Skip to content

[copilot-opt] Copilot session telemetry (events.jsonl/conversation logs) missing for all 50 recent sessions #55465

Description

@github-actions

Problem

Per-session performance telemetry (events.jsonl, tool-call timing, payload sizes, validation timing, context-size, model-switch events) required by this optimizer's Phase 2 analysis was completely absent for all 50 Copilot sessions in the last-14-day window, blocking detection of slow tool calls, oversized responses, late validation failures, oversized instruction payloads, and prompt drift.

Evidence

  • Analysis window: 2026-08-10 to 2026-08-24 (last 14 days)
  • Sessions enumerated: 50 (from /tmp/gh-aw/agent/session-data/sessions-list.json)
  • /tmp/gh-aw/agent/session-data/logs/ contained 0 files — no *-conversation.txt, no events.jsonl, and no fallback logs for any of the 50 in-scope sessions
  • The shared/copilot-session-data-fetch.md import that is supposed to populate this directory ran without error but produced an empty logs/ folder, while the parallel shared/copilot-pr-data-fetch.md import correctly populated /tmp/gh-aw/agent/pr-data/copilot-prs.json with 1000 records
  • This is a systemic gap: every run of this optimizer workflow with the current data-fetch configuration will silently skip Phase 2 (tool latency, payload size, validation timing, context size, model-switching, prompt-drift analysis) because there is no source data to analyze

Proposed Change

Investigate why shared/copilot-session-data-fetch.md is not writing conversation/event logs to /tmp/gh-aw/agent/session-data/logs/ (e.g., check the session IDs/branch filter used to fetch logs, confirm the correct artifact or API endpoint is targeted, and confirm write permissions/paths match what this workflow expects). Add an explicit non-fatal warning step immediately after the fetch that fails loudly (or posts a missing_data signal) when the logs directory is empty, so future runs surface the gap instead of silently producing partial reports.

Expected Impact

  • Restores Phase 2 performance analysis (tool latency, payload size, validation timing, context size, model-switching, prompt drift) for all future runs of this optimizer, which is currently 100% blocked.
  • Prevents future optimizer runs from silently under-reporting findings without any indication that a whole analysis phase was skipped.

Notes

  • Distinct root cause category: tool payload/latency optimization infrastructure (session telemetry availability), which blocks the underlying analysis this recommendation depends on
  • Data quality caveat: this issue itself is evidence of the gap — no session-level timing/context/drift metrics could be computed this run because the source data was missing.

Generated by ⚡ Copilot Opt · copilot · auto · 46.5 AIC · ⌖ 13.5 AIC · ⊞ 8.3K ·

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions