Sangeetha-Grantha

Metadata Value
Status Active
Version 1.1.1
Last Updated 2026-09-10
Author Sangeetha Grantha Team
Document Type Evidence record

AI-native SDLC artifacts (TRACK-134)


[!NOTE] Historical evidence: results, counts, commands, and observations below belong to the original work described here. The editorial update date is not a new test or corpus verification. For present behavior, use current feature map.

Close the gaps around agent Build speed: committed Intent / Spec / Plan on conductor tracks, review policy, session verification, edit-time hooks, and deterministic evals of the files that steer agents.

Conductor remains the source of truth. There is no parallel intent/ directory.

Artifacts

Artifact Location
Intent / Spec / Plan template .agents/workflows/conductor-track-manager.md
/spec-from-track .claude/commands/spec-from-track.md (Cursor symlink under .cursor/commands/)
/plan-from-spec .claude/commands/plan-from-spec.md
Review policy REVIEW.md
Session verification CLAUDE.md → Verifying your work
Edit-time hooks .claude/hooks/*.py + .claude/settings.json (Claude Code) and .cursor/hooks.json (Cursor). Same scripts; Cursor also gets JSON { "permission": "deny" }.
Agent-config evals evals/check.py, evals/cases/, make agent-evals, CI job agent-evals

Gates

  1. Intent Status Accepted (human) before /spec-from-track.
  2. Spec Status Accepted (human) before /plan-from-spec. Import/seed/lyric specs also get a carnatic-musicologist report.
  3. Plan Status Accepted (human) before product-code edits.
  4. PR review uses REVIEW.md. Findings inform; a human still approves.
  5. Committed database/migrations/V*.sql files are immutable at edit time (Write/Edit/StrReplace/Delete); new VNN__*.sql and R__*.sql are allowed. Reads of V__ files stay allowed. .env reads and shell dumps that actually target a secret file (cat/head/source/< .env redirection, open('.env'), grep PATTERN .env, …) are denied; the path merely appearing as text (commit messages, PR bodies) or as a search pattern (grep 'Read(.env' file) is allowed. Committed env templates (.env.example, *.env.example, .env.*.example such as config/.env.auto-approval.example) stay readable.
  6. Committed test files (*Test.kt, *.test.ts, *.spec.ts, test_*.py) are immutable at edit time. Fixtures, Vitest setup.ts, and test helpers are not tests. SANGITA_ALLOW_TEST_EDITS=1 (or true/yes) overrides; 0/false do not.

Evals

evals/check.py is deterministic (no model API). It asserts required policy strings still exist in CLAUDE.md, skills, commands, REVIEW.md, .cursor/hooks.json, and .gitignore, and smokes the edit-time hooks (including Cursor payload shapes and secret-file shell bypasses). LLM-in-CI evals from the playbook stay deferred until TRACK-109 budget/secrets work.

Proof

python3 evals/check.py
python3 conductor/check-registry-sync.py

Ref: ADR-013, Domain Model §6.


Section index · Documentation home · Feature status