Skip to content

v9.16.0 — Session analytics for agents-consilium

Choose a tag to compare

View all tags
@rodion-m rodion-m released this 16 Sep 14:15
· 9 commits to main since this release
v9.16.0
This tag was signed with the committer’s verified signature.
rodion-m Rodion Mostovoi
SSH Key Fingerprint: 1lMBOnLDZ4i0tgRrkl+OM8XuLW37/DE3PgMxSfyKyeI
Verified
Learn about vigilant mode.
29f63a1
This commit was signed with the committer’s verified signature.
rodion-m Rodion Mostovoi
SSH Key Fingerprint: 1lMBOnLDZ4i0tgRrkl+OM8XuLW37/DE3PgMxSfyKyeI
Verified
Learn about vigilant mode.

Summary

  • consilium sessions gains a reflection layer: turns (one row per human-triggered turn with duration/models/tool/error/usage metrics), flags (deterministic anti-patterns — retry_loop, search_loop, edit_without_read, correction(_burst), abandoned, permission_friction, context_pressure, interrupted, error_burst, failed_run — each with {seq, locator} evidence), stats (grouped aggregates by model/harness/cwd/day with honest coverage).
  • Fragments carry error / usage / interrupted where native records prove them; layered stores dedupe to canonical identity (codex rollout > state > catalog; claude-desktop by cliSessionId).
  • Fixes found by dogfooding (Devin agents ran real reflection queries through the CLI): codex response_item.user misclassification without an event_msg mirror, retry/search loop false positives on fan-out batches, devin call_id inflation, silent --around scope misses.
  • Codex session listing now reads file heads only — list -a codex went from minutes to ~1.4s.

Test plan

  • sessions_test.py 30/30 (incl. step-collapsing + unparseable-input regressions)
  • Full suite run.sh: 955 passed, 0 failed
  • Live runs: grok/claude-code turns+flags+stats, codex catalog/history/day stats, flag evidence → show --around verified

Generated with Devin

Assets 2
Loading