You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Adds Concentrate (concentrate) as a first-class, opt-in, BYOK provider — the OpenAI Responses-compatible gateway at https://api.concentrate.ai/v1 — inside the existing provider authorities (no parallel secret store, router, or runtime). Mirrors the Eden AI aggregator addition across the registry, config tables, env overrides, secrets, TUI wiring, docs, web facts, and goldens.
Identity: ProviderKind::Concentrate (+ concentrate-ai/concentrate_ai/concentrateai), hand-written impl Provider with WirePolicy::Fixed(Responses), default model deepseek-v4-pro, env CONCENTRATE_API_KEY / CONCENTRATE_BASE_URL / CONCENTRATE_MODEL, own secret-store slot.
Model ids pass through: plain catalog id → the gateway picks the upstream; provider/model pins it; only the gateway's own concentrate/ namespace is stripped so concentrate/auto reaches its router (Codewhale's bare auto stays the resolver sentinel).
Wire: documented fields only (model, input, stream, max_output_tokens, tools/tool_choice/parallel_tool_calls, reasoning.effort); system prompt as a leading system input item; typed response.* SSE via the existing parser, ending on response.completed (no [DONE]).
Catalog: unauthenticated GET /v1/models through the existing named-gateway live-catalog path (rows provider-scoped, unclaimed).
Commercial boundary preserved: Concentrate's Terms of Service forbid resale / white-label / service-bureau use without written consent; its AUP forbids key sharing; pricing says "No platform markup on tokens". This PR ships no Codewhale-owned key, no stored customer key, no default/managed routing, and no markup. A saved/env key is bound to the official base URL and is never sent elsewhere (a custom endpoint gets a key only when base_urlandapi_key are both in [providers.concentrate]). Hosted-lane activation stays gated on written consent, terms, and billing approval (ops evidence concentrate-gateway-20260829/CHECKLIST.md).
Keyless dogfood: scripts/concentrate-selftest.sh boots scripts/concentrate-stub.py (the documented contract on loopback) and drives the real codewhale exec --auto --output-format stream-json path through it, asserting URL, bearer header, verbatim model, only-documented fields, system-first input, the completed-turn receipt, and the wrong-key 401. No network call leaves the machine.
No-Issue: founder-requested provider addition; no tracking issue exists.
Testing
Local, hosted-equivalent flags (RUSTFLAGS=-Dwarnings, RUST_MIN_STACK=16MiB, cargo nextest … --all-features --locked --profile ci), summary in the ops evidence:
cargo fmt --all -- --check: clean
codewhale-config: 624 run / 624 passed / 1 skipped — incl. new concentrate_resolves_named_responses_gateway_and_environment_overrides (aliases, metadata, fixed Responses wire, secret slot, env + config resolution, and the credential-scope rule that an env key is never sent to a non-official base URL) and concentrate_passes_ids_through_and_strips_only_its_own_namespace (plain / provider/model verbatim, concentrate/auto → auto, bare auto → provider default, all on the Responses protocol); goldens regenerated (golden_route_ids.txt, providers-export.golden.json); count assertions 47→48 / 42→43; expected-wire tables updated.
codewhale-secrets: 62 run / 62 passed — incl. concentrate_env_aliases_resolve (own slot; no cross-provider bleed).
codewhale-tui focused (concentrate | error_taxonomy:: | client::responses::tests:: | llm_client:: | config provider tests | edenai|telecomjs|opencode_zen_responses|catalog): 440 run / 440 passed — incl. concentrate_responses_request_matches_the_documented_contract (wiremock: POST /v1/responses, Bearer header, verbatim provider/model, no store/include/instructions/messages, system item first, typed-event SSE without [DONE] assembles text + usage 12/5, official URL maps to /v1/responses), concentrate_error_bodies_surface_verbatim_and_classify (401 → Authentication with "Invalid API key", 402 → quota with "insufficient credits", 400 → InvalidInput with "Invalid model name"), concentrate_live_catalog_is_provider_scoped_and_marks_the_default, concentrate_responses_body_sends_only_documented_fields, insufficient_credits_classifies_as_rate_limit_not_auth, flat_error_and_message_body_surfaces_both_halves.
Keyless self-test PASS ×3 (scripts/concentrate-selftest.sh with the debug CLI): concentrate/auto (wire auto), openai/gpt-5.6-sol, deepseek-v4-pro — health 200, catalog without a key, exactly one bearer-authenticated POST /v1/responses with the verbatim model, stream:true, system item first, only documented fields; stream-json receipt content…turn_usage, session_capture, metadata, done; wrong key → documented 401 → exit 1. Defeat evidence: the first run (env key only) reached the stub with an empty Authorization header, which is the credential-scope rule working; the recipe now writes base_url + api_key into [providers.concentrate] as a BYOK user pointing at a local gateway would.
scripts/check-provider-registry.py PASS; cargo clippy for config+secrets and tui (--all-targets --all-features, CI allow list) clean; node web/scripts/derive-facts.mjs (providers 46) + check-facts OK; dead-code budget PASS (444/444); git diff --check clean.
Not done: no live Concentrate canary (needs a key and founder-gated spend — the live gateway was never contacted beyond the public, unauthenticated /v1/models read used to pin the catalog shape); Windows-target compile is hosted-CI-only on this macOS host.
cargo fmt --all -- --check
cargo clippy on the changed crates, all targets, all features, CI allow list
cargo test --workspace --all-features --locked (hosted CI; locally the changed crates' suites passed as above)
Medium Risk
Touches shared provider registry, route resolution, Responses request shaping, and credential scoping across config/TUI/secrets; mistakes could mis-route keys or send wrong wire fields, but scope is additive with extensive tests and no managed keys shipped.
Overview
Adds Concentrate as an opt-in, BYOK catalog provider (concentrate) wired through the same registry, config, secrets, route contract, and TUI paths as other gateways—not a parallel stack.
The provider is fixed on the OpenAI Responses wire (POST /v1/responses) with defaults https://api.concentrate.ai/v1 and deepseek-v4-pro, auth via CONCENTRATE_API_KEY and [providers.concentrate] (plus aliases). The route resolver passes model ids through verbatim and strips only the concentrate/ prefix so concentrate/auto hits the gateway router while Codewhale’s bare auto stays the provider-default sentinel.
TUI/runtime behavior is tailored to Concentrate’s documented API: Responses bodies omit store, include, and top-level instructions (system prompt as a leading system input item); Chat Completions reasoning fields are not injected; unauthenticated GET /v1/models drives live catalog refresh but does not count as API-key verification. Env/config keys are scoped to the official base URL unless both base_url and api_key are set in config for a custom endpoint.
Supporting changes include golden route/provider export updates, docs and web facts, registry check allowance for a hand-written Concentrate provider, error handling for flat error+message JSON and HTTP 402 “insufficient credits” as quota, and keyless loopback dogfood via scripts/concentrate-stub.py + scripts/concentrate-selftest.sh.
Reviewed by Cursor Bugbot for commit 9af679b. Bugbot is set up for automated code reviews on this repo. Configure here.
…ateway
Adds `concentrate` (aliases `concentrate-ai`, `concentrate_ai`,
`concentrateai`) inside the existing provider authorities — no parallel
secret store, router, or runtime:
- Identity/metadata: `ProviderKind::Concentrate`, hand-written
`impl Provider` with `WirePolicy::Fixed(WireFormat::Responses)` (the
gateway documents the Responses API as its production surface),
default base URL `https://api.concentrate.ai/v1`, default model
`deepseek-v4-pro`, env `CONCENTRATE_API_KEY` (+ `CONCENTRATE_BASE_URL`,
`CONCENTRATE_MODEL`), its own secret-store slot, credential help.
- Routing: aggregator-class pass-through. A plain catalog id lets the
gateway choose the upstream provider, `provider/model` pins one, and
only the gateway's own `concentrate/` namespace is stripped so
`concentrate/auto` reaches its `auto` router while Codewhale's bare
`auto` stays the resolver sentinel (provider default).
- Wire: the Responses body carries only documented fields — `model`,
`input`, `stream`, `max_output_tokens`, `tools`/`tool_choice`/
`parallel_tool_calls`, `reasoning.effort` — with the system prompt as a
leading `system` input item (`instructions`, `store`, `include`, and
`reasoning.summary` are absent from the gateway's parameter reference).
Streaming rides the existing typed `response.*` SSE parser and ends on
`response.completed` without a `[DONE]` sentinel.
- Catalog: the unauthenticated `GET /v1/models` (OpenAI list shape) joins
the named-gateway live-catalog path; rows stay provider-scoped and
unclaimed.
- Errors (provider-neutral): a 402 "insufficient credits/funds" body now
classifies as quota (RateLimit) instead of falling through, and a flat
`{"error":"<class>","message":"<detail>"}` body surfaces both halves
in the TUI sanitizer instead of the class alone.
- Registry parity: `ProviderKind::ALL` 42→43, registry 47→48, golden
route ids + providers-export golden regenerated,
`scripts/check-provider-registry.py` manual-impl allowlist, web facts
label maps + `facts.generated.ts` (providers 45→46), docs rows in
PROVIDERS.md (+ a Concentrate Notes section) and CONFIGURATION.md,
CHANGELOG.
Commercial boundary, preserved in code and docs: BYOK only. Concentrate's
Terms of Service forbid resale, white-label, and service-bureau use
without written consent and its AUP forbids key sharing, so there is no
Codewhale-owned key, no stored customer key, no default or managed
routing, and no markup; any hosted lane is gated on written consent,
terms, and billing approval (ops evidence
concentrate-gateway-20260829/CHECKLIST.md). A saved or environment
Concentrate key is bound to the official base URL and is never sent to
any other endpoint — a custom endpoint receives a key only when
`base_url` and `api_key` are both written into `[providers.concentrate]`.
Keyless dogfood: `scripts/concentrate-selftest.sh` boots
`scripts/concentrate-stub.py` (the documented contract on loopback:
`/v1/responses/health`, unauthenticated `/v1/models`, typed-SSE
`/v1/responses`, documented error bodies) and drives the real
`codewhale exec --auto --output-format stream-json` path through it,
asserting the URL, bearer header, verbatim model, only-documented fields,
system item first, the completed-turn receipt, and the wrong-key 401.
No network call leaves the machine; no account exists in the loop.
Contract sources (fetched 2026-08-29):
https://concentrate.ai/docs/api-reference/introductionhttps://concentrate.ai/docs/api-reference/endpoint/request-parametershttps://concentrate.ai/docs/api-reference/endpoint/streaminghttps://concentrate.ai/docs/api-reference/endpoint/errorshttps://concentrate.ai/docs/api-reference/endpoint/list-modelshttps://concentrate.ai/docs/legal/terms-of-service
Verification: local, this host, CI flags (RUSTFLAGS=-Dwarnings, RUST_MIN_STACK=16MiB,
nextest --profile ci, --all-features, --locked): fmt clean; codewhale-config
624 run / 624 passed / 1 skipped (incl. new concentrate_resolves_… and
concentrate_passes_ids_through_and_strips_only_its_own_namespace, regenerated
goldens); codewhale-secrets 62/62; codewhale-tui focused (concentrate +
error_taxonomy + client::responses + llm_client + provider config + catalog
filters) 440 run / 440 passed; debug CLI built; scripts/concentrate-selftest.sh
PASS ×3 (concentrate/auto → wire `auto`, openai/gpt-5.6-sol, deepseek-v4-pro);
scripts/check-provider-registry.py PASS; clippy config+secrets and tui
(--all-targets --all-features, CI allow list) clean; web facts regenerated
(providers 46) + check-facts OK; dead-code budget PASS; git diff --check
clean.
Not done: no live Concentrate canary (needs a key and founder-gated spend);
Windows-target compile is hosted-CI-only on this host.
No-Issue: provider addition requested by the founder; no tracking issue exists.
Signed-off-by: CodeWhale Bot <bot@codewhale.net>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Well-scoped, mirrors the existing Eden AI aggregator pattern closely and stays inside the existing provider authorities (registry, secrets, resolver, client) rather than adding a parallel system. I could not execute cargo test/cargo clippy in this sandbox (no tool permission for cargo), so this is a source-level review; the PR body's own local run numbers are the executed evidence.
Findings
1. Stale doc count in docs/PROVIDERS.md:45 (minor correctness/docs bug)
The canonical provider IDs are the 42 entries of `ProviderKind::ALL`
ProviderKind::ALL was bumped to 43 entries by this PR (crates/config/src/provider_kind.rs:250, confirmed by the PR's own new test crates/config/src/tests.rs:5165: assert_eq!(ProviderKind::ALL.len(), 43)), and the ID list right below the sentence on line 45 was correctly updated to include concentrate — but the "42" count itself wasn't bumped to 43. Same pattern to double check: docs/CONFIGURATION.md:424-426 lists the id but doesn't state a count, so it's unaffected.
2. Minor: codex_responses_reasoning_effort reused as-is for Concentrate's reasoning.effort (crates/tui/src/client/responses.rs:936-941)
Concentrate shares OpenAI Codex's low/medium/high/xhigh effort mapper rather than having its own tier table (unlike DeepSeek, which gets deepseek_effort_tier_or_default). Reasonable reuse, but worth a one-line comment confirming Concentrate's docs actually accept xhigh/medium as spelled — the PR's contract-source list doesn't include a reasoning-tier reference. Not blocking; low risk since it degrades to a documented value either way.
What's solid
Route/provider-selection safety (EPIC v0.8.65 EPIC: Separate provider facts, model facts, offerings, and route resolution #2608 invariant): respected. The concentrate/ prefix stripping in crates/config/src/route/resolver.rs:346-356 only runs inside scope_selector, which is reached after provider_kind is already fixed by explicit user/config selection — it never infers a provider switch from a provider/model-shaped string. Upstream prefixes (openai/gpt-5.6-sol) are left verbatim and only pin the upstream inside the already-selected gateway, matching the Eden AI precedent.
Credential-scope rule is generic, not new/duplicated logic — provider_base_url_is_official (crates/config/src/lib.rs:4880) and the existing resolve_runtime_options env-key-scoping path handle Concentrate the same way as every other provider; the new concentrate_resolves_named_responses_gateway_and_environment_overrides test (crates/config/src/tests.rs) exercises it directly and the assertion that an env key doesn't leak to a non-official base URL checks out against the shared implementation.
Generic error-body fix (extract_json_error_message, crates/tui/src/llm_client/mod.rs:790-805) is scoped correctly — it only fires when both /error and /message are top-level strings and differ, so it can't hijack the existing nested {"error":{"message":...}} shape used by other providers (verified by the added flat_error_and_message_body_surfaces_both_halves test, which also checks the nested case is untouched).
402 → RateLimit reclassification (crates/tui/src/error_taxonomy.rs:370-376) is a reasonable generic heuristic addition (prepaid-gateway quota condition), with keyword ordering that doesn't collide with the existing InvalidInput/Authentication branches for the documented Concentrate bodies.
concentrate-stub.py / concentrate-selftest.sh: loopback-only, isolated HOME/config, no real credentials, no network egress — good BYOK-safe test harness design.
Registry, secrets, config env-override, and TUI wiring changes are mechanical, one-to-one mirrors of the Eden AI addition; no dead code or needless duplication spotted.
The reason will be displayed to describe this comment to others. Learn more.
Do not treat the public catalog as key verification
For Concentrate, GET /v1/models is explicitly unauthenticated (docs/PROVIDERS.md:670), so this request returns 2xx even when api_key is invalid. The guided provider setup consequently records a successful observed probe and advances to model selection for any nonempty key, only failing on the first POST /responses; special-case this provider as an unobserved probe or validate against an authenticated endpoint.
The reason will be displayed to describe this comment to others. Learn more.
Fixed in 0f09427: api_provider_skips_models_probe now includes Concentrate. GET /v1/models is unauthenticated, so a 2xx is not key proof. provider_api_key_verification_is_observed(Concentrate) is false; health_check issues no request. Pinned by concentrate_health_check_does_not_treat_unauthenticated_models_as_key_proof (3 passed; 0 failed with the sibling Concentrate contract tests).
The reason will be displayed to describe this comment to others. Learn more.
Fail the self-test when reply content is missing
If the Responses parser or stream-json renderer drops every content event while still emitting a done receipt, this branch merely prints a note and the script exits successfully. That means the advertised end-to-end assertion that the stub reply reached CLI output can regress while CONCENTRATE SELFTEST PASS is still reported; make the missing expected text an assertion failure.
The reason will be displayed to describe this comment to others. Learn more.
Fixed in 0f09427: scripts/concentrate-selftest.sh now exits 1 when the stub reply text is missing from content events, instead of printing a note and exiting 0.
The reason will be displayed to describe this comment to others. Learn more.
Classify 402 exhaustion in the typed HTTP path
A real Concentrate 402 is converted by send_with_retry through LlmError::from_http_response before this string classifier is consulted, and has_explicit_quota_phrase recognizes neither “insufficient funds” nor “insufficient credits.” The resulting downcastable LlmError::Other takes precedence in envelope_for_llm_error, so the UI reports an Internal, recoverable failure instead of the quota category promised here; add these phrases to the typed quota-evidence path (the new test currently hides this by reclassifying only the formatted string).
Hosted `Test (ubuntu-latest)` on #5725 failed exactly one test out of
13,958: `cli_provider_helpers_follow_config_metadata` still asserted the
pre-Concentrate registry sizes (47 full / 42 catalog). The config crate's
twin assertion was updated in the parent commit; this mirrors it. The rest
of the test (env vars and secret-store slot per registry kind) already held
for `concentrate`. Local: codewhale-cli 276 run / 276 passed.
Signed-off-by: CodeWhale Bot <bot@codewhale.net>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Hosted Test (ubuntu-latest) at 2860a5710 failed one test of 13,958: codewhale-cli tests::cli_provider_helpers_follow_config_metadata still asserted the pre-Concentrate registry sizes (47 full / 42 catalog). Fixed in 689f52a61e (48 / 43, mirroring the config crate's twin assertion). Local: codewhale-cli 276 run / 276 passed; fmt clean. Hosted CI re-runs at that head.
Check what changed in the latest commit (689f52a) relative to the prior review round
Verify whether previously flagged issues were addressed
Review the new commit's diff for correctness/safety/reuse/tests/security
Post final review summary
Summary
This synchronize only added one commit, 689f52a, which is a mechanical, correct fix: hosted CI's cli_provider_helpers_follow_config_metadata test still asserted the pre-Concentrate registry sizes (47/42); this bumps them to 48/43 to match the config crate's twin assertion from the parent commit. Nothing else changed. I re-verified this diff plus re-checked the three issues raised in the prior Codex review round — none of them were touched by this push, so they still stand:
1. (Unresolved, real bug) 402 quota classification never reaches the typed error path — crates/tui/src/llm_client/mod.rs:759 (has_explicit_quota_phrase) and the is_explicit_quota_code/marker checks it's paired with don't recognize "insufficient credits"/"insufficient funds"/402, so has_explicit_quota_evidence returns false for Concentrate's documented 402 body. LlmError::from_http_response (llm_client/mod.rs:462) then falls through its match status with no 402 arm to _ => LlmError::Other(...), which error_taxonomy.rs:325 maps to Internal/Error/recoverable: true — the opposite of the quota/RateLimit category the PR claims. The new insufficient_credits_classifies_as_rate_limit_not_auth test (error_taxonomy.rs:603) only exercises classify_error_message (the string heuristic added at error_taxonomy.rs:373-376), never the typed LlmError::from_http_response boundary that send_with_retry actually uses (client.rs:2816, 2897), so the test passes while the real request path still misclassifies. Fix by adding a 402 if has_explicit_quota_evidence(body) is false arm (or teaching has_explicit_quota_phrase/is_explicit_quota_code the "insufficient credits"/"insufficient funds" phrasing) so a real Concentrate 402 lands in QuotaExhausted, not Other.
2. (Unresolved) Public, unauthenticated /v1/models used as key verification — crates/tui/src/client.rs:1818-1821 still treats a Concentrate GET /v1/models 2xx as a successful key-verification probe. Per docs/PROVIDERS.md:670, that endpoint doesn't require auth, so any nonempty (including wrong) key currently "verifies" during guided setup and only fails on the first real POST /responses.
3. (Unresolved) scripts/concentrate-selftest.sh:136-138 still degrades a missing reply-text match to a printed note: instead of failing the script, so a regression that drops all content events (while still emitting done) would still print CONCENTRATE SELFTEST PASS.
4. (Unresolved, minor) docs/PROVIDERS.md:45 still says "the 42 entries of ProviderKind::ALL" — the array is now 43 (matching this same commit's own ProviderKind::ALL.len(), 43 assertion).
None of these are regressions from 689f52a — they were flagged before this push and remain open. The count-bump commit itself is uncontroversial and matches the failing hosted-CI test it targets.
Not a real issue: the earlier Codewhale-agent "hardcoded secret" flags on crates/config/src/tests.rs:4991,5047 (concentrate-config-key, concentrate-env-key, stub-key) are synthetic test fixtures, consistent with the existing convention used for every other provider's tests (e.g. nim-env-key, mistral-config-key, go-config-key, telecom-config-key elsewhere in the same file).
· Branch: feat/concentrate-provider-20260829
Hosted Test (macos-latest) at 689f52a61: 13,964 run, 13,958 passed, 6 failed — all six are the inherited sandbox::read_guard macOS failures on origin/main20e957406 (symlink_chains_resolve_to_the_denied_target, symlink_pointing_into_a_denied_tree_is_refused_by_its_target, symlinked_parent_directory_is_refused, relative_read_from_inside_a_denied_tree_is_refused, dot_dot_through_a_symlink_is_applied_after_symlink_resolution, denial_message_names_the_rule_without_echoing_the_resolved_secret_path). Nothing Concentrate-related failed. #5724 fixes exactly these and is green/CLEAN at c4bbb357; after it lands I'll re-merge main here for a clean exact-head macOS leg. Test (ubuntu-latest) passed at this head.
Matrix settled at 689f52a61: Test (windows-latest) 13,641 run / 13,640 passed, the one failure is the inherited sandbox::read_guard::tests::root_parent_traversal_does_not_escape_above_root from origin/main20e957406 (fixed by #5724). Net at this head: Ubuntu green; macOS 6 inherited failures; Windows 1 inherited failure; Lint / Safety gate / Version drift / Integrations / npm smoke / DCO / GitGuardian / Analyze ×4 / CodeQL / OHOS check / Claude, Devin, Codewhale reviews all green. Zero Concentrate-related failures. Next step after #5724 lands: re-merge main here.
Review any new diff content for correctness/safety/reuse/tests/security
Post final review summary
Summary
This synchronize is the merge commit 823cf7b (Merge remote-tracking branch 'origin/main' into HEAD), following the plan stated in the PR thread ("Next step after #5724 lands: re-merge main here"). I diffed 689f52a (the last commit reviewed) against 823cf7b restricted to the Concentrate-touching paths (crates/tui/src/client.rs, crates/tui/src/error_taxonomy.rs, crates/tui/src/llm_client/mod.rs, scripts/concentrate-selftest.sh, crates/config/src/route/resolver.rs, crates/config/src/provider.rs, crates/config/src/lib.rs, crates/config/src/provider_defaults.rs, docs/PROVIDERS.md) — zero lines of Concentrate-related code changed. Everything that changed is unrelated upstream work pulled in from main: Moonshot CN base-URL/web-search handling (provider.rs, lib.rs, provider_defaults.rs, resolver.rs, a docs/PROVIDERS.md edit to the Moonshot row only), plus large unrelated surfaces (localization, sandbox hardening, subagent gating, TUI views/goldens, cli/src/update.rs).
Since nothing Concentrate-specific moved, the four issues raised in the prior review round (00:04 UTC, against 689f52a) are still open at this head, verified directly against the current file contents:
1. (Unresolved) 402 quota classification never reaches the typed error path — crates/tui/src/llm_client/mod.rs (has_explicit_quota_phrase) still doesn't recognize "insufficient credits"/"insufficient funds"/402, so a real Concentrate 402 still falls through LlmError::from_http_response's match status (no 402 arm) to LlmError::Other, which error_taxonomy.rs maps to Internal/recoverable — the opposite of the quota category the PR claims. The added insufficient_credits_classifies_as_rate_limit_not_auth test only exercises the string heuristic (classify_error_message), not the typed boundary send_with_retry actually uses.
2. (Unresolved) Public, unauthenticated /v1/models still used as key verification — crates/tui/src/client.rs around line 1818 still treats a 2xx from Concentrate's documented-unauthenticated GET /v1/models as a successful key-verification probe, so guided setup "verifies" any nonempty (including wrong) key and only fails on the first real POST /responses.
3. (Unresolved) scripts/concentrate-selftest.sh still degrades a missing reply-text match to a printed note: rather than a failing assertion — a regression that drops all content events while still emitting done would still print CONCENTRATE SELFTEST PASS.
4. (Unresolved, minor) docs/PROVIDERS.md:45 still reads "the 42 entries of ProviderKind::ALL"; the array is 43 as of this same PR's own test assertion. (Confirmed the merge only touched the unrelated Moonshot row in this file, not line 45.)
Not re-raising: the two codewhale-agent "hardcoded secret" blockers on crates/config/src/tests.rs:4991,5047 are synthetic test fixture strings (concentrate-config-key, concentrate-env-key, stub-key), consistent with every other provider's test convention in the same file (e.g. nim-env-key, mistral-config-key) — not real credentials.
No new correctness, provider/route-safety, reuse, test-coverage, or security findings from this merge itself — it's a clean, uneventful re-sync with main.
The reason will be displayed to describe this comment to others. Learn more.
Codewhale review
Adds Concentrate as a first-class, opt-in, BYOK OpenAI Responses-compatible provider across the registry, config, secrets, TUI, docs, web facts, and scripts. The implementation is comprehensive and includes focused tests for wire contract, model passthrough, error classification, and credential scoping. Review found a missing serde alias for concentrateai in the config table definitions and a redundant env override block.
Findings
[WARNING] Missing concentrateai alias on Concentrate config table serde attributes (crates/config/src/lib.rs:475) ProviderKind::Concentrate, the provider aliases() list, and secrets::env_for all accept concentrateai as an alias, but the ProvidersToml.concentrate field in crates/config/src/lib.rs and the ProvidersConfig.concentrate field in crates/tui/src/config.rs only declare concentrate-ai and concentrate_ai. A user selecting provider = "concentrateai" and placing credentials under [providers.concentrateai] will have that table silently ignored (serde default ignores unknown fields), leaving the provider unconfigured. Add the missing alias to both serde attributes and add a regression test that parses [providers.concentrateai].
[INFO] Duplicate base URL override block for Concentrate (crates/tui/src/config.rs:8580)
In crates/tui/src/config.rs, the apply_env_overrides_unlocked function already sets the Concentrate base URL through the provider_env_base_url_override match around line 8316. A second inline if matches!(config.api_provider(), ApiProvider::Concentrate) && ... block starting around line 8580 repeats the same assignment with the same env var. This is redundant and can be removed to reduce maintenance burden.
Suggestions
crates/config/src/lib.rs:475 — Add the concentrateai alias to the ProvidersToml.concentrate serde attribute so [providers.concentrateai] tables are recognized and match the accepted provider aliases.
alias = "concentrate-ai",
alias = "concentrate_ai",
alias = "concentrateai"
crates/tui/src/config.rs:3919 — Add the concentrateai alias to the ProvidersConfig.concentrate serde attribute so the TUI config table accepts the same alias as the provider enum and secrets layer.
#[serde(default, alias = "concentrate-ai", alias = "concentrate_ai", alias = "concentrateai")]
Assessment
The PR is generally high quality with thorough testing and careful adherence to Concentrate's documented contract. The missing concentrateai config alias is a correctness gap that could silently drop user configuration and should be fixed before merge; the redundant env override block is minor cleanup. With those addressed, the addition looks solid.
Advisory review by Codewhale (codewhale review --pr 5725 --post, head 823cf7b1866425884f74cee2ddc805b5bc0f86d2). Line-specific findings are also posted as inline review comments; mechanical fixes arrive as committable suggestions you can apply from the Files tab. CODEOWNERS approval still governs merge.
The reason will be displayed to describe this comment to others. Learn more.
[WARNING] Missing concentrateai alias on Concentrate config table serde attributes
ProviderKind::Concentrate, the provider aliases() list, and secrets::env_for all accept concentrateai as an alias, but the ProvidersToml.concentrate field in crates/config/src/lib.rs and the ProvidersConfig.concentrate field in crates/tui/src/config.rs only declare concentrate-ai and concentrate_ai. A user selecting provider = "concentrateai" and placing credentials under [providers.concentrateai] will have that table silently ignored (serde default ignores unknown fields), leaving the provider unconfigured. Add the missing alias to both serde attributes and add a regression test that parses [providers.concentrateai].
The reason will be displayed to describe this comment to others. Learn more.
Fixed in 0f09427: concentrateai serde alias is on both ProvidersToml.concentrate and ProvidersConfig.concentrate. [providers.concentrateai] now deserializes onto the concentrate table — pinned in concentrate_resolves_named_responses_gateway_and_environment_overrides (1 passed; 0 failed).
The reason will be displayed to describe this comment to others. Learn more.
[INFO] Duplicate base URL override block for Concentrate
In crates/tui/src/config.rs, the apply_env_overrides_unlocked function already sets the Concentrate base URL through the provider_env_base_url_override match around line 8316. A second inline if matches!(config.api_provider(), ApiProvider::Concentrate) && ... block starting around line 8580 repeats the same assignment with the same env var. This is redundant and can be removed to reduce maintenance burden.
The reason will be displayed to describe this comment to others. Learn more.
Add the concentrateai alias to the ProvidersToml.concentrate serde attribute so [providers.concentrateai] tables are recognized and match the accepted provider aliases.
The reason will be displayed to describe this comment to others. Learn more.
Add the concentrateai alias to the ProvidersConfig.concentrate serde attribute so the TUI config table accepts the same alias as the provider enum and secrets layer.
Suggested change
pub edenai:ProviderConfig,
#[serde(default, alias = "concentrate-ai", alias = "concentrate_ai", alias = "concentrateai")]
…ftest
[providers.concentrateai] was silently ignored on both config tables.
GET /v1/models is unauthenticated, so a 2xx must not count as key proof —
the probe is unobserved and health_check issues no request. Missing stub
reply text now fails the selftest instead of exiting 0. Duplicate
CONCENTRATE_BASE_URL assignment removed; retired launch_screen row dropped
after the main merge.
Proven: concentrate_resolves_named_responses_gateway_and_environment_overrides
1 passed; concentrate_health_check_does_not_treat_unauthenticated_models_as_key_proof
plus two sibling Concentrate tests 3 passed; 0 failed.
Co-authored-by: Cursor <cursoragent@cursor.com>
Review new diff content for correctness/safety/reuse/tests/security
Post final review summary
Summary
0f09427 is a focused leftovers commit. I re-verified each item against the current file contents (not just the diff) rather than trusting the commit message.
Fixed, confirmed correct:
concentrateai alias — both ProvidersToml.concentrate (crates/config/src/lib.rs:475-477) and ProvidersConfig.concentrate (crates/tui/src/config.rs:3968-3973) now carry alias = "concentrateai". New regression test crates/config/src/tests.rs:4999-5015 parses [providers.concentrateai] and confirms the key is no longer silently dropped.
Unauthenticated /v1/models no longer counted as key verification — api_provider_skips_models_probe (crates/tui/src/client.rs:1806-1815) now includes ApiProvider::Concentrate, matching the existing DeepseekAnthropic precedent: verify_provider_api_key returns Ok(()) optimistically without a network call, and the separate live-catalog fetch path (client.rs:2525-2544) is untouched, so unauthenticated catalog listing still works. New test concentrate_health_check_does_not_treat_unauthenticated_models_as_key_proof (client.rs:6555-6570) asserts zero requests were recorded.
Selftest silent-degrade — scripts/concentrate-selftest.sh:135-138 now does sys.exit(1) with a stderr message on missing reply text instead of printing a note: and passing.
Duplicate CONCENTRATE_BASE_URL block — removed from crates/tui/src/config.rs:8641-8655; a comment explains provider_env_base_url_override already serves it.
The unrelated launch_screen config-row removal (crates/tui/src/tui/views/mod.rs) is a clean, well-formed edit (verified the surrounding braces directly, not just the diff) — it's leftover fallout from the main merge, not a Concentrate change, but it's correct.
Still open — not touched by this commit (verified against current code, not assumed from the diff):
(Real bug, correctness) 402 quota classification still never reaches the typed HTTP error path.has_explicit_quota_phrase (crates/tui/src/llm_client/mod.rs:759-789) still has no "insufficient credits" / "insufficient funds" / "payment required" phrase, only error_taxonomy.rs:373-376's separate string heuristic (classify_error_message) got those. LlmError::from_http_response (llm_client/mod.rs:462-467) gates its 402 quota check on has_explicit_quota_evidence, which is false for Concentrate's documented body (scripts/concentrate-stub.py:135: {"error":"Insufficient funds","message":"Your account has insufficient credits..."}), so it falls through to _ => LlmError::Other(...) (llm_client/mod.rs:539), which error_taxonomy.rs:325-331 maps to ErrorCategory::Internal, recoverable — the opposite of the quota/rate-limit category the PR claims. I confirmed the real request path uses this same typed constructor: client.rs:2870 and client.rs:2951 both call LlmError::from_http_response_with_retry_after, which delegates to from_http_response (llm_client/mod.rs:605). The insufficient_credits_classifies_as_rate_limit_not_auth test only exercises classify_error_message directly, never this boundary, so it can't catch the regression. Fix: add the "insufficient credits"/"insufficient funds"/402 phrasing to has_explicit_quota_phrase (or add an explicit 402 arm to the match status in from_http_response) so a real Concentrate 402 lands in QuotaExhausted.
(Minor, docs) docs/PROVIDERS.md:45 still reads "the 42 entries of ProviderKind::ALL"; the array is [Self; 43] (crates/config/src/provider_kind.rs:250), matching this same PR's own ProviderKind::ALL.len(), 43 assertion (crates/config/src/tests.rs:5182).
(Minor, test coverage) The tui crate's ProvidersConfig.concentrate alias fix has no direct regression test parsing [providers.concentrateai] under that crate — only the config crate got one. Not blocking; the fix itself is a mechanical, symmetric serde attribute.
None of these three are regressions introduced by 0f09427 — items 1 and 2 have been flagged across three prior review rounds (two Codex, one Claude) and remain untouched; item 3 is new-but-minor observation from this pass.
Not a real issue (repeating from prior rounds for completeness): the codewhale-agent "hardcoded secret" blockers on crates/config/src/tests.rs are synthetic test fixtures (concentrate-config-key, concentrate-alias-table-key, stub-key), consistent with the file's existing convention for every other provider.
The reason will be displayed to describe this comment to others. Learn more.
Stale comment
Left a non-blocking comment and assigned reviewers. Cursor Bugbot was present on this head but completed as skipped (usage limit, incomplete analysis), so this update is not approved and needs human review.
Sent by Cursor Approval Agent: Pull Request Router and Approver
The reason will be displayed to describe this comment to others. Learn more.
🟡 Changes recommended
docs/PROVIDERS.md still states ProviderKind::ALL has 42 entries even though this PR adds Concentrate (now 43), which should be corrected for accuracy.
Once you've addressed the issues Copilot identified, you can request another Copilot review.
Pull request overview
Adds Concentrate (concentrate) as a first-class, opt-in BYOK provider wired through Codewhale’s existing provider registry/config/secrets/TUI/client plumbing, using a fixed OpenAI Responses wire against https://api.concentrate.ai/v1.
Changes:
Introduces ProviderKind::Concentrate (and aliases), defaults, secret/env handling, route resolution behavior (pass-through model ids; strip only concentrate/).
Implements Concentrate-specific Responses request shaping (documented-field subset, system prompt as leading system input item) plus error taxonomy improvements (flat {error,message} surfacing; 402 → quota).
Adds docs + web facts updates and a keyless local stub + self-test script for end-to-end verification without hitting the live gateway.
File summaries
File
Description
web/scripts/facts-lib.mjs
Adds Concentrate to provider label/env map used by web facts generation.
web/lib/facts.generated.ts
Regenerates facts to include Concentrate and updated generation timestamp.
web/lib/facts-drift.ts
Updates drift checker’s provider derivation mapping for Concentrate.
scripts/concentrate-stub.py
Adds a local HTTP stub implementing the documented Concentrate surface for offline testing.
scripts/concentrate-selftest.sh
Adds an end-to-end self-test that drives the real CLI against the local stub.
scripts/check-provider-registry.py
Extends registry consistency checks to include Concentrate’s handwritten Provider impl.
docs/public-surface-facts.json
Updates public-surface provider count for the new provider.
docs/PROVIDERS.md
Documents Concentrate provider selection, env vars, defaults, and operational/commercial notes.
docs/CONFIGURATION.md
Documents Concentrate config/env overrides and default wire behavior.
crates/tui/src/tui/views/mod.rs
Adds clarification about a retired setting behavior in the config view area touched by the PR.
crates/tui/src/tui/ui/session_state.rs
Ensures saved API key mirroring supports Concentrate provider config.
The reason will be displayed to describe this comment to others. Learn more.
Codewhale review
Adds Concentrate as an opt-in BYOK OpenAI Responses gateway across provider registry, config, secrets, TUI, docs, web facts, and scripts. The change is extensive and well-tested, but a stale docs count, a broad error-formatting change, and missing CI wiring should be addressed.
Findings
[WARNING] PROVIDERS.md canonical provider count is stale (docs/PROVIDERS.md:51)
The prose above the provider list still says "The canonical provider IDs are the 42 entries of ProviderKind::ALL", but ProviderKind::ALL now has 43 entries after adding Concentrate. This is a user-facing documentation drift.
[WARNING] Flat error body extraction is global, not Concentrate-scoped (crates/tui/src/llm_client/mod.rs)
extract_json_error_message now returns error: message for every provider that returns a flat {error, message} body, before checking nested OpenAI error paths. This intentionally changes user-facing error text for providers other than Concentrate, but the new tests only cover the Concentrate/fixture cases. Add cross-provider regression coverage or constrain the new behavior to avoid unintended error-format regressions.
[INFO] Concentrate self-test is not wired into CI (scripts/concentrate-selftest.sh:1)
scripts/concentrate-selftest.sh is a valuable keyless contract test that drives the real codewhale exec path through a local stub, but it is only run manually. Without CI integration, wire/body contract drift can go unnoticed.
Suggestions
docs/PROVIDERS.md:51 — Update the canonical provider count from 42 to 43 to match ProviderKind::ALL after the Concentrate provider was added.
The canonical provider IDs are the 43 entries of `ProviderKind::ALL`
Assessment
Solid provider addition with strong test coverage and careful BYOK credential scoping. The main actionable fix is the stale documentation count; the global error formatter change should either be scoped or get broader regression coverage, and the self-test should ideally run in CI.
Advisory review by Codewhale (codewhale review --pr 5725 --post, head 0f0942740e6935f2ecbf44e9e08055e6184c58a9). Line-specific findings are also posted as inline review comments; mechanical fixes arrive as committable suggestions you can apply from the Files tab. CODEOWNERS approval still governs merge.
The reason will be displayed to describe this comment to others. Learn more.
[WARNING] PROVIDERS.md canonical provider count is stale
The prose above the provider list still says "The canonical provider IDs are the 42 entries of ProviderKind::ALL", but ProviderKind::ALL now has 43 entries after adding Concentrate. This is a user-facing documentation drift.
The reason will be displayed to describe this comment to others. Learn more.
[INFO] Concentrate self-test is not wired into CI
scripts/concentrate-selftest.sh is a valuable keyless contract test that drives the real codewhale exec path through a local stub, but it is only run manually. Without CI integration, wire/body contract drift can go unnoticed.
Check what changed in the newest commit (e4e8174, "Merge origin/main into the Concentrate provider slice") relative to the last reviewed head (0f09427)
Verify Hmbown's 13:00 claim that the 402-quota typed-path issue is "already classified — no code change" against current source
Verify docs/PROVIDERS.md count and other previously-flagged items are still open
Post final review summary
Summary
e4e8174 is a pure sync merge from origin/main (git diff 0f09427..e4e8174 restricted to every Concentrate-touching path — client.rs, error_taxonomy.rs, llm_client/mod.rs, concentrate-selftest.sh, route/resolver.rs, provider.rs, lib.rs, provider_defaults.rs, docs/PROVIDERS.md, tests.rs, tui/config.rs — is empty). Everything in the 6,736-line diff is unrelated main work (Operate/OperateRecord TUI, #5703, automation routing, history, phase strip, etc.). Nothing Concentrate-specific to review here.
However, re-verifying the open items against current source turned up a correction to Hmbown's 12:59:41 reply:
1. (Still a real bug) The 402-quota typed-path issue is not resolved — the 12:59:41 reply cited the wrong function.
Hmbown's reply pointed at error_taxonomy.rs:370-376's classify_error_message, but that function operates on an already-formatted display string and is never consulted for LlmError::Other. I traced the actual real-request path end to end at the current head:
client.rs:2870 / client.rs:2951 construct errors via LlmError::from_http_response_with_retry_after, which delegates to LlmError::from_http_response (llm_client/mod.rs:462).
from_http_response only special-cases 402 through has_explicit_quota_evidence (llm_client/mod.rs:463), which checks: a JSON /error/code|/type|... pointer against is_explicit_quota_code (llm_client/mod.rs:729-744, a fixed code list — no insufficient_credits/insufficient_funds), a "provider error code:" marker, or has_explicit_quota_phrase (llm_client/mod.rs:759-789). I grepped the whole file for insufficient|payment required|http 402 — the only hit is the normalized code "insufficientquota" inside is_explicit_quota_code, which requires a code/type JSON field. None of these match Concentrate's documented flat body.
Concentrate's actual 402 body (scripts/concentrate-stub.py:135) is {"error": "Insufficient funds", "message": "Your account has insufficient credits..."} — a flat two-string object with no code/type field at all, so has_explicit_quota_evidence returns false.
With no explicit 402 arm in the match status block (llm_client/mod.rs:469-540), it falls to _ => LlmError::Other(...) (line 539), which error_taxonomy.rs:325-331 maps straight to ErrorCategory::Internal, recoverable: true — the opposite of the quota category the PR claims.
The test Hmbown cited, explicit_400_402_and_429_quota_responses_are_typed_and_non_retryable (crates/tui/src/llm_client/tests.rs:97-128), only exercises a structured 402 body ({"error":{"code":"billing_hard_limit_reached","message":"Payment required"}}) — an OpenAI/Anthropic-style nested shape, not Concentrate's flat one — so it can't catch this. The phrases added at error_taxonomy.rs:373-376 ("insufficient credits"/"insufficient funds"/"payment required"/"http 402") live in classify_error_message, a different, later-stage string classifier that from_http_response never calls.
Net: a real Concentrate 402 still lands in LlmError::Other → Internal/recoverable, not QuotaExhausted. Fix (as flagged across five prior review rounds) is to teach has_explicit_quota_phrase the flat-body wording, or add an explicit 402 arm to from_http_response's match status, so the quota check actually gates on the phrasing Concentrate returns.
2. (Still open, minor) docs/PROVIDERS.md:45 still reads "the 42 entries of ProviderKind::ALL"; the array is [Self; 43]. Flagged by three tools now (Claude, Codex/Copilot, Codewhale) across six rounds and still unfixed — a one-line change.
3. Not re-raised: the repeating codewhale-agent "hardcoded secret" blockers are synthetic test fixtures (concentrate-config-key, concentrate-env-key, concentrate-alias-table-key, stub-key), consistent with every other provider's test convention in the same file.
All other items from prior rounds (concentrateai alias, unauthenticated /v1/models no longer used as key proof, selftest silent-degrade, duplicate base-URL override block) remain fixed as of 0f09427 and untouched by this merge.
The reason will be displayed to describe this comment to others. Learn more.
Stale comment
Left a non-blocking comment and assigned reviewers. Cursor Bugbot was present on this head but completed as skipped, so this update is not approved and needs human review.
Sent by Cursor Approval Agent: Pull Request Router and Approver
The reason will be displayed to describe this comment to others. Learn more.
Codewhale review
Adds Concentrate as a first-class BYOK OpenAI Responses gateway across the registry, config, secrets, TUI, docs, scripts, and goldens. The change is largely additive and well-tested, with a couple of minor alias/documentation gaps.
Findings
[WARNING] Concentrate provider alias concentrate-ai is not recognized in subagent provider key matching (crates/tui/src/config.rs:514)
In crates/tui/src/config.rs, subagent_provider_key_matches lowercases the key but only matches "concentrate" | "concentrate_ai" | "concentrateai". Serde accepts provider = "concentrate-ai" for this provider, so a subagent using that alias may be treated as a different provider. Add the hyphenated alias.
[INFO] PROVIDERS.md provider count still says 42 (docs/PROVIDERS.md)
The introduction in docs/PROVIDERS.md still reads "The canonical provider IDs are the 42 entries of ProviderKind::ALL", but the enum now has 43 catalog entries. Update the count to 43 to avoid misleading readers.
Suggestions
crates/tui/src/config.rs:514 — Add concentrate-ai to the recognized subagent keys so the serde alias is honored here too.
Well-structured additive provider implementation with comprehensive tests. The identified issues are minor and should be fixed before merge, but the core Responses-wire path, credential scoping, catalog, and error classification are solid.
Advisory review by Codewhale (codewhale review --pr 5725 --post, head e4e817449943fdfbd986377d1ac5b6b96f3a613b). Line-specific findings are also posted as inline review comments; mechanical fixes arrive as committable suggestions you can apply from the Files tab. CODEOWNERS approval still governs merge.
The reason will be displayed to describe this comment to others. Learn more.
[WARNING] Concentrate provider alias concentrate-ai is not recognized in subagent provider key matching
In crates/tui/src/config.rs, subagent_provider_key_matches lowercases the key but only matches "concentrate" | "concentrate_ai" | "concentrateai". Serde accepts provider = "concentrate-ai" for this provider, so a subagent using that alias may be treated as a different provider. Add the hyphenated alias.
Check what changed in the latest commit (9af679b, "Merge remote-tracking branch 'origin/main'") relative to the last reviewed head (e4e8174)
Re-verify previously flagged unresolved issues (402 typed-path quota classification, docs count, new subagent-alias finding, global error-body extraction)
Review any new diff content for correctness/safety/reuse/tests/security
Post final review summary
Summary
git diff e4e8174..HEAD --stat shows only CHANGELOG.md, crates/tui/CHANGELOG.md, and crates/tui/src/automation_manager.rs (5 lines, unrelated automation-run-file fix from main) changed — zero Concentrate-related code moved in this sync. So this round is re-verification against current source, not a new-diff review.
1. (Still open, real bug) 402 quota classification never reaches the typed HTTP error path. Traced end-to-end again at the current head:
has_explicit_quota_phrase (crates/tui/src/llm_client/mod.rs:759-789) still has no "insufficient credits"/"insufficient funds"/"payment required"/bare-402 phrasing.
explicit_quota_code (llm_client/mod.rs:713-726) only matches quota codes nested under /error/code, /error/type, /code, /type, etc. — it can't match Concentrate's flat body where /error itself is the string "Insufficient funds" (scripts/concentrate-stub.py:135), because a string has no /error/code child to point into.
So has_explicit_quota_evidence (llm_client/mod.rs:707-711) returns false, from_http_response's matches!(status, 400 | 402 | 429) && has_explicit_quota_evidence(body) guard (llm_client/mod.rs:463) is skipped, and with no 402 arm in the match status block (llm_client/mod.rs:469-538) it falls to _ => LlmError::Other(...) (line 538) → error_taxonomy.rs maps that to ErrorCategory::Internal, recoverable — the opposite of quota.
The only 402 regression test, explicit_400_402_and_429_quota_responses_are_typed_and_non_retryable (crates/tui/src/llm_client/tests.rs:97-128), still exercises a nested OpenAI/Anthropic-shaped 402 body ({"error":{"code":"billing_hard_limit_reached",...}}), never Concentrate's flat {"error":"...","message":"..."} shape, so it can't catch this. This is the same conclusion reached in the prior review round (which corrected an earlier claim from the PR author that this was already handled) — still accurate at this head. Fix: teach has_explicit_quota_phrase the flat-body wording, or add an explicit 402 arm to from_http_response.
2. (Still open, minor) docs/PROVIDERS.md:45 still reads "the 42 entries of ProviderKind::ALL"; the array is [Self; 43] (crates/config/src/provider_kind.rs:250), matching this PR's own assert_eq!(ProviderKind::ALL.len(), 43) (crates/config/src/tests.rs:5182). Flagged by four tools across seven rounds now.
3. Correcting a false positive from the 17:50:55 codewhale-agent review, which claimed subagent_provider_key_matches (crates/tui/src/config.rs:516-519) doesn't recognize "concentrate-ai". It does: normalize_subagent_provider_key (config.rs:478-488) maps -/_/./ all to _ before comparison, so "concentrate-ai" normalizes to "concentrate_ai", which is already in the match arm at line 518. No fix needed here.
4. Minor, non-blocking observation on the global error-body extraction change (crates/tui/src/llm_client/mod.rs:791-805, flagged by codewhale-agent at 17:19:50 as provider-unscoped): the new flat error+message branch only fires when both are top-level JSON strings and differ — it can't match the pre-existing nested {"error":{"message":...}} shape (Value::as_str() on an object returns None), so it's low risk for other providers even though it isn't Concentrate-scoped. Worth a cross-provider regression test if this is touched again, but not a defect as written.
Not re-raised: the repeating codewhale-agent "hardcoded secret" blockers on crates/config/src/tests.rs remain synthetic test fixtures, consistent with the file's convention for every other provider.
Net: nothing regressed by this merge; item 1 (402 typed-path quota misclassification) is the one actionable, unresolved correctness bug across all seven review rounds and is worth fixing before merge. Item 2 is a one-line docs fix.
The reason will be displayed to describe this comment to others. Learn more.
Left a non-blocking comment and assigned reviewers. Cursor Bugbot was present on this head but completed as skipped (usage limit, incomplete analysis), so this update is not approved and needs human review.
Sent by Cursor Approval Agent: Pull Request Router and Approver
The reason will be displayed to describe this comment to others. Learn more.
Codewhale review
Adds Concentrate as an opt-in, BYOK, OpenAI Responses-compatible provider across registry, config, secrets, TUI, docs, and scripts. The change is mostly additive and well-tested, but a few consistency and contract risks remain.
Findings
[WARNING] Subagent provider key matcher misses the concentrate-ai alias (crates/tui/src/config.rs:518)
The provider serde aliases include concentrate-ai, and secrets::env_for recognizes it, but subagent_provider_key_matches only matches concentrate, concentrate_ai, and concentrateai. A subagent table using [providers.concentrate-ai] would not be treated as belonging to the Concentrate provider, potentially causing configuration to be ignored or misrouted.
[INFO] Stale provider count in docs/PROVIDERS.md (docs/PROVIDERS.md)
The canonical provider list still says 'The canonical provider IDs are the 42 entries of ProviderKind::ALL', but ProviderKind::ALL now has 43 entries after adding concentrate.
[WARNING] Concentrate request body may still include fields outside the documented parameter reference (crates/tui/src/client/responses.rs)
The PR summary says only model, input, stream, max_output_tokens, tools/tool_choice/parallel_tool_calls, and reasoning.effort are documented for Concentrate. However, build_responses_body_for_provider only suppresses store, include, instructions, and reasoning.summary. Generic Responses fields such as temperature, top_p, text, routing, cache_control, and prompt_cache_options can still be emitted if present in the resolved request. The stub and unit test whitelist these fields as documented, but the PR description omits them, so either the live gateway does not accept them and the builder should filter them, or the contract description is incomplete.
[WARNING] TUI credential-scoping rule for custom Concentrate base URLs may not be enforced or tested (crates/tui/src/config.rs:8646)
The config-crate test verifies that a saved or environment Concentrate key is never sent to a non-official base URL unless both base_url and api_key are set in [providers.concentrate]. The TUI apply_env_overrides_unlocked does not appear to implement the same clearing rule for CONCENTRATE_API_KEY when CONCENTRATE_BASE_URL points elsewhere, and no TUI test covers that scenario. A TUI session with the env key and a custom base URL could therefore send the bearer key to an arbitrary endpoint.
Suggestions
crates/tui/src/config.rs:518 — Add the hyphen alias so concentrate-ai provider tables are recognized by subagent key matching.
The provider addition is well structured and extensively tested, but the subagent alias omission should be fixed, the request-body field contract should be reconciled with the implementation, and the TUI credential-scoping behavior should be verified before merge.
Advisory review by Codewhale (codewhale review --pr 5725 --post, head 9af679b2f41fb0313f689c6f5d71fef342af2732). Line-specific findings are also posted as inline review comments; mechanical fixes arrive as committable suggestions you can apply from the Files tab. CODEOWNERS approval still governs merge.
The reason will be displayed to describe this comment to others. Learn more.
[WARNING] Subagent provider key matcher misses the concentrate-ai alias
The provider serde aliases include concentrate-ai, and secrets::env_for recognizes it, but subagent_provider_key_matches only matches concentrate, concentrate_ai, and concentrateai. A subagent table using [providers.concentrate-ai] would not be treated as belonging to the Concentrate provider, potentially causing configuration to be ignored or misrouted.
The reason will be displayed to describe this comment to others. Learn more.
[WARNING] TUI credential-scoping rule for custom Concentrate base URLs may not be enforced or tested
The config-crate test verifies that a saved or environment Concentrate key is never sent to a non-official base URL unless both base_url and api_key are set in [providers.concentrate]. The TUI apply_env_overrides_unlocked does not appear to implement the same clearing rule for CONCENTRATE_API_KEY when CONCENTRATE_BASE_URL points elsewhere, and no TUI test covers that scenario. A TUI session with the env key and a custom base URL could therefore send the bearer key to an arbitrary endpoint.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
![codewhale-agent[bot]](https://avatars.githubusercontent.com/u/101357273?s=60&v=4){kind=small}
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
Summary
Adds Concentrate (
concentrate) as a first-class, opt-in, BYOK provider — the OpenAI Responses-compatible gateway athttps://api.concentrate.ai/v1— inside the existing provider authorities (no parallel secret store, router, or runtime). Mirrors the Eden AI aggregator addition across the registry, config tables, env overrides, secrets, TUI wiring, docs, web facts, and goldens.ProviderKind::Concentrate(+concentrate-ai/concentrate_ai/concentrateai), hand-writtenimpl ProviderwithWirePolicy::Fixed(Responses), default modeldeepseek-v4-pro, envCONCENTRATE_API_KEY/CONCENTRATE_BASE_URL/CONCENTRATE_MODEL, own secret-store slot.provider/modelpins it; only the gateway's ownconcentrate/namespace is stripped soconcentrate/autoreaches its router (Codewhale's bareautostays the resolver sentinel).model,input,stream,max_output_tokens,tools/tool_choice/parallel_tool_calls,reasoning.effort); system prompt as a leadingsysteminput item; typedresponse.*SSE via the existing parser, ending onresponse.completed(no[DONE]).GET /v1/modelsthrough the existing named-gateway live-catalog path (rows provider-scoped, unclaimed).{"error","message"}bodies surface both halves.base_urlandapi_keyare both in[providers.concentrate]). Hosted-lane activation stays gated on written consent, terms, and billing approval (ops evidenceconcentrate-gateway-20260829/CHECKLIST.md).scripts/concentrate-selftest.shbootsscripts/concentrate-stub.py(the documented contract on loopback) and drives the realcodewhale exec --auto --output-format stream-jsonpath through it, asserting URL, bearer header, verbatim model, only-documented fields, system-first input, the completed-turn receipt, and the wrong-key 401. No network call leaves the machine.Contract sources (fetched 2026-08-29): introduction, request parameters, streaming, errors, list models, health.
No-Issue: founder-requested provider addition; no tracking issue exists.
Testing
Local, hosted-equivalent flags (
RUSTFLAGS=-Dwarnings,RUST_MIN_STACK=16MiB,cargo nextest … --all-features --locked --profile ci), summary in the ops evidence:cargo fmt --all -- --check: cleancodewhale-config: 624 run / 624 passed / 1 skipped — incl. newconcentrate_resolves_named_responses_gateway_and_environment_overrides(aliases, metadata, fixed Responses wire, secret slot, env + config resolution, and the credential-scope rule that an env key is never sent to a non-official base URL) andconcentrate_passes_ids_through_and_strips_only_its_own_namespace(plain /provider/modelverbatim,concentrate/auto→auto, bareauto→ provider default, all on the Responses protocol); goldens regenerated (golden_route_ids.txt,providers-export.golden.json); count assertions 47→48 / 42→43; expected-wire tables updated.codewhale-secrets: 62 run / 62 passed — incl.concentrate_env_aliases_resolve(own slot; no cross-provider bleed).codewhale-tuifocused (concentrate | error_taxonomy:: | client::responses::tests:: | llm_client:: | config provider tests | edenai|telecomjs|opencode_zen_responses|catalog): 440 run / 440 passed — incl.concentrate_responses_request_matches_the_documented_contract(wiremock:POST /v1/responses,Bearerheader, verbatimprovider/model, nostore/include/instructions/messages, system item first, typed-event SSE without[DONE]assembles text + usage 12/5, official URL maps to/v1/responses),concentrate_error_bodies_surface_verbatim_and_classify(401 → Authentication with "Invalid API key", 402 → quota with "insufficient credits", 400 → InvalidInput with "Invalid model name"),concentrate_live_catalog_is_provider_scoped_and_marks_the_default,concentrate_responses_body_sends_only_documented_fields,insufficient_credits_classifies_as_rate_limit_not_auth,flat_error_and_message_body_surfaces_both_halves.scripts/concentrate-selftest.shwith the debug CLI):concentrate/auto(wireauto),openai/gpt-5.6-sol,deepseek-v4-pro— health 200, catalog without a key, exactly one bearer-authenticatedPOST /v1/responseswith the verbatim model,stream:true, system item first, only documented fields; stream-json receiptcontent…turn_usage, session_capture, metadata, done; wrong key → documented 401 → exit 1. Defeat evidence: the first run (env key only) reached the stub with an empty Authorization header, which is the credential-scope rule working; the recipe now writesbase_url+api_keyinto[providers.concentrate]as a BYOK user pointing at a local gateway would.scripts/check-provider-registry.pyPASS;cargo clippyfor config+secrets and tui (--all-targets --all-features, CI allow list) clean;node web/scripts/derive-facts.mjs(providers 46) +check-factsOK; dead-code budget PASS (444/444);git diff --checkclean.Not done: no live Concentrate canary (needs a key and founder-gated spend — the live gateway was never contacted beyond the public, unauthenticated
/v1/modelsread used to pin the catalog shape); Windows-target compile is hosted-CI-only on this macOS host.cargo fmt --all -- --checkcargo clippyon the changed crates, all targets, all features, CI allow listcargo test --workspace --all-features --locked(hosted CI; locally the changed crates' suites passed as above)🤖 Generated with Claude Code
Note
Medium Risk
Touches shared provider registry, route resolution, Responses request shaping, and credential scoping across config/TUI/secrets; mistakes could mis-route keys or send wrong wire fields, but scope is additive with extensive tests and no managed keys shipped.
Overview
Adds Concentrate as an opt-in, BYOK catalog provider (
concentrate) wired through the same registry, config, secrets, route contract, and TUI paths as other gateways—not a parallel stack.The provider is fixed on the OpenAI Responses wire (
POST /v1/responses) with defaultshttps://api.concentrate.ai/v1anddeepseek-v4-pro, auth viaCONCENTRATE_API_KEYand[providers.concentrate](plus aliases). The route resolver passes model ids through verbatim and strips only theconcentrate/prefix soconcentrate/autohits the gateway router while Codewhale’s bareautostays the provider-default sentinel.TUI/runtime behavior is tailored to Concentrate’s documented API: Responses bodies omit
store,include, and top-levelinstructions(system prompt as a leadingsysteminput item); Chat Completions reasoning fields are not injected; unauthenticatedGET /v1/modelsdrives live catalog refresh but does not count as API-key verification. Env/config keys are scoped to the official base URL unless bothbase_urlandapi_keyare set in config for a custom endpoint.Supporting changes include golden route/provider export updates, docs and web facts, registry check allowance for a hand-written
Concentrateprovider, error handling for flaterror+messageJSON and HTTP 402 “insufficient credits” as quota, and keyless loopback dogfood viascripts/concentrate-stub.py+scripts/concentrate-selftest.sh.Reviewed by Cursor Bugbot for commit 9af679b. Bugbot is set up for automated code reviews on this repo. Configure here.