fix(install): write opencode MCP config in official schema

[fdcp]

code-review-graph install --platform opencode was writing {repo_root}/.opencode.json with a Cursor-shaped mcpServers payload (command + args + type: stdio). OpenCode's documented schema is different: project config lives in opencode.json / opencode.jsonc, MCP servers go under the mcp key, local servers use type: local with a single array-form command. The previous file was never loaded by OpenCode, so install --platform opencode was effectively a no-op. Issue #388 (closed as fixed by PR #198 in 2026-05, but that PR only added a hook plugin — not the MCP config fix) was reopened in practice.

Changes:

• PLATFORMS["opencode"]: config_path now resolves opencode.jsonc if it already exists, otherwise opencode.json, otherwise creates opencode.jsonc (matches OpenCode's documented convention); key: mcp; needs_type: false
• _build_server_entry emits a single dict for OpenCode with array-form command ([command, *args]) and type: local; no separate args / env array. All other platforms unchanged
• _warn_legacy_opencode_config prints a one-line hint when an old .opencode.json with a code-review-graph entry is still present (no auto-delete — safer than clobbering). The hint now derives the "new config is at …" path from the same _opencode_config_path helper used by config_path, so the message stays consistent with the actual write path
• _strip_jsonc_comments helper — replaces the previous regex call in install_platform_configs and _warn_legacy_opencode_config. The earlier r'(^|\s)//.*$' regex still ate // sequences inside quoted strings when preceded by whitespace (e.g. "text": "\thello // world"). The new helper is a small stateful scanner that tracks string boundaries, stops at either \r or \n (so CRLF and old-Mac CR-only files are handled), and only strips // outside quoted regions. Found and fixed while adding .jsonc support
• Tests: existing test_install_opencode_config rewritten for the new schema; new tests cover the legacy-file warning (presence + absence), merge-into-existing behavior for both opencode.json and opencode.jsonc, JSONC comment stripping (string preservation, full-line and inline stripping, CRLF/CR line endings), and _opencode_config_path precedence
• docs/USAGE.md updated to show opencode.json / opencode.jsonc

Review status (Copilot, 3 rounds): 8 of 10 threads resolved. The 2 remaining are intentionally left open:

• docs row wording (L33 uses a → arrow that could be misread as a migration; the code only does path selection, not migration) — left as-is to keep the diff minimal; happy to refine if the maintainer prefers.
• Trailing-comma stripping regex at skills.py:416 (re.sub(r',(\s*[}\])])', …)) is pre-existing on main and is string-unaware in the same class as the JSONC-comment bug. Out of scope for this PR; worth a follow-up issue if the maintainer agrees.

Fixes #388.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Updates OpenCode platform support to use the newer mcp schema and prefer opencode.jsonc/opencode.json, including merge behavior and a warning for legacy .opencode.json files.

Changes:

• Switch OpenCode config target from .opencode.json + mcpServers to opencode.jsonc/opencode.json + mcp.
• Update OpenCode server entry format to type: local and a list-based command.
• Add a warning when a legacy .opencode.json containing code-review-graph is detected; expand tests/docs accordingly.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
tests/test_skills.py	Updates/extends tests to validate new OpenCode config paths, schema, merge behavior, and legacy warning output.
docs/USAGE.md	Documents OpenCode config file locations as opencode.json / opencode.jsonc.
code_review_graph/skills.py	Implements OpenCode config path selection, new entry format, legacy warning, and adjusts JSONC comment stripping.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

[fdcp]

Friendly ping, @tirth8205 👋

Just flagging this is still on the table — issue #388 has 3+ users (PhilYang09, dragonabyss, Dhruv-Garg79) confirming install --platform opencode is broken on v2.3.2 / v2.3.3 / v2.3.5. This PR addresses it directly.

Summary:

• 3 commits, 1275 tests pass, ruff clean
• Copilot 3 rounds of review: 8/10 threads resolved, 2 intentionally left open with rationale in the PR body
• No new external dependencies
• Backwards-compatible: existing .opencode.json users get a deprecation hint, not a silent overwrite

Happy to iterate on review feedback or split into smaller PRs if that helps. If you're not actively maintaining right now, no worries — at minimum the diff here gives anyone hitting #388 a known-good reference. Just let me know either way so I can stop pinging 🙂

[tirth8205]

Thanks @fdcp — and apologies for the slow response on this and #388. I applied the branch and verified it end to end: the schema target is right (confirmed against opencode.ai/docs/mcp-servers and the published config.json — opencode.json/.jsonc, top-level "mcp", type "local", array command), the legacy-hint-without-delete call is the right one, and _strip_jsonc_comments is a genuine fix over the old regex. Full suite: 1282 passed; mypy clean.

Two changes to get this merged:

1. Drop "cwd" from the OpenCode entry. McpLocalConfig in the official schema is additionalProperties: false (only type/command/environment/enabled/timeout), so we'd be writing an out-of-spec key in a spec-compliance PR. Append "--repo", str(repo_root) to the command array instead — serve --repo already feeds _default_repo_root — and assert "cwd" not in entry in test_install_opencode_config.
2. Fix the new E501 at tests/test_skills.py:913 (103 > 100 chars).

Agreed the trailing-comma regex is out of scope — please file that follow-up issue. Happy to merge once those land.

[fdcp]

Both changes are pushed (commits rebased onto current main, head b7ec0ef):

1. cwd removed from the OpenCode entry. --repo <path> is now appended to the command array — serve consumes it via _default_repo_root (cli.py:652 → main.py:1038). Schema-compliant: entry now contains only type and command.
   • Assertion added: assert "cwd" not in entry in test_install_opencode_config.
2. E501 fixed at tests/test_skills.py:913 by splitting the long line into named locals.

Verification:

• pytest tests/test_skills.py → 150 passed (was 149)
• ruff check code_review_graph/skills.py tests/test_skills.py → 3 errors, all pre-existing on main (I001 import sort at L3 + L18, docstring >100 chars at L1472). No new errors introduced by this commit.

Follow-up issue filed per your request: #553 (trailing-comma regex at skills.py:417 doesn't track string boundaries — same bug class as the _strip_jsonc_comments one fixed here).

One thing I can't do from a fork: the CI and PR Review checks are showing as action_required (0 jobs) on head b7ec0ef. That's the fork-PR workflow-approval gate. Could you approve them in the Checks tab? Once they go green, this should be ready to merge.

[fdcp]

Follow-up: I re-ran the full verification on the rebased head after pushing b7ec0ef (not just the touched file). Results:

• python -m pytest -q → 1385 passed, 8 skipped, 2 xpassed (23 warnings)
• ruff check . → 49 errors, which matches origin/main exactly (same pre-existing baseline; no new lint introduced by this branch)

So the branch now has:

• the two requested fixes (drop cwd / add --repo, plus the E501 cleanup)
• full-suite green on the rebased branch
• no new repo-wide ruff regressions

Still waiting on the fork-PR workflow approval gate in Checks (CI / PR Review showing action_required). Once those are approved and rerun, this should be merge-ready from my side.

[fdcp]

Small follow-up pushed on top: 55db06e (fix(tests): clean up pre-existing lint in test_skills). I cleaned the 3 pre-existing ruff findings in the touched file (tests/test_skills.py): 2x I001 import ordering + 1x stale long docstring.

Updated verification:

• ruff check tests/test_skills.py → clean
• python -m pytest tests/test_skills.py -q → 150 passed
• python -m pytest -q → 1385 passed, 8 skipped, 2 xpassed
• ruff check . baseline is now 46 on this branch vs 49 on origin/main (i.e. this branch removes 3 pre-existing lint issues rather than introducing any)

So the branch now has the requested install/schema fixes plus a small lint cleanup in the touched test file.

[fdcp]

Quick nudge, @tirth8205: the code side is done on head 55db06e, but the PR is still blocked on the fork-PR workflow gate. Could you hit Approve and run for the CI / PR Review checks in the Checks tab when you get a moment? Once those run, this should be ready to merge from my side.