Bootstrap

strata:bootstrap

Onboarding pass. Walk existing docs, dispatch each one to a worker Subagent, aggregate the results.

Posture: act, don't narrate

• No upfront questions about scope unless the candidate count is > 50.
• No tables, no recommendations, no "two-step plan" prose before starting work.
• Source file content stays out of your context — that's the whole point of the workers. You see only the scan listing and the one-line worker summaries.
• Final summary: one line plus counts. The user can /strata:review for detail.

Workflow

1. Scan

bootstrap_scan(unprocessed=true, bucket="fresh", verify=true,
               min_freshness=0.6, max_group_size=12)

Default pass: fresh docs only (<90d), claims verified, freshness ≥ 60%, dense parent dirs split into sub-groups of at most 12 files (by date prefix when filenames are dated). For a wider net, drop min_freshness to 0.4, or omit bucket so aging + old appear too.

max_group_size is what stops dense planning folders (a docs/plans/ with 40+ unrelated initiatives) from collapsing into 3-6 consolidated Notes. Pass max_group_size=12 for typical mixed-initiative folders; Omit it for folders where siblings genuinely describe one thing.

If after filtering the candidate count is > 50, ask the user once to Narrow scope (e.g. by directory). Otherwise proceed.

2. Get the vault dir

memory_status()

Note the vault_dir path printed in the response. You'll pass it to Each worker.

3. Dispatch by parent-dir GROUP, not by single file

The scan's JSON output (when --json is on) includes a dispatch_groups map keyed by parent directory. Sibling files in the same folder (e.g. .planning/auth-rewrite/PLAN.md + .../CONTEXT.md + .../SPEC.md) almost always describe the same Initiative, dispatch them to ONE worker that handles the set together and writes one consolidated note, not three near-duplicates. (Previous "one worker per file" dispatch produced the duplicate ADRs we saw on the earliest bootstrap runs.)

Batch the groups into runs of 5 groups at a time. For each batch, invoke strata:bootstrap-worker once per group in a single Message with multiple Agent tool calls so they run concurrently.

Per worker, the prompt to send:

file_paths:
  .planning/auth-rewrite/PLAN.md
  .planning/auth-rewrite/CONTEXT.md
  .planning/auth-rewrite/SPEC.md

vault_dir: /Users/gd/StrataVault/myrepo
plugin_root: ${CLAUDE_PLUGIN_ROOT}

Group freshness: <average freshness% across the group, when verify was on>
Group age: <oldest bucket across the group>

Follow your standard procedure. Return one line per note written + one
per skipped file.

Hand the worker only metadata, never paste source file content into the prompt. The worker reads each file itself.

When verify was on and any file in the group had paths_missing, mention that in the prompt so the worker downgrades the status (proposed instead of accepted for decisions; framed as a lesson when group freshness < 50%).

4. Aggregate and continue

Each worker returns 1-3 lines (one per write + one per skipped sibling in the group). Collect and print them grouped by kind, no commentary:

DECIDE decisions/2026-05-24-use-postgres-tenant-data.md from .planning/auth/PLAN.md,.planning/auth/SPEC.md
SKIP - from .planning/auth/CONTEXT.md (generic preamble)
DOMAIN domain/auth-tokens.md from .planning/auth/TOKENS.md
DOMAIN domain/scheduling-tenant.md from .planning/scheduling/PLAN.md
SAVE pr-context/feat-x/...handoff.md from .planning/x/HANDOFF.md
ERROR .planning/oversized/PLAN.md — write failed: ...

Then dispatch the next batch of groups. Stop when the group list is Exhausted or you've processed ~20 groups (then ask the user whether to continue, not because of context pressure, but to let them Eyeball quality mid-flight).

5. Final report

Bootstrapped N files. D domain, A decisions, L lessons. Skipped S. E errors.
State: <vault_dir>/.bootstrap-state.json

That's the whole report.

Why subagents (and why grouped)

Each worker has its own context window, your context stays clean. Five parallel workers cut wall time by ~5x. A bad file fails one Worker without poisoning the run.

Grouping by parent dir is the dedup mechanism: workers see all Siblings together, so the three-files-of-the-same-initiative case (PLAN.md + CONTEXT.md + SPEC.md) produces one consolidated note Instead of three near-duplicates. Early bootstrap runs exposed this when un-grouped dispatch wrote 3 near-identical "extract visit Aggregate" ADRs from PLAN/CONTEXT/SPEC.

Trade-off that remains: no shared context across workers, so Wikilinks between notes written in the same batch may be unresolved. That's fine, /strata:review surfaces them later and the user can [[link]] them by hand.

Don't do

• Don't read source files yourself. Dispatch to a worker.
• Don't dispatch one worker per file. Use dispatch_groups from the scan output and dispatch one worker per parent-dir group.
• Don't ask the user to confirm each file. Dispatch the batch.
• Don't expand a batch beyond 5 groups in flight — concurrency past that starts costing more than it saves and risks rate limits.
• Don't auto-invoke forget / promote-to-pr / export-to-repo / archive during bootstrap, those are user-driven only.

Idempotency

<vault>/<repo>/.bootstrap-state.json tracks {path: {sha256, processed_at}}. A file resurfaces if its SHA changes. To re-process A file manually, edit it or delete its entry from the state file. The worker handles the mark step itself; you don't need to.

Tuning what gets scanned

Out of the box, the scan skips build dirs, plugin meta-config (.claude/, .github/, etc.), our own outputs, and common AI-tool Configs at root. To tune per repo, drop a .strataignore at the Project root using gitignore syntax, see examples/.strataignore in the plugin for a starter. .ignore (the ripgrep/fd cross-tool Convention) is also honoured. Precedence, low → high: defaults → .ignore → .strataignore. Negate a default with !pattern.