1. Executive Summary

Chimera is a multi-model deliberation engine that orchestrates large language models (LLMs) across a directed acyclic graph (DAG) of stages. It employs a 3‑tier provider architecture:

• PAYG Foreman – pay‑as‑you‑go models used for initial stages (e.g., deepseek-foreman).
• Prepaid Worker Buckets – pre‑purchased model capacity (e.g., opencode-go, zai-glm).
• Backup / Overflow – fallback providers (e.g., ollama-cloud, openrouter) with cost‑optimized routing.

Competitive differentiators:

• Tactical deliberation – Stage 0 dictates downstream contracts: each stage declares a JSON Schema for validation and a model routing strategy. (commit 0241610)
• Validation‑as‑code – mechanical jsonschema.validate() ensures every aggregated output conforms; failed stages trigger re‑iteration. (commit f488ef2)
• Progressive prompting – wait_messages and trigger fields control multi‑step reasoning within a single stage. (commit 2b6d5d1)

Note: Google Gemini and Anthropic Claude are classified as external PAYG (direct API billing), distinct from the internal foreman tier.

2. Architecture

The deliberation engine is implemented as a DAG where each node is a stage. The lifecycle follows a closed‑loop iteration:

    +-----------+       +---------+       +-------------+       +-------+       +--------+
    |Dispatcher|----->| Workers  |----->| Aggregator |----->| Audit |----->| Refine |
    +-----------+       +---------+       +-------------+       +-------+       +--------+
         |                                                                          |
         +----  _RE_ITERATION_SIGNAL -----------------------------------+
  

Dispatcher parses the DAG definition, resolves model routing per stage, and fans out to worker nodes. Workers execute model calls in parallel waves via asyncio.create_task. (commit 469eec4) Aggregator collects results, merges according to stage schema, and passes to Audit. Audit runs jsonschema.validate(aggregated_output, stage_schema); on failure it emits _RE_ITERATION_SIGNAL back to Dispatcher. Refine adjusts prompt parameters (temperature, max_tokens) before the next iteration.

The model catalog contains 36 models across 9 providers. Each model is scored on 32 hierarchical category axes (0–100) for capability, latency, cost, and availability. (commit 85a3a11)

3. Features

Formations

Pre‑defined deliberation patterns:

• auto – Default; 3-stage DAG (foreman, worker, aggregator) with automatic model selection.
• speed – Optimized for low latency; single-stage parallel workers, no refinement.
• spec-writer – Multi-stage DAG that produces structured specifications; includes audit loop.
• debate – Two opposing worker groups with a judge aggregator stage.
• audit – Focus on validation; multiple audit stages with strict schema enforcement.

(commit 0241610)

Model Selection & Cost‑Weighted Sensitivity

Effectiveness is calculated as quality / (cost ^ sensitivity). Sensitivity (0–1) is configurable per request. Default sensitivity = 0.5. (commit b7cca8d)

Per‑Model Enable Toggle

Each model in the catalog has an enabled: true/false flag. Disabled models are excluded from routing. (commit 5dadfc9)

Provider Auto‑Discovery

Chimera fetches models.dev/api.json from each provider and caches locally for 5 minutes. New models are added automatically. (commit caca56b)

Configurable 3‑Tier Timeouts

• Code default – 30 s per stage
• Admin ceiling – 120 s (set via env CHIMERA_MAX_TIMEOUT)
• Request header – X-Chimera-Timeout overrides per request (clamped to admin ceiling)

(commit 0241610)

REST API + MCP Server

Primary API is REST (see §5). A MCP (Model Context Protocol) server is also available via stdio. (commit a2ae192)

Structured Output

json_schema strict mode is used at the model provider level (when supported). Chimera enforces mechanical jsonschema.validate() after aggregation. (commit f488ef2)
Note: The output_schema field is required for strict mode validation; if omitted, only free-form output is allowed without mechanical validation.

Progressive Prompting

Stages can define wait_messages (list of messages to inject before model answers) and trigger (a Python regex pattern evaluated against the aggregated text output). Example: {"wait_messages": ["Think step by step."], "trigger": ".*therefore.*"}. (commit 413979b)

Test Coverage

416 tests, 61 integration skips (due to provider unavailability). (commit f488ef2)

4. Provider Palette

Provider	Tier	Models
deepseek-foreman	PAYG Foreman	deepseek-v4-pro, deepseek-v4-flash
opencode-go	Prepaid Worker Bucket	deepseek-v4-pro, deepseek-v4-flash
zai-glm	Prepaid Worker Bucket	glm-5.2
minimax	Prepaid Worker Bucket	MiniMax-M3
kimi-for-coding	Prepaid Worker Bucket	kimi-k2.7
xai-oauth	Prepaid Worker Bucket	grok-4.5
ollama-cloud	Backup / Overflow	various open‑source
openrouter	Backup / Overflow	31+ models incl. GPT‑5.6 family, Claude Opus 4.8, Gemini 3.1 Pro
Google Gemini	External PAYG	gemini‑3.1‑pro, gemini‑3.1‑flash
Anthropic Claude	External PAYG	claude‑opus‑4.8, claude‑sonnet‑4.5

All provider entries validated as of July 2026. (commit 85a3a11)

5. API Surface

Endpoints

Method	Path	Description
POST	/v1/deliberate	Execute a deliberation job. Accepts JSON: formation, overrides, output_schema (required for strict mode), dag (only allowed when allow_custom_dag=true). (commit 0241610)
GET	/health	Returns status and model catalog freshness. (commit 1d32721)
GET	/v1/models	List all models with current enabled/disabled, costs, quality scores. (commit 85a3a11)
GET	/v1/formations	List available formations and their DAG definitions. (commit 0241610)
GET	/v1/trace/{request_id}	Retrieve full trace of a completed deliberation (stages, timings, outputs). (commit 1bf73ea)
WS	/ws/trace/{request_id}	Stream trace updates in real‑time. (commit a2ae192)

Note on allow_custom_dag: When this parameter is true, the request may supply a custom dag object; otherwise only predefined formations are permitted.

6. Roadmap

• Speed optimizations – Reduce per‑stage overhead by batching schema validation. (commit b785711)
• Clustering formation – Multi‑fabric clustering for ensemble voting. (commit 1bf73ea)
• Streaming SSE responses – Deliver partial aggregations via text/event-stream. (commit 1bf73ea)
• Provider SDK migration – Replace LiteLLM with direct provider SDKs (OpenAI, Anthropic, Google). (commit b785711)
• Langfuse observability – Trace every deliberation step to Langfuse. (commit 0241610)
• Helm chart for Kubernetes – One‑click deployment with autoscaling. (commit b785711)
• Chimera Gateway – Multi‑tenant API key management, rate limiting, and billing ingestion. (commit b785711)