[copilot-session-insights] Daily Copilot Agent Session Analysis — 2026-09-27

[github-actions[bot]]

🤖 Copilot Agent Session Analysis — 2026-09-27

Executive Summary

• Sessions Analyzed: 50 (all workflow-run entries in the fixed trailing window; metadata-only, see data-gap note below)
• Analysis Period: 2026-09-26 23:20 UTC → 2026-09-27 05:28 UTC
• Completion Rate: 48.0%
• Average Duration: 4.46 min (raw, session-level); 8.26 min mean / 5.93 min median across the 27 non-zero entries
• Experimental Strategy: None (roll=95/100, standard run)

⚠️ Data-gap notice: Conversation transcript logs ({run_id}-conversation.txt) were completely absent for the 35th consecutive recorded day — find .../logs -name "*-conversation.txt" returned 0 files against 50 listed sessions. This blocks turn-by-turn tool-usage, token-efficiency, loop-detection, and prompt-quality analysis (the "true behavioral analysis" this workflow is designed to produce). Everything below is derived from run metadata only (timestamps, conclusion, workflow name, branch) plus live gh api queries against PRs and Actions runs. Flagged via missing_data.

Key Metrics

Metric	Value	Trend
Total Sessions	50	→
Successful Completions	24 (48.0%)	↑ (+12.0pts vs 09-26's 36.0%)
Failed / Action-Required	26 (52.0%)	↓
Average Duration	4.46 min	↑ (vs 3.06 min 09-26; no cascade/outlier inflation today)
Loop Detection Rate	n/a — no conversation logs to detect loops	—
Context Issues	n/a — no conversation logs	—
Orphaned-Branch Rate	0.0% (0/17 open PRs)	→ (35th consecutive 0.0% day)

Success Factors ✅

1. True-agentic "Addressing comment on PR" runs: 7/7 succeeded (100%) today, restarting the recovery streak at day 1 after 09-26 broke a 5-day streak (7/8 = 87.5%, one cancelled by its own branch merging mid-run).
2. Code scanning AI findings runs: 8/8 succeeded (100%) — this bot-driven gate remains the most reliable single workflow in the window, consistent with 30+ days of history.
3. Isolated (non-burst) sessions outperform burst-fired ones: 66.7% (4/6) vs 45.5% (20/44), a 1.47x gap — narrower than the historical 3–8x range but the same direction seen on nearly every prior recorded day, reinforcing burst-clustering as a cheap proxy for "gate-stub noise" vs a real task push.
4. No cascade-inflated duration today: longest genuine run was 23.9 min (a successful "Addressing comment" run); no >100-min status-resync artifact, unlike several recent days.

Failure Signals ⚠️

1. branch_level_stuck_gate (persistent): Squad 0/6, Agentic Commands 0/5, Squad Implement Worker 0/2, Doc Build - Deploy 0/2 — all 0% success — vs. Code scanning (8/8) and Addressing comment (7/7) at 100%. Same bimodal split observed for weeks; these gate workflows appear structurally blocked rather than task-specific failures.
2. merge_invalidation_cascade ×2 independent events: (a) PR Track skill usage in audit artifacts #63688 merged 23:41:44Z → CWI/CGO/CJS all flipped to failure at 23:41:45Z (1s later, 3 runs invalidated); (b) PR Report MCP payload sizes in audit and aggregate analysis #63687 merged 00:34:21Z → Squad Implement Worker went action_required at 00:34:27Z (6s later, 1 run invalidated). Event (b) is the 4th confirmed instance of the "merge → Squad Implement Worker +6–7s" sub-signature (first seen 09-19/09-24, recurred 3× on 09-26 alone) — this now looks like a routine per-merge side effect, not a rare edge case.
3. provenance_inversion, sub-band for the 3rd time in 5 days: only 70.8% of today's successes (17/24) were bot-driven (Code scanning + CGO/CWI/CJS), just below the historical 72–86% band. Combined with 09-23 (58.3%) and 09-26 (61.1%), 3 of the last 5 days have now dipped below the band — worth watching as a possible band shift rather than a one-off.
4. Branch-noise concentration: copilot/dynamic-checkouts-github-action produced 20/50 (40%) of all sessions but only 20.0% success — the heaviest gate-noise branch in today's window, tied to PR Support dynamic checkout sets from GitHub Actions expressions #63241 which has been open since 09-24.

Prompt Quality Analysis 📝

Per-Prompt Breakdown

No prompt text or task-description content is available in run metadata — session-list entries expose only workflow name, branch, timestamps, and conclusion. Meaningful prompt-quality scoring (high/medium/low characteristics, example prompts) requires the conversation transcripts described in the data-gap notice above and cannot be produced from this window's inputs. This has now been true for 35 consecutive recorded days.

Orphaned Branch Escalation Alerts 🚨

Branches with ≥5 simultaneous gate firings and no Copilot agent assigned for >2 hours.

Summary

• Orphaned Branches Today: 0 out of 17 open PRs (0.0%)
• Historical Baseline: 0.0% orphaned rate (30-day mean over 30 recorded days)
• Status: NORMAL (today's rate does not exceed baseline + 20pts)

Escalation Candidate Details

Escalation Candidates

✅ No orphaned branches exceed the escalation threshold today. Only 2 in-progress Actions runs existed repo-wide in the trailing 6 hours, both on main (this workflow + Tidy) — no branch approached the ≥5-gate threshold.

CI Waste Estimate

• Orphaned gate-hours today: 0 — no escalation candidates, so no orphaned CI capacity to recover.

Notable Observations

Loop Detection and Session Diagnostics

Loop Detection

• Sessions with loops: Not measurable — requires conversation transcripts, which were absent (35th consecutive day).

Tool Usage

• Not measurable from metadata-only input; no tool.execution_* log lines were available.

Branch / Workflow Concentration (metadata proxy for behavior)

• 4 unique branches active: copilot/dynamic-checkouts-github-action 20/50 (40%, 20.0% success), copilot/task-...87cd5693 15/50 (30%, 73.3% success, merged as PR Report MCP payload sizes in audit and aggregate analysis #63687), copilot/task-...ca3f0885 9/50 (18%, 55.6% success, merged as PR Track skill usage in audit artifacts #63688), copilot/preserve-review-provenance-marker 6/50 (12%, 66.7% success).
• Per-workflow success: Addressing comment on PR 7/7 (100%), Code scanning AI findings 8/8 (100%), CGO 3/6 (50%), CWI 3/6 (50%), CJS 3/8 (37.5%), Squad 0/6 (0%), Agentic Commands 0/5 (0%), Squad Implement Worker 0/2 (0%), Doc Build - Deploy 0/2 (0%).

Context Issues

• Not measurable from metadata alone.

Experimental Analysis

Standard run only (roll=95/100, threshold <30 for experimental) — no experimental strategy this run.

📈 Session Trends Analysis

Completion Patterns

Today's 48.0% completion rate is the 2nd-highest reading of the last two weeks (after 09-24's 44.0%) and sits well above the ~31% 30-day mean, continuing the sawtooth pattern this workflow has shown for over a month rather than any sustained trend in either direction. The failed/action-required series remains persistently high in absolute count even on good days, reflecting the branch-level stuck-gate pattern (Squad/Agentic Commands/Doc Build-Deploy) rather than genuine task failures.

Duration & Efficiency

Today's average duration (4.46 min) sits in the low range with no cascade- or status-resync-driven spike, unlike the 75.8-minute outlier around 09-20 and the 40-minute one around 09-16. The bar overlay for "sessions with loops" specified in the standard chart template was omitted here rather than fabricated, since loop detection requires conversation-log data that has been unavailable for 35 consecutive days.

Actionable Recommendations

For Users Writing Task Descriptions

No prompt text is visible in this data source, so prompt-level recommendations cannot be generated this run. General guidance from prior runs' findings still applies: reference specific files/paths and state expected outcomes explicitly.

For System Improvements

1. Restore conversation-log collection for the copilot-session-data-fetch module: Highest-impact fix. This is now a 35-consecutive-day-old gap that prevents the workflow from doing the turn-by-turn behavioral analysis it was built for (tool usage, token efficiency, loop detection, error recovery). Every other finding in this report is a metadata-only proxy.
2. Investigate the Squad / Agentic Commands / Doc Build - Deploy gate workflows: These have shown ~0% success for weeks straight while true-agentic and code-scanning workflows sit near 100% — worth confirming whether this is expected gating behavior (blocked pending review) or a genuine regression.
3. Track push-level (not just session-level) completion rate: The 09-20 push_event_reattribution experiment found push-clustered completion (66.7%) more than double the raw session rate (30.0%) — promoting this as a standing secondary metric would give a clearer picture than the noisy raw rate.

For Tool Development

1. Conversation-transcript availability check: A pre-flight check that fails fast (or explicitly flags via missing_data) when logs/ is empty would save analysis time — currently this is discovered fresh each day via manual find.

Historical Trends and Statistical Summary

Trends Over Time

• Completion rate trend: Sawtooth pattern continues — no sustained upward or downward trend over the last 30 days (range 4–78%, today 48.0% vs 31.7% 30-day mean).
• Average duration trend: No outlier today; duration outliers over the last 30 days have consistently traced back to merge/close cascades or status-resync artifacts rather than genuinely long task sessions.
• Data availability: Conversation logs empty for the 35th consecutive recorded day — the single largest standing gap in this analysis pipeline.

Statistical Summary

Total Sessions Analyzed:     50
Successful Completions:      24 (48.0%)
Failed Sessions:              3 (6.0%)
Action-Required (blocked):   23 (46.0%)

Average Session Duration:    4.46 min (raw) / 8.26 min (27 nonzero entries)
Median Session Duration:     3.68 min (raw) / 5.93 min (27 nonzero entries)
Longest Session:             23.92 min (Addressing comment on PR #63687, success)
Shortest Session:            0.0 min (23 entries, instant action_required gate stubs)

Loop Detection:              not measurable (0 conversation logs available)
Context Issues:              not measurable (0 conversation logs available)
Tool Failures:                not measurable (0 conversation logs available)

Prompt Quality Scoring:      not measurable (no prompt text in metadata)

Next Steps

• Review recommendations with team
• Prioritize restoring conversation-log collection (35-day-old gap)
• Investigate persistent 0% success on Squad / Agentic Commands / Doc Build - Deploy gates
• Consider promoting push-level completion rate as a secondary headline metric
• Schedule follow-up analysis tomorrow

---

Analysis generated automatically on 2026-09-27
Run ID: 36301452080
Workflow: Copilot Session Insights

Generated by 📊 Copilot Session Insights · claude · agent · 196.1 AIC · ⌖ 13.6 AIC · ⊞ 18.6K · ◷

• expires on Sep 27, 2026, 11:20 PM UTC-08:00

[github-actions[bot]]

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

A newer discussion is available at Discussion #63958.