feat: customization tool + clearer skill, ready for v0.7.0

[devinoldenburg]

Makes Goal Mode customization actually easy and powerful, and cuts the release to v0.7.0.

What's better

• goal-config tool (scripts/goal-config.mjs, bin opencode-goal-mode-config, npm run config): the agent/user can now discover → apply → verify instead of guessing.
  • list — every key with type, default, env var, and what it does.
  • explain <key> — the exact opencode.json + env snippet to set it, and how to verify.
  • recipes / recipe <name> — paste-ready presets (safe-yolo, full-yolo, allow-commands, block-extra, quiet, no-auto-review, strict).
  • effective '<json>' [--diff] — preview the resolved config before committing.
• CONFIG_DOCS is the single doc source for every key; a test keeps it in lockstep with DEFAULT_CONFIG so docs can never drift.
• Rewrote the skill + /goal-mode-customize command into a tight, tool-driven discover → choose → apply → verify loop with a request→key map and the YOLO precedence rule spelled out — far easier for the agent to act on than the previous prose.
• README points at the tool.

Release

• Bumped to v0.7.0 with a CHANGELOG entry (YOLO + customization tooling + the ~190-issue hardening sweep since 0.6.12); package-lock synced.

Verification

• node --test "tests/*.test.mjs" → 668 pass, 0 fail (8 new goal-config tests).
• npm run lint clean; validate-opencode-config.mjs passes; release-notes.mjs renders the v0.7.0 section; check-npm-publish-ready --skip-registry ready at 0.7.0; npm pack --dry-run ships the tool, skill, and command.

Does not touch Issue #2.

[devinoldenburg]

Review: APPROVE (recommend merge)

I reviewed this rigorously and ran the full verification suite locally on feat/v0.7.0-customization-tooling. No blocking defects found. (Posting as a comment since GitHub blocks --approve on one's own PR — this is an approving recommendation.)

What I verified

1. goal-config.mjs correctness — all subcommands work, exit codes correct

• list, explain <key>, recipes, recipe <name>, effective '<json>' [--diff] all produce correct output.
• Exit non-zero on bad input: explain badkey → 1, recipe nope → 1, effective '{not json' → 1, unknown subcommand → 1. Bare effective (no opts) and bare invocation print help and exit 0 (intended).
• effective --diff correctly shows only changed-from-default; cross-checked against resolveConfig: full-yolo {yolo,allowDestructive} → blockDestructive/blockNetworkExec/enforceCompletion/restrictSubagents/toastOnBlock all false; safe-yolo keeps blockDestructive=true. Matches resolveConfig's relax() set exactly (config.js:270-278).
• Printed opencode.json/env snippets are valid and reproduce the claimed effect — re-ran each emitted effective verify command and the GOAL_GUARD_ALLOW_COMMANDS="...,..." env-string form; both resolve as advertised.
• Global-bin path: the ../plugins/goal-guard/config.js import resolves against the script's own URL, not cwd — ran the bin from /tmp (exit 0). Both plugins/goal-guard/config.js and scripts/goal-config.mjs are in files/npm pack output, so the relative import holds in an installed global package.

2. CONFIG_DOCS no-drift — covers exactly DEFAULT_CONFIG (30 keys); the test (goal-config.test.mjs:11-19) is a genuine independent cross-check, not tautological. envVarFor is consistent with every fromEnv mapping; spot-checked sessionTtlMs→GOAL_GUARD_SESSION_TTL_MS, blockDestructive, allowCommands.

3. Skill/command accuracy — recipes and the request→key map are correct against real semantics:

• YOLO precedence claim is true: effective '{"yolo":true,"blockNetworkExec":true}' --diff does NOT relax blockNetworkExec (explicit override wins, config.js:267-269).
• allowCommands vs extraDestructive match guard.js:582-586 — allow-list bypasses the guard entirely (!allowlisted gates both block paths), deny-list extends the destructive classification. Skill wording ("never blocked, whatever the analyzer thinks" / "treated as destructive") is accurate.
• full-yolo = {yolo:true, allowDestructive:true} correctly documented as the only way to drop destructive guarding.

4. Release bump — package.json and package-lock.json both at 0.7.0. node scripts/release-notes.mjs reads package.json version, finds ## v0.7.0, emits a 37-line non-empty section, exits 0. CHANGELOG heading is ## v0.7.0. exports/bin/validator contract intact — validate-opencode-config.mjs passes; the new opencode-goal-mode-config bin points at a packed file.

5. Tests + tooling green

• node --test: 669 pass / 0 fail (full suite, including the new goal-config.test.mjs and unchanged config.test.mjs).
• npm run lint (biome): exit 0; the 3 changed JS files lint clean. The warnings/infos shown are pre-existing across the tree, not introduced here.
• npm pack --dry-run: exit 0, ships goal-config.mjs + config.js.

Nits (non-blocking)

• scripts/goal-config.mjs:60 exampleValue for list keys hardcodes a sample (["^docker compose ", "^rm -rf \\./tmp/"]) rather than echoing the key's actual default — harmless since list defaults are all [], and a concrete example is arguably more useful. No action needed.
• The strict recipe resolves to "No effective change from defaults" under effective --diff (it spells out the defaults). Correct and intentional, but a reader running the verify line might briefly expect output. Optional: a one-word note in the recipe blurb. Non-blocking.

Solid, well-tested PR. Recommend merge.