pi-agent-integration

Implement Pi consumers against the latest published API. Keep streaming stable, queue behavior explicit, and wrapper code small.

Step 1: Classify the request

Load each reference that the task needs. Keep guidance Pi-specific unless the user asks about a consuming product.

Step 2: Apply integration guardrails

1. Check npm latest for @earendil-works/pi-agent-core before relying on a contract. Treat the published package as authoritative. Treat upstream main as unreleased evidence.
2. Pass a streamFn when constructing Agent. Also pass it as the last argument to low-level loop functions.
3. Use Agent when event handling must finish before run settlement. Use agentLoop only when an observational event stream is enough.
4. Stream user-visible text only from message_update when assistantMessageEvent.type === "text_delta".
5. Preserve assistant message boundaries when forwarding multi-message output.
6. Do not call prompt(), continue(), or reset() while an agent is active. Queue mid-run input with steer() or followUp().
7. Continue a normal run only from a non-empty user or toolResult tail. An assistant tail can only drain queued steering or follow-up messages.
8. Keep streamFn, convertToLlm, transformContext, getApiKey, queue providers, and loop hooks no-throw for expected failures. Return safe values or encode the failure in stream events.
9. Keep tool calls, progress, results, thinking deltas, and provider payloads internal unless the product exposes them on purpose.
10. Since 1.0.0, the package contains only Agent, the agent loop, the proxy stream, and their types. The harness, sessions, compaction, skills, and prompt-template helpers are gone. Do not recommend them.

Step 3: Implement with minimal surface

1. Prefer Pi options over wrapper state machines: streamFn, getApiKey, sessionId, thinkingBudgets, transport, maxRetryDelayMs, identity, onPayload, onResponse, onProviderStreamEvent, beforeToolCall, afterToolCall, finishTurn, prepareRequest, prepareNextTurnWithContext, toolExecution, steeringMode, and followUpMode.
2. Mutate Agent through agent.state and public runtime options. agent.state.systemPrompt is read-only: append a system message to change the prompt. Call reset() only when idle.
3. Use transformContext for message-level pruning or injection. Use convertToLlm for provider-compatible role conversion and filtering.
4. Set queue modes to "one-at-a-time" or "all" when ordering or batching matters.
5. Use streamFn with streamProxy-style behavior for server-proxied model access.
6. Use toolExecution, per-tool executionMode, beforeToolCall, afterToolCall, thrown errors or isError: true results, and terminate before adding a custom tool runner.
7. Keep timeout and abort paths observable. Make sure streams and iterables settle.

Step 4: Verify behavior

1. Verify the event bridge emits only text deltas, preserves intended boundaries, and closes on success, error, and abort.
2. Verify active-run handling for prompt(), continue(), and reset(). Verify queued steer() and followUp() behavior.
3. Verify continue() with empty history and with user, toolResult, and assistant tails.
4. Verify custom messages remain in agent state while convertToLlm emits only provider-compatible messages.
5. Verify streamFn encodes expected provider failures instead of throwing or rejecting.
6. Verify parallel and sequential tool ordering, hook blocking and patches, progress updates, usage, agent.state.tools changes, and termination.
7. Verify Agent.subscribe() listener settlement and waitForIdle() with async listeners.

Step 5: Keep version discipline

1. Target npm latest only.
2. Re-check package metadata, declarations, implementation, README, and changelog before a material API update.
3. Do not present unreleased main behavior as published behavior.
4. Do not add compatibility shims or old package guidance unless the user asks for a migration.