Skip to content

[bug] Shell output cap drops final summaries after 64 MiB #247

Closed
#253
Closed
[bug] Shell output cap drops final summaries after 64 MiB#247
#253

Description

@AetherAI3
AetherAI3
opened on Oct 1, 2026

Problem

The runner's absorber stops retaining every chunk after cumulative output exceeds 64 MiB. At close it calls capHeadTail on that retained prefix. The apparent “tail” is therefore not the actual final output. A verbose build/test can lose its final failure summary while preserving the process exit code.

Verification

Isolated reproduction of the exact absorber logic: a 64 MiB ASCII chunk followed by FINAL SUMMARY: 1 failed retains 67,108,864 bytes and loses the summary. This is an algorithm reproduction, not a full repository test run.

Fix / acceptance

  • Maintain bounded head and rolling tail buffers while continuing to drain both pipes; preserve the actual last bytes regardless of total output.
  • Decode UTF-8 incrementally so split multibyte characters remain valid. Report omitted bytes/truncation explicitly without inventing a summary.
  • Test a final marker beyond the cap, a single oversized chunk, interleaved stderr/stdout, split UTF-8, cancellation and timeout. The real final marker must survive and memory retention must stay bounded.
  • Preserve the authoritative exit status and cancellation/timeout distinctions.

Priority: P1. Independent of the new console shell features.

Activity

  1. AetherAI3
    closed this as completedin #253on Oct 1, 2026

  2. AetherAI3 commented on Oct 1, 2026

    @AetherAI3
    AetherAI3
    on Oct 1, 2026 – with ChatGPT Codex Connector
    OwnerAuthor

    Re-reviewed the original acceptance requirements and completed the fix in merged PR #253 (e4845eb323038cddcbf68a860a885823925b487c).

    • Fixed bounded head/rolling-tail buffers continue draining both pipes and preserve the actual final output after 64 MiB. The same fix covers persistent user/model shell execution and explicit /shell-result sharing.
    • Independent incremental UTF-8 decoders and code-point-safe elision; omitted decoded UTF-8 bytes are reported explicitly. No summary is fabricated.
    • Real child regressions stream 66 MiB and preserve initial + final stdout/stderr markers and exit 7. Direct oversized append, rolling-memory bounds, interleaved/split Unicode, post-cap cancellation 130 and timeout 124, reset/refusal, and next-command isolation are covered.
    • Original seven integration regressions failed before the fix and passed afterward. Independent code reviews found no blockers; 5,000 randomized Unicode capture cases also passed.

    Final local evidence: npm test exited 0; the complete bounded full-suite report confirms 2,930 tests, 2,919 passed, 11 skipped, 0 failed, 0 cancelled. Typecheck, generated docs, production package verification, and four independent Python contract verifiers passed. Two stale fixed-delay console routing fixtures were replaced with observed completion checks while retaining all original assertions.

    Acceptance mapping: docs/issue-247-validation.md. Merged source tree matches the locally tested tree 2ac4ecfbebd4a2a8aaca8b5bab7210bae56e2b95.

    Hosted limitation: CI and CodeQL jobs were rejected before execution because GitHub reports an account billing lock. These are not passing hosted checks; Windows CI remains unverified. No workflow/protection bypass was added.

    All #247 functional acceptance items are satisfied; closing as completed by the merged fix.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions