Bootstrap agent-daemon in a project

Full end-to-end initialization: scaffold files, read the project, populate memory with real context, verify.

Phase 0 — Scaffold

1. Run ad init in the project root. If .agent-daemon/ already exists, it will report "already initialized."
2. Check if memory files contain {{ placeholders:

   grep -l '{{' .agent-daemon/memory/*.md 2>/dev/null

3. If no placeholders found → memory is already populated. Report "already bootstrapped" and STOP.
4. If placeholders exist → proceed to Phase 1.

Phase 1 — Read the project

Read ALL sources that exist. Skip silently those that don't. Do NOT ask the user for each one.

Collect all this into working notes before writing any files.

Phase 2 — Populate memory files

Write each file using extracted content. Replace the template entirely — don't patch placeholders.

projectbrief.md (keep under 1KB)

Extract from manifest + README:

• What this project is — one paragraph from README or package description
• Who uses it — user roles inferred from README or codebase (routes, UI screens)
• Core problem — from README or infer from project purpose
• Non-goals — mark "TBD — will be refined over sessions" if not obvious
• Stage — infer: has CI + production config = "shipped"; few commits = "early-prototype"; else "beta"

techContext.md (keep under 4KB)

Extract from manifest + README + CLAUDE.md + env files:

• Stack table — language, framework, DB, hosting from dependencies
• Dev setup — commands from README or package.json scripts
• Env vars — from .env.example or CLAUDE.md
• Verification commands — from package.json scripts or CI config (typecheck, lint, test, build)
• Gotchas — from CLAUDE.md "gotchas" or "important" sections. If none found: "No gotchas captured yet — the digest pipeline will populate from session learnings."
• External services — from README/CLAUDE.md or mark "TBD"

systemPatterns.md (keep under 5KB)

Extract from directory structure + CLAUDE.md + code conventions:

• Architecture — from CLAUDE.md architecture section, or infer from directory layout
• Key directories — from ls output, describe what each contains
• Reusable utilities — scan for lib/, utils/, helpers/, hooks/, shared/ directories. List key exports if found
• Recurring patterns — from CLAUDE.md conventions or code patterns
• ADRs — leave empty: "No ADRs recorded yet."

productContext.md (keep under 2KB)

Extract from README + CLAUDE.md:

• Before/after workflow — from README or infer from project purpose. Mark "TBD" if unclear
• Success metrics — infer from README or mark "TBD"
• Key product decisions — from CLAUDE.md or README design rationale
• Constraints — from CLAUDE.md directives or compliance mentions
• Stakeholders — mark "TBD" unless explicitly stated

activeContext.md (keep under 5KB)

Extract from git state:

• Current focus — from git log -5 messages + git status (what's being worked on now)
• Recent decisions — leave empty: "Digest pipeline will populate from session learnings."
• In-flight branches — from git branch -a, list non-main branches with last commit subject
• Active blockers — leave empty
• Known temporary state — leave empty

progress.md (keep under 5KB)

Extract from git history:

• What works — infer from git log: features that appear shipped (merged PRs, "feat:", "add:" commits)
• In progress — from active branches
• Planned — "Check issue tracker" or leave empty
• Recently fixed — from git log: commits with "fix:" in last 30 days
• Known issues / tech debt — leave empty: "Will be populated by review-slice and audit-runner skills."

Phase 3 — User profile

Check if ~/.agent-daemon/user.md exists and contains {{ placeholders.

• If already populated (no placeholders) → skip entirely
• If has placeholders → ask the user these 3 questions in ONE message:
  1. "What's your name and role?" (e.g., "Asha, full-stack developer")
  2. "Any strong preferences for how I work?" (e.g., "be terse", "always show diffs", "Hindi-English mix is fine")
  3. "Anything I should never do?" (e.g., "don't auto-commit", "don't touch production files")

Then auto-detect installed tools:

node --version 2>/dev/null; python3 --version 2>/dev/null; go version 2>/dev/null; rustc --version 2>/dev/null; docker --version 2>/dev/null; gh --version 2>/dev/null

Write ~/.agent-daemon/user.md with the user's answers + detected tools.

Phase 4 — Verify

1. Run ad doctor — all checks should pass
2. Show a 3-line summary:

   Daemon bootstrapped for [project name]:
   - 6 memory files populated from [sources used]
   - User profile: [populated / already set / skipped]

3. Tell the user: "Memory is live. The digest pipeline will keep it updated as you work."

What NOT to do

• Do NOT ask for confirmation before reading each source file — just read them all
• Do NOT exceed the size limits noted for each memory file
• Do NOT populate gotchas or tech debt speculatively — leave them for the digest pipeline
• Do NOT overwrite memory files that are already populated (no {{ placeholders)
• Do NOT skip ad init — always run it first for idempotent scaffolding
• Do NOT put code samples in memory files — they're context docs, not code references