OpenCode Goal Mode

Strict Goal Mode for OpenCode: a primary goal agent, a matrix of specialized review subagents, slash commands, and a goal-guard plugin that enforces review discipline, blocks destructive shell commands, and preserves goal state across compaction and restarts.

npm install -g opencode-goal-mode && opencode-goal-mode-install --global

Install · Why it's different · Benchmarks · TUI integration · Configuration · Releasing · Architecture

See ARCHITECTURE.md for the design and research/ for the platform reference, comparison, and threat model.

Why it's different

Most "goal mode" / agentic setups are prompt-only: the model is asked to review its work and to keep going until done. Goal Mode adds a guard plugin that makes that discipline mechanical at the harness layer — the model cannot declare Goal Completed until the required reviews actually passed, and it is blocked from the benchmarked destructive-command bypasses that a regex guard would miss.

Compared to Claude Code and OpenAI Codex (full analysis, with citations and honest caveats, in research/goal-mode-comparison.md):

• It is the only one of the three that mechanically blocks a premature completion claim by default. Goal Mode intercepts the finished message and rewrites Goal Completed → Goal Not Completed unless every required reviewer gate has a fresh PASS and the claimed Review cycles: N matches the recorded counter. Claude Code can do this only via a user-authored Stop hook; Codex's code review is advisory.
• An edit automatically invalidates prior approvals. A reviewer gate counts only when its PASS is newer (by a monotonic integer sequence) than the last edit — so any change forces the relevant reviews to re-run. The public Claude Code and Codex docs reviewed do not describe this stale-review invariant.
• Required specialist reviews are auto-selected and enforced (security, api, data, performance …) from the goal text, contract, and changed files — not left to the model's discretion.
• Destructive commands are blocked by a real shell tokenizer, not a regex. Claude Code's own docs call Bash argument-matching "fragile".

Benchmarks (honest edition)

The headline number is measured on commands the analyzer was never fitted to: 704 real example commands from tldr-pages (common/linux/osx), authored by hundreds of contributors who have never seen this guard. Ground-truth labels come from a deliberately simple, analyzer-independent rule (see build-external-corpus.mjs). Reproduce with npm run bench or node benchmarks/external.mjs.

Honest caveats, because the point of this rewrite was to stop overclaiming:

• The ~7 remaining "misses" are almost all un-flagged single-target rm <file>, which the guard intentionally permits (plain rm is common and the guard blocks rm -r/rm -f, $(rm …), bash -c, interpreters, etc.). Under a strict every-rm-is-destructive labeling those count against it.
• The single counted false positive (git filter-repo …) actually is a history-rewriting command, so the real-world false-positive rate is effectively zero. node benchmarks/external.mjs --json lists every miss and false positive so you can audit the disagreements yourself.

Two curated fixture sets also ship — and they are explicitly fixtures, not an unbiased benchmark. They define the patterns the analyzer must catch and guard against regressions, so they pass by construction; do not read the 100%/0% there as measured accuracy:

• benchmarks/corpus.mjs — 71 destructive patterns (incl. $(…), bash -c, sudo -u, /bin/rm, git -C … reset --hard, curl | sh, interpreter deletes) and their safe look-alikes (git checkout -b, echo "rm -rf /").
• benchmarks/completion-corpus.mjs — 9 completion-claim policy cases (missing review-cycle line, stale review after edit, missing contextual gate, inactive session, custom marker). npm run bench:truthfulness prints them.

The analysis costs ~1µs per command (hundreds of thousands of classifications per second) — negligible for a per-tool-call guard:

Requirements

• Node.js 20.11 or newer.
• OpenCode configured to load local agents, commands, and plugins.

What it adds

• A primary goal agent that owns implementation but delegates research, discovery, verification planning, and reviews to subagents. goal is the only user-selectable agent — every specialist (security, diff, verifier, …) is a mode: subagent that the Goal agent invokes via the task tool; the user never picks one directly. They surface with friendly names (e.g. "Security Reviewer", "API Reviewer") rather than raw ids.
• Strict review gates for prompt compliance, diff review, verification, security, UX, operations, data, API, performance, tests, docs, quality, and final audit.
• Slash commands: /goal, /goal-contract, /goal-review, /goal-evidence-map, /goal-status, /goal-repair, /goal-final.
• The goal-guard plugin:
  • Quote-aware shell analysis that blocks destructive and remote-exec commands (including ones that evade naive regexes — $(rm -rf …), bash -c "…", /bin/rm, git -C … reset --hard, curl | sh) without false-positiving harmless commands like git checkout -b.
  • Completion enforcement: a premature Goal Completed is rewritten to Goal Not Completed with the exact missing review gates.
  • Contextual gating: the goal text and changed files determine which specialist reviewers are required.
  • Reviewer Memory: blocking reviewer findings are carried across cycles, surfaced in status/system context, and marked resolved by fresh PASS verdicts.
  • Disk persistence: review ledgers and Reviewer Memory survive OpenCode restarts.
  • Custom tools: goal_contract, goal_evidence, goal_evidence_map, goal_reviewer_memory, goal_status, goal_reset.
  • Live state injection into the system prompt so the model always knows what the guard requires.
  • TUI toasts: a toast on each review verdict (PASS/FAIL), with the reviewer's friendly name, and a single "completion unlocked" toast the moment the last required gate clears.
• An experimental companion TUI plugin (plugins/goal-sidebar.js) that shows the active goal as a shining-yellow banner in the sidebar with a compact gate status line. See TUI integration.
• A test suite validating the analyzer, plugin hooks, state store, install safety, and config compatibility.

TUI integration

Goal Mode is a plugin pair: the server-side goal-guard plugin owns enforcement and writes its state to disk, and an experimental TUI plugin (plugins/goal-sidebar.js) reads that same state to render a live banner.

• Sidebar goal banner (experimental). The current goal renders in shining yellow in the sidebar (sidebar_content slot), with a passing/total gates · dirty/ready status line, and updates as reviews land. When a task is running but no goal is set, it shows a clean grey No goal and nothing else. It requires a TUI-plugin-capable OpenCode (one exposing api.slots.register); on any older runtime it silently no-ops, so it can never break your TUI. Set sidebarBanner: false (or GOAL_GUARD_SIDEBAR_BANNER=0) to disable, sidebarColor to recolour the goal, or sidebarMutedColor for the "No goal" line. It is rendered-and-asserted headlessly by the visual test (npm run test:visual); still worth a glance in your own TUI.
• Toasts. Review verdicts and completion-unlock events surface as toasts (toastOnReview), and blocked destructive commands / premature completions toast as before (toastOnBlock).

Install

From npm (recommended)

npm install -g opencode-goal-mode
opencode-goal-mode-install --global    # installs into ~/.config/opencode

Then restart OpenCode (it loads agents, commands, and plugins at startup). In the agent picker you will see only the goal agent — the specialist reviewers are subagents the Goal agent drives for you; they are never selectable by the user.

Install into a single project instead of globally:

npm install -D opencode-goal-mode
npx opencode-goal-mode-install         # writes to ./.opencode

Upgrade later by re-running the same install command after npm install -g opencode-goal-mode@latest; the installer replaces only the files it owns and leaves your local edits alone (see Installer options).

From source

git clone https://github.com/devinoldenburg/opencode-goal-mode
cd opencode-goal-mode
npm ci
npm run validate
npm run install:global                 # or: npm run install:local

Installer options

node scripts/install.mjs --dry-run
node scripts/install.mjs --target /path/to/opencode-config
node scripts/install.mjs --global --force
node scripts/install.mjs --global --uninstall

The installer records a manifest of the files it writes. On upgrade it replaces files it owns but refuses to clobber files you have locally modified unless --force is passed. --uninstall removes only the files it installed and leaves your local edits in place.

Configuration

The guard works with zero configuration. To tune it, add options in opencode.json:

{
  "plugin": [
    ["./plugins/goal-guard.js", { "blockDestructive": true, "contextualGates": true }]
  ]
}

Or via environment variables (GOAL_GUARD_*):

Custom tools

The plugin registers six tools the model can call directly:

• goal_contract — record the Goal Contract (requirements, non-goals, acceptance criteria). Activates enforcement and fixes the required gates.
• goal_evidence — record a verification command and result.
• goal_evidence_map — return the acceptance-criteria evidence map with reviewer status, gaps, and next actions.
• goal_reviewer_memory — return unresolved and recently resolved reviewer findings.
• goal_status — return the authoritative gate/dirty/completion status.
• goal_reset — clear the session's goal state (requires confirm: true).

Use /goal-evidence-map when you need a read-only matrix of each acceptance criterion against recorded evidence, reviewer status, gaps, and the next required action. The command is backed by the goal_evidence_map tool, so it uses persisted Goal Guard state rather than relying on transcript memory.

Validation

npm test
npm run validate
npm run audit
npm run publish:check

npm run validate runs the test suite, the structural config validator, the publish readiness check, and an npm pack --dry-run.

Models

Agents do not pin a provider-specific model, so they inherit the model OpenCode is configured to use. To give a particular agent a specific model, add a model: (and optional variant:) line to that agent's frontmatter in your installed copy.

Safety

The installer copies only agents/*.md, commands/*.md, and the plugins/ tree — never auth files, session files, tokens, or personal provider config.

The guard blocks destructive shell commands, marks real file mutations dirty, keeps read-only inspection from dirtying the session, preserves goal state during compaction and across restarts, and blocks premature Goal Completed responses when review gates are missing or stale.

Releasing

Releases are fully automated and version-synced: one pushed tag publishes to npm and creates the matching GitHub Release. The pipeline lives in .github/workflows/publish.yml (Node 24).

npm version patch        # bumps package.json + package-lock.json and creates the vX.Y.Z tag
git push --follow-tags   # pushes main + the tag → the Release workflow runs

On a vX.Y.Z tag push the workflow:

1. installs and runs the full CI gate (npm run ci — tests, audit, structural validation, npm pack --dry-run);
2. runs npm run publish:check, which fails if the tag does not match package.json or if that version already exists on npm;
3. publishes with npm publish --access public using the NPM_TOKEN repository secret;
4. creates the GitHub Release for the tag with auto-generated notes.

So the git tag, the package.json version, the npm version, and the GitHub Release version are always identical. A manual workflow_dispatch is available and defaults to a safe npm publish --dry-run.

One-time setup: add a publish-scoped npm token as the NPM_TOKEN repository secret (gh secret set NPM_TOKEN). Treat that token as sensitive — never commit it.

Goal Completion Contract

Goal Completed is allowed only when:

• All acceptance criteria are mapped to evidence.
• Required verification passed or is credibly accounted for.
• No edit is newer than the latest required review cycle.
• Required reviewers return Verdict: PASS.
• The final answer includes an accurate Review cycles: N.