Skip to content

[deep-report] push_repo_memory silently deletes repo-memory checkout while reporting false success #60773

Description

@github-actions

Description

The prior DeepReport cycle (discussion #60674, 2026-09-13 ~18:34Z) reproduced a data-loss bug in the push_repo_memory safe-output tool twice in one run: calling it (once with a malformed argument, once with a clean, documented invocation) deleted/staged-for-deletion the entire repo-memory working tree under /tmp/gh-aw/repo-memory/default/ (6+ tracked files), while the tool still returned {"result":"success", ...: "validation passed"}. Both times the loss was recovered manually via git restore --staged --worktree . against the last committed baseline.

This almost certainly explains a multi-cycle mystery flagged repeatedly in this workflow's own repo-memory: last_analysis_timestamp.md and sibling memory files had not been updated since roughly 2026-09-07/08 despite 6+ newer DeepReport briefings existing in the interim — this run independently confirmed the memory files were still stuck at the 2026-09-08 ~06:00Z cycle as of 2026-09-14, i.e. the bug (or a variant of it) has silently discarded every intended memory update for about 6 days across multiple runs.

Per the Lockfile Statistics report, 36 of 299 workflows use push_repo_memory, so any workflow relying on it for cross-run continuity is at risk of silently losing state while its own logs claim success.

Expected Impact

Fixing this restores cross-cycle continuity for every workflow that uses push_repo_memory for repo-memory (currently ~36 workflows), and specifically unblocks this DeepReport/deep-report workflow's own trend-tracking, which has been operating on stale data for about a week without any error signal to indicate it.

Suggested Agent

New agent, or whoever owns the safe-outputs tool implementation (push_repo_memory handler) — likely in pkg/workflow or the actions/setup JS handlers for repo-memory.

Estimated Effort

Medium (1-4 hours) — needs root-causing why validation/staging deletes tracked files instead of just checking size, plus a regression test that asserts existing repo-memory files survive a push_repo_memory call.

Data Source

DeepReport Intelligence Briefing analysis run, 2026-09-14. Root cause and reproduction steps captured verbatim from the prior cycle's discussion #60674 (not filed at the time because the dedup-search gate was down); re-verified this cycle that no existing open issue covers it and that the memory-staleness symptom is still present today.

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

  • api.anthropic.com

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "api.anthropic.com"

See Network Configuration for more information.

Generated by 🔬 Deep Report · claude · agent · 370 AIC · ⌖ 10.9 AIC · ⊞ 12.8K · ◷

  • expires on Sep 15, 2026, 11:05 PM UTC-08:00

Activity

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

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions