
   and inbox/.gitkeep. Never in the plugin's own checkout. A re-run removes an
   .weside/optimization/inbox/*.md merge=union line from .gitattributes and says so (why:
   optimization-store.md § Inbox entry).

2. Authoring rule: copy ${CLAUDE_PLUGIN_ROOT}/references/instruction-authoring.md to
   .claude/rules/instruction-authoring.md, body unchanged, with paths: derived from what the
   repo holds. Candidates: .claude/rules/**, .claude/skills/**/SKILL.md, .claude/agents/*.md,
   **/AGENTS.md, **/CLAUDE.md, and <dir>/skills/**/SKILL.md plus <dir>/agents/*.md for each
   plugin directory (one holding .claude-plugin/). List each candidate's matches with
   python3 ${CLAUDE_PLUGIN_ROOT}/scripts/load-rules.py --files-matching '<glob>'. Keep a candidate
   that matches a file outside optimization.exclude; drop one that matches none. A candidate that
   also matches files under exclude is replaced by narrower globs over the kept files' directories,
   because paths: has no negation. A re-run refreshes the body after asking and keeps the paths:.

3. Gate as pre-commit hook, when .pre-commit-config.yaml exists: add this local hook. It resolves
   the gate from the plugin cache and passes with a notice where the plugin is absent. The gate is a
   local hook and a /we:optimize step, not a CI job: the repo vendors no copy of the plugin's scripts.

   - repo: local
     hooks:
       - id: instruction-budget
         name: Instruction budget (we plugin)
         entry: bash -c 'g=$(ls -d ~/.claude/plugins/cache/weside-ai/we/*/scripts/check-instruction-budget.py 2>/dev/null | sort -V | tail -1); if [ -z "$g" ]; then echo "we plugin absent - instruction budget not checked"; exit 0; fi; python3 "$g"'
         language: system
         pass_filenames: false
         files: '(^|/)(\.claude/rules/.*|SKILL\.md|agents/[^/]*\.md|AGENTS\.md|CLAUDE\.md|references/.*\.md|\.weside/config\.json)$'

   Run it once (pre-commit run instruction-budget --all-files) and report the count; a red first
   run is the repo's backlog for /we:optimize, not a setup failure.

4. Config keys: optimization.target_model and target_effort (default: this session's model at
   medium), exclude for trees that hold content rather than instructions, and workspace: the
   sibling git repos of this one, one per remote URL, each the main checkout
   (git rev-parse --git-dir equals --git-common-dir). Ask which siblings to include.

7. Next

/we:story <KEY> (plan) → /we:orchestrate <KEY> (build to a green PR; auto-merge unless a stop
applies) → /we:merged.