Cognee performance report

Use the bundled Python 3.9+ evaluator, ported from team-skills PR 3. It reads local plugin logs and Claude transcripts and produces aggregates. It does not send data anywhere unless an explicit export or server option is requested.

For “how did Cognee perform?”, report the current session and the last seven days. Default to seven days when no range is supplied; use the user's range when given.

python3 "${CLAUDE_PLUGIN_ROOT}/skills/cognee-report/scripts/plugin_eval.py" report --since 7d --compare

Use --session <cognee-session-id> --since all for the current session. Resolve the session from the current hook context or the host's entry in ~/.cognee-plugin/claude-code/sessions/<host-session-id>.json; do not assume a machine-wide last-recall marker belongs to this terminal. When present, read this host session's sessions/<host-session-id>/stats.json for exact injected/deduped counters. --format json exposes the complete scorecard. --no-transcripts is faster but cannot measure transcript-only latency and token usage. markers lists available date ranges and update times; --since last-update --compare evaluates an update.

Present the computed hit rate, recall and SessionStart latency, saves, improves, injected characters, and grade. Quote evidence from the output, including sample counts. Mark absent or low-sample metrics as unmeasured. Grades are withheld when gates fail or evidence is insufficient; never convert that into a positive grade. Token counts are estimates at four characters per token. Older audit logs are clipped; say so if exact context counters are unavailable. A recalled or injected item is not proof that it changed an answer.

For controlled comparisons of memory behavior, use the separate outcomes command and OUTCOMES.md. It runs repeated matched task trials and reports task success, citation correctness, false-memory adoption, tokens/cost and P50/P95 latency. Keep the 45-second cloud recall limit. Distinguish the offline smoke adapter from live model results; missing usage is unmeasured, not zero.

Keep recommendations tied to measured failures. Compare equal windows when possible. Do not infer seven-day post-release results from a short smoke test. Do not use export flags unless the user asks to export the report.

Exact current-session and recent session totals are also available without reading transcripts:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/_session_stats.py" --session <host-session-id> --days 7

These totals cover sessions that started within the window; use the evaluator's event timestamps for exact calendar-window recall metrics.

For why a specific memory was or was not injected on the last prompt, inspect its body-free receipt:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/cognee-evidence.py" explain --session-key <host-session-id>

Report per-item reasons (injected, already_in_context, over_budget, own_session, or source availability status), retrieval rank and priority. Versioned citations can be expanded through the search skill. Unaddressable legacy/guidance entries have hashes only; do not present those as source citations. The receipt describes delivery selection, not whether an item improved the answer.

The default report leads with What Cognee did for you, Where it helped, Quality, then the unchanged grade. Omit the latency bar and the entire Needs attention section. Keep latency measurements and gate results in the scorecard, including low-n markers. Show the generated bars rather than inventing scores. Use --ascii for #/- bars or --format markdown for fenced bars. --full restores the detailed layout and all recommendation why/do lines. Session hashes are evidence references, never memory content; new-token counts from clipped historical bodies are estimates. Delivery quality is not outcome quality, even when a last-turn receipt exists.