Chimera — Dynamic Multi-Model Deliberation Gateway

One API call. A team of models. One answer.

Chimera takes your prompt, dispatches it to a hand-picked team of LLMs (each with a custom subtask scoped to their strengths), and an aggregator merges their outputs using dispatcher-written instructions. One dispatcher model call designs the entire deliberation at once.

Quickstart

# Install (choose one)
pip install chimera-deliberation[full]        # full CLI + server in a venv
pipx install chimera-deliberation[full]       # same, isolated in its own env
pip install chimera-deliberation[server]      # API server only

# Configure — the repo ships NO chimera.yaml. Your live config is local-only
# (gitignored, never committed), so a fresh clone has none — create it:
chimera config init                           # copies chimera.yaml.example → chimera.yaml
# (from a repo checkout: cp chimera.yaml.example chimera.yaml works too)
# Add your API keys (at minimum: DEEPSEEK_API_KEY)

# Run
chimera "What is the capital of France?"      # CLI deliberation
chimera run "Compare React and Vue"           # explicit `run` subcommand (same as above)
chimera --quiet "Compare React and Vue"       # stdout = the raw answer only
chimera --json "Compare React and Vue"        # stdout = one JSON object (answer + trace)
chimera --formation simple "Compare …"        # pick a formation (an unknown name exits 2)
chimera --version                             # print the package version
chimera serve                                 # REST API + web UI (see Server & MCP)
chimera-mcp                                   # MCP tools for agents
python -m chimera --version                   # fallback when the `chimera` script is not on PATH (fresh clone, non-activated venv, bare wheel)

Machine-readable output

--quiet and --json are mutually exclusive group flags — place them BEFORE the subcommand. Both work for the implicit-prompt form and the explicit run form, and in both modes stdout carries exactly ONE payload (diagnostics and logs stay on stderr):

chimera --quiet "Summarize this changelog"        # stdout: the answer + "\n"
chimera --quiet run "Summarize this changelog"    # same, explicit run form
ANSWER=$(chimera --quiet "Name one HTTP status code for 'not found'")

chimera --json "Compare React and Vue" | jq .answer
chimera --json run "Compare React and Vue" | jq '.trace.total_tokens'

--json writes one object per invocation (span fields abridged here for readability — the real trace is the complete DeliberationTrace dump):

{
  "answer": "merged output from multiple models",
  "trace": {
    "request_id": "a71b3f2c...",
    "formation": "auto",
    "source": "auto",
    "dispatch": {"stage_id": "dispatch", "kind": "dispatch", "model": "deepseek/deepseek-v4-flash", "latency_ms": 1234},
    "stages": [{"stage_id": "worker_1", "kind": "worker", "model": "anthropic/claude-sonnet-4.6", "tokens_input": 300, "tokens_output": 1200}],
    "aggregator": {"stage_id": "aggregator", "kind": "aggregator", "model": "deepseek/deepseek-v4-flash"},
    "answer_stage_id": "aggregator",
    "total_tokens": 12345,
    "total_cost": 0.012,
    "total_duration_ms": 15234,
    "worker_failures": [],
    "dispatch_note": null,
    "dispatch_repairs": []
  }
}

The trace value is the COMPLETE trace serialization (model_dump(mode="json")) — the same object the REST API returns. Unicode answers are preserved (ensure_ascii=False), so café stays café.

When the dispatcher had to repair a malformed auto DAG (an edges entry pointing at a stage missing from stages, or a worker-only DAG with no aggregator), trace.dispatch_note carries the human-readable summary and trace.dispatch_repairs carries the same fact in machine-readable form — one record per repair with kind, action, stage_ids, depends_on and reason. Clean dispatches and fallbacks serialize [].

Operational warnings are never suppressed: dropped-worker and dispatch-degradation/repair warnings (which the human mode prints next to the panel) move to stderr under --quiet / --json, so chimera --json "..." > out.json cannot be corrupted. Combining both flags is a usage error (exit code 2); --verbose is ignored in the machine modes. The flags cover deliberation output — chimera models / chimera formations tables are unchanged (use GET /v1/models / GET /v1/formations for a machine-readable catalog).

Open http://localhost:8765/web/ for the web UI with live DAG visualization.

Python:

# If chimera.yaml is missing, create it before calling load_config():
chimera config init        # finds the template: local copy, wheel copy, or repo checkout
# Missing template? Copy chimera.yaml.example from a repo checkout or reinstall:
# error: chimera.yaml.example not found — copy the template from a chimera-v2 repo checkout (chimera.yaml.example at the repo root) or reinstall chimera-deliberation to restore the packaged template.

load_config() finds the config in this order: explicit path argument, then the CHIMERA_CONFIG env var, then a walk-up from the current directory.

import asyncio

from chimera import Engine, LiteLLMGateway, load_config

config = load_config()
engine = Engine(config, LiteLLMGateway(config))
result = asyncio.run(engine.deliberate("Explain quantum computing."))
print(result.answer)  # merged output from multiple models

OpenAI-compatible:

curl -X POST http://localhost:8765/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "auto", "messages": [{"role": "user", "content": "Hello"}]}'

Configuration

chimera.yaml is found in this order: an explicit path argument, then the CHIMERA_CONFIG env var, then a walk-up from the current directory.

${VAR} environment-variable substitution

Any string in chimera.yaml may reference an environment variable as ${VAR}. Tokens are substituted at load time from the process environment, so no API key has to be written into the YAML:

api_keys:
  deepseek: "${DEEPSEEK_API_KEY}"
  openrouter: "${OPENROUTER_API_KEY}"

A token is resolved from, in order:

1. the process environment;
2. ~/.hermes/.env (Hermes' dotenv file), as a fallback for variables that are not in the process environment — so cron/agent sessions resolve keys they never exported.

A variable that resolves nowhere is replaced with the empty string: the config still loads, but every call to that provider fails auth. Substitution is recursive — it applies to nested mappings and list items, not just top-level scalars.

.env auto-load

Before substitution, load_config() loads a .env file sitting next to chimera.yaml into the process environment. It is best-effort: only variables that are not already set are populated (the real environment keeps precedence), and a missing file, a missing python-dotenv, or a malformed file is tolerated (logged, never fatal). Full key resolution order:

process environment  >  repo .env (next to chimera.yaml)  >  ~/.hermes/.env

Environment-variable overrides

Common knobs can be set from the environment instead of editing the YAML (env vars win):

• CHIMERA_CONFIG — config file path
• CHIMERA_HOST, CHIMERA_PORT — server bind address
• CHIMERA_DISPATCHER, CHIMERA_WORKER, CHIMERA_AGGREGATOR — model defaults
• CHIMERA_LOG_LEVEL — observability log level
• CHIMERA_AUTH_ENABLED, CHIMERA_RATE_LIMIT_ENABLED, CHIMERA_AUTO_ALLOW_ALL_CATALOG — "true" / "1" to enable
• DEEPSEEK_KEY / DEEPSEEK_API_KEY, OPENROUTER_KEY / OPENROUTER_API_KEY, OPENAI_*, XAI_*, ZAI_*, ANTHROPIC_*, GEMINI_* — populate api_keys.* for that provider

First-run remedies

Symptom	The message names	Do this
error: No chimera.yaml found. … (one line, exit 2, no traceback)	chimera config init	Run chimera config init — it bootstraps chimera.yaml from the shipped chimera.yaml.example (local copy, repo checkout, or the copy inside the installed wheel, so a bare pip install works). --force overwrites an existing file.
A provider rejects the call (401, Missing credentials, invalid api key)	the provider and its own env var, e.g. provider 'deepseek' rejected the credentials or none were found: set DEEPSEEK_API_KEY	Set that variable (or your providers.<name>.api_key_env) and re-run.

The env var is resolved from the provider that actually served the model — providers.<name>.api_key_env if set, else the canonical name in the table above (google → GEMINI_API_KEY), else <PROVIDER>_API_KEY. LiteLLM's own error prose is provider-blind (a DeepSeek call is served through the OpenAI SDK and mentions OPENAI_API_KEY), so Chimera puts its accurate remedy in front of it on every surface: CLI, REST body, MCP tool result, trace and logs. Keyless loopback providers (lmstudio, ollama) keep their raw error text — Chimera never names a key variable for an endpoint that needs none. A custom base_url provider follows the same rule: a failure from the Hermes gateway names API_SERVER_KEY, because that is what providers.hermes.api_key_env declares — not OPENAI_API_KEY, even though the call itself is served over the OpenAI SDK.

See docs/CONFIG.md for the full chain.

Documentation

Detailed guides live in docs/:

Guide	Contents
docs/CONFIG.md	Full configuration reference (chimera.yaml, categories, model selection)
docs/USAGE.md	CLI usage and examples
docs/INTEGRATION.md	Integrate Chimera into your app (deployment, clients, auth, errors)
docs/OPENAI_API.md	OpenAI-compatible API contract
docs/SECURITY.md	Security model and credential handling
docs/EDGE_CASES.md	Edge cases and behavior notes
docs/FAILURE_RESILIENCE.md	Failure handling and resilience design
docs/RESILIENCE.md	Resilience guarantees and degradation semantics
docs/LOCAL_CI.md	Run the CI suite locally with act (ruff + tests, same workflow as CI)

Full docs index, including historical artifacts and dogfood logs: docs/README.md.

Architecture

flowchart TB
    subgraph Client
        A[User Prompt]
    end

    subgraph Dispatcher["Dispatcher (1 call)"]
        B[Designs DAG<br/>Picks models by category<br/>Writes custom prompts<br/>Writes merge instructions]
    end

    subgraph Workers["Workers (parallel)"]
        C1[Worker A<br/>domain-scoped task]
        C2[Worker B<br/>domain-scoped task]
        C3[Worker C<br/>domain-scoped task]
    end

    subgraph Aggregation
        D[Aggregator<br/>merges with dispatcher instructions]
    end

    A --> B
    B --> C1
    B --> C2
    B --> C3
    C1 --> D
    C2 --> D
    C3 --> D
    D --> E[Final Answer]

Formation Types

flowchart LR
    subgraph Simple["Simple (2 workers)"]
        S1[W1] --> SA[Aggregator]
        S2[W2] --> SA
    end

    subgraph Debate["Debate (3 workers + merge)"]
        D1[W1] --> DA1[Agg 1]
        D2[W2] --> DA1
        D2 --> DA2[Agg 2]
        D3[W3] --> DA2
        DA1 --> DM[Merge]
        DA2 --> DM
    end

    subgraph Custom["Custom DAG (client-defined)"]
        C1[Researcher] --> C2[Critic]
        C2 --> C3[Polisher]
        C3 --> C4[Final]
    end

Flow: Request to Answer

sequenceDiagram
    participant Client
    participant Engine
    participant Dispatcher
    participant Workers
    participant Aggregator

    Client->>Engine: POST /v1/chat/completions
    Engine->>Dispatcher: Design formation
    Dispatcher->>Dispatcher: Pick models by category weights
    Dispatcher->>Dispatcher: Write per-worker prompts + merge instructions
    Dispatcher-->>Engine: DAG + prompts + instructions

    par Workers (parallel)
        Engine->>Workers: Worker A (custom prompt)
        Engine->>Workers: Worker B (custom prompt)
        Engine->>Workers: Worker C (custom prompt)
        Workers-->>Engine: responses
    end

    Engine->>Aggregator: Merge with dispatcher instructions
    Aggregator-->>Engine: Final answer

    Engine-->>Client: JSON response + trace

Server & MCP

chimera serve needs the [server] extra (fastapi/uvicorn) — it is part of the [full] extra too. On a bare pip install chimera-deliberation the command does not crash with a traceback: it exits 2 with the exact remedy (pip install chimera-deliberation[server], or [full] for the web UI and MCP extras). chimera serve --help works either way.

chimera serve starts the REST API plus the web UI:

• http://localhost:8765/v1/chat/completions — OpenAI-compatible endpoint
• http://localhost:8765/docs — OpenAPI docs
• http://localhost:8765/web/ — web UI with live DAG visualization

Health probes (all return JSON):

• http://localhost:8765/health — alias of the liveness probe (always 200 when the process is alive)
• http://localhost:8765/v1/health — health check (healthy / degraded); see below
• http://localhost:8765/v1/health/ready — readiness probe (200 when ≥1 provider reachable, 503 otherwise)
• http://localhost:8765/v1/health/live — liveness probe (always 200 when alive)

/v1/health status semantics. healthy means every configured provider's live probe succeeded. degraded means at least one provider probe failed — missing credentials, an auth error (401/403), a quota/billing failure (429, insufficient balance, "no resource package"), an API error, a probe timeout, or an exception inside the check itself. The endpoint always answers HTTP 200 and reports the verdict in the status field (plus a details object with per-provider healthy / error / model_tested), so monitoring must read status, not the HTTP code. Any probe failure — including "all providers failed" — reports degraded; an unhealthy status value is not emitted.

Machine-readable degraded reason (DF-CHIMERA-V2-14). status alone tells you that something is wrong, not what:

• the top-level unhealthy_providers list names the providers whose probe did not succeed, sorted. It is present in every response and is [] exactly when status is healthy, so a client can branch on it instead of walking details.providers and string-matching error.
• every failed entry under details.providers.<name> carries an error_class: missing_credentials (no resolvable key — no live call was made), slow (the probe never answered inside server.health_timeout_s — see below), timeout (a probe-side timeout: the call itself failed with a timeout), auth (401/403 or auth prose), quota (429, insufficient balance, "no resource package"), or api (anything else). A healthy provider has no error_class; treat an unrecognized value as opaque.
• top-level slow_providers names the providers whose probe never answered inside the budget. Such a provider is UNMEASURED, not proven broken, so it does NOT make status degraded and is NOT in unhealthy_providers — slow_providers: ["hermes"] with status: "healthy" is the honest reading for a gateway that is alive but slower than the probe budget (DF-CHIMERA-V2-27). A provider that fails a different way (connection refused, 5xx, auth, quota) still degrades status exactly as before.
• if the probe itself raises, the response is still degraded with every configured provider named in unhealthy_providers — none of them was proven healthy — and the underlying reason in details.error.

A probe is one upstream attempt per model (DF-CHIMERA-V2-16). The health probe does not use the retry ladder a normal completion uses: it makes exactly one call per model and reports whatever the provider answers. A provider that replies with a fast 429/401 is therefore reported as its real class (quota / auth) carrying the provider's own message — a quota reset timestamp, say — instead of being retried with backoff until the shared server.health_timeout_s budget expires and the verdict turns into a fabricated timeout. Retrying an answer that is already final buys nothing and destroys the signal; the non-probe retry policy (config.retry) is unchanged.

Both are additive: status, details.config_loaded, details.models_configured, details.providers_configured, details.providers_discovered, details.commit, and the per-provider healthy / error / model_tested / note fields keep their names and meanings, and the HTTP status stays 200. GET /v1/health/ready carries the same unhealthy_providers list in its 200 body.

Configured vs discovered providers (CH-GAP-053). details.providers_configured counts only the providers explicitly declared in chimera.yaml — not the full provider map. Provider auto-discovery (on by default, provider_discovery: true) merges extra providers from models.dev into config.providers at load time, and providers_configured used to silently include them. The discovery-added names are now reported separately in the additive details.providers_discovered list (sorted, names only — no credentials). The per-provider map under details.providers still lists every provider, both declared and discovered; a config built programmatically (no YAML load) keeps the old behavior where everything in the map counts as configured and providers_discovered is [].

The provider probe costs money. Each /v1/health (and /v1/health/ready) call runs a real completion per configured provider — gateway.complete(model, "ping", max_tokens=1) — so it consumes a (small) number of tokens on every provider it can reach, and takes real latency. Details worth knowing before you poll it in a tight loop:

• providers with no resolvable credentials are reported immediately (missing-credentials) with no live call;
• on a non-timeout failure up to 3 models of that provider are tried before it is marked unhealthy (one blocked model does not condemn the provider); a timeout is terminal for that provider after one model;
• probes run concurrently, bounded by server.health_timeout_s (default 10.0 s); providers still pending at that bound are reported unhealthy with a timeout: error — after one bounded server.health_probe_grace_s window (default 1.0 s, DF-CHIMERA-V2-17) that lets a merely-cold probe land and report its real verdict instead; set it to 0 for cancel-at-deadline. Slow tail latency is therefore indistinguishable from a dead provider — raise server.health_timeout_s if your provider is merely slow.
• each of those probes is one upstream attempt per model — the retry ladder is not used (DF-CHIMERA-V2-16), so a provider that answers with a fast 429/401 is reported as quota/auth with the provider's own message (including a quota reset time) rather than retried into the timeout budget.

chimera-mcp runs the MCP server over stdio so AI agents (Hermes, Claude Code, etc.) can call Chimera directly. Both read the same chimera.yaml configured in Quickstart.

Live smoke test

/v1/health reporting healthy does not prove a deliberation works (billing, auth and prompt-format regressions only surface on a real call). Run the live smoke test after a deploy:

python scripts/smoke_live.py --base-url http://localhost:8765     # default base URL
python scripts/smoke_live.py --base-url http://myhost:8765 --formation auto
CHIMERA_API_KEY=... python scripts/smoke_live.py --base-url http://myhost:8765   # if auth is enabled

The flag is --base-url (there is no --port); CHIMERA_BASE_URL is the env default. Other flags: --formation (default simple), --prompt, --api-key, --timeout. The script checks liveness + running commit (warning when the deployed commit diverges from local HEAD), probes /v1/health, then POSTs a real /v1/deliberate and prints the merged answer. Exit 0 = merged answer received, 1 = failure with an actionable message, 2 = usage/config error. Stdlib-only.

The provider-health block is class-aware (DF-CHIMERA-V2-4), because a fresh install and a real degradation used to print the same words:

• providers: <healthy>/<total> healthy — counted from the payload's provider map, never from providers_configured (that counts providers explicitly declared in the config, excluding auto-discovery additions — see providers_discovered for those — and says nothing about whether any of them answered);
• INFO: providers without a configured API key (expected on a fresh install): … — the providers whose error_class is missing_credentials. Expected on a fresh install, so it is information rather than a warning: the deliberation below is the real proof that the deployment works;
• WARNING: providers reported unhealthy by /v1/health: <name> [<class>] … — every other class (timeout, auth, quota, api, unknown), naming each provider with its class and surfacing the provider's own message (a quota reset time, for example);
• INFO: providers slower than the probe budget (not a failure — the probe never landed; a real call may still work): <name> [slow]: … — the providers the server classified slow (DF-CHIMERA-V2-27), with the measured wait. Information, not a warning: catching a standing slow condition is useful, but it is not a degradation the operator must act on, and the server does not report it as one either;
• INFO: providers whose live health probe is disabled (health_probe: false): … — the providers explicitly opted out of probing. Nothing was measured, so the line exists to keep the omission visible.

None of these lines changes the exit code: the provider block never fails the run — only the deliberation decides pass/fail.

Model Selection

The dispatcher picks models using category-weighted scoring:

Category	What it measures
code	Programming, debugging, software engineering
analysis	Data analysis, research, evaluation
reasoning	Logic, math, complex problem-solving
design	Creative work, UX, content generation
audit	Fact-checking, safety, correctness review

Each model in the catalog has a percent score (0–100) per category, on the same scale as the shipped chimera.yaml.example catalog and the categories map in GET /v1/models. A catalog written on the historical 0.0–1.0 scale is accepted and rescaled ×100 at load (with a warning); a score outside 0–100 or a non-numeric score is rejected at load. The dispatcher matches task domains to model strengths. You can override any model choice per request:

{
  "model": "auto",
  "messages": [{"role": "user", "content": "..."}],
  "dispatcher_model": "deepseek/deepseek-v4-flash",
  "aggregator_model": "zai-coding-plan/glm-5.2",
  "worker_model": "deepseek/deepseek-v4-pro",
  "allowed_models": ["deepseek/deepseek-v4-pro", "zai-coding-plan/glm-5.2"],
  "stage_models": {"worker_1": "anthropic/claude-sonnet-4.6"}
}

• dispatcher_model — model for the planning/dispatch call itself.
• aggregator_model — forces every aggregator/merge/audit stage.
• worker_model — forces every worker stage.
• stage_models — per-stage overrides; these win over the global ones above.
• Config locks (defaults.lock_aggregator: true / defaults.lock_dispatcher: true) force the configured default for that role and discard the request override. A discarded override is never silent: it is recorded in the trace's dispatch_note (e.g. override discarded: aggregator_model='X' ignored (lock_aggregator=true)).

Custom DAGs

Send your own formation structure at request time. The example below is the complete request body for POST /v1/chat/completions with "model": "custom" — that endpoint is the custom-DAG contract (POST /v1/deliberate does not accept formation custom):

{
  "model": "custom",
  "allow_custom_dag": true,
  "dag": {
    "stages": [
      {"id": "researcher", "kind": "worker", "model": "anthropic/claude-sonnet-4.6"},
      {"id": "critic", "kind": "aggregator", "model": "zai-coding-plan/glm-5.2", "depends_on": ["researcher"]},
      {"id": "polisher", "kind": "worker", "model": "deepseek/deepseek-v4-pro", "depends_on": ["critic"]},
      {"id": "final", "kind": "aggregator", "model": "anthropic/claude-sonnet-4.6", "depends_on": ["polisher"]}
    ],
    "edges": [["researcher","critic"], ["critic","polisher"], ["polisher","final"]]
  },
  "messages": [{"role": "user", "content": "..."}]
}

The dispatcher writes custom prompts for each stage but uses YOUR structure exactly.

Interfaces

Interface	Endpoint	Use
REST API	POST /v1/chat/completions	OpenAI-compatible drop-in
REST API	POST /v1/deliberate	Full control (DAG, overrides, trace)
REST API	GET /v1/models	OpenAI-shaped model list (data) + chimera catalog map
REST API	GET /v1/formations	Available formation presets
CLI	chimera run	Command-line usage (add --quiet / --json for machine-readable output, --version for the package version)
MCP	chimera_deliberate	Run a deliberation from an agent
MCP	chimera_formations	List formation presets
MCP	chimera_models	List models with weights

Response Trace

Every deliberation returns a full trace:

request_id: a71b3f2c...
formation: auto
source: auto        ← not "fallback" — dispatcher designed it
total_tokens: 12345
total_cost: $0.012
total_duration_ms: 15234

dispatch: V4 Flash (1,234ms, 800+420 tok)
workers:
  worker_rust: Claude Sonnet 4 (4,500ms, 300+1,200 tok)
  worker_go: Kimi K2.7 (8,200ms, 250+800 tok)
aggregator: V4 Flash (2,100ms, 3,000+500 tok)

Providers

Chimera uses LiteLLM under the hood. Supported providers:

Provider	Direct API	Via OpenRouter	Models
DeepSeek	✅	✅	v4-flash, v4-pro, r1
Anthropic	—	✅	Sonnet 4, Opus 4.7/4.8, Haiku 4.5
OpenAI	—	✅	GPT-5.5, GPT-5.1
xAI	—	✅	Grok 4.20
Google	—	✅	Gemini 3.5 Flash, 3.1 Pro, 2.5 Flash
Z.AI	✅ (direct)	✅	GLM-5.2
MoonshotAI	—	✅	Kimi K2.7 Code, K2.6
MiniMax	—	✅	M3
Meta	—	✅	Llama 4 Maverick
Hermes gateway	✅ (local)	—	any model the running Hermes serves, addressed as hermes/<model>
9router fleet gateway	✅ (tailnet)	—	~269 models across 15 upstream prefixes, addressed as router9/<upstream id>

Custom OpenAI-compatible endpoints

Any provider Chimera does not route natively is called through its own base_url, over the OpenAI-compatible SDK, using the provider's resolved key. Exactly one leading <provider>/ segment is stripped before the request goes out, so hermes/glm-5.3-flash reaches the endpoint as the bare glm-5.3-flash — the same convention the zai provider uses.

Endpoints that serve namespaced model ids of their own keep the inner slashes: the catalog id router9/ds/deepseek-v4-flash reaches 9router as ds/deepseek-v4-flash (not deepseek-v4-flash), and router9/openrouter/x-ai/grok-4.6 as openrouter/x-ai/grok-4.6.

That is how the local Hermes gateway (:8642, OpenAI-compatible) plugs in as a first-class provider:

providers:
  hermes:
    base_url: http://127.0.0.1:8642/v1
    api_key_env: API_SERVER_KEY   # the var name in ~/.hermes/.env — never inline the key

models:
  hermes/glm-5.3-flash:
    provider: hermes
    cost_tier: budget

Two things to plan for with the Hermes gateway: it adds its own large system prompt to every request (~41k tokens of input), so the effective context budget is the model's window minus that; and /v1/health probes providers with a real completion under server.health_timeout_s (default 10.0s), so raise that bound if your gateway is slower than that on a 1-token call. (The 9router gateway adds no prompt of its own — its lanes are the upstream providers' own endpoints.)

…and the same seam carries the 9router fleet gateway — one OpenAI-compatible endpoint (http://master001:20128/v1, reachable on the tailnet as master001) in front of ~269 models across 15 upstream prefixes (ds/, mmx/, minimax/, kimi/, openrouter/, xai/, ollama/, …):

providers:
  router9:
    base_url: http://master001:20128/v1
    api_key_env: ROUTER9_API_KEY   # the var name in ~/.hermes/.env — never inline the key

models:
  router9/ds/deepseek-v4-flash:
    provider: router9
    cost_tier: budget

Each upstream lane is an account of its own, so an exhausted lane answers 429/403 for every model behind it while the gateway itself is healthy — a provider-side quota condition, not a Chimera defect. Point a formation at another prefix's model rather than re-dispatching at the same one.

See docs/CONFIG.md for the full contract.