Description
The push_repo_memory job in agentic workflows (confirmed on Deep Report, .github/workflows/deep-report.lock.yml) silently treats a missing repo-memory-<id> artifact as "nothing to push" and exits success, instead of surfacing it as a warning/failure. Live job logs from run 36612476935 (job push_repo_memory, id 109557111748→109563491196, 2026-09-29T18:31Z, overall run conclusion: success) show:
##[error]Unable to download artifact(s): Artifact not found for name: repo-memory-default
...
Memory directory not found in artifact: /tmp/gh-aw/repo-memory/default
The step then completes in under 1 second with conclusion: success. Because of this, the memory/deep-report branch's last real commit is from workflow run 34195380008 (~2026-09-08), even though dozens of subsequent Deep Report runs — including many with overall conclusion: success — have executed since then. The repo-memory store used by Deep Report (/tmp/gh-aw/repo-memory/default/deep-report/*.md) has effectively been frozen for ~3-4 weeks while still being treated by downstream cycles as current.
Root cause is likely upstream: the agent job isn't always uploading a repo-memory-<id> artifact (e.g., when the agent decides there's nothing new to persist, or when the upload step itself is skipped/fails without failing the job), and push_repo_memory doesn't distinguish "artifact legitimately absent, no-op expected" from "artifact missing unexpectedly, memory continuity broken."
Expected Impact
Restores reliable long-term memory continuity for push_repo_memory-based workflows (confirmed affected: Deep Report; likely others using the same safe-output). Prevents agents from silently re-deriving context from scratch every cycle while believing memory is current, and gives maintainers a visible signal (warning annotation or job-level notice) when memory writes aren't landing.
Suggested Agent
Existing: whichever agent/maintainer owns the push_repo_memory safe-output implementation (pkg/workflow / the corresponding .cjs action script referenced in logs as push_repo_memory.cjs). Note: Lockfile Stats discussion #64836 (2026-10-01) observed push_repo_memory usage fleet-wide fell 40→35 workflows in 2 days, with new ledger_append/ledger_request_compaction safe-output types appearing — there may already be an in-progress migration to a ledger-based mechanism that resolves this; worth checking before investing in a push_repo_memory-specific fix.
Estimated Effort
Medium (1-4 hours) — needs a decision on whether to (a) fail/warn loudly when the artifact is unexpectedly missing, (b) have the agent job always upload an artifact (even empty) when repo-memory was requested, or (c) fold this into the ledger migration.
Data Source
DeepReport Intelligence Briefing analysis run 36948800971, 2026-10-02. Live-verified against GitHub Actions job logs (run 36612476935) and the memory/deep-report branch's own git history, not just discussion text.
Generated by 🔬 Deep Report · claude · agent · 283.5 AIC · ⌖ 7.58 AIC · ⊞ 7.3K · ◷
Description
The
push_repo_memoryjob in agentic workflows (confirmed on Deep Report,.github/workflows/deep-report.lock.yml) silently treats a missingrepo-memory-<id>artifact as "nothing to push" and exitssuccess, instead of surfacing it as a warning/failure. Live job logs from run 36612476935 (jobpush_repo_memory, id 109557111748→109563491196, 2026-09-29T18:31Z, overall run conclusion: success) show:The step then completes in under 1 second with
conclusion: success. Because of this, thememory/deep-reportbranch's last real commit is from workflow run34195380008(~2026-09-08), even though dozens of subsequent Deep Report runs — including many with overallconclusion: success— have executed since then. The repo-memory store used by Deep Report (/tmp/gh-aw/repo-memory/default/deep-report/*.md) has effectively been frozen for ~3-4 weeks while still being treated by downstream cycles as current.Root cause is likely upstream: the
agentjob isn't always uploading arepo-memory-<id>artifact (e.g., when the agent decides there's nothing new to persist, or when the upload step itself is skipped/fails without failing the job), andpush_repo_memorydoesn't distinguish "artifact legitimately absent, no-op expected" from "artifact missing unexpectedly, memory continuity broken."Expected Impact
Restores reliable long-term memory continuity for
push_repo_memory-based workflows (confirmed affected: Deep Report; likely others using the same safe-output). Prevents agents from silently re-deriving context from scratch every cycle while believing memory is current, and gives maintainers a visible signal (warning annotation or job-level notice) when memory writes aren't landing.Suggested Agent
Existing: whichever agent/maintainer owns the
push_repo_memorysafe-output implementation (pkg/workflow/ the corresponding.cjsaction script referenced in logs aspush_repo_memory.cjs). Note: Lockfile Stats discussion #64836 (2026-10-01) observedpush_repo_memoryusage fleet-wide fell 40→35 workflows in 2 days, with newledger_append/ledger_request_compactionsafe-output types appearing — there may already be an in-progress migration to a ledger-based mechanism that resolves this; worth checking before investing in apush_repo_memory-specific fix.Estimated Effort
Medium (1-4 hours) — needs a decision on whether to (a) fail/warn loudly when the artifact is unexpectedly missing, (b) have the agent job always upload an artifact (even empty) when repo-memory was requested, or (c) fold this into the ledger migration.
Data Source
DeepReport Intelligence Briefing analysis run 36948800971, 2026-10-02. Live-verified against GitHub Actions job logs (run 36612476935) and the
memory/deep-reportbranch's own git history, not just discussion text.