Pause goals after repeated transport failures

[ruizkinio]

Summary

Align automatic goal continuation with native Codex-style pause semantics when the provider stops answering.

• cap consecutive unresolved transport/no-response attempts at three
• pause the goal on the third failure instead of scheduling a fourth retry
• reset the failure episode only after substantive progress or explicit /goal resume
• suppress duplicate idle events and recursive watchdog re-arming
• keep quota, interruption, context-overflow, and other terminal errors on their existing paths
• preserve the upstream OpenTUI function API and max_turn_time behavior

Why

A live connection outage caused one goal session to issue 4,066 automatic continuations until a downstream alias limit was exhausted. Goal mode should stop safely when the model cannot answer and wait for an explicit resume, rather than retrying indefinitely.

Validation

• 109 tests pass
• lint passes
• typechecks against OpenCode 1.17.17 and 1.18.x pass
• build and npm package dry-run pass
• regression coverage includes exact live socket/header-timeout errors, duplicate idle events, watchdog ceilings, incomplete tool shells, substantive progress resets, and explicit resume
• independent review found no remaining issues

[danyel117]

Thanks for tackling the runaway continuation problem. The bounded transport/no-response recovery is valuable, and the branch passes its local gate for me: lint, both typechecks, 109 tests, build, and package dry-run.

I do not think the PR is ready to merge yet. Please rebase it onto current main and narrow it to the transport/no-response fix. The following issues need to be addressed:

1. Retries can stall with the default minimum interval. After an unresolved attempt, the immediate retry can be rejected by the default 3-second min_continue_interval_seconds, but no timer schedules the retry for the remaining delay. The existing bounded-retry tests use interval 0 and do not cover the default. Please schedule the deferred retry and add coverage using the default interval.

2. Direct promptAsync exceptions do not preserve the existing terminal error paths. The send catch classifies the error, but quota, interruption, context-overflow, and other terminal errors still flow through generic continuation-failure handling instead of the same usageLimited, interruption, compaction, or blocked behavior used for session.error. Please unify these paths and add direct-send tests for 429/quota, abort/interruption, context overflow, and terminal provider errors.

3. The no-response failure episode is process-local. continuationFailureStreaks is lost on plugin restart, while a successful prompt submission resets the persisted continuationFailures before substantive response progress is known. Please make the failure count and pause reason durable, and add restart tests after unresolved attempts.

4. Persisted continuation reservations can survive a crash and permanently block future reserveContinuation calls. Please add stale-reservation recovery/expiry and a restart regression test.

5. Please extend duplicate-idle coverage to reordered events, especially a delayed duplicate idle after a busy event or repeated observation of the baseline assistant message.

6. The branch is behind current main and currently conflicts in README.md, src/prompts.ts, and dist/server.js. The rebase must preserve the cache-stable reminder behavior added by PR Fix prompt cache invalidation for active goals #17; active system reminders must not reintroduce live token/time/checkpoint counters.

Please leave the following changes out of scope for this PR:

• package/repository/plugin rebranding and release-workflow changes
• default state-directory changes
• public tool and status redesign, including unmet to blocked and completion-evidence changes
• removal of the existing default auto-turn and no-progress safety caps
• SQLite cross-process locking and the broader message-accounting redesign
• TUI command/action redesign
• replacement of the project prompt set with native Codex text
• compatibility-matrix, community-template, conduct/security-contact, and unrelated documentation rewrites

A focused PR should contain the bounded transport/no-response episode tracking, only the reservation/watchdog deduplication required for that ceiling, focused regression tests, and a regenerated dist/server.js from current main. Please also update the PR description with the AI model and agent harness used, or state that the changes were made manually, per the current contribution guidelines.

[danyel117]

Superseded by #20, which carries forward the focused bounded transport/no-response retry work on current main and is green in CI.

Thank you @ruizkinio for identifying the runaway continuation problem and for the original recovery design and live transport cases. The replacement commit credits Sam as co-author, and the PR description documents the adaptation.