🤖 fix: stop idle auto-compaction loop on repeated/model-not-found failures

[ammar-agent]

Summary

Stops the hourly idle auto-compaction loop from re-attempting a workspace that keeps failing: it now gives up immediately when an attempt fails with a model-not-found error, and otherwise after two consecutive failures of any kind. Previously a persistently failing workspace (e.g. a misconfigured/typo'd compaction model) was re-queued every hour forever, since a failed compaction neither marks the workspace compacted nor refreshes recency, so it stayed eligible indefinitely.

Background

IdleCompactionService re-checks eligibility every hour and re-queues any idle workspace. There was no failure tracking, so a workspace whose compaction always fails would be retried on every cycle with no cap. The most common cause is a compaction model the provider rejects.

A key subtlety: the failure can surface in two different places.

• Pre-stream failures (invalid model string, model unavailable, provider/runtime issues) come back synchronously as the sendMessage Result that executeIdleCompaction already inspects.
• Mid-stream failures (the common model_not_found case for a well-formed but non-existent Anthropic/OpenAI model) are not returned by sendMessage — they arrive later as an async "error" stream event and are only persisted transiently to partial.json, so they can't be recovered from chat history on the next check.

Implementation

• IdleCompactionService now tracks per-workspace consecutive failures and a sticky suppressedWorkspaceIds set. checkEligibility short-circuits suppressed workspaces (reason: "suppressed_after_failures"). A new recordOutcome(workspaceId, outcome) applies the policy: success resets the streak; model_not_found suppresses immediately; any other failure suppresses on the second consecutive occurrence.
• WorkspaceService reports terminal idle-compaction outcomes through a listener:
  • pre-stream failure → reported from executeIdleCompaction (model-not-found classified via formatSendMessageError),
  • mid-stream failure → reported from the existing aiService.on("error") handler when the workspace is mid idle-compaction (read before stopStreamingStatus clears the marker),
  • success → reported from handleStreamCompletion.
  • error / stream-end / stream-abort are mutually exclusive per turn, so a failed turn never also reports success. Aborts are intentionally neutral (not a failure).
• ServiceContainer wires the listener to IdleCompactionService.recordOutcome.

Suppression is in-memory and sticky for the service lifetime; restarting the app (e.g. after fixing the configured compaction model) clears it.

Validation

• New unit tests in idleCompactionService.test.ts: suppress after two consecutive failures, suppress immediately on model_not_found, success resets the streak, and checkAllWorkspaces no longer queues a suppressed workspace.
• New workspaceService.test.ts tests: executeIdleCompaction reports modelNotFound: true for an invalid compaction model and false for a generic pre-stream failure.
• make static-check and targeted tests pass.

Risks

Low. Behavior is additive and scoped to the background idle-compaction loop. The outcome listener is optional (no-op until wired), and normal interactive sends are untouched. Worst case of a wrong suppression is that a workspace stops auto-compacting on idle until restart, which a user can already trigger manually.

---

Generated with mux • Model: anthropic:claude-opus-4-8 • Thinking: xhigh • Cost: $1.48

[ammar-agent]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8a65f175c2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ammar-agent]

Addressed in 083e71d.

The idle-compaction outcome is no longer reported from the provider stream-end. Instead, CompactionHandler.handleCompletion() now reports the terminal outcome via a new onIdleCompactionOutcome(success) callback:

• success is reported only after the compacted summary is actually persisted (right after onCompactionComplete),
• failure is reported on the post-stream persistence paths (performCompaction returns Err, empty summary, raw-JSON summary).

WorkspaceService.createSession forwards that to reportIdleCompactionOutcome (a persistence failure → { success: false, modelNotFound: false }), so these post-stream failures now count toward the suppression threshold. The unconditional success report in handleStreamCompletion was removed. Mid-stream provider errors (no stream-end/no persistence) are still reported from the "error" handler, and the three terminal events are mutually exclusive per turn, so each idle attempt reports exactly one outcome.

Added CompactionHandler tests: success → [true], empty summary → [false], manual (non-idle) compaction → not called.

[ammar-agent]

@codex review

Please take another look — addressed the post-stream persistence concern.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 083e71d465

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ammar-agent]

Addressed in c1a1b76.

The requireIdle busy-skip ("Workspace is busy; idle-only send was skipped.") is now treated as neutral — executeIdleCompaction no longer calls recordOutcome for it, so normal user-interaction races can't accumulate toward suppression. I factored the message into a shared IDLE_ONLY_BUSY_SKIP_MESSAGE constant (used by the requireIdle return, the early busy throw, and the skip check) to keep the contract DRY. Extended the busy-skip test to assert no outcome is reported.

[ammar-agent]

@codex review

Please take another look — the requireIdle busy-skip is now treated as neutral.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c1a1b76bc4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ammar-agent]

Addressed in 866a615.

recordOutcome now also clears suppressedWorkspaceIds on success, so a later in-flight retry that actually persists a compaction lifts the suppression and the workspace becomes eligible again (self-healing). Added a test: two failures suppress → a subsequent success makes the workspace eligible again.

[ammar-agent]

@codex review

Please take another look — suppression is now lifted on a later success.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Breezy!

Reviewed commit: 866a615c91

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".