Cold start

<!-- Copyright (c) 2026 Kunal Suri (CEA LIST). All rights reserved. -->

Cold-start the knowledge layer

The contract: read-and-write-docs-only — touch no source code, tag everything [inferred], and stop at an AUDIT TODO for the human. You draft the map; a human verifies it before any feature is built.

1. Load the facts

Read ai/repo-profile.json (deterministic output of the kit's orient step). Treat its stack facts as given; VERIFY its build/test commands against the real config files before writing them anywhere as confirmed. If the profile says this repo is a fork, separating OUR code from FROZEN upstream code is the single most important output of this pass. If a humanContext block is present, use it: developer.skillLevel/codebaseFamiliarity calibrate how conservative your Stability guesses are; stack.kind tells you whether to map a frontend/backend split; record branch/firstRunAt in PROJECT_OVERVIEW.md.

2. Absorb prior config (Process 2 repos only)

If repo-profile.json → maturity.process is 2, CLAUDE_bkp_*.md / AGENTS_bkp_*.md exist at the repo root with the team's previous AI configuration. Read them, extract KNOWLEDGE — not raw text — (build commands, conventions, architecture notes, gotchas, module descriptions) and merge it into ai/guide/CONVENTIONS.md, ai/guide/ARCHITECTURE.md, ai/guide/MODULE_MAP.md (seed rows), and ai/guide/PROJECT_OVERVIEW.md. Tag every extract [inferred — from prior config].

3. Explore cheaply, read-only

List the tree 2 levels deep, read build manifests (not source), scan the last ~30 commit subjects for active areas. Prefer grep + line counts over whole-file reads. Modify nothing in this step.

4. Fill MODULE_MAP

ai/guide/MODULE_MAP.md: one row per module/package — directory, a one-line responsibility, the entry-point file, and a Stability GUESS. Use the ACTUAL names you found; do not assume names. Each row has FIVE cells — Directory, Responsibility, Entry point, Stability guess, Status — and the Status cell is where the [inferred] tag goes (the tooling reads provenance from the LAST cell only).

5. Correct the stamped commands

If the build/test commands you verified in config files differ from the Build/Test lines stamped into CLAUDE.md and AGENTS.md, correct exactly those lines (and the Definition-of-done line in ai/guide/CONVENTIONS.md), tag the correction [inferred], and change nothing else in those files. This is the sanctioned exception to the No-Churn rule.

6. Draft diagrams

Into ai/analysis/diagrams/ as Mermaid .mmd: package-deps.mmd (dependency graph), domain-core.mmd (core types), seam.mmd (the main boundary — note the protocol as a question if unverified).

7. Note features, update the guides

Note candidate features in ai/guide/FEATURE_MAP.md using its template. Update ai/guide/ARCHITECTURE.md and ai/guide/PROJECT_OVERVIEW.md only where you VERIFIED something in code or config.

8. Hard rules for this pass

• Tag EVERYTHING you write [inferred]. You are guessing; say so.
• Separate OBSERVED facts (file A imports B) from INFERENCES (A is "the domain layer").
• On a fork: mark anything inherited as Stability frozen and FLAG it "UNSURE — needs human" rather than asserting it.
• Where you cannot tell, write ? and "UNSURE — needs human". Never guess confidently.
• Do NOT modify any source file. Only write inside ai/.
• Re-run safety: leave rows already carrying [verified] unchanged; only populate rows still ? or holding placeholder text like <fill in>.

9. Stop and report

Print an "AUDIT TODO" table (columns: # · Location · What to verify · Why uncertain) covering every row still ?, every frozen guess needing confirmation, and any protocol/command assumption still [inferred]. The human sets Stability and flips [inferred] → [verified]. Do NOT proceed to building features. Tell the user they can run /review-agent-config (structural check of CLAUDE.md/AGENTS.md) and /post-cold-start-verification (deep semantic audit) once their audit is done.