#17 refactor: self-updating model catalog, then trim agent-notes to its mission

[verkligheten]

Implements epic #17 — make the model catalog self-updating, then trim agent-notes back to its mission: a shared template of skills, agents, and a role→model→effort map.

144 files changed, +3,852 / −5,816 (net −1,964 LOC). Suite green: 1780 passed, 15 deselected.

Three live bugs found and fixed

These were not in the original scope. They surfaced while verifying the plan's assumptions.

1. Claude Opus 5 was billed at 3× — $15/$75 instead of $5/$25. Two defects compounded: normalize_model only rewrote dashed pairs, so claude-opus-5 passed through untouched and then hit the *opus* catch-all in the legacy pricing row. No warning fired, because something did match. Verified against real session files, where claude-opus-5 is the exact string reaching the pricer.
2. claude-opus-4-10 and 4-11 priced as legacy — surfaced only once model IDs were stored in a single canonical form. Fixed by anchoring the globs so *opus-4-1 cannot swallow 4-10.
3. claude-fable-5 was unreachable as a default — it declared model_class: fable, but no role declares typical_class: fable. Now classified into the frontier class and marked never_default, so it stays selectable by pin but is never auto-assigned. Fable is a frontier model, but too expensive to hand an agent by default.

What changed

Catalog (#18, #20, #21, #22) — data/models/*.yaml (14 hand-maintained files) is replaced by three layers: a committed seed.json snapshot of what providers publish, a rules.yaml judgment layer (classes, families, deprecation, capabilities), and an optional user override. agent-notes models refresh fetches the two provider endpoints and writes a cache; models freeze promotes that cache to the seed for committing. Refresh → review diff → freeze → commit is now the whole model-update workflow — no code change per provider release.

build never touches the network. It reads ~/.cache/agent-notes/catalog.json, falling back to the committed seed. Only refresh fetches, via a lazily-imported stdlib urllib.request — no new dependency.

The catalog swap was gated: test_catalog_equivalence.py asserts the generated catalog matches the previous 14 models field-for-field. The YAMLs were deleted only after it was green.

Scope narrowed to anthropic + openai — github-copilot aliases dropped. #29 (adopt models.dev) was closed as not-planned on the same grounds: every install in state.json is claude only, so the Bedrock/Vertex/Copilot alias tables it uniquely provided are not reachable today.

Mission trim (#19, #23, #24, #25, #26) — deleted the wiki memory backend (~2,500 LOC across 8 modules, plus the wiki-compiler agent), dead code (tiers, Model.pricing, strip_memory_section, doctor_checks.py), and three skills that Superpowers ships longer versions of (tdd, brainstorming, writing-great-skills). Memory and cost-report were decoupled from rendering.py, hook.py and installer.py and colocated under agent_notes/memory/ and agent_notes/cost/.

A state.json with memory.backend: "wiki" now raises a clear migration error — never a silent switch to another backend.

Credential guard (#27) — fixed false positives that were blocking work on this repo: template files (credentials.example.yml, secrets.sample.json), source files (services/credentials.py), and keyword-only commands. A template marker must be its own dotted component, and hard-secret extensions (.enc, .key, .pem, …) are never exempt, so foo.example.key stays blocked. A security audit of the first implementation found six bypasses — bash -c, dd if=, python -c, example.env, secret/config.py, and --flag=path — all confirmed empirically and closed. data/rules/safety.md was updated so the shipped doc matches the guard.

Dependencies (#28) — dropped the tomli conditional, the project's only version-conditional dependency, and raised the Python floor to 3.11 so tomllib covers the single TOML read path. tomli-w stays: it is the Codex writer, and hand-rolling TOML escaping on a path that carries agent prompts and secret material is not worth removing a 10 KB pure-Python dependency.

Verification

• uv run pytest tests/ -v → 1780 passed, 15 deselected
• agent-notes build compared under a pinned XDG_CONFIG_HOME against a git worktree built at the branch point — dist changes are limited to the deliberate ones (wiki agent and three skills removed, safety.md rules text)
• Cache isolation regression-tested: a developer's ~/.cache/agent-notes/catalog.json can no longer silently drive the test suite

Follow-up

#30 remains open — reinstall from this branch to clear the stale pre-#27 guard, then remove the two workarounds it forced (an orphaned test module with its collect_ignore entry, and a dead import shim in the credentials service).

Closes #18, #19, #20, #21, #22, #23, #24, #25, #26, #27, #28

Note: this PR targets develop, so GitHub will not auto-close those issues on merge — they close when develop reaches main.