feat(tui): #3083 /provider readiness dashboard — capability/metadata badges

[Hmbown]

Closes #3083

Supersedes the draft #3504 (its reasoning-readiness commit is cherry-picked here with Hunter's authorship preserved) and completes the remaining /provider dashboard acceptance criteria.

What's included

1. Reasoning readiness (from [codex] feat(tui): show provider reasoning readiness #3504, rebased onto main): ProviderReasoningSummary (support / controls / stream visibility / selected control) on each row, with the 4-region provider_picker.rs conflict resolved (struct now carries both reasoning and maturity; compact_hint carries both).
2. GLM seed → bundled catalog (refactor): the reasoning projection no longer hand-seeds a single GLM-5.2 row. It sources from codewhale_config::catalog::bundled_catalog_offerings() (the single bundled Models.dev snapshot, same source v0.8.65: Provider-owned live catalogs and secret-free model cache #3385 uses), so every bundled provider with reasoning facts is covered and the row can't drift from the catalog.
3. Capability + metadata badges (feat): wires the existing resolved_capability_profile() projection (no new capability logic, no wire-id substring inference) into each row:
   • metadata: context window + max output, humanized (ctx:1M out:128K), ? when unknown;
   • capability: tools / json / stream / cache — tri-state (y / n / ?) so unknown stays distinct from unsupported and is never silently omitted;
   • model-origin: origin:default | saved | custom;
   • self-hosted hint next to the base URL for local runtimes (Ollama/vLLM/SGLang).

Badges are computed in the row projection (testable without rendering — acceptance criterion) and ordered after route/base so provider identity survives narrow-width truncation.

Deferred (flagged, non-blocking)

• Live "test connection" action: deferred to a follow-up to honor the "opening /provider must not block on live network I/O" constraint — a clean backgrounded/cached probe is its own design. Question for Hunter: OK to land v0.8.65: /provider readiness dashboard from route/catalog projections #3083 with this deferred, or treat it as a close-blocker?
• model-origin live-catalog / static split: not distinguishable until the v0.8.65: Provider-owned live catalogs and secret-free model cache #3385 live-fetch layer lands; the variant is reserved.

Verification

• cargo fmt; cargo clippy --workspace --all-targets --locked -- -D warnings — clean.
• cargo test -p codewhale-tui -- provider_dashboard provider_picker route_resolver pricing usage_meter — green.
• Full codewhale-tui bin suite: 5278 passed; only the known environmental/flaky papercuts failed (config_command_allow_shell_*, run_verifiers_background_*, run_tests_*), reproduced on pristine main.

New tests: provider_dashboard_row_surfaces_capability_and_metadata_badges, provider_dashboard_row_classifies_model_origin, model_origin_classifier_covers_default_saved_custom, self_hosted_provider_row_marks_self_hosted_in_hint, humanize_token_count_is_compact_and_marks_unknown (plus #3504's reasoning tests, rebased).

🤖 Generated with Claude Code

[claude]

Claude finished @Hmbown's task in 2m 39s —— View job

---

PR Review — feat(tui): #3083 provider dashboard capability/metadata badges

Tasks

• Gather context (PR metadata, changed files)
• Read the full diff of provider_picker.rs
• Review for correctness bugs, provider/model safety, reuse, tests, security
• Post final review

---

Summary

Solid, well-structured work. Capability/metadata badges are correctly wired through resolved_capability_profile() without duplicating inference logic; the catalog-backed reasoning projection fixes the GLM drift issue cleanly; tests cover the new row properties end-to-end. Two correctness issues below; the rest is minor or advisory.

---

1. Self-hosted hint misses SGLang and vLLM (bug)

File: provider_picker.rs, ~line 415

let self_hosted = if self.auth_status == ProviderAuthStatus::Local {
    " (self-hosted)"
} else {
    ""
};

ProviderAuthStatus::Local is returned only for Ollama. SGLang and vLLM without a credential return ProviderAuthStatus::Optional, so they silently omit the (self-hosted) hint — contradicting the PR's stated goal of surfacing self-hosted posture for local runtimes. The test self_hosted_provider_row_marks_self_hosted_in_hint only exercises Ollama so it doesn't catch this.

Fix: widen the predicate:

let self_hosted = if matches!(
    self.auth_status,
    ProviderAuthStatus::Local | ProviderAuthStatus::Optional
) {
    " (self-hosted)"
} else {
    ""
};

Add a parallel test assertion for ApiProvider::Sglang (no-key path) to lock this in.

Fix this →

---

2. default_reasoning_stream_visibility inconsistency: SGLang in, vLLM out

File: provider_picker.rs, ~line 672

| ApiProvider::Sglang
// ...but ApiProvider::Vllm is absent → falls through to Unknown

Both SGLang and vLLM are OpenAI-compatible self-hosted runtimes. If the rationale for Sglang → StructuredThinking is "commonly serves DeepSeek-R1-class reasoning models," the same logic applies to vLLM. Either add ApiProvider::Vllm to the StructuredThinking arm or document why they diverge. This only fires when the catalog says reasoning: Some(true), so blast radius is limited to display, but the inconsistency will confuse a future maintainer adding providers.

---

3. humanize_token_count K-band truncates silently

File: provider_picker.rs, ~line 232

Some(v) if v >= 1_000 => format!("{}K", v / 1_000),

1_999 → "1K", 999_999 → "999K". For context windows in the 1K–999K range this is visually lossless (windows are almost always power-of-two multiples of 1024 or round numbers) but the test only exercises 131_072 → "131K", which happens to be exact. Worth noting because a 196K window (196_608) would display as "192K" (196_608 / 1_000), not "197K". Consider v / 1_024 with a Ki suffix if accuracy matters more than compactness, or leave as-is with a comment acknowledging truncation.

---

4. Good choices worth preserving

• Provider/model safety (v0.8.65 EPIC: Separate provider facts, model facts, offerings, and route resolution #2608): No code infers a provider or model change from a wire-id prefix or freeform text. resolved_capability_profile(provider, wire_model) takes an explicit ApiProvider, fully respecting the invariant.
• OnceLock<CatalogSnapshot> for bundled_reasoning_catalog(): Lazy-init, contention-free, correct for a read-only bundled snapshot.
• eq_ignore_ascii_case in reasoning_catalog_offering: Correctly handles mixed-case model IDs (e.g. GLM-5.2 vs glm-5.2 in user config).
• Tri-state SupportState propagated through the badge cluster: ? staying distinct from n was the right call; the comment in the code explaining why is appropriate.
• Test model_origin_classifier_covers_default_saved_custom covers all four combinations including the Custom + has_saved_model override case cleanly.

---

[Hmbown]

Status: CI green. Completes the remaining #3083 dashboard acceptance criteria.

• Supersedes draft [codex] feat(tui): show provider reasoning readiness #3504 — its reasoning-readiness commit is cherry-picked here with Hunter B's authorship preserved; close [codex] feat(tui): show provider reasoning readiness #3504 when this lands.
• GLM reasoning rows now source from bundled_catalog_offerings() (single bundled snapshot) instead of a hand-seeded OnceLock.
• Capability/metadata badges + model-origin + self-hosted hint are projected from resolved_capability_profile() — tri-state (unknown ≠ unsupported), no wire-id substring inference.

Maintainer calls: (a) confirm it belongs in the v0.8.65 cleanup; (b) the live "test connection" probe is intentionally deferred to honor "opening /provider must not block on live I/O" — accept as a follow-up, or treat as a close-blocker for #3083?

[Hmbown]

Addressed the provider-dashboard review nits in dad6695:

• SGLang/vLLM no-key rows now show the (self-hosted) hint via ProviderAuthStatus::Optional.
• vLLM now uses the same default structured-thinking stream visibility as SGLang.
• Added regression coverage for both cases.

Verification:

• cargo fmt --all
• cargo test -p codewhale-tui --bin codewhale-tui --locked provider_dashboard — 10 passed
• cargo test -p codewhale-tui --bin codewhale-tui --locked self_hosted — 5 passed

[claude]

Claude finished @Hmbown's task in 1m 47s —— View job

---

Follow-up Review — feat(tui): #3083 provider dashboard

Tasks

• Gather context (trigger: PR synchronize after Hmbown addressed review nits in dad6695)
• Read the updated diff to verify fixes for the two correctness issues
• Check for any new issues introduced by the fix
• Post findings

---

Verification of dad6695

Both issues from the first review are correctly resolved.

Issue 1 — Self-hosted hint for SGLang/vLLM: FIXED

The predicate widening in compact_hint() (line 415–418) is exactly right:

let self_hosted = if matches!(
    self.auth_status,
    ProviderAuthStatus::Local | ProviderAuthStatus::Optional
) {

Importantly, ProviderAuthStatus::Optional is only returned by auth_status_for() for ApiProvider::Sglang | ApiProvider::Vllm on the no-key path (line 706–711). No cloud provider returns Optional — they fall through to Configured or Missing. So the widened predicate is correctly scoped and cannot misclassify a cloud provider as self-hosted.

This is also internally consistent: the existing readiness check at line 773 already uses Local | Optional => LocalReady, so the hint now aligns with the readiness posture rather than diverging from it.

The test correctly exercises both the auth_status value and the compact_hint() string for SGLang. vLLM's hint is implicitly covered by the same auth_status_for() branch, so a redundant vLLM-specific hint assertion isn't needed.

Issue 2 — vLLM reasoning stream visibility: FIXED

ApiProvider::Vllm is added to the StructuredThinking arm (line 691), and the new test self_hosted_reasoning_visibility_covers_vllm locks in both SGLang and vLLM together.

No new issues introduced. The fix commit is minimal (26 lines, single file, no behavioral changes beyond the two targeted paths).

---

The earlier advisory note on humanize_token_count K-band truncation stands as a low-priority cosmetic point, not a blocker.

PR looks clean. Both correctness issues from the first pass are resolved with appropriate tests. Ready to land when Hunter confirms the deferred "test connection" probe is acceptable as a follow-up.