Inference and routing

Inference

Signet’s shared inference control plane is configured under the top-level inference key in agent.yaml.

Background workloads resolve only through this control plane. Configure the memoryExtraction workload binding explicitly (the Dashboard writes it with the selected model). The internal session_synthesis operation inherits the same route. Signet no longer compiles memory.pipelineV2.extraction or memory.pipelineV2.synthesis into an implicit provider, and a missing or invalid binding is reported as a routing error rather than silently falling back to llama-cpp.

Use the same inference control plane to choose models across harnesses, accounts, APIs, and local runtimes per turn or per subtask.

Example:

inference:
  enabled: true
  defaultPolicy: auto

  accounts:
    claude-dot:
      kind: subscription_session
      providerFamily: anthropic
      label: Dot Claude Connected
      sessionRef: CLAUDE_DOT_SESSION
    openrouter-main:
      kind: api
      providerFamily: openrouter
      credentialRef: OPENROUTER_API_KEY
    codex-subscription:
      kind: subscription_session
      providerFamily: openai-codex

  targets:
    opus:
      executor: claude-code
      account: claude-dot
      models:
        opus46:
          model: opus-4.6
          reasoning: high
          toolUse: true
          streaming: true
    sonnet:
      executor: openrouter
      account: openrouter-main
      privacy: remote_ok
      endpoint: https://openrouter.ai/api/v1
      models:
        default:
          model: anthropic/claude-sonnet-4-6
          reasoning: medium
          toolUse: true
          streaming: true
          costTier: medium
    local:
      executor: ollama
      endpoint: http://127.0.0.1:11434
      privacy: local_only
      models:
        gemma4:
          model: gemma4
          reasoning: medium
          streaming: true
          costTier: low
    codex-direct:
      executor: openai-codex
      account: codex-subscription
      models:
        default:
          model: gpt-5.4
          reasoning: high

  policies:
    auto:
      mode: automatic
      defaultTargets:
        - opus/opus46
        - sonnet/default
        - local/gemma4

  taskClasses:
    casual_chat:
      reasoning: medium
      preferredTargets:
        - sonnet/default
    hard_coding:
      reasoning: high
      toolsRequired: true
      preferredTargets:
        - opus/opus46
    hipaa_sensitive:
      privacy: local_only
      preferredTargets:
        - local/gemma4

  workloads:
    interactive:
      policy: auto
      taskClass: casual_chat
    memoryExtraction:
      policy: auto
      taskClass: casual_chat

  agents:
    rose:
      defaultPolicy: auto
      roster:
        - opus/opus46
        - sonnet/default
        - local/gemma4
      pinnedTargets:
        hard_coding: opus/opus46

inference.accounts

Named account or credential identities used by targets.

Field	Type	Description
kind	string	subscription_session or api
providerFamily	string	pi-ai provider id, for example anthropic, openai-codex, github-copilot, or openrouter
label	string	Human-readable account label
credentialRef	string	Secret name or env var name for API-backed targets
sessionRef	string	Session identifier for subscription-backed targets
usageTier	string	Optional account tier label

For an OAuth subscription, set kind: subscription_session, use the pi-ai OAuth provider id as providerFamily, and omit credentialRef. Connect the account through /api/inference/oauth/login/:id. Signet stores the resulting OAuth credential in its encrypted secret store and asks pi-ai to refresh it at request time. For a conventional API-key account, use kind: api and set credentialRef; environment variables continue to take precedence over the encrypted secret with the same name.

inference.targets

Executable route targets. A target can be a local runtime, API backend, subscription-backed CLI session, or gateway.

Field	Type	Description
executor	string	acpx, a local compatibility executor, or a provider id returned by /api/inference/catalog
kind	string	Optional explicit target kind. Inferred when omitted
account	string	Account id from inference.accounts
endpoint	string	Optional base URL override
command	object	Command executor config with bin, optional args, cwd, and env
agent	string	For executor: acpx, the ACPX adapter command to run, for example codex, claude for Claude Code, opencode, or kimi for Kimi CLI. Signet normalizes legacy claude-code values to ACPX’s claude command.
acpxVersion / version	string	Optional ACPX package version. Defaults to Signet’s pinned ACPX version
mode	string	Optional ACPX execution mode. Defaults to one-shot exec
cwd	string	Optional working directory for harness execution
session	string	Optional ACPX session identifier when a persistent session is desired
permissions	string	Optional ACPX permission policy passed through to the harness
hooks	string	Set to disabled for sterile/background execution (SIGNET_NO_HOOKS=1 and SIGNET_ENABLED=false)
terminal	boolean	For ACPX, set false to pass --no-terminal
allowedTools	array	Optional ACPX allowed-tool list
format / outputFormat	string	ACPX output format. quiet is the default; json parses ACPX JSON events and extracts the final response
captureEvents	boolean	When true, defaults ACPX to JSON output and enables the provider event-capture path
maxCapturedEvents	number	Maximum number of JSON events delivered to the provider-side event callback; defaults to 200
emptyResponseRetries	number	Fresh-session retries after an exit-0 empty response. Defaults to 1 and is capped at 3; retries run only for sterile deny-all targets with hooks and tools disabled
modelSelection	string	acp passes the routed model through ACPX negotiation; agent lets the ACP agent’s native configuration choose it. Defaults to agent for OpenCode and acp for other agents
timeoutMs	number	Per-call ACPX subprocess deadline
extraArgs	array	Additional ACPX CLI args appended after Signet-managed args
privacy	string	remote_ok, restricted_remote, or local_only
models	map	Named model entries for this target

Example ACPX background target (see also docs/ACP-INTEGRATION.md for the architecture and current limitations):

inference:
  targets:
    background-codex:
      executor: acpx
      agent: codex
      hooks: disabled
      terminal: false
      format: json
      captureEvents: true
      timeoutMs: 120000
      models:
        mini:
          model: gpt-5.4-mini
          reasoning: medium
          toolUse: true

Direct Codex CLI targets use the codex binary in the daemon runtime. The Docker image includes the CLI; mount a logged-in Codex config directory only when you opt into this executor:

SIGNET_CODEX_HOME="$HOME/.codex" docker compose -f compose.yml -f compose.codex.yml up -d

inference:
  targets:
    codex-cli:
      executor: codex
      models:
        default:
          model: gpt-5.4-mini
          reasoning: medium
  workloads:
    memoryExtraction:
      target: codex-cli/default
      taskClass: memory_extraction

Signet marks the target unavailable unless codex --version works and either CODEX_HOME/auth.json is mounted or OPENAI_API_KEY is present. The auth cache is copied into a sterile temporary home for each call; the mounted Codex config stays read-only.

Model fields:

Field	Type	Description
model	string	Provider-native model identifier
label	string	Optional display label
reasoning	string	low, medium, or high
contextWindow	number	Maximum prompt tokens the model can accept
toolUse	boolean	Whether tool use is supported
streaming	boolean	Whether streaming is supported
multimodal	boolean	Whether multimodal input is supported
costTier	string	low, medium, or high
averageLatencyMs	number	Optional routing latency hint

inference.policies

Named routing policies that agents and workloads reference.

Field	Type	Description
mode	string	strict, automatic, or hybrid
allow	array	Route refs allowed by the policy
deny	array	Route refs denied by the policy
defaultTargets	array	Ordered preferred target refs
taskTargets	map	Task-class specific preferred target refs
fallbackTargets	array	Explicit fallback refs
maxLatencyMs	number	Hard latency ceiling used by routing
costCeiling	string	Hard cost ceiling used by routing

Policies are optional: when targets exist but no policies (and no defaultPolicy) are configured, the router synthesizes an implicit default policy (mode: automatic, defaultTargets/fallbackTargets over all configured target refs) so every generation path keeps working. signet route list shows it like any other policy. Add an explicit policy to pin routing deterministically; signet route doctor warns when agent.yaml relies on the synthesized policy.

inference.taskClasses

Task-family hints for automatic routing.

Field	Type	Description
reasoning	string	Required reasoning depth
toolsRequired	boolean	Require tool use support
streamingPreferred	boolean	Prefer or require streaming support
multimodalRequired	boolean	Require multimodal support
privacy	string	Hard privacy tier, including local_only
maxLatencyMs	number	Task latency budget
costCeiling	string	Task cost ceiling
expectedInputTokens	number	Prompt-size hint
expectedOutputTokens	number	Output-size hint
preferredTargets	array	Preferred target refs
keywords	array	Lightweight classifier keywords

inference.workloads

Binds Signet-owned workloads to router policies or explicit targets.

Supported workload keys:

• interactive
• memoryExtraction

Each workload can define:

Field	Type	Description
policy	string	Named policy id
taskClass	string	Default task class for this workload
target	string	Explicit target/model pin

inference.agents

Per-agent routing overrides.

Field	Type	Description
defaultPolicy	string	Default policy for that agent
roster	array	Allowed target refs for that agent
preferredTargets	map	Task-class target preferences
pinnedTargets	map	Hard pins, usually managed by signet route pin