fix(workspaces): recover from watcher startup failures

[EivindKjosbakken]

Description

Recover the workspace registry when a native filesystem watcher cannot start.

Previously, a rejected WatchHandle.ready() promise escaped as an unhandled rejection and terminated the workspace-registry worker. After the generic supervision budget was exhausted, new workspace activation stayed unavailable until the entire desktop app was restarted.

This change:

• observes watcher startup failures and keeps the polling freshness floor active;
• retries failed native watches with bounded exponential backoff;
• supervises the workspace-registry worker indefinitely with a 30-second maximum delay;
• clears all retry state when targets disappear or the scheduler is disposed.

Related issues

No existing issue found.

Testing

• pnpm --filter @emdash/core run format
• pnpm --filter @emdash/core run typecheck
• pnpm --filter @emdash/core run lint
• pnpm --filter @emdash/core exec vitest run --maxWorkers=1 (203 files, 1,369 tests)
• pnpm run build (9 projects)
• packaged an arm64 macOS Canary build
• local Codex review: clean after addressing its polling-fallback finding

Screenshot/Recording (if applicable)

Not applicable; this is worker recovery behavior.

Checklist

• I kept this PR small and focused
• I ran a self-review before opening this PR
• I ran the relevant local checks or explained why not
• I updated docs when behavior or setup changed, or docs were not required
• I added or updated tests when behavior changed, or explained why not
• I only added comments where the logic is not obvious
• I used Conventional Commits for the commit message and PR title

[greptile-apps]

Greptile Summary

The PR makes workspace-registry recovery resilient to native watcher startup failures and exhausted worker restart budgets.

• Observes watcher readiness failures while retaining periodic polling.
• Retries native watcher creation with capped exponential backoff.
• Clears watcher retry timers and attempt state when targets disappear or the scheduler is disposed.
• Gives the workspace-registry worker indefinite failure supervision with a 30-second maximum delay.
• Adds focused tests for polling fallback, watcher reattachment, and worker supervision behavior.

Confidence Score: 5/5

The PR appears safe to merge, with watcher recovery, polling fallback, retry cleanup, and worker shutdown behavior remaining coherent.

Failed watcher creations are released and retried through fresh resource-cache entries while polling continues independently, and indefinite worker supervision remains cancellable during shutdown.

Important Files Changed

Filename	Overview
packages/core/src/runtimes/workspace-registry/node/scan/scheduler.ts	Adds guarded watcher readiness handling, bounded retry state, and complete timer cleanup without disrupting the independent polling floor.
packages/core/src/runtimes/workspace-registry/node/scan/scheduler.test.ts	Adds coverage for polling during watcher failure and event delivery after a failed watch is successfully reattached.
packages/core/src/runtimes/workspace-registry/node/worker-spec.ts	Replaces bounded default worker supervision with intentional indefinite on-failure retries capped at 30 seconds.
packages/core/src/runtimes/workspace-registry/node/worker-spec.test.ts	Verifies the worker’s restart policy, initial delay, and indefinitely repeated maximum delay.

Sequence Diagram

sequenceDiagram
  participant S as Scan Scheduler
  participant W as Watch Service
  participant P as Polling Floor
  S->>W: watch(target)
  W-->>S: ready() rejects
  S->>W: release failed handle
  S->>S: schedule bounded backoff
  P->>S: request freshness scan
  S->>W: retry watch(target)
  W-->>S: ready() resolves
  W-->>S: filesystem events resume

Loading

_{Reviews (1): Last reviewed commit: "fix(workspaces): recover from watcher st..." | Re-trigger Greptile}

[jschwxrz]

closing as resolved