[copilot-session-insights] Daily Copilot Agent Session Analysis — 2026-05-25

[github-actions[bot]]

🤖 Copilot Agent Session Analysis — 2026-05-25

Executive Summary

• Sessions Analyzed: 50
• Analysis Period: 2026-05-25 (05:33–05:43 UTC, a 10-minute burst)
• Completion Rate: 0.0% (0 successful, 2 failures, 48 action_required)
• Average Duration: 0.31 min (median 0s, max 8m09s)
• Experimental Strategy: None this run (standard analysis only)
• Data Quality: ⚠️ Metadata-only — conversation logs fetch failed for the 2nd consecutive day (OAuth token missing)

Today is the worst day in the 6-day observation window: zero successful completions, 96% of sessions ended in action_required, and the only two sessions that did real work both crashed inside Running Copilot cloud agent. The 2026-05-23 recovery is now confirmed as an outlier — completion has trended down for two consecutive days.

Key Metrics

Metric	Today (05-25)	Yesterday (05-24)	Trend
Total Sessions	50	50	→
Successful Completions	0 (0%)	1 (2%)	↓
Failed Sessions	2 (4%)	0 (0%)	↑
Action-required / abandoned	48 (96%)	49 (98%)	→
Average Duration	0.31 min	0.15 min	↑ (driven by 2 failed cloud-agent runs)
Median Duration	0.00 min	0.00 min	→
Loop Detection (≥20 min)	0	0	→
Unique Branches	4	5	↓

Trend Charts

Completion Patterns

The chart confirms the recovery-regression oscillation has resolved into a downward slide: after the 44% peak on 05-23, completion fell to 2% on 05-24 and 0% today. Action-required volume is staying pinned near the top of the chart (~48 of 50 sessions each day), with today's 2 failures the only non-gating signal.

Duration & Efficiency

Average duration ticked up from 0.15 min → 0.31 min, but the increase is misleading: it is entirely driven by two failed Running Copilot cloud agent runs (7m23s and 8m09s). Median remained at 0s and no sessions exceeded 20 minutes, so there is no productive iteration today.

Success Factors ✅

With zero successful completions today, no success patterns can be derived from this run. Drawing from the 6-day history:

1. Long-running sessions correlate with completion: On 2026-05-23, 9 sessions exceeded 20 minutes and 44% of all sessions completed successfully. Today's median of 0s and zero long sessions explain the 0% completion.
2. Running Copilot cloud agent is the work-doing workflow: The only sessions that ran more than 1 second were Running Copilot cloud agent invocations. Other workflows (Doc Build, CGO, Smoke CI, CJS, Agentic Commands, Q) all gated on action_required in <1s.

Failure Signals ⚠️

1. action_required saturation (96%): 48 of 50 sessions terminated immediately with action_required and zero duration. This is the same activation/permission-gating profile seen on 2026-05-20, 05-22, and 05-24. Workflows are firing but never reaching the agent.

   • Affected workflows: Agentic Commands (14), Q (13), Smoke CI (6), Doc Build (4), CGO (4), AI Moderator (3), CJS (2), Content Moderation (2).

2. Running Copilot cloud agent 100% failure rate today: Both cloud-agent invocations failed.

   • PR Update model alias and multiplier inventories for 2026-05-25 #34585 copilot/model-inventory-update — 7m23s, conclusion failure
   • PR Break logger↔timeutil import cycle causing CGO/fuzz workflow failures #34584 copilot/cgo-fuzz-workflow-failure — 8m09s, conclusion failure
   • These are the only sessions where the agent did meaningful work, and both crashed. Distinct from the gating pattern — these are real agent failures.

3. Concentrated branch activity, zero throughput: 4 branches account for all 50 sessions — copilot/model-inventory-update (18), copilot/cgo-fuzz-workflow-failure (17), copilot/aw-sg18a1-fix-graphql-response-assertion (11), copilot/migrate-gemini-to-gravity-cli (4). Heavy retry/iteration with no closure — same pattern as 05-23 (27/50 on one branch) and 05-24 (20/50 on one branch).

Prompt Quality Analysis 📝

⚠️ Prompt-level analysis is unavailable today because conversation logs could not be fetched (OAuth token missing) for the second consecutive day. Only infrastructure metadata is available, so the agent's prompts and reasoning cannot be inspected. Logging this as a recurring data-quality risk (see conversation_log_fetch_failure pattern in cache memory).

Orphaned Branch Escalation Alerts 🚨

Summary

• Orphaned Branches Today: 0 out of 13 active branches (0%)
• Historical Baseline: ~40% orphaned rate
• Status: ✅ NORMAL (well below baseline)

Escalation Candidates

✅ No orphaned branches exceed the escalation threshold today.

Detection note: Only 2 in-progress workflow runs were active during the analysis window, both on main (Outcome Collector and Copilot Session Insights itself). No PR branch carried ≥5 simultaneous gate firings, so no escalation candidates surfaced. This mirrors 2026-05-24 — gate activity is centred on main, not on Copilot PR branches.

CI Waste Estimate

• Orphaned gate-hours today: 0
• Recoverable capacity: N/A — no orphans detected

Notable Observations

Loop Detection

• Sessions ≥20 min: 0
• Average loop count: 0
• Comment: No iteration loops today. The 2 cloud-agent failures lasted 7–8 minutes each, below the loop-proxy threshold.

Tool Usage

• Conversation logs unavailable → no tool-usage analysis possible
• Workflow distribution (proxy for tool surface): Agentic Commands (14, 28%), Q (13, 26%), Smoke CI (6, 12%), CGO (4), Doc Build (4), AI Moderator (3), CJS (2), Content Moderation (2), Running Copilot cloud agent (2)

Context Issues

• Cannot be measured without conversation logs
• Inference from outcomes: 96% action_required suggests the agent never reaches a context-loading step — gating happens upstream

Experimental Analysis

This run was not experimental. Standard analysis only.

Rationale: Two consecutive days of metadata-only data quality (log fetch failure) makes experimental strategies low-value — most novel analyses (semantic clustering, prompt analysis, code quality from logs) require the conversation transcripts that are currently missing. Recommend deferring experimental runs until log fetch is restored.

Actionable Recommendations

For Users Writing Task Descriptions

Cannot be derived from today's data (no logs). General guidance from prior runs remains in effect.

For System Improvements

1. Investigate Running Copilot cloud agent failures on model-inventory-update and cgo-fuzz-workflow-failure (High impact): Both cloud-agent invocations failed today after 7–8 minutes of work. These are the only sessions doing real work — failures here are direct lost productivity, not gating noise. Worth inspecting the run logs for the two specific run IDs (26385103314, 26385093013) to identify the failure mode.

2. Restore OAuth token for copilot-session-data-fetch (High impact): Conversation log retrieval has now failed on two consecutive days (05-24, 05-25). This blocks behavioral analysis, prompt quality assessment, tool usage tracking, and most experimental strategies. The workflow remains operational but is producing infrastructure-only reports. Treat as a recurring workflow risk in cache memory.

3. Audit action_required gating root cause (Medium impact): For 4 of the last 6 days, ≥86% of sessions terminate with zero duration in action_required. This is consistent with permission/activation gating happening before the agent runs. Worth confirming whether this is intentional (e.g., manual approval gates on certain workflows for copilot/* branches) or an unintended configuration regression.

For Tool Development

• Conversation log retrieval reliability: A retry/fallback mechanism for the log fetch step would prevent the metadata-only data-quality degradation seen on 05-24 and 05-25.

Trends Over Time

6-day rolling history (from cache memory)

Date	Sessions	Success%	Avg dur (min)	≥20 min	action_required%	Notes
2026-05-20	50	0%	0.009	0	92%	Activation gating
2026-05-21	50	12%	1.526	0	86%	First partial recovery
2026-05-22	50	2%	0.364	0	98%	Regression to gating
2026-05-23	50	44%	8.543	9	28%	Outlier recovery — 27/50 on one branch
2026-05-24	50	2%	0.153	0	98%	Logs fetch failed (OAuth)
2026-05-25	50	0%	0.311	0	96%	Logs fetch failed (OAuth) + 2 cloud-agent failures

• Completion rate: Trending down (44% → 2% → 0%) over the last 3 days. The 05-23 peak is now confirmed as an outlier rather than a sustained turn.
• Average duration: Stuck near 0 min for 4 of the last 6 days; today's small uptick is from failed (not productive) cloud-agent runs.
• Quality improvement: None observable in current data.

Statistical Summary

Total Sessions Analyzed:     50
Successful Completions:      0 (0.0%)
Failed Sessions:             2 (4.0%)
Action-required:            48 (96.0%)

Average Session Duration:   0.311 min (18.6s)
Median Session Duration:    0.000 min
Longest Session:            8.150 min (Running Copilot cloud agent, failure)
Shortest Session:           0.000 min

Loop Detection (≥20 min):   0 sessions
Sessions w/ non-zero dur:   2 sessions (both failures)

Unique Branches:            4
Top Branch:                 copilot/model-inventory-update (18/50 = 36%)

Next Steps

• Restore OAuth token for conversation log fetch (highest priority — 2 consecutive days lost)
• Investigate the two Running Copilot cloud agent failures: runs §26385103314 and §26385093013
• Confirm whether the persistent action_required saturation is intentional gating or unintended
• Re-evaluate the recovery_regression_oscillation pattern after 3 more days of data — current trajectory suggests a sustained low-completion regime, not oscillation
• Resume experimental strategies once conversation logs are available again

---

References:

• §26390330462 — this analysis run
• §26385103314 — cloud-agent failure on copilot/model-inventory-update
• §26385093013 — cloud-agent failure on copilot/cgo-fuzz-workflow-failure

Generated by 📊 Copilot Session Insights · opus47 13.6M · ◷

• expires on May 26, 2026, 8:30 AM UTC

[github-actions[bot]]

This discussion has been marked as outdated by Copilot Session Insights.

A newer discussion is available at Discussion #34901.