fix(opencode): recover stuck stop boundaries

[timigod]

Summary

• classify OpenCode stop-boundary start failures and retry the unsent prompt exactly once
• isolate the replacement OpenCode runtime generation so closing the old session cannot cancel replacement work
• atomically register the replacement before closing the old session so Stop never loses agent ownership during recovery
• revalidate the original registered agent immediately before replacement install so a concurrent close cannot be resurrected
• close a stale replacement and preserve the durable closed record when reload loses ownership
• make old-session close and recovery cancellation abortable and bounded for Stop, close, and daemon shutdown
• fence asynchronous OpenCode turn startup with a generation canceled by Stop or shutdown before provider submission
• reject pending local startup when autonomous OpenCode activity wins slash-command discovery
• classify derived OpenCode providers through typed base-provider provenance rather than their visible provider ID
• preserve current event-stream reconnect and provider-status reconciliation behavior

Testing

• npx vitest run packages/server/src/server/agent/providers/opencode-agent.test.ts --bail=1 (92 passed)
• npx vitest run packages/server/src/server/agent/agent-manager.test.ts --bail=1 (165 passed)
• npx vitest run packages/server/src/server/agent/provider-registry-wrap.test.ts --bail=1 (2 passed)
• held-storage regression for close before reload replacement commit
• successful follow-up prompt after Stop during recovery handoff
• named shutdown coverage during asynchronous slash-command discovery
• npm run build:server
• npm run typecheck
• npm run lint
• npm run format
• npm run format:check
• git diff --check

[greptile-apps]

Greptile Summary

This PR hardens OpenCode stop-boundary recovery and concurrent lifecycle handling.

• Retries an unsent prompt once after a typed OpenCode stop-boundary rejection.
• Adds atomic replacement-session registration, ownership revalidation, bounded cleanup, and cancellation handling.
• Fences asynchronous OpenCode startup and supports derived OpenCode providers through base-provider provenance.
• Expands regression coverage for Stop, close, shutdown, reload, and startup races.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains within the scope of the available follow-up review threads.

Important Files Changed

Filename	Overview
packages/server/src/server/agent/agent-manager.ts	Adds coordinated OpenCode recovery, replacement registration, cancellation admission, bounded cleanup, and lifecycle-race handling.
packages/server/src/server/agent/agent-run-state.ts	Extends pending-run tracking to stage startup events and settle admitted turns safely.
packages/server/src/server/agent/agent-sdk-types.ts	Adds typed startup rejection, admission, close, resume, and provider-provenance contracts.
packages/server/src/server/agent/provider-registry.ts	Preserves typed base-provider provenance when wrapping derived provider sessions.
packages/server/src/server/agent/providers/opencode-agent.ts	Classifies stop-boundary failures and fences asynchronous startup against cancellation and autonomous activity.
packages/server/src/server/agent/agent-manager.test.ts	Adds broad behavioral coverage for recovery, replacement ownership, cancellation, and shutdown races.
packages/server/src/server/agent/providers/opencode-agent.test.ts	Covers OpenCode startup generation fencing, stop-boundary classification, and slash-command discovery races.
packages/server/src/server/agent/provider-registry-wrap.test.ts	Verifies wrapped sessions retain their derived identity and typed base-provider provenance.
packages/server/src/server/agent/providers/opencode/test-utils/test-opencode-harness.ts	Extends the OpenCode test harness to coordinate and observe asynchronous startup scenarios.

Sequence Diagram

sequenceDiagram
  participant C as Client
  participant M as AgentManager
  participant O as Old OpenCode session
  participant R as Replacement session

  C->>M: Submit prompt
  M->>O: startTurn(prompt)
  O-->>M: previous_turn_still_stopping
  M->>R: Resume isolated provider generation
  M->>M: Revalidate ownership
  M->>M: Atomically register replacement
  M->>O: Close old session (bounded)
  alt Stop, close, or shutdown before admission
    C->>M: Cancel
    M->>M: Abort recovery
    M-->>C: turn_canceled
  else Retry admitted
    M->>R: Retry original prompt once
    R-->>M: Turn events
    M-->>C: Stream turn result
  end

Loading

_{Reviews (2): Last reviewed commit: "Merge latest main into OpenCode stop rec..." | Re-trigger Greptile}