providers-config.json
Discovery URL: https://ccs.the1studio.org/providers
Providers
Section titled βProvidersβRaw JSON (ground truth)
{ "_comment": "Model catalog for model-router. FLAT since 2026-10-21 (USER DECISION): there is no backend dimension left. Every model is served by the single `endpoint` below through one Anthropic-compatible surface, routed by model id β the old <retired-block-key>.<name>.models.<id> nesting existed to let hops name different hosts, and no hop has for some time. Backend-block-level `budget`, `standbyOnly`, `_disabled_reason`, `type` and `enabled` are gone with the map: `budget` was live code whose every daily limit was null (never fired), and `standbyOnly` described a backend block that was enabled:false and absent from the pipe. `_mergedFrom` was REMOVED 2026-10-28 (#1185): it recorded which old backend block(s) contributed a model id, and every display site joined it with `+` and printed the two retired block names joined by `+` as if the router had observed who served a request. It never was β the wire payload is `{model}` + auth + session headers, so the gateway alone knows the serving upstream. A flat catalog has ONE transport, and `lib-catalog.cjs backendLabel()` reports it (`ccs`) for every model; real per-request attribution would need the gateway's `x-cpa-trace-id` / `/cache-rate` surface (separate follow-up).", "auth": "gh-token", "capabilityPipes": { "_comment": "REPOINTED 2026-09-24 (id correction): `deepseek-flash` is RETIRED (enabled:false, supersededBy deepseek-v4.1-flash) β the id never existed in opencode-go's models.dev namespace and the gateway silently resolved it to a V4.0-class model. Every pipe position and every comment reference below that still says `deepseek-flash` means what is now `deepseek-v4.1-flash`. | Ordered preference for capability-gated selection, consumed by mr-task-interceptor.cjs pickFromCandidates(): when a task requires a capability listed here, ENABLED candidates are sorted by their index in that pipe (models not listed fall back to the tier/quality rules; disabled models are simply never candidates). vision pipe = value-first order from the empirical benchmark 2026-06-01 (all 4 scored 3/3 correct; cheap+diverse-upstream first). Shipped-enabled vision models are gpt-5.4-mini + kimi-k2.6; gpt-5.4/gpt-5.5 are listed for when consumers enable them. glm-5.1 deliberately excluded (0/3, confabulates). image-generation is NOT consumed by pickFromCandidates() (#92 β no detectRequiredCapabilities() path ever adds 'image-generation' to requiredCaps, and every model here deliberately omits 'text' from its own capabilities, so an ordinary Task delegation can never route here regardless); it is read directly by mr-imagine.sh as the default --model when the caller doesn't pass one. Order is verification-strength, not a benchmark (none run yet, unlike vision) -- gpt-image-1.5 has two independent live confirmations (2026-08-06, 2026-08-18) against gpt-image-2's one; grok-imagine-image(-quality) are listed disabled for when a consumer re-enables them. UPDATED 2026-09-11: `deepseek-flash` gained `vision` from a live measurement (see its own _comment β two images HTTP 200, exact transcription, 2.0s/3.1s) and took the pipe's FIRST position. It displaces kimi-k2.6 to #2, which is now an enabled-but-not-a-routing-target backstop rather than the primary. Rationale is the same value-first rule the pipe already used: identical correct answer on both images, roughly half kimi-k2.6's latency, on the model that is already the shipped default cheap target β so a vision task and a text task land on the same backend. The moonshot pool's CCS weekly usage-limit wall (kimi-k2.6's own _comment) makes that reordering strictly safer than the reverse. NOTE this does NOT trip the one-per-family whitelist: deepseek-flash and kimi-k2.6 are different families, and the operator constraint that excludes the GPT-family entries is untouched. UPDATED 2026-09-07: gpt-5.4-mini disabled (retired server-side, see its own _comment) and kimi-k2.6 re-enabled + moved to the FIRST position (was 2nd) so the pipe's declared preference order does not lead with a disabled model β tests/test-mr-delegate-contract.sh Contract 46 enforces this. kimi-k2.6 is now the SOLE enabled vision candidate (gpt-5.4/gpt-5.5 both stay disabled per the same USER DECISION β see t1k-config-mr.json failover._pipeComment). `long-context` pipe gained kimi-k2.6 as its new FIRST entry, ahead of the now-disabled deepseek-v4-flash β order matters here (pipeForCaps is consulted before the context_window DESC sort, see deepseek-v4-flash's own _comment); kimi-k2.6's 256K window makes it the effective long-context target now, also the sole enabled candidate. CORRECTED 2026-09-13 β the preceding two claims were BOTH false when written, and the error was load-bearing: `deepseek-flash` is enabled:true with a 1000000 context window and declares `long-context` in its own capabilities, so kimi-k2.6 was neither the effective target nor the sole enabled candidate. The sentence conflated `deepseek-flash` (DeepSeek V4.1 Flash, enabled, 1M) with the RETIRED `deepseek-v4-flash` (DeepSeek V4.0 Flash, enabled:false, supersededBy deepseek-flash) β an easy conflation to make, because the id that LOOKS versioned is the older one. deepseek-flash now takes the FIRST position: it more than quadruples the usable window (1M vs 256K) on the axis this pipe exists to serve, and it is the shipped default target already carrying ~99.5% of routed traffic. WHY THIS MATTERED, not merely tidiness: pipeForCaps() is consulted BEFORE the context_window DESC sort, and detectRequiredCapabilities() adds `long-context` on any prompt over 50000 chars (mr-task-interceptor.cjs:1368) β a routine size for a large brief. So while kimi-k2.6 led this pipe, every long prompt routed FIRST into the quota-exhausted moonshot pool, and the 2026-09-08 failover.pipe repoint did NOT close that path because capability selection runs ahead of the static pipe. kimi-k2.6 remained enabled at #2 as a backstop at the time of that correction, consistent with its own _comment's \"catalog-eligible but NOT a routing target\" note β that is no longer true as of 2026-09-18, see the RETIRED paragraph below. RETIRED 2026-09-18 (user decision, this change): `kimi-k2.6` is removed from BOTH this vision pipe and `long-context` β it is now `enabled: false` with `supersededBy: deepseek-flash`, and a retired id left in a pipe is worse than dead config: the pipe's declared ORDER is a claim about preference, and an entry the router can never candidate (or, before this change, could candidate only to have every hop die on the moonshot pool's CCS weekly usage-limit wall) makes that order a lie. The vision pipe is therefore deepseek-flash FIRST with no second live route: gpt-5.4-mini / gpt-5.4 / gpt-5.5 are all GPT-family and stay disabled under the standing operator constraint, so in practice deepseek-flash must carry vision ALONE. They remain LISTED on purpose and are NOT the defect this change targets β they are disabled but NOT retired (no `supersededBy`), so a consumer who re-enables one gets the pipe order already accounted for; the new assertion in tests/test-capability-pipe-retired-models.sh fails only on a RETIRED id (non-empty `supersededBy`) appearing in any pipe, never on a merely-disabled forward-compat listing like these or `minimax-m3`. | PRUNED 2026-10-22 to the two-model catalog: `image-generation` was DELETED from this object (and its _comment with it) because no surviving model declares that capability, so nothing could ever satisfy it; leaving an empty [] would have advertised a capability nothing serves. `vision` and `long-context` each reduced to their single surviving entry, deepseek-v4.1-flash. Every statement above about image-generation, gpt-image-*, grok-imagine-*, kimi-k2.6, minimax-m3 and the GPT-family listings describes models this catalog NO LONGER SHIPS and is retained only as history.", "long-context": [ "deepseek-v4.1-flash" ], "vision": [ "deepseek-v4.1-flash" ] }, "discoveryUrl": "https://ccs.the1studio.org/providers", "endpoint": "https://ccs.the1studio.org", "models": { "deepseek-v4.1-flash": { "_comment": "ADDED 2026-09-25 as `cc-deepseek-v4.1-flash`, RENAMED 2026-09-28 to the gateway's pooled id. DeepSeek V4.1 Flash β the studio default cheap/reasoning/tool-use target (modelMapping sonnet/haiku, failover.pipe hop 1, capabilityPipes.long-context head). The gateway now serves ONE id for this model across both upstreams (Command Code and OpenCode Go) and picks the upstream per request from its per-key quota weights, so the backend dimension no longer distinguishes the two routes; opencode-go declares the same id for that reason.", "capabilities": [ "text", "reasoning", "tool-use", "long-context", "vision" ], "context_window": 1000000, "enabled": true, "tier": "budget" }, "glm-5.3-flash": { "_comment": "Live-verified 2026-09-15: HTTP 200 via both localhost:8316 and public ccs.the1studio.org, real completion with usage tokens. Tool-use CONTRACT: PASS 10/10, 0 violations. context_window from https://docs.z.ai/guides/llm/glm-5.3-flash (verified GET 200): 1M context, 128K max output. USER DECISION: the MAIN model for this backend β paired with opencode-go/deepseek-flash in the 2026-09-15 quota-weighted loadBalance split (t1k-config-mr.json loadBalance._weightNote). Tier matches deepseek-flash's 'budget' role. REMOVED FROM loadBalance.targets 2026-09-24 for latency (measured 30.7 s/turn p50 vs deepseek-flash 10.1 s/turn) β the model stays enabled in this catalog; see t1k-config-mr.json loadBalance.targets[0]._glmRemovedComment for the numbers and the re-add condition. | ROLE TARGET 2026-10-09 (plans/261008 phase 8): enabled with routable:false so only modelRouter.roles can name it, never tier or loadBalance selection; video role chain[1] (native video_url 4 of 4 full score, p50 2.8 s). The `video` tag is not in any capabilityPipes entry.", "capabilities": [ "text", "tool-use", "video" ], "context_window": 1000000, "enabled": true, "tier": "budget" } }}