Verdict for item 6 of the 2026-07-13 board build plan: per-tool usage over a ≥3-week post-board window, then merge/trim the MCP tool surface under the pre-registered decision rule (merge recall_episodes into recall if it draws <5% of recall-family calls).
Call counts (recall_metrics, model-facing source='mcp-tool')
Full telemetry window 2026-06-20 → 2026-07-25 (~5 weeks):
| kind |
calls |
| recall |
313 |
| remember |
220 |
| episodes (recall_episodes) |
12 |
| fetch |
7 |
| timeline |
2 |
recall_episodes share: 12 / 325 recall+episodes calls = 3.7% (3.6% of all 334 recall-family reads). Sub-window since 2026-07-01: 7 / 259 = 2.7% — usage is falling, not recovering.
- Board (
kind='board', push-only serve): 799 serves since 2026-07-13.
Decision
Under the pre-registered 5% threshold → merge. recall_episodes retires as a standalone tool; recall gains mode="overview"|"turns", with "turns" routing to the same engine drill-down path so telemetry keeps kind='episodes' and the historical series stays comparable. Shipped in PR #121.
Secondary findings
Scope already shipped before this audit (not redone)
Verdict for item 6 of the 2026-07-13 board build plan: per-tool usage over a ≥3-week post-board window, then merge/trim the MCP tool surface under the pre-registered decision rule (merge
recall_episodesintorecallif it draws <5% of recall-family calls).Call counts (recall_metrics, model-facing
source='mcp-tool')Full telemetry window 2026-06-20 → 2026-07-25 (~5 weeks):
recall_episodesshare: 12 / 325 recall+episodes calls = 3.7% (3.6% of all 334 recall-family reads). Sub-window since 2026-07-01: 7 / 259 = 2.7% — usage is falling, not recovering.kind='board', push-only serve): 799 serves since 2026-07-13.Decision
Under the pre-registered 5% threshold → merge.
recall_episodesretires as a standalone tool;recallgainsmode="overview"|"turns", with"turns"routing to the same engine drill-down path so telemetry keepskind='episodes'and the historical series stays comparable. Shipped in PR #121.Secondary findings
e:0in all six post-envelope rows; one older row predates the envelope). The unifiedn:path from mcp: tool-surface consolidation — remove query_graph/list_projects, hide issue_machine_token, unified fetch (e:/n:), timeline telemetry #58 works in prod and is fetch's dominant real use. (This corrects the 2026-07-24 board-review note that said the calls were all episode ids.)n:. Root cause: recall never servesn:ids (only the board does), and recall_feedback's description said to report "only ids the recall actually served". PR Merge recall_episodes into recall(mode="turns") + wire n: ids into the feedback loop (item 6) #121 records board/fetch note serves as"n:N"in the served_ids envelopes and rewrites the description to invite board/fetch note ids. No new DDL.recall_timelinedrew only 2 direct calls, but the timeline leg also serves insiderecall()on temporal intent; no action, keep watching.Scope already shipped before this audit (not redone)
query_graph+list_projects, hidissue_machine_tokenfrom listings, unifiedfetchwithe:/n:prefixes.