Bootstrap

One-time migration from existing planning docs into the vault. Use when you install Strata on a repo that already has a docs/, .planning/, or similar directory full of markdown.

Source files never enter your main context. Workers do the reading in their own context windows, return one-line summaries, and the aggregator tallies them.

The model

Bootstrap is not the recurring way knowledge enters the vault. After installation, engineers write directly into the vault as they work, through /strata:save, /strata:decide, /strata:domain during conversation. Bootstrap exists for the chicken-and-egg case: an existing repo with months of planning docs that need migration before Strata has anything to retrieve.

Run it once per major initiative. Then .gitignore (or archive) the planning subdir.

How it works

/strata:bootstrap

That’s the whole invocation. Behind the scenes:

1. Scan the repo for markdown candidates (respects .gitignore, .strataignore, .ignore)
2. Group by parent directory. .planning/auth-rewrite/PLAN.md + .../CONTEXT.md + .../SPEC.md go to one worker together. Siblings about the same initiative produce one consolidated note, not three near-duplicates
3. Dispatch 5 groups in parallel to strata:bootstrap-worker subagents. Each worker:
   • Reads its group’s files
   • Calls plan_correlate to cross-check claims against git history + code graph
   • Calls code_map to understand mentioned symbols
   • Classifies → domain / decide / save / skip
   • Writes the note(s)
   • Marks the source files processed (SHA-tracked, idempotent)
4. Aggregate one-line summaries; print final tally

Source file contents never enter your main Claude context. The workers do the reading; you see only the summary lines.

Classification

Each group’s docs are pulled into one of four shapes:

Kind	When	Lands in
domain	Defines vocabulary, invariants, conventions (“what is X”)	domain/<slug>.md
decide	Locked-in choice with reasoning + alternatives	decisions/YYYY-MM-DD-<slug>.md
save	Retrospective, “we considered X”, historical context	lessons/YYYY-MM-DD-<slug>.md
skip	Generic README, TOC, autogenerated — mark only	(no file written)

The classification is informed by plan_correlate’s verdict:

• ≥80% completion (paths exist + symbols resolve) → decide accepted, domain stable
• 40–80% → decide proposed (caveats inline)
• under 40% → save framed as retrospective (“we considered this in 2026”), not authoritative
• No testable claims → classify by content alone

What gets scanned

Default scope: every *.md under the repo root, minus:

• Build / cache dirs (node_modules/, dist/, .venv/, __pycache__/, etc.)
• Plugin meta-config (.claude/, .github/, .vscode/, .idea/, .zed/, .agents/)
• Strata outputs (graphify-out/, .strata/)
• AI-tool config files (.cursor.md, .impeccable.md, .windsurf.md, …)
• Low-signal root files (CHANGELOG.md, LICENSE.md)

Tune per repo with a .strataignore at the project root (gitignore syntax). .ignore (ripgrep/fd convention) also honoured.

.strataignore

docs/legacy/          # skip an old subtree
!RUNBOOK.md           # re-include a default-excluded file

Tuning the pass

Default pass is fresh docs only (under 90 days old) with claims verified at ≥60% freshness. Override via the MCP bootstrap_scan tool:

bootstrap_scan(
  unprocessed=true,
  bucket="aging",          # 90-365 days
  verify=true,
  min_freshness=0.4,       # accept partial-evidence plans
)

Or skip verification entirely (faster, less precise):

bootstrap_scan(unprocessed=true, verify=false)

Idempotency

<vault>/<repo>/.bootstrap-state.json tracks {path: {sha256, processed_at}}. Re-running the bootstrap skips already-processed files unless their SHA changed. To force a re-process, edit the source or delete its entry.

After bootstrap

Two options for the planning subdir:

• Keep + archive: git mv .planning/auth-rewrite .attic/auth-rewrite. Historical reference, no longer scanned.
• Delete: git rm -r .planning/auth-rewrite. The vault now holds what should survive.

Either way: future planning happens directly in the vault. The .planning/ workflow was the migration source, not the steady state.

---

Next: Code graph — Graphify integration and how code_map cuts token budget.