Chaos Mode

🌍 View in other languages

Chaos Mode

Dashboard: Chaos Mode (sidebar) → /dashboard/chaos
API: GET / PUT /api/chaos/config · POST /api/chaos/run (dashboard session) · POST /api/skills/collect/chaos (API key)
Source: src/lib/chaos/chaosExecutor.ts, src/lib/chaos/chaosConfig.ts

Chaos Mode sends one task to several providers at once — every participating provider contributes one model instance, and you get all the answers side by side (or chained). It is a multi-model execution surface, not a routing strategy: your normal /v1/chat/completions traffic is never affected by it.

Disambiguation — three different things ship with "chaos" in the name:

Thing	What it is	Where documented
Chaos Mode	The dashboard page + API described here: fan one task out to many providers (parallel or collaborative).	This guide
auto/chaos	Auto-Combo model id: parallel fan-out, one model per provider, one upstream call each. Not fault injection (details).	AUTO-COMBO.md
Chaos combo config	A persisted combo with config.chaos.enabled fans out the same way (API-only); judgeModel only picks the final answer, no synthesis call.	open-sse/services/autoCombo/chaosEngine.ts

auto/chaos: parallel fan-out

auto/chaos is not a fault-injection or resilience-testing knob. Requesting model: "auto/chaos" on /v1/chat/completions:

1. Builds a panel of one model per provider: the first candidate of each connected provider, in candidate-pool order, up to 5 members (OMNIROUTE_CHAOS_MAX_PANEL, capped at 10) (open-sse/services/autoCombo/virtualFactory.ts). The chaos-mode weight pack only sets each member's weight; the fan-out does not read it.
2. Sends the same request to every panel member in parallel, so one request costs one upstream call per panel member (open-sse/services/autoCombo/chaosEngine.ts, dispatched from open-sse/services/combo.ts).
3. Streams one status line per panel member as it lands: an SSE comment (: chaos <index> ok|fail <model>) by default, plus an omni-chaos-part event (model, index, ok, error) when the request sets stream_options.include_chaos_parts: true. These carry no answer text.
4. Sends one panel answer as the final OpenAI-style chunk: the first panel member's (auto/chaos sets it as judgeModel) when it succeeds, otherwise the last successful member's. The other panel answers are not returned, so you pay for N calls and receive one completion.

Setup

1. Open Dashboard → Chaos Mode (/dashboard/chaos).

2. Turn it on — Chaos Mode ships disabled by default (enabled: false in src/lib/chaos/chaosConfig.ts). While disabled, POST /api/chaos/run answers 400 — "Chaos Mode is not enabled. Enable it in Dashboard → Chaos Mode.".

3. Pick the participants and defaults (persisted per instance via the settings store):

   Field	Meaning	Default / limits
   enabled	Master switch	false
   defaultMode	parallel or collaborative (see below)	parallel
   providerOverrides	Per-provider participation (providerId, optional modelId, enabled)	empty = every active provider, max 200
   systemPrompt	Override for the built-in Chaos system prompt	optional, max 10 000 chars
   timeoutMs	Max time per model call	120000 (5 000–600 000)
   maxTokens	max_tokens per model call	4096 (256–128 000)

4. Run a test from the page itself — the results panel shows each provider's answer, status and duration.

Execution modes

• parallel — every model gets the same task simultaneously; you receive all answers independently.
• collaborative — models run in a chain: each one sees the previous model's output and is asked to refine, extend, critique or offer an alternative. The response's summary field concatenates the successful outputs in chain order (parallel runs have no summary).

API

POST /api/chaos/run — dashboard session

Cookie-authenticated (the management session — see MANAGEMENT-AUTH.md); used by the dashboard page.

// body
{
  "task": "Compare approaches to X",   // required
  "providers": ["glm", "kimi"],         // optional filter
  "mode": "parallel",                   // optional — overrides defaultMode
  "systemPrompt": "…",                  // optional override
  "maxTokens": 4096                      // optional override
}

POST /api/skills/collect/chaos — API key

Bearer-token variant for external callers. The key must carry the Chaos Mode permission (chaosModeEnabled), which is off by default — enable it per key in Dashboard → API Manager → edit key → permissions → Chaos Mode. Same body as above.

curl -X POST http://localhost:20128/api/skills/collect/chaos \
  -H "Authorization: Bearer $OMNIROUTE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"task":"Compare approaches to X","mode":"parallel"}'

Both endpoints return the same shape:

{
  "task": "…",
  "mode": "parallel",
  "startedAt": "2026-09-01T00:00:00.000Z",
  "totalProviders": 3,
  "totalResults": 3,
  "models": [
    { "providerId": "glm", "providerName": "GLM", "modelId": "glm-4.7",
      "status": "success", "content": "…", "durationMs": 3210 }
  ],
  "summary": "…" // collaborative mode only
}

Troubleshooting

• 400 Chaos Mode is not enabled — step 2 above: the global switch is off.
• API key gets rejected on /api/skills/collect/chaos — the key lacks the per-key chaosModeEnabled permission (off by default; this is a setting, not an error).
• A provider you expected is missing from the results — check providerOverrides on the Chaos Mode page (a disabled override excludes it) and whether the provider connection is active.