feat(config): drive effort capabilities from provider config (#2963)

[Bernardxu123]

Summary

Close #2963: drive /effort capability from provider config (instead of hard-coding the level list in internal/config/effort.go).

Two new optional [[providers]] keys land everywhere (TOML, Go struct, Go renderer, desktop settings panel, i18n, example toml):

• supported_efforts = ["low", "medium", "high", ...] — non-empty list means /effort <level> is accepted for this provider (the literal auto is always prepended to the level set, matching the existing UI/UX).
• default_effort = "medium" — applied when the user has not set an effort explicitly; falls back to supported_efforts[0] if blank or not in the list, and is required to be one of the supported levels.

Nothing changes for the bundled presets. mimo-pro / mimo-flash keep no supported_efforts, so they remain rejected by /effort (preserving TestEffortCommandRejectsUnsupportedProvider). DeepSeek and anthropic keep their hard-coded capability paths — users can still opt-in to a custom list by TOML-overriding, but they don't have to.

Why a v2

The earlier draft attached supported_efforts to the mimo-* presets in Default(). That made /effort start accepting mimo-pro (which the test in #2963 was specifically designed to reject), so v1 had to be reworked.

v2 instead treats supported_efforts as a pure opt-in knob:

• The bundled Default() provider list is unchanged — the existing rejection test stays green.
• Users with a real MiMo API get a documented opt-in path (a commented [[providers]] block in reasonix.example.toml).
• RenderTOML only emits the new keys when non-empty, so existing ~/.config/reasonix/config.toml files round-trip byte-for-byte.

Commits (5)

1. config: add SupportedEfforts/DefaultEffort to ProviderEntry — struct + TOML tags.
2. config(effort): support config-driven SupportedEfforts/DefaultEffort — capability + normalize branches in effort.go; prepending auto; default-effort fallback to supported_efforts[0] when blank or unknown.
3. config(render): emit supported_efforts/default_effort in TOML output — round-trip output; only emit when non-empty.
4. config: cover config-driven effort paths in tests — 4 new tests: custom supported, custom normalize, custom default fallback, empty-slice behavior.
5. desktop+ui: expose supported_efforts/default_effort in settings panel — desktop ProviderView + 2 fields, Settings() mapping, SaveProvider round-trip; TS interface + 2 fields; 2 inputs in SettingsPanel (en + zh locales); mock data updated; commented example in reasonix.example.toml.

MiMo opt-in story

User adds to ~/.config/reasonix/config.toml:

[[providers]]
name = "mimo-pro"
kind = "openai"
base_url = "https://api.xiaomimimo.com/v1"
api_key = "..."
models = ["mimo-v2-flash"]
supported_efforts = ["low", "medium", "high"]
default_effort = "medium"

/effort low|medium|high|auto then works for mimo-pro. Without supported_efforts set, the existing rejection path is unchanged.

Verification

• go vet ./... (root + desktop) — exit 0
• go test ./... (root, REASONIX_RELEASE_CACHE_GUARD=1) — exit 0
• go test ./... (desktop) — exit 0
• pnpm build (desktop frontend) — 0 TS errors
• gofmt -l on changed Go files (in LF form) — clean
• New tests: TestEffortCapabilityCustomSupportedEfforts, TestNormalizeEffortCustomSupportedEfforts, TestNormalizeEffortCustomDefaultEffort, TestEffortCapabilityEmptySupportedEffortsNotConfigurable — all pass
• Pre-existing tests unchanged: TestEffortCommandRejectsUnsupportedProvider still passes

Risk register

• i18n — only en.ts and zh.ts are updated; if a third locale exists it would lose the new keys. The repo currently only ships en + zh, but a TODO is left in code review for any future locale authors.
• Settings round-trip — SaveProvider now writes 2 new fields; users with provider entries that previously had effort = "high" would round-trip to effort = "high" + (empty default_effort), which RenderTOML skips. No data loss.
• Lockfile — pnpm-lock.yaml is intentionally not committed in log-frame: migrate plan-resumed role from ink to frame #5; an unrelated pnpm 11 housekeeping diff (deprecated overrides, new libc entries) was reverted.

Files

desktop/frontend/src/components/SettingsPanel.tsx | 14 ++++
desktop/frontend/src/lib/bridge.ts                |  4 +-
desktop/frontend/src/lib/types.ts                 |  2 +
desktop/frontend/src/locales/en.ts                |  6 ++
desktop/frontend/src/locales/zh.ts                |  6 ++
desktop/settings_app.go                           | 32 ++++----
internal/config/config.go                         |  9 +++
internal/config/edit_test.go                      | 94 +++++++++++++++++++++++
internal/config/effort.go                         | 25 ++++++
internal/config/render.go                         |  6 ++
reasonix.example.toml                             | 15 ++++
11 files changed, 198 insertions(+), 15 deletions(-)

Test plan

• CI green on all 5 jobs: test (ubuntu/macos/windows), race (ubuntu), desktop (ubuntu), lint (golangci-lint), coverage
• Manual: open desktop Settings → Providers → edit entry → enter supported_efforts = ["low","medium","high"] and default_effort = "medium" → save → reopen config.toml → see the keys present → start CLI session → /effort high accepted
• Manual: same flow on mimo-pro without supported_efforts set → /effort high rejected with the same error message as before

Ref: #2963

[esengine]

Running the e2e suite to confirm this does not regress cache-hit % (watching the regression in #3091 before we merge any cache-sensitive change).

/e2e

[github-actions]

🤖 Reasonix e2e benchmark

Accuracy: 3/4 (75%) · Cache hit: 78% · Tokens: 131,517 (prompt 129,698 / completion 1,819) · Compactions: 0 · Cost: ¥ 0.0336

Task	Result	Steps	Prompt	Completion	Cache hit	Compact	Cost
compaction	❌ fail	6	73,942	595	63%	0	¥ 0.0293
fix-add-bug	✅ pass	3	18,346	234	98%	0	¥ 0.0011
fizzbuzz	✅ pass	3	18,633	389	99%	0	¥ 0.0013
palindrome	✅ pass	3	18,777	601	99%	0	¥ 0.0018

_{Real provider run. Cache-hit % is cached prompt tokens / total prompt tokens.}

agent: PR head (1255687) · triggered by @esengine

[esengine]

Re-running the suite — the previous run showed 3/4 (the compaction task failed) and a 63% cache hit on that task; checking whether that was a stochastic single-sample miss before we judge this PR.

/e2e

[github-actions]

🤖 Reasonix e2e benchmark

Accuracy: 4/4 (100%) · Cache hit: 79% · Tokens: 167,392 (prompt 165,120 / completion 2,272) · Compactions: 0 · Cost: ¥ 0.0412

Task	Result	Steps	Prompt	Completion	Cache hit	Compact	Cost
compaction	✅ pass	8	102,855	938	68%	0	¥ 0.0365
fix-add-bug	✅ pass	3	18,373	263	98%	0	¥ 0.0012
fizzbuzz	✅ pass	4	25,225	546	99%	0	¥ 0.0019
palindrome	✅ pass	3	18,667	525	99%	0	¥ 0.0017

_{Real provider run. Cache-hit % is cached prompt tokens / total prompt tokens.}

agent: PR head (1255687) · triggered by @esengine