v0.8.0

What's new in 0.8.0

Added

• OpenAI-compatible chat-completions HTTP face (amplifier-agent serve chat-completions) — /v1/models + /v1/chat/completions over HTTP with bearer-token auth. Streams responses, multi-provider routing via served-models registry, workspace correlation through X-Client-Session-Id header. Enables direct integration with opencode via the amplifier-app-opencode wrapper and any OpenAI-compatible client. (#65)

• amplifier-agent auth subcommand — set/list/remove/status/clear actions over ~/.amplifier-agent/credentials.json (mode 0600). Resolution chain is env-first: shell env vars (ANTHROPIC_API_KEY, …) win over the file so existing shell-rc workflows are unchanged. "Set once, works everywhere" UX matching claude login / gh auth login / aws configure without the OAuth ceremony. (#65)

• Host-tool delegation — tools declared by the host (host_config.json:host_tools) are surfaced to the model with stub schemas; on invocation, server emits a signal tool_call back to the client (same chunk_id), client executes host-side, returns result for the model to continue. Lets the host own filesystem, shell, browser, or any custom tool without bundling into amplifier-agent. (#65)

• Model routing matrix integration — per-role provider/model preferences resolved per turn. (#64)

• cost_usd in chat-completions usage envelope — non-standard field carrying the actual dollar cost provider modules computed, accumulated across sub-turns, serialized as a string for Decimal precision. Standard OpenAI clients ignore the field; cost-aware clients render the real per-turn $$. (#68)

• host_config.providers (plural) registry — declares which providers the server-mode lifespan loads and how to instantiate each. Schema: providers: {<provider_id>: {module?: str, config?: dict}}. Module defaults to provider_id when omitted; each provider's config is passed as extra_config into list_provider_models(). (#69)

• amplifier-agent serve status / stop / restart subcommands — operational lifecycle for the chat-completions HTTP server. Status reports running/stale/not-running, where it's reachable, and how many models from which providers it's serving (self-cleans stale state files when the PID is gone). Stop sends SIGTERM with a configurable graceful-exit window (--timeout), escalating to SIGKILL on expiry or on --force. Restart performs an identity-restart using the args stored at original launch. State is tracked in ~/.amplifier-agent/state/serve.json (mode 0600, parent dir 0700; api_key is sensitive and never logged). (#69)

Changed

• Breaking (server mode only): amplifier-agent serve chat-completions now requires host_config.providers to be a non-empty dict. Any provider declared there that cannot initialize (missing credentials, module not installed, list_models() raises, returns 0 models) causes the server to exit 2 with a structured error listing every problem. The previous behavior — iterating a hardcoded KNOWN_PROVIDERS list, silently skipping unreachable providers, and falling back to an unusable placeholder amplifier model — is gone. Single-turn mode (amplifier-agent run) is unaffected. (#69)

• POST /v1/chat/completions now validates model against the served registry. Requests with an unknown model return HTTP 400 {"error": {"code": "unknown_model", ...}} immediately, instead of being silently routed to whichever provider loaded first and failing 4 seconds later with an upstream not_found_error embedded in delta.content. (#69)

• stream: false is now honored. Requests with that flag return a single JSON body; only stream: true (or absent) uses SSE. (#69)

• Upstream errors raised before any content chunks are emitted now surface as HTTP 502 with a structured OpenAI-shape error envelope, instead of being embedded inside delta.content of a 200 SSE response. (Mid-stream errors after the first chunk remain embedded in delta.content — once SSE starts, the status line is committed.) (#69)

• /v1/models no longer falls back to a placeholder {"id": "amplifier", ...} entry. The lifespan guarantees served_models_registry is non-empty (or the server exits at boot), so the fallback was unreachable in practice. (#69)

• Lifespan provider iteration now reads from host_config.providers instead of the hardcoded KNOWN_PROVIDERS catalog. The CLI's KNOWN_PROVIDERS constant stays defined for single-turn mode and other CLI commands. (#69)

• /v1/models response surfaces a _provider tag per model so OpenAI-compatible clients can see which provider serves each entry. (#65)

• Usage-counter telemetry in chat-completions responses correctly reflects the provider that actually served the turn. (#65)

Wire protocol

Unchanged at 0.3.0 — no wrapper bump. TypeScript wrapper stays at 0.7.0, Python wrapper stays at 0.3.0.

Migration

Existing server-mode users (anyone running amplifier-agent serve chat-completions on a 0.7.x or pre-#69 0.8.0 commit): Add a providers block to your host_config.json. Minimum to keep working with just Anthropic:

{
  "providers": {
    "anthropic": {}
  }
}

Multi-provider example:

{
  "providers": {
    "anthropic": {},
    "openai":    {"config": {"base_url": "https://api.openai.com/v1"}}
  }
}

Without host_config.providers, the server will exit at boot with a clear error message rather than running in a broken half-state.

No other breaking changes. Existing CLI (run, models list), JSON-RPC wire, single-turn mode, and HTTP clients reading standard OpenAI fields all continue to work.

See CHANGELOG.md [0.8.0] for full details.