[Bug]: opencode provider fails at startSession — opencode.mcp.add returns 500 UnknownError

[Bug]: opencode provider fails at startSession — opencode.mcp.add returns 500 UnknownError

Before submitting

• I searched existing issues and did not find a duplicate (ProviderAdapterProcessError + opencode hits are #3571 model-id, #2668 codex, #2163 windows PATH, #3094 remote URL — none is mcp.add 500).

Area

apps/server (opencode provider adapter)

Steps to reproduce

1. T3 service running on Linux (Orange Pi 5, systemd user service), provider = local OpenCode 1.18.29.
2. ~/.config/opencode/opencode.jsonc has 5 Cloudflare remote MCPs enabled (4 of them in needs_auth, only cloudflare-docs connected).
3. Open any thread with the OpenCode provider and send a prompt.
4. Turn fails immediately with the error below.

Expected behavior

Session starts; T3 registers its MCP endpoint on the OpenCode server.

Actual behavior

ProviderAdapterProcessError: Provider adapter process error (opencode) for thread 53e10d8c-c496-448b-8f7d-8b7872317be4: Unexpected server error. Check server logs for details.
...
[cause]: Error: {"body":{"name":"UnknownError","data":{"message":"Unexpected server error. Check server logs for details.","ref":"err_90bd82fd"}},"status":500}

Trace chain (server.trace.ndjson, trace 979074f06a4e769ff881a415b253c991, 2026-09-07T09:00:59Z):
McpSessionRegistry.issue Success → opencode.global.health Success →
verifyOpenCodeServerVersion Success → opencode.mcp.add Failure (500, ref err_90bd82fd) →
startSession Failure → ensureSessionForThread Failure.

Notably the OpenCode server log (~/.local/share/opencode/log/opencode.log) contains
no ERROR line at 09:00:59Z for this request — the 500 is silent server-side
(the only err_90bd82fd matches anywhere are T3's own trace lines).

Side effects of the never-started session:

• ProviderService.stopSession: Cannot route thread ... because no persisted provider binding exists.
• ThreadSettlementReactor: settlement skipped for the thread.

Version / Environment

• t3 v0.0.38 (service), also present on 0.0.39 changelog check: no mcp.add fix listed
• opencode 1.18.29 (/home/sunsilk/.opencode/bin/opencode, satisfies >=1.14.19)
• Linux ARM64 (Orange Pi 5 Max), Node v26.5.0, systemd user service
• Model in other sessions: opencode / muse-spark-1.3-contributor-free
  (free-tier rate_limit_exceeded also observed same day in unrelated sessions)

Logs

Attached / available on request:

• trace-thread-53e10d8c.jsonl — all T3 trace spans mentioning the thread (73 lines)
• trace-turn-failure.jsonl — the 9-span failure chain for the turn
• opencode-log-0900-UTC.txt — OpenCode server log slice around the failure
• opencode.jsonc — MCP config (5 Cloudflare remote MCPs)
• Full stacks redacted to the two at frames above; complete traces in the files.

Root cause (confirmed 2026-09-07, opencode 1.18.29 source + local DB)

Project.migrateProjectId (packages/opencode/src/project/project.ts) runs
UPDATE workspace SET project_id=..., but the local
~/.local/share/opencode/opencode.db still has the ancient
workspace(id, provider, binding, created_at, last_used_at) shape — no
project_id column — even though migration 20260808023530_workspace_domain
is recorded as applied. Every Project.fromDirectory(/home/sunsilk/infra)
therefore throws SQLiteError: no such column: project_id
(opencode.log refs err_1cc751ba, err_48f24f83, err_db9fd44c,
err_cfd939de), and T3's opencode.mcp.add surfaces it as 500
UnknownError (err_90bd82fd, then err_e76414f8 on t3 0.0.39).
Likely trigger: git remote add origin in ~/infra changed the computed
project ID, activating the previous → new migration path.
Verified fixed locally by ALTER TABLE workspace ADD COLUMN project_id TEXT
(backup: opencode.db.bak in this folder) + POST /session for
/home/sunsilk/infra returning a live session. Upstream issue belongs to
sst/opencode (migration recorded-but-not-applied / missing-column guard),
with a T3-side hardening ask below.

Suspected cause / questions

1. Does T3's mcp.add payload conflict... → answered: no, it's the workspace schema mismatch (above).
2. Should a failing T3 MCP registration fail the whole startSession, or degrade (session without T3 tools) with a warning?
3. Should a failing T3 MCP registration fail the whole startSession, or degrade (session without T3 tools) with a warning?
4. OpenCode returns UnknownError 500 with a ref but logs nothing — may need an upstream sst/opencode report with the ref.

Workaround tried

Restarting the T3 service; error persists per-thread on new turns.

Start in apps/server at the OpenCode provider adapter and trace startSession through opencode.mcp.add; compare this with the reported OpenCode Project.migrateProjectId path and missing workspace.project_id schema. The issue is done only when the upstream migration problem and the T3-side choice about failing versus degrading MCP registration are resolved and session-start behavior is verified.