intent-extract

name: intent-extract description: Use during bob's Step 4.5 (after UNIT_TESTED) or evo's INTENT_MAPPED phase (per design §6 state machine) to extract per-component functional intent from a project's source code. Anchor-and-expand on contract-map components plus 1-hop call-graph neighbourhoods; per-component LLM pass at temperature=0; structured YAML output conforming to functional-intent.v1 with closed-set function_class enum, REQUIRED evidence_edges that resolve in static.jsonl, two-arm verification on prose claims; content-hash cached. Requires a bob-issued claim token, heartbeats it every 60s, emits transition requests for INTENT_MAPPED. Skill NEVER writes claim files, NEVER writes the integration ledger, NEVER traverses the call graph beyond 1-hop. CB4-compliant.

intent-extract (v1)

Functional-intent extraction for the EVO agent ecosystem (S032).

Status: PRODUCTION (S032 WP-2 ship). Design document: /path/to/project/docs/plans/2026-05-13-evo-agent-design.md §4.1 + §5.1. Schema: ~/.claude/skills/_meta/schemas/functional-intent.v1.json (frozen WP-1).

What this skill is for

This skill takes a frozen contract-map and a static-wiring snapshot (.wiring/latest.json) and produces one functional-intent.v1 YAML file per component, describing what the component does in both mechanically-derived structural terms (entry points, side effects, error paths, all with required evidence_edges into static.jsonl) and LLM-interpretive prose (responsibilities, assumptions, invariants, unknowns) that NEVER feeds gates or test generation in its raw form.

Two output forms per component:

1. Cache file at .wiring/intent-cache/<sha256>.yaml — content-addressable.
2. Per-run symlink at .wiring/runs/<run_id>/intent/<component>.yaml.

Invocation contract

The skill is always invoked by evo (or, in S033+, by other orchestrators) with a bob-issued claim UUID. It NEVER issues its own claim. It heartbeats every 60s. If the lease expires, the skill exits cleanly and bob's transition-request queue rejects the stale request.

python3 ~/.claude/skills/intent-extract/scripts/run.py \
  --project-root <abs path> \
  --run-id <uuid> \
  --claim-uuid <uuid> \
  --workspace-tree-hash <40-hex> \
  --components <component-id>[,<component-id>...] \
  --model-id claude-opus-4-7 \
  [--mode <intent-map-only|version-upgrade|cve-fix>] \
  [--two-arm <strict|skip>]  # default strict

What lives in scripts/

• run.py — CLI entry. Parses args, holds claim, dispatches per-component.
• anchor_expand.py — given a component id + static.jsonl, returns the 1-hop symbol neighbourhood, evidence edge ids, file contents.
• prompt_template.py — locked LLM prompt template; hashed into output.
• llm_call.py — Anthropic API wrapper at temperature=0, with retry and budget checks (EVO_MAX_TOKENS_PER_RUN soft-fail).
• two_arm_verify.py — cold-context second pass for prose-field claims, semantic similarity scoring (≥0.95 required for confidence_level:grounded).
• cache.py — content-hash cache key derivation + read/write.
• schema_validate.py — validates output against functional-intent.v1.
• manifest.py — writes per-run intent-manifest.json (hit/regenerated/failed/gap).

What lives in references/

• anchor-and-expand.md — algorithm spec for the 1-hop expansion strategy.
• two-arm-verification.md — cold-context second-pass methodology.
• closed-enum-reference.md — function_class / entry_point.kind / side_effect.kind / error_path.kind enumerations with examples.
• caching-and-determinism.md — cache key derivation, hit ratios, determinism budget.

What lives in templates/

• prompt-base.txt — the canonical LLM prompt template (template_hash is sha256 of this exact file).

Hard rules (HARD-RULE alignment per design §13)

1. NEVER writes .ledger/claims/ — bob owns claim files (CB4).
2. NEVER writes the integration ledger — emits transition requests under .ledger/requests/<request_id>.request.yaml (CB4).
3. NEVER traverses the call graph beyond 1-hop — auto-traversal is an anti-pattern; declared flows only.
4. NEVER returns confidence_level: grounded without two-arm verification — single-arm output is interpretive; pure heuristic fallback is degraded.
5. NEVER calls LLM if cache hit — content-addressable cache short-circuits.
6. NEVER blocks > EVO_MAX_TOKENS_PER_RUN — soft-fail with cached fallback + advisory in intent-manifest.json (HARD-RULE 8 budget honesty).
7. Heartbeats every 60s — calls claims.heartbeat_claim(claim_uuid).

What feeds what downstream

Determinism contract

Re-running on the same workspace_tree_hash + content_hash + extractor_version

• model_id MUST produce ≥0.95 semantic similarity output (cached_interpretive class). Cold runs on fresh content are fresh_interpretive. Pure-mechanical fields (entry_points, side_effects, error_paths) are deterministic.

How to invoke for testing

# Validate a generated functional-intent file against schema
python3 ~/.claude/skills/intent-extract/scripts/schema_validate.py \
    --intent .wiring/runs/<run_id>/intent/auth-service.yaml