You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
[deep-report] Actually land the DeepReport repo-memory path fix from #51172 (closed without merged change; still 2-level path, still 1 commit
[Content truncated due to length] #51809
This is a reopening with a concrete fix location, not a rediscovery. #51172 diagnosed the exact bug (slashless file-glob: ["*.md"] compiles to matchSubfolderRoot, which only matches paths exactly one directory deep) and was closed 2026-08-07 — but the GitHub API timeline shows it was closed directly with no linked merged PR (closed event, source: null), and the source file was never actually changed.
Confirmed still broken today (2026-08-10)
.github/workflows/deep-report.md (current main) still tells the agent to read/write at:
Line 165: `/tmp/gh-aw/repo-memory/default/memory/deep-report/` (read)
Line 217: memory/deep-report/processed-discussions.json (read)
Line 247: `/tmp/gh-aw/repo-memory/default/memory/deep-report/` (write)
This is two directory levels under the repo-memory root (memory/ + deep-report/), which the slashless *.md glob (file-glob: ["*.md"] at line 76) does not match — it only matches one level (<dir>/<file>.md).
The memory/deep-report git branch (where this repo-memory is actually stored) has exactly one commit, e50c816, dated 2026-04-03, "Update repo memory from workflow run 23950425855" — despite this workflow running daily since then and multiple cycles (including 2026-08-07s DeepReport Intelligence Briefing - 2026-08-07 #51116) explicitly claiming to have "rewritten all four memory files with fresh data."
Fix
In .github/workflows/deep-report.md:
Change the read/write paths at lines 165, 217, 247 (and the mining-step reference at 227) from /tmp/gh-aw/repo-memory/default/memory/deep-report/ to /tmp/gh-aw/repo-memory/default/deep-report/ (drop the redundant memory/ prefix — the git branch is already named memory/deep-report via the repo-memory.branch-name frontmatter at line 74; the local folder path doesnt need to repeat it).
Add "*.json" to file-glob at line 76: file-glob: ["*.md", "*.json"].
Recompile: gh aw compile (or equivalent) to regenerate deep-report.lock.yml.
Verify next cycle: confirm the memory/deep-report branch gets a new commit dated after this fix lands — that is the only reliable signal this is actually fixed (see companion process issue on requiring verified evidence before closing this class of bug).
Data source
DeepReport Intelligence Briefing, 2026-08-10. Confirmed via direct read of .github/workflows/deep-report.md on main, git log on the checked-out memory/deep-report branch (1 commit total), and gh api repos/github/gh-aw/issues/51172/timeline (closed with no source PR).
This is a reopening with a concrete fix location, not a rediscovery. #51172 diagnosed the exact bug (slashless
file-glob: ["*.md"]compiles tomatchSubfolderRoot, which only matches paths exactly one directory deep) and was closed 2026-08-07 — but the GitHub API timeline shows it was closed directly with no linked merged PR (closedevent,source: null), and the source file was never actually changed.Confirmed still broken today (2026-08-10)
.github/workflows/deep-report.md(currentmain) still tells the agent to read/write at:`/tmp/gh-aw/repo-memory/default/memory/deep-report/`(read)memory/deep-report/processed-discussions.json(read)`/tmp/gh-aw/repo-memory/default/memory/deep-report/`(write)memory/+deep-report/), which the slashless*.mdglob (file-glob: ["*.md"]at line 76) does not match — it only matches one level (<dir>/<file>.md).file-globis still["*.md"]only — Fix DeepReport repo-memory persistence: depth-2 paths silently dropped by file glob #51172 said it would add"*.json"(needed forprocessed-discussions.json/extracted-tasks.json, referenced at lines 217/227) but that change is also not present.memory/deep-reportgit branch (where this repo-memory is actually stored) has exactly one commit,e50c816, dated 2026-04-03, "Update repo memory from workflow run 23950425855" — despite this workflow running daily since then and multiple cycles (including 2026-08-07s DeepReport Intelligence Briefing - 2026-08-07 #51116) explicitly claiming to have "rewritten all four memory files with fresh data."Fix
In
.github/workflows/deep-report.md:/tmp/gh-aw/repo-memory/default/memory/deep-report/to/tmp/gh-aw/repo-memory/default/deep-report/(drop the redundantmemory/prefix — the git branch is already namedmemory/deep-reportvia therepo-memory.branch-namefrontmatter at line 74; the local folder path doesnt need to repeat it)."*.json"tofile-globat line 76:file-glob: ["*.md", "*.json"].gh aw compile(or equivalent) to regeneratedeep-report.lock.yml.memory/deep-reportbranch gets a new commit dated after this fix lands — that is the only reliable signal this is actually fixed (see companion process issue on requiring verified evidence before closing this class of bug).Data source
DeepReport Intelligence Briefing, 2026-08-10. Confirmed via direct read of
.github/workflows/deep-report.mdonmain,git logon the checked-outmemory/deep-reportbranch (1 commit total), andgh api repos/github/gh-aw/issues/51172/timeline(closed with no source PR).Suggested Agent:
deep-reportmaintainer / general code-quality agentEstimated Effort: Quick (< 1 hour) — 3-line frontmatter/prompt change + recompile