feat(skills): the expense-reporter skill and its eval cases

[IBJunior]

Slices C1 and C2 of v3 phase 2 — the real skill, and the evals that grade it. Completes the phase: the loader (#26), the renderer (#28), the tool (#29), and now the instructions plus their guard.

The skill

Replaces the placeholder with four steps: whether to chart at all → which type → how to shape the query → how to read it back, plus what each of render_chart's refusals means.

The restraint half is the part that matters. Nothing in code stops the agent charting a single figure, and a skill that only encourages charts would make answers worse — so the skill says a single number is a sentence and a short ranking is a table.

The eval cases

eval/cases/skills.cases.mts — a defect class the suite did not cover: a capability the agent must choose to consult, and choose not to overuse. A unit test proves load_skill returns the body; only a model can show whether the agent reaches for it, reads it before acting, or leaves it alone.

Case	Grades
chart-request-loads-skill	activation
skill-read-before-charting	order — toolCalledBefore(load_skill, render_chart)
trend-question-picks-line	chart-type selection (a closed enum, so gradeable without a judge)
single-figure-needs-no-chart	the over-triggering guard

The fixture needed no changes: the seeding formula already spreads expenses across nine months with varying totals, so the trend case has real shape and no second source of truth for expected figures.

What the evals found

A prompt defect. The skills section said "when a task matches a skill" without defining matching, and haiku fell back to lexical overlap — loading the skill for "show me a chart" but never for "chart my spending": 0/5, deterministic, identical trajectory every run. Sonnet loaded it either way, so the instruction was carrying the difference, not the model.

The fix states the principle (judge the KIND of work, not the vocabulary) and blocks the easy-task escape. Deliberately no example phrasings — listing synonyms would have made the eval pass by feeding it the answer. Activation went 0/5 → ~6/8 on haiku with no vocabulary shared between the prompt and the case.

A harness bug, older than this work. Only approval and config cases reset the fixture, so an approved log_expense left a $4.20 row behind and the next case asserting a Dining total read $2,737.20 against an expected $2,733.00. Three existing cases assert that same total and passed only because they run before the approval cases. Case order was load-bearing and silent.

A stale case. no-double-prompt-on-mutation claims "nothing legitimate to ask for", but the currency feature made that false — with no currency established the agent correctly asks before writing, so the case failed for the inverse of its own reason. Pre-seeding the currency makes the premise true again.

Verified

18/18 eval cases on claude-haiku-4-5 (2 legacy scripted twins still skipped), 185 unit tests, tsc and typecheck:eval clean.

🤖 Generated with Claude Code