fix(task-session): fill false-completed task parts with child session real output (#863)

[Jiajun0413]

Problem

When a foreground task's primary model halts on a non-retryable error (e.g. 403 quota exhausted), the orchestrator receives a completed task tool part with an empty <task_result>, even though the fallback model continues on an orphan runLoop and produces the real result minutes later. The orchestrator reads the empty output, mis-judges the task as failed/empty, and self-amplifies by launching redundant replacement tasks (#863).

Verified root cause (DB + opencode 1.18.5 source)

Complete chain, anchored in opencode core:

1. Model 1 returns 403 isRetryable:false → opencode halt (processor.ts:599) sets ctx.assistantMessage.error, publishes session.error, sets status idle
2. process returns "stop" (processor.ts:680: if (ctx.assistantMessage.error) return "stop")
3. prompt loop breaks → ops.prompt returns the last assistant message, which is empty (finish=undefined, parts=[]) because halt produced no content
4. omos session.error handler triggers tryFallback (non-abort path) → promptAsync re-prompts with model 2
5. opencode runTask (task.ts:213): result.parts.findLast(text)?.text ?? "" → empty string
6. settle Exit.succeed("") (background-job.ts:148) → status:"completed", output:""
7. parent task part written as completed with empty <task_result> — 12s after launch
8. omos tryFallback's promptAsync takes effect only after model 1's runLoop ends → model 2 runs on an orphan runLoop, producing 9346 chars over 855s — but the parent part already says completed+empty

Root cause is in opencode runTask's ?? "" fallback, which treats a halt-empty assistant message as success. omos cannot patch this as a plugin (no hook to intercept runTask settle). This is the false-complete counterpart to the false-cancel bug in PR #880.

DB evidence (reproduced session)

	Time	State
B-1 launch	+0s	—
parent part completed + empty output	+12s	<task_result></task_result>
B-1 child continues	+12s → +867s	70 parts, 9346-char audit report
orchestrator mis-judges	+645s	"Councillor B errored/empty — retry once"
orchestrator relaunches B-2/B-3	+692s/+693s	from scratch, missing the 9346 chars

Fix

In the experimental.chat.messages.transform hook, after stabilizeRunningTaskParts, scan task tool parts with status:"completed" whose parsed <task_result> is empty. For each, read the child session's real assistant text via the existing extractSessionResult helper and in-place rewrite the part output with renderTaskCompletedWithText (mirrors opencode renderOutput task.ts:64-76).

Idempotent and non-blocking: if the child session has not yet produced non-empty text (fallback model still running), extractSessionResult returns empty and the part is left unchanged — the next transform turn re-evaluates naturally, so no explicit wait/poll is needed. The transform hook only mutates the in-memory output.messages (not the persisted DB), so the rewrite is per-turn but the real output eventually lands in history once the child session completes and opencode writes it itself.

Gated: only acts on completed task parts whose parsed result is empty. True completions with real text are preserved unchanged. Non-task tool parts are not touched.

Tests

New false-complete-fallback.test.ts (5 cases):

• Fills empty completed output with child session real assistant text
• Does not rewrite when completed output already has real text
• Does not rewrite when child session has no assistant text yet (still running)
• Idempotent across consecutive transforms
• Does not touch non-task tool parts

Full suite: 1719 pass, 0 fail. Typecheck: pass.

Relation to other work

PR/Issue	Scope	Relation
#863 (closed, #870 merged)	self-amplify; #870 added reconcile_task tool (operation layer)	This PR fixes the root cause (#870 only mitigated the operation layer)
#606 (OPEN, Fixes #595)	background task: prevent fallback from aborting background task child sessions (preventive)	Complementary, not conflicting — #606 is preventive for background tasks; this PR is corrective for foreground tasks. File-level changes don't overlap (#606: foreground-fallback + event/options; this: board-injection + transform/import)
#595 (open)	fallback detached from lifecycle	Same family; #606 targets the background-task sub-case, this targets the foreground-task false-complete sub-case
#880 (my OPEN PR)	false-cancel: error+cancelled parts rewritten from board record	Mirror counterpart — #880 handles false-cancel, this handles false-complete. Shared render template shape

AI assistance

Co-developed-with:

• oracle (GLM-5.2) — read full omos + opencode source, confirmed precise root cause chain (runTask ?? "" on halt-empty assistant), recommended this approach over idle-reconcile竞态 alternatives

Note: the analysis council (councillor-reviewer-a/b/c) could not access the omos source tree from their workspace, so their input was based on the root-cause summary rather than direct source verification. The oracle's analysis is the source-verified basis for this fix.

[greptile-apps]

Greptile Summary

This PR repairs false-completed foreground task results using output recovered from the child session.

• Adds a transform-time reconciliation pass for completed task parts with empty results.
• Adds an XML-like renderer for recovered task output.
• Adds tests covering recovery, preservation, delayed output, idempotence, and non-task tools.

Confidence Score: 2/5

The PR should not merge until child-session lookup failures are contained and recovered text cannot prematurely terminate the task-result wrapper.

The new transform path allows session read errors to abort valid parent turns, while its renderer permits literal closing tags in assistant output to truncate the recovered task result.

Files Needing Attention: src/hooks/task-session-manager/board-injection.ts, src/hooks/task-session-manager/index.ts, src/utils/task.ts

Important Files Changed

Filename	Overview
src/hooks/task-session-manager/board-injection.ts	Adds false-completion reconciliation, but an unhandled child-session read failure can abort the parent messages transform.
src/hooks/task-session-manager/index.ts	Integrates the asynchronous reconciliation pass into every managed messages transform without an error boundary.
src/utils/task.ts	Adds a completed-task renderer whose unescaped assistant text can prematurely terminate the task-result wrapper.
src/hooks/task-session-manager/false-complete-fallback.test.ts	Covers normal reconciliation behavior but not child-session request failures or recovered text containing task-result delimiters.

Sequence Diagram

sequenceDiagram
    participant Parent as Parent transform
    participant Board as Background job board
    participant Child as Child session
    Parent->>Board: Resolve empty completed task ID
    Board-->>Parent: Job metadata
    Parent->>Child: Read assistant messages
    alt Non-empty child output
        Child-->>Parent: Recovered result
        Parent->>Parent: Rewrite task output in memory
    else No output yet
        Child-->>Parent: Empty result
        Parent->>Parent: Preserve part for later turn
    else Request fails
        Child--xParent: Error propagates and aborts transform
    end

Loading

_{Reviews (1): Last reviewed commit: "fix(task-session): fill false-completed ..." | Re-trigger Greptile}

[greptile-apps]

Child lookup aborts transforms

When the child-session messages request fails because of a transient client error, unavailable service, or missing session, the uncaught exception propagates through the new reconciliation pass and aborts the parent message transform instead of preserving the empty result for a later retry.

[greptile-apps]

Embedded tags truncate recovered results

When recovered assistant text contains a literal </task_result>—for example in code or an audit report—this unescaped interpolation closes the wrapper early, and parseTaskResultFromOutput extracts only through that first closing tag, causing the remaining task result to be omitted.

[Jiajun0413]

Superseded by #913, which merges this false-complete fix (including fail-open / tag sanitize / board-running gate hardenings) with the false-cancel path from #880 into a single transform-path PR.

Please review: #913