feat: Goal Mode works for non-code agentic goals (research, analysis, explanation)

[devinoldenburg]

Problem (observed in the field)

Goal Mode failed a non-code agentic goal — producing an explanation of a PR — because its review gates are hard-wired for code. The goal-prompt-auditor blocked with "No working-tree changes... git diff is empty," and the always-required goal-diff-reviewer / goal-verifier gates can never pass when the deliverable is a text answer + evidence rather than a diff. Every goal was treated as a code-implementation goal.

Fix — task-aware gating

• Gate selection (gates.js): a goal is "code-bearing" only once it is dirty, edits a file, or records changed files. The code-only gates (goal-diff-reviewer, goal-verifier) are required only for a code-bearing goal. A research / analysis / explanation / planning goal is gated on its recorded evidence by the always-on reviewers (goal-prompt-auditor, goal-reviewer, goal-final-auditor), so an empty git diff no longer blocks it. An edit instantly restores the code gates (lastEditSeq is monotonic, so edit-then-revert cannot game it).
• New config requireCodeReview (auto | always | never, default auto) + GOAL_GUARD_REQUIRE_CODE_REVIEW to override the detection; documented in CONFIG_DOCS, the goal-config tool, and the README.
• Reviewer prompts: the five reviewers now treat a clean working tree as expected for a non-code goal and evaluate the recorded evidence/answer against the acceptance criteria instead of failing on a missing diff.
• Goal agent: goal.md makes non-code goals first-class and tells the agent to record its produced answer as evidence (with the criteria it covers) so coverage is checkable.

Verification

• node --test "tests/*.test.mjs" → 672 pass, 0 fail (new tests: code vs non-code gate selection, the requireCodeReview override, and a non-code goal completing on the always-on gates alone; updated two tests that assumed a fresh state needs all five gates).
• npm run lint, validate-opencode-config.mjs, and the goal-config no-drift test all green.

[devinoldenburg]

Adversarial review: APPROVE (no enforcement weakening found)

Verified by running probes against gates.js/events.js/state.js and the full suite (node --test tests/*.test.mjs → 672 pass; npm run lint exit 0; validate-opencode-config.mjs passes).

1. No bypass on code goals — CONFIRMED

isCodeBearing (gates.js:27) is true if dirty OR lastEditSeq>0 OR changedFiles.length>0. The lastEditSeq>0 term is the lock: markEdit (events.js:79) sets it via store.nextSeq() which is seq+=1 (store.js:136) — strictly monotonic, restored as Math.max across persistence (store.js:202), and never reset. markEdit is the only writer of lastEditSeq and is called on every edit tool / bash mutation / file.edited event.

Edit-then-revert/commit scenario probed directly: after one edit lastEditSeq=1; then forcing dirty=false, changedFiles=[] (simulating a clean tree) → isCodeBearing STAYS true and requiredGates STILL returns goal-diff-reviewer + goal-verifier. A code goal cannot shed the code gates by cleaning its tree.

2. Non-code completion still gated — CONFIRMED

The 3 always-on gates (goal-prompt-auditor, goal-reviewer, goal-final-auditor) are never dropped — only CODE_GATES are conditionally filtered (gates.js:88). Probed a non-code goal: zero verdicts → blocked (3 missing); all FAIL → blocked; 2/3 PASS → blocked. So a non-code goal with no real deliverable/evidence cannot complete; the reviewers gate it.

3. Override correctness — CONFIRMED

gates.js:86-87: always→force on, never→force off, anything else→isCodeBearing (auto). Default requireCodeReview:"auto" (config.js). Coerced via coerceStr (config.js:269, string-typed default) — a number/object/null passed programmatically is String()-coerced to an unknown mode and safely falls through to auto; no crash (probed 123, {}, null, "AUTO", "Always" → all keep code gates for a code goal).

4. Reviewer prompts — no rubber-stamping

The prompt-auditor/reviewer/final-auditor/verifier additions only stop treating a clean tree as an automatic FAIL; they still direct the reviewer to judge the deliverable's correctness/completeness against goal_evidence_map and acceptance criteria. The one auto-PASS instruction (goal-diff-reviewer.md:45, PASS on empty git diff) is narrow: an empty working-tree diff was already produced by any committed work pre-PR, and committed code is still substantively reviewed by goal-reviewer ("actual files and behavior", git log access) plus the final-auditor — so it is not a new escape hatch for code goals.

5. Tests — meaningful, not weakened to pass

tests/gates.test.mjs adds real assertions (non-code drops only CODE_GATES but keeps the 3 always-on; editing restores code gates; override matrix; completion-on-always-on). tests/programmatic-review.test.mjs >=5→>=3 is correct: an evidence-only goal legitimately needs only the 3 always-on reviewers, and the test still asserts review runs AND uses a PASS-stubbed reviewer (completion reflects real passes, not an evaporated gate).

6. CI — green

Tests 672/672, lint exit 0 (only pre-existing warnings/infos; the 2 in changed test files are a pre-existing unused-var helper), config validation passes.

Nits (non-blocking): goal-diff-reviewer.md auto-PASS on empty diff is a pre-existing diff-tooling limitation, not introduced here, and is backstopped by the always-on reviewers — fine to leave.