Fix forked workers exiting immediately and false memory-persistence failures

[jamiepine]

Two bugs from the worker-reliability work, both found by testing #633 against a live instance.

Forked workers never ran

Forking channel history by default populated prior_history for every new worker. run_inner read that field as "this worker already ran and relayed its result", so every forked worker skipped its task loop entirely and returned an empty result in about a millisecond:

05:11:42.605543  worker starting      task=Clone the GitHub repository...
05:11:42.605854  resuming interactive worker with prior history  prior_messages=2
05:11:42.606735  worker completed

No LLM call, no tools, nothing. The agent got empty results back and re-spawned the same worker four times trying to verify work that never happened.

prior_history was carrying two meanings: a resumed idle worker, and a fresh worker seeded with channel context. Only the first has already relayed a result, and the state machine already distinguishes them — Worker::resume is the only pre-run site that sets WaitingForInput:

let resuming = self.state == WorkerState::WaitingForInput;

Second, worker_runs.transcript was read as Vec<u8>. A NULL blob decodes to an empty vec rather than erroring, so NULL became Some([]), failed the gzip header read with unexpected end of file, and suppressed the live-transcript fallback — which is why worker transcripts came back empty in the UI and via worker_inspect.

Memory persistence reported failure after succeeding

A persistence run called memory_persistence_complete, then took one more LLM turn with nothing left to say. The provider answers that with an empty message, the Responses parser rejects it, and the branch reported:

Branch failed: CompletionError: ResponseError: empty or unsupported response from
OpenAI ChatGPT Responses API ... received output types: message

Every memory had already been persisted. Channel turns already terminate after reply/skip for exactly this reason; the branch's terminal tool never got the same treatment. The hook now terminates once a valid terminal outcome is recorded, branches conclude from that outcome, and ingestion treats the signal as a completed chunk. Tool errors still continue, so the contract-retry path is unchanged.

This also drops one LLM call per persistence run and per ingestion chunk.

Testing

Two hook tests cover the new termination: a valid terminal outcome stops the loop with the right reason (both saved and no_memories), and a tool error does not. Full lib suite passes at 1100.

The worker fix has no test — nothing constructs a Worker outside channel_dispatch with live AgentDeps, so covering resume-vs-fork needs a test fixture that doesn't exist yet. Worth adding separately, since this is the second bug in that path.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: c2b1cd97-8f5d-410b-9904-6ba302ecc7e4

📥 Commits

Reviewing files that changed from the base of the PR and between a95b32e and c79fcc4.

📒 Files selected for processing (5)

• src/agent/branch.rs
• src/agent/ingestion.rs
• src/agent/worker.rs
• src/conversation/history.rs
• src/hooks/spacebot.rs

---

Walkthrough

Memory-persistence completion now records terminal outcomes, stops hook processing, and propagates completion through ingestion and branch execution. Worker resumption and empty transcript handling also receive state-specific updates.

Changes

Memory persistence completion

Layer / File(s)	Summary
Hook completion contract src/hooks/spacebot.rs	SpacebotHook adds a completion reason and predicate. Valid memory_persistence_complete results record terminal outcomes and stop the hook loop. Tests cover saved, no-memory, and error outcomes.
Agent completion propagation src/agent/branch.rs, src/agent/ingestion.rs	Branch execution clears the active contract and formats the terminal outcome. Ingestion treats the completion cancellation as successful chunk processing.
Worker and transcript state handling src/agent/worker.rs, src/conversation/history.rs	Worker resumption detection uses WaitingForInput. Fresh workers log seeded forked messages. Empty transcript blobs use the missing-transcript fallback.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• spacedriveapp/spacebot#522: Shares the memory_persistence_complete flow but addresses event types and persistence validation.
• spacedriveapp/spacebot#524: Expands the event taxonomy and schema used by memory-persistence completion handling.
• spacedriveapp/spacebot#612: Wires the memory-persistence contract into SpacebotHook and ingestion handling.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch jamiepine/worker-fork-fixes

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}