Upgrading

Breaking change. This release replaces Signet’s hand-rolled inference providers with two backends: Pi (@earendil-works/pi-ai) for every direct API call, and ACPX for harness subprocess calls. Existing installs must verify their background pipeline still functions after updating.

If you are upgrading from a release before this change, read this entire page.

What changed

Before this release, Signet shipped six hand-rolled LLM providers in a 4,000-line provider.ts: Claude Code (subprocess), OpenCode (HTTP server), Codex (subprocess), a generic command-line provider, and direct HTTP for Anthropic / OpenAI-compatible / Ollama / llama.cpp / OpenRouter.

After this release there are exactly two backends behind the same LlmProvider interface:

• Pi (pi-ai) — every direct API call: Anthropic, OpenAI (incl. Codex), Google, Bedrock, Mistral, Azure, Cloudflare, Copilot, OpenRouter, and all OpenAI-compatible local servers (LM Studio, Ollama, llama.cpp).
• ACPX — the retained harness-subprocess backend, now a bundled dependency (acpx@0.12.0) and a first-class peer endpoint in the routing registry. It drives Claude Code, Codex, OpenCode, Gemini, and the rest of the ACP agent family, plus a --agent <binary> escape hatch for custom harnesses.

The following routing executors have been removed and will now fail with a structured error pointing at this page:

Removed executor	Replace with
claude-code	acpx (agent: claude), or anthropic for direct API
codex	acpx (agent: codex), or openrouter/openai-compatible for direct API
opencode	acpx (agent: opencode), or an OpenAI-compatible direct target
command	acpx with --agent <binary>, or re-implement as an OpenAI-compatible target
anthropic	unchanged (now Pi-backed; same config)
openrouter	unchanged (now Pi-backed)
ollama	unchanged (now Pi-backed via OpenAI-compatible)
llama-cpp	unchanged (now Pi-backed via OpenAI-compatible)
openai-compatible	unchanged (now Pi-backed)

memory.pipelineV2.extraction.provider: command is also retired. Configure a canonical inference.workloads.memoryExtraction target instead; the daemon will reject the retired command configuration rather than silently falling back to another provider.

Action required after updating

1. Check the daemon log on first start. If your agent.yaml references a removed executor, the daemon logs a structured error naming the target and the executor to replace. Background extraction/synthesis will not run until you fix it.
2. Reconfigure the target. In nearly every case the replacement is a one-line edit: change executor: claude-code to executor: acpx (and add acpx: { agent: claude }), or executor: codex to executor: acpx (with acpx: { agent: codex }). Direct-API targets (anthropic, openrouter, ollama, llama-cpp, openai-compatible) keep working as-is — only the underlying engine changed.
3. Verify a background call. Trigger a Dreaming pass or session summary (signet memory ingest ..., or send a session through the pipeline) and confirm it completes.
4. Verify aggregate recall if you use it: signet recall "<query>" --aggregate. It now flows through pi-ai; latency is unchanged (verified at ~200–500ms for synthesis-sized calls on a local model).

Behavioral changes to expect

These are intentional consequences of the cutover, not bugs:

• OpenRouter reasoning controls. The { enabled, maxTokens } reasoning block is now translated by pi-ai into its own { effort } abstraction (omitted entirely when reasoning is disabled) rather than forwarded verbatim. pi-ai owns the reasoning wire format now.
• Keyless local servers (LM Studio, Ollama, llama.cpp) receive a placeholder Authorization header. pi-ai’s credential resolver requires a non-empty API key; local servers ignore it. No real credential is sent.
• Availability probes. Each target now does a /models reachability check so the router can skip unreachable targets before attempting a real call (this also makes fallback attribution correct).

ACPX as a harness replacement

ACPX is bundled with the daemon (no separate install, no npx fetch). It replaces the removed harness subprocess providers. Verified to drive each:

• acpx claude — replaces executor: claude-code
• acpx codex — replaces executor: codex
• acpx opencode — replaces executor: opencode
• acpx --agent <binary> — replaces executor: command

Configure it as a routing target:

inference:
  targets:
    claude-harness:
      executor: acpx
      acpx:
        agent: claude        # or codex, opencode, gemini, ...
      models:
        default:
          model: claude-sonnet-4

daemon-rs removed

The experimental Rust daemon rewrite (platform/daemon-rs) and its shadow proxy, CLI runtime switch (SIGNET_DAEMON_RUNTIME), and route-parity harness have been removed. The TypeScript/Bun daemon is the sole runtime. Rust accelerators (@signet/native) are unaffected.

Need help

If a background pipeline operation does not resume after updating, run signet doctor and check the daemon log for the structured “folded executor” error — it names the exact target and executor to change.