Fix the Copilot CLI post-result termination — the agent keeps issuing bash commands after producing its result, the 20s inactivity watchdog kills it with exit 1, and all 4 retries fail identically, so the daily reports never post.
Two scheduled report workflows fully failed in the last 6h with zero safe-output items emitted. Firewall diff shows no network anomalies and the harness explicitly rules out CAPI/model errors — this is a Copilot CLI (v1.0.70) clean-shutdown defect, distinct from #44298 (hang-on-exit) and #46326 (CAPI 400 model).
Affected workflows and runs
| Workflow |
Run |
Failed step |
Result |
| Daily Security Observability Report |
§29651417530 (representative) |
Execute GitHub Copilot CLI |
exit 1, 30.2m, 119.2 AIC, 0 safe outputs |
| Daily Issues Report Generator |
§29648548749 (co-occurring, same engine/step) |
Execute GitHub Copilot CLI |
exit 1 |
Comparator (nearest prior run, same workflow): §29595547699.
Probable root cause
After the agent produced its work (uploaded chart assets via upload_asset), the model kept issuing exploratory bash commands (agenticworkflows --help, agenticworkflows logs --firewall ...) instead of terminating. The copilot-harness arms a post-result watchdog (inactivityTimeout=20000ms); the session never signals clean completion and the process exits 1, classified failureClass=partial_execution with hasOutput=true. The harness retries the entire run 4×; every attempt reproduces the same partial_execution exit, retries are exhausted, and no final report safe-output is emitted.
Evidence (agent-stdio.log, run 29651417530)
[copilot-harness] attempt 1 failed: exitCode=1 failureClass=partial_execution isCAPIError400=false isModelNotSupportedError=false isHTTP400ResponseError=false isSDKSessionIdleTimeoutError=false hasOutput=true retriesRemaining=3
[copilot-harness] attempt 1: post-result watchdog armed inactivityTimeout=20000ms
... (attempts 2,3,4 identical) ...
[copilot-harness] attempt 4 failed: exitCode=1 failureClass=partial_execution ... retriesRemaining=0
[copilot-harness] all 3 retries exhausted — giving up (exitCode=1)
[copilot-harness] done: exitCode=1 totalDuration=11m 48s
Process exiting with code: 1
audit safe_output_summary: total_items: 0 (no report posted)audit session_analysis: timeout_detected: falseaudit-diff vs 29595547699 firewall: has_anomalies: false, 0 new/removed/status-changed domains → not a network/model regression- Last agent action before death:
agenticworkflows logs --firewall true --start_date "-7d" --count 100
Proposed remediation
- Treat
partial_execution + hasOutput=true where safe-output items were already produced as a success-with-late-activity, not a retryable failure — finalize the run on the emitted outputs instead of retrying the whole session.
- When the post-result inactivity watchdog fires, terminate the copilot SDK session gracefully with exit 0 if valid safe outputs exist, rather than propagating exit 1.
- Tighten the prompt/stop-conditions for these report workflows so the agent emits its final safe-output report and stops, rather than running additional
agenticworkflows inspection commands after its deliverable is ready.
Success criteria
- Daily Security Observability Report and Daily Issues Report Generator complete with
conclusion: success and emit their report safe-output (>=1 item).
- No
all retries exhausted — giving up on runs that already produced safe outputs.
- AIC per run returns to baseline (the 119 AIC here reflects 4 full re-runs).
Parent: #46171
Related to #46171
Generated by 🔍 [aw] Failure Investigator (6h) · 250.5 AIC · ⌖ 39.3 AIC · ⊞ 5.2K · ◷
Fix the Copilot CLI post-result termination — the agent keeps issuing bash commands after producing its result, the 20s inactivity watchdog kills it with exit 1, and all 4 retries fail identically, so the daily reports never post.
Two scheduled report workflows fully failed in the last 6h with zero safe-output items emitted. Firewall diff shows no network anomalies and the harness explicitly rules out CAPI/model errors — this is a Copilot CLI (v1.0.70) clean-shutdown defect, distinct from #44298 (hang-on-exit) and #46326 (CAPI 400 model).
Affected workflows and runs
Comparator (nearest prior run, same workflow): §29595547699.
Probable root cause
After the agent produced its work (uploaded chart assets via
upload_asset), the model kept issuing exploratory bash commands (agenticworkflows --help,agenticworkflows logs --firewall ...) instead of terminating. The copilot-harness arms apost-result watchdog (inactivityTimeout=20000ms); the session never signals clean completion and the process exits 1, classifiedfailureClass=partial_executionwithhasOutput=true. The harness retries the entire run 4×; every attempt reproduces the same partial_execution exit, retries are exhausted, and no final report safe-output is emitted.Evidence (agent-stdio.log, run 29651417530)
auditsafe_output_summary:total_items: 0(no report posted)auditsession_analysis:timeout_detected: falseaudit-diffvs 29595547699 firewall:has_anomalies: false, 0 new/removed/status-changed domains → not a network/model regressionagenticworkflows logs --firewall true --start_date "-7d" --count 100Proposed remediation
partial_execution+hasOutput=truewhere safe-output items were already produced as a success-with-late-activity, not a retryable failure — finalize the run on the emitted outputs instead of retrying the whole session.agenticworkflowsinspection commands after its deliverable is ready.Success criteria
conclusion: successand emit their report safe-output (>=1 item).all retries exhausted — giving upon runs that already produced safe outputs.Parent: #46171
Related to #46171