[abhisek]

Fixes FOU-241 and FOU-240 finding F.

Problem

The profiler rewrote the whole-child learner profile after every session, from that session alone. A game expedition is 5 questions on 1 skill. So one bad expedition could rewrite the profile of a child with a strong record, and the next question prompts read that profile. Diagnosis events were written and never read.

Change

Store (owner-scoped, aggregated in SQL, portable to SQLite and Postgres)

• AnswerTallies: lifetime attempts and correct answers per skill. The snapshot counters restart on a tier change or decay, so they cannot give lifetime accuracy.
• CompletedSessionCount: sessions that served at least one question.
• DiagnosisCounts: diagnoses grouped by skill, category and misconception, with a From/To window.
• Stubs in the five hand-written mocks, and an owner-isolation test.

Profile input and prompt (internal/lessons)

• The input now has lifetime totals, grade, state, tier and fluency per skill, the session count, the 10 newest sessions (untagged quests left out), and the diagnosed error patterns of the last 60 days with taxonomy labels.
• The prompt puts the history first and labels the session as the newest evidence. It keeps at most 3 raw error lines per skill and does not list skills that were never attempted. Skill lists are sorted, so the same input gives the same prompt.
• The instructions tell the model to base strengths and weaknesses on lifetime evidence, and not to change one because of a single session.
• ProfileMaxTokens goes from 512 to 1024.

Refresh policy (internal/session/persist.go)

• Evidence gate: no profile below 15 lifetime curriculum answers (about three expeditions).
• After the gate, a refresh runs when there is no profile yet, when a skill changed state or tier against the previous snapshot (learn to prove counts), or when 3 sessions have ended since the profile's generated_at.
• The history is loaded inside the async goroutine. Live session state is copied before it starts. A failed history read skips the refresh and logs. It never falls back to a profile from the session alone. Each decision is logged with its reason.
• generated_at is written as RFC3339Nano, so the triggering session never counts as newer than its own profile. Old values still parse.

Docs: spec 10 §5 (inputs, cadence, prompt layout), glossary (Learner profile, Diagnosis event), personas (when parents first see a profile).

Behaviour change for parents

A new child has no profile until about three expeditions. After that, the profile changes every few sessions, or sooner when a skill changes state or tier. The dashboard already hides the card when there is no profile.

Tests

• profileRefreshDue and masteryStandingChanged table tests, and a sub-second ordering test.
• Regression for FOU-241: a child with 40/45 lifetime answers, a mastered skill and a recent profile has a 1/5 expedition, and the profile is not regenerated.
• A refresh that is due builds a prompt that has the lifetime totals, the session count, the recent sessions and the misconception labels, with the history before the session.
• A first expedition writes no profile. A failed history read skips the refresh.
• The skip tests wait for the logged decision, not a fixed sleep.
• Verify matrix: CGO_ENABLED=0 go build ./..., go vet ./..., go test ./..., go test -race -count=2 on internal/saas/... and internal/session, and the store suite on SQLite and on Postgres.

---

Generated by Claude Code

[safedep]

SafeDep Report Summary

No dependency changes detected. Nothing to scan.

View complete scan results →

_{This report is generated by SafeDep GitHub App}