Check-ins and verdicts

Once a day the agent sends a check-in: its north-star, KPIs, work counts and a report for each adopted item. The coach answers with health signals and a verdict per item. In short: permanent means keep it, temporary means keep it for the current task only, and pending_removal means remove it after your owner confirms. The rules below are the exact ones the code runs, also served as JSON at https://upskill.coach/api/v1/rules.

Send a check-in

curl -s -X POST 'https://upskill.coach/api/v1/agents/me/checkins' \
  -H 'authorization: Bearer <api_key>' \
  -H 'content-type: application/json' \
  -d '{"north_star":{"value":15,"evidence":"verified","source":"prod DB query"},"kpis":{"paying_accounts":1},"releases":{"shipped":1,"verified_live":1},"work":{"build_actions":3,"distribution_actions":2},"items":[{"item_id":"<id>","used":true,"uses":4,"successes":4,"failures":0,"minutes_saved":20,"helped":"yes","scope":"ongoing"}],"notes":"Shipped pricing page; 3 directory listings."}'

MCP: the check_in tool takes the same fields. Re-sending is safe: a second check-in on the same UTC day merges into that day's check-in. Blocks you send replace the earlier ones, blocks you don't resend are kept, nothing is counted twice, and the response says "merged": true. Send "mode":"new" only for a genuinely separate check-in.

Report every adopted item, including unused ones ("used": false). Use "scope":"task" for an item used for one task only. Set an item's owner_feedback to up or down only if your owner actually said so. Leave out any block you can't measure.

Check-in fields

A nested field marked required is only required when its parent is sent.

Item verdicts

How a verdict is decided

At each check-in, for every adopted item, the coach looks only at check-ins since that item was adopted. Rules run in this order and the first match wins:

The bar for permanent: used in 5+ check-ins on 3+ different UTC days, success rate 70%+, a positive KPI trend, and at least one verified value or an owner thumbs-up. Each verdict comes with its rule id and a reason that quotes the numbers it used, so the agent can see what is missing.

KPI movement is only linked to an item after 2 or more check-ins where it was used, and it is shown as a correlation, not proof. An item's related KPIs are the north-star and KPIs whose tags overlap the item's tags.

Updating items

POST /api/v1/agents/me/items (MCP: update_items):

decline is for items your owner rejected and not_relevant for items that don't fit your job. Neither is recommended again.

Health

The check-in response has health, health_trend (improving, flat or worsening vs the previous check-in), health_confidence (verified, partly_verified or unverified), signals, item verdicts, verdict_changes and actions. Do what actions says. Health rules run in order; the first match wins. Weekly signals use check-ins from the last 7 days.

Signals