[copilot-session-insights] Daily Copilot Agent Session Analysis — 2026-09-28

[github-actions[bot]]

🤖 Copilot Agent Session Analysis — 2026-09-28

Executive Summary

• Sessions Analyzed: 50
• Analysis Period: 2026-09-26 23:26 UTC → 2026-09-27 16:28 UTC (~17h window)
• Completion Rate: 50.0% (25/50)
• Average Duration: 3.84 min raw / 7.68 min (25 non-instant sessions)
• Experimental Strategy: none — standard run (roll=84 ≥ 30 threshold)

Key Metrics

Metric	Value	Trend
Total Sessions	50	→ (fixed window)
Successful Completions	25 (50.0%)	↑ +2.0pts vs 09-27 (48.0%)
Failed / Cancelled	0 (0%)	↓ down from 3 on 09-27
Action Required (gate-blocked)	25 (50.0%)	↓ down from 26 on 09-27
Average Duration	3.84 min raw (7.68 min excl. zero-duration)	→ roughly flat vs 4.46 min on 09-27
Loop Detection Rate	N/A — no transcript data	→ (35th+ day without logs)
Context Issues	N/A — no transcript data	→ (35th+ day without logs)

Today is the first fully binary day in the recent record: exactly 25 success + 25 action_required, with zero failure, cancelled, or in_progress outcomes — the prior four days (09-24 through 09-27) each had at least one non-binary outcome. This also makes today the 4th-highest completion rate of the last 33 recorded days (behind only 78.0% on 09-02, 58.0% on 09-18, and 52.0% on 09-15), and +18.1pts above the 32-day mean of 31.9% (range 4–78%).

Success Factors ✅

1. True-agentic "Addressing comment on PR" runs: 9/9 succeeded (100%).

   • Success rate: 100%
   • Continues the streak restarted on 09-27 (7/7) after the 5-day streak broke on 09-26 (87.5%, 1 cancelled by its own branch merging mid-run) — this is now day 2 of the new streak.

2. Isolated (single-fire) event workflows: "Addressing comment on PR" + "Code scanning AI findings on PR" = 17/17 succeeded (100%).

   • Success rate: 100%
   • These fire once per PR event rather than in repeated gate sweeps, and continue to outperform burst-fired gate workflows by a wide margin (see Burst vs. Isolated below).

3. Partial gate-workflow recovery: unlike most recent weeks' strict 0%-or-100% bimodal split, CJS (3/9=33.3%), CWI (3/6=50.0%), and CGO (2/5=40.0%) all posted genuine mixed success today rather than a clean 0%.

   • Success rate: 33–50%
   • Example: these three gates converted on the copilot/task-*-87cd5693 and copilot/preserve-review-provenance-marker branches specifically, while remaining at 0% on copilot/dynamic-checkouts-github-action.

4. Branch copilot/task-...-87cd5693: 11/12 sessions succeeded (91.7%) — the highest single-branch success rate seen in recent days, on a branch that fired a large (12-session) volume.

   • Success rate: 91.7%

5. Zero orphaned branches, zero escalations: 36th consecutive healthy day (see Orphaned Branch Alerts below).

Failure Signals ⚠️

1. branch_level_stuck_gate (persists, partially): Squad (0/4), Agentic Commands (0/4), and Doc Build - Deploy (0/3) remained at a strict 0% — all firings on the heaviest-volume branch, copilot/dynamic-checkouts-github-action.

   • Failure rate: 100% (0/11 combined)
   • Example: all 11 firings are action_required (gate re-fires), none resolve to success.

2. provenance_inversion band shift (now 4 of the last 6 days below band): only 64.0% of today's successes (16/25: 8 Code scanning + 2 CGO + 3 CWI + 3 CJS = 16) are bot/CI-gate-driven, vs the historical 72–86% band.

   • Failure rate (band deviation): -8 to -22pts below band floor
   • Example: joins 09-23 (58.3%), 09-26 (61.1%), 09-27 (70.8%) as sub-band readings — now a majority of the last 6 days, reinforcing this looks like a structural shift rather than noise, as flagged in the 09-27 report.

3. Single-branch concentration remains heavy: copilot/dynamic-checkouts-github-action produced 26/50 (52%) of all sessions, but converts only 23.1% (6/26) to success — well below the 91.7%/66.7% rates on the other three branches.

   • Failure rate: 76.9% (20/26 action_required)
   • Example: this branch alone accounts for 20 of today's 25 action_required outcomes (80%).

4. Recurring single-fire "merge → Squad Implement Worker" sub-signature: 2 more isolated instances today (23:41:51Z on ...ca3f0885, 00:34:27Z on ...87cd5693), consistent with the pattern seen 4x across 09-26/09-27 — but neither triggered a multi-run cascade this time (no failure conclusions occurred at all today).

   • Failure rate: N/A (single re-fires, not full cascades)

5. Conversation transcript logs remain empty — 0 files in the logs directory for the 36th+ consecutive recorded day.

   • Failure rate: 100% (0/50 sessions have transcript data)
   • Impact: no turn-by-turn tool-usage, loop-detection, or prompt-quality signal is available; this analysis remains metadata-only (session status/timing/branch/workflow-name only).

Prompt Quality Analysis 📝

Per-Prompt Breakdown

No prompt or task-description text is available — conversation transcripts are empty for the 36th+ consecutive day (see Notable Observations). This section is necessarily metadata-only: workflow trigger name and branch are the only proxies available for "prompt quality" this cycle.

Workflow-name Proxy Observations

• Addressing comment on PR and Code scanning AI findings on PR (single-fire, event-triggered) — 100% success (17/17) today, as in most prior recorded days.
• CJS / CWI / CGO / Squad / Agentic Commands / Doc Build - Deploy (repeated gate-sweep triggers) — 24.2% success (8/33) today.
• No actual prompt text, file references, or acceptance-criteria phrasing can be assessed without transcript data.

Example available signal (workflow name only, sanitized):

Addressing comment on PR #63687 (branch: copilot/task-...-87cd5693) → success, 23.92 min

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 21 open PRs (0.0%)
• Historical Baseline: 0.0% orphaned rate (30-day mean over 30 recorded days)
• Status: NORMAL

Escalation Candidate Details

Escalation Candidates

✅ No orphaned branches exceed the escalation threshold today.

Only 2 in-progress workflow runs repo-wide in the trailing 6 hours at snapshot time (both on main: this workflow + Tidy) — no branch had any gate concentration to evaluate. Both actively-worked copilot/* PRs today (#63496, #63241) carry Copilot in their assignee list.

CI Waste Estimate

• Orphaned gate-hours today: 0 (no candidates)
• Recoverable capacity: N/A — 36th consecutive day with nothing to recover

📈 Session Trends Analysis

Completion Patterns

Completion rate has climbed for three straight days (36.0% → 48.0% → 50.0%), pulling well clear of the 32-day mean (31.9%) and marking the 4th-highest single-day reading on record. The successful/failed count lines show today closing right at parity (25/25) rather than the wide oscillation seen through most of September.

Duration & Efficiency

Average and median duration both stayed low and close together today (3.84 min / 1.48 min raw), continuing the 2nd consecutive clean day without a multi-hour GitHub status-resync artifact (last seen 437.73 min on 09-26). The longest genuine run was 23.92 min (Addressing comment on PR, copilot/task-...-87cd5693).

Notable Observations

Loop Detection and Session Diagnostics

Loop Detection

• Sessions with loops: N/A — cannot be measured without turn-by-turn transcript data.
• Average loop count: N/A
• Common loop patterns: N/A

Tool Usage

• Most used tools: N/A — no tool-call data in session metadata.
• Tool success rates: N/A
• Missing tools: N/A — flagged via missing_data below.

Context Issues

• Sessions with confusion: N/A
• Common confusion points: N/A
• Clarification requests: N/A

Experimental Analysis

Standard analysis only — no experimental strategy this run (roll=84, threshold <30).

Actionable Recommendations

For Users Writing Task Descriptions

1. Reduce re-triggers on the concentration-leader branch: copilot/dynamic-checkouts-github-action drove 52% of all sessions today but converts only 23.1% to success — batching pushes on high-churn branches would cut redundant action_required gate re-fires.
2. Treat the provenance-inversion band shift as real, not noise: with 4 of the last 6 days below the historical 72–86% bot-driven band, task descriptions relying on bot/CI-gate auto-resolution may need more explicit human/agent follow-up steps going forward.

For System Improvements

1. Gate re-fire suppression on stuck branches: Squad, Agentic Commands, and Doc Build - Deploy remain at a strict 0% on the busiest branch — potential impact: High (would materially raise the concentration-leader branch's conversion rate).
2. Conversation transcript log capture: still 0 files after 36+ consecutive recorded days — potential impact: High (blocks all turn-by-turn/loop/tool-usage/prompt-quality analysis; this is the single largest standing tooling gap in this report series).

For Tool Development

1. Conversation transcript export: needed in every session analyzed (50/50) since the gap was first recorded — use case: turn-by-turn tool-call and reasoning analysis for loop detection, error recovery, and prompt-quality scoring.

Historical Trends and Statistical Summary

Trends Over Time

• Completion rate trend: 3-day rising streak (36.0% → 48.0% → 50.0%), now +18.1pts above the 32-day mean.
• Average duration trend: stable and low for 2 consecutive days (3.06 → 4.46 → 3.84 min raw), no resync-artifact inflation.
• Quality improvement: not assessable — no prompt-quality transcript data available.

Statistical Summary

Total Sessions Analyzed:     50
Successful Completions:      25 (50.0%)
Failed Sessions:             0 (0.0%)
Abandoned/Cancelled:         0 (0.0%)
Action Required (gate):      25 (50.0%)

Average Session Duration:    3.84 min (raw) / 7.68 min (25 non-instant)
Median Session Duration:     1.48 min (raw) / 5.72 min (25 non-instant)
Longest Session:             23.92 min (Addressing comment on PR, task-87cd5693 branch)
Shortest Session (nonzero):  2.95 min

Loop Detection:               N/A (no transcript data)
Context Issues:               N/A (no transcript data)
Tool Failures:                N/A (no transcript data)

High-Quality Prompts:          N/A (no transcript data)
Medium-Quality Prompts:        N/A (no transcript data)
Low-Quality Prompts:           N/A (no transcript data)

Next Steps

• Review recommendations with team
• Investigate whether transcript/conversation log export can be restored (36+ day gap)
• Watch provenance-inversion band shift for a 5th sub-band day before renaming the pattern
• Schedule follow-up analysis tomorrow (daily cadence)

---

References:

• §36390312866 — this analysis run
• §36280262918 — longest session today (23.92 min), Addressing comment on PR Report MCP payload sizes in audit and aggregate analysis #63687
• §36333388876 — most recent session in window, Addressing comment on PR Preserve durable provenance for superseded PR reviews #63496

Generated by 📊 Copilot Session Insights · claude · agent · 290.7 AIC · ⌖ 11.5 AIC · ⊞ 18.6K · ◷

• expires on Sep 28, 2026, 11:43 PM UTC-08:00