Wire opencode interactive through its agent abstraction

[willwashburn]

Summary

Opencode interactive was broken on two counts. Personas would launch but their system prompt got dumped into the TUI input box as pending user text, and the declared model silently fell back to opencode's default regardless of what the persona specified.

Root causes, per opencode's own CLI help and https://opencode.ai/config.json schema:

1. --prompt = pre-fill the chat input with a user message. Not the agent's instructions.
2. -m/--model wants full provider/model form (e.g. opencode/gpt-5-nano). Harness-kit was stripping the opencode/ prefix, leaving a bare model name opencode could not resolve.

The correct pathway is opencode's agent abstraction: define agent.<id>.{ model, prompt, mode } in an opencode.json under cwd, select it at launch with --agent <id>.

Changes

• packages/harness-kit — extend InteractiveSpec with configFiles: InteractiveConfigFile[] (relative path + contents) so the opencode branch can emit the per-session opencode.json. Builder now takes personaId as the agent name. Claude/codex return an empty array.
• packages/cli — materialize spec.configFiles into the mount dir via onBeforeLaunch so opencode finds opencode.json at its cwd. The non-mount opencode path (only reachable under --install-in-repo) would pollute the user's real repo root, so it degrades: stripAgentFlag removes --agent <id> from argv, warns, and launches opencode with its default agent. Documented trade-off: --install-in-repo cannot apply the persona's prompt; the default (mount) path handles it cleanly.

Known scope limits

• --install-in-repo with opencode loses persona prompt application in this PR. Fix is possible (either write opencode.json into a scratch dir we still resolve from, or accept writing into the user's repo with cleanup on exit) but orthogonal to the primary bug and deferred.
• Claude and codex paths are unchanged beyond adding configFiles: [] and threading personaId through.

Test plan

• corepack pnpm run check green — harness-kit 26, workload-router 37, cli 47
• Interactive smoke: npm run dev:cli agent frontend-implementer (opencode tier) — confirm (a) TUI shows no pending user text, (b) persona's declared model appears in the model indicator rather than opencode's default, (c) agent responds using the persona's system prompt
• npm run dev:cli --install-in-repo agent frontend-implementer — confirm warning fires and opencode still launches

🤖 Generated with Claude Code

---

[barryollama]

LGTM 🚀

Clean fix addressing the root causes of both opencode interactive bugs:

What this fixes:

1. ✅ was incorrectly pre-filling TUI as user message instead of setting agent instructions
2. ✅ Stripped model names caused silent fallback to opencode defaults

Architecture:

• Clean separation of concerns: harness-kit defines WHAT config files needed, CLI decides WHERE to write them
• opencode.json structure correctly follows https://opencode.ai/config.json schema
• Degraded path for with clear warning to users

Edge cases handled well:

• Config files hidden from mount-mirror in both directions (prevents masking existing + polluting user's repo)
• stripAgentFlag handles all argv positions correctly
• Trailing without value preserved as defensive measure
• Dynamic ignoredPatterns keeps fix generic for future configFile producers

Tests:

• All 26 harness-kit tests passing, new coverage for opencode.json shape and stripAgentFlag behavior

Minor suggestion (non-blocking): Consider adding a test case for multiple flags in stripAgentFlag, though unlikely in practice.

Great work on the thorough commit messages and inline documentation too.

[barryollama]

LGTM

Clean fix addressing the root causes of both opencode interactive bugs:

• --prompt was incorrectly pre-filling TUI as user message instead of setting agent instructions
• Stripped model names caused silent fallback to opencode defaults

Architecture: Clean separation of concerns - harness-kit defines WHAT config files needed, CLI decides WHERE to write them. opencode.json structure correctly follows https://opencode.ai/config.json schema. Degraded path for --install-in-repo with clear warning to users.

Edge cases handled well: Config files hidden from mount-mirror in both directions (prevents masking existing + polluting user's repo). stripAgentFlag handles all argv positions correctly. Trailing --agent without value preserved as defensive measure. Dynamic ignoredPatterns keeps fix generic for future configFile producers.

Tests: All 26 harness-kit tests passing, new coverage for opencode.json shape and stripAgentFlag behavior.

Minor suggestion (non-blocking): Consider adding a test case for multiple --agent flags in stripAgentFlag, though unlikely in practice.

Great work on the thorough commit messages and inline documentation too.

[barryollama]

LGTM. Clean fix for both opencode interactive bugs: --prompt misuse and stripped model names. Great test coverage and edge case handling.

[barryollama]

Code Review: Wire opencode interactive through its agent abstraction

Summary

This PR correctly fixes two significant bugs in opencode interactive mode:

1. System prompt dumped as user input — Previously --prompt was used, which opencode treats as a user message pre-fill, not agent instructions
2. Silent model fallback — Stripping the opencode/ provider prefix left an unresolvable bare model name

The solution using opencode's agent abstraction (opencode.json + --agent <id>) is the architecturally correct approach.

---

🔴 Requested Changes

1. Missing Exhaustiveness Check in buildInteractiveSpec (Line 214, harness.ts)

The switch statement on the Harness union type lacks a final default case or exhaustiveness check. While all three current harnesses are handled, TypeScript's strict mode won't catch future additions to the Harness union.

Suggestion: Add a default case that throws:

default: {
  const _exhaustiveCheck: never = harness;
  throw new Error(`Unhandled harness: ${_exhaustiveCheck}`);
}

This provides compile-time safety when new harnesses are added.

2. Comment/Documentation Staleness (Lines 93-98, harness.ts)

The JSDoc comment for buildInteractiveSpec still describes the old opencode behavior:

"The opencode branch uses opencode's own --prompt flag..."

This is now outdated and misleading. Please update to reflect the new agent-based approach.

---

⚠️ Observations (Non-blocking)

3. Degraded User Experience for --install-in-repo

The PR acknowledges this trade-off, but I want to flag it: users running --install-in-repo with opencode personas will see a warning and lose the persona's system prompt. This is acceptable for now, but consider documenting this prominently in the CLI help text or adding a "known limitations" section.

4. Test Coverage

The new tests for stripAgentFlag and the opencode configFiles generation are thorough. However, there's no test for the CLI-level integration — specifically the warning emission when degradeConfigFiles triggers. Consider adding a test in cli.test.ts that verifies the warning is printed when opencode + --install-in-repo is used.

---

✅ Looks Good

• Clean separation between harness-kit (pure data) and CLI (I/O)
• The configFiles abstraction is generic and will support future harnesses
• stripAgentFlag is well-tested with edge cases (trailing --agent, surrounding args)
• The dynamic ignoredPatterns push for configFiles (lines 587-589 in cli.ts) prevents sync-back pollution elegantly

---

Reviewed by Hermes Agent