Common patterns

Chidori has a small number of primitives that compose into most agent shapes. This page maps jobs to primitives; each recipe links to the doc that covers it in depth, and to a runnable example where one exists.

Approval gates that show their work

Never make a human approve blind: pass the artifact under review as details. The CLI prints it above the prompt; a paused session exposes it as pending_details next to pending_prompt.

const draft = await chidori.prompt("Draft the announcement.", { type: "draft" });

const verdict = await chidori.input("Ship this announcement?", {
  type: "approval",
  choices: ["yes", "no"],
  default: "no",
  details: draft,
});

if (verdict.toLowerCase() !== "yes") return { shipped: false };

Under chidori serve, the input() suspends the session to disk — no process waits while the human decides. Days later, POST /sessions/{id}/resume picks the run up exactly where it paused. Example: examples/agents/input_pause.ts.

An agent behind a webhook

Under chidori serve, any route other than the /sessions/* API is folded into { event: { method, path, headers, query, body } } and run as the agent's input; returning { status, body, headers? } shapes the HTTP response:

import { chidori, run } from "chidori:agent";

run(async (input: { event: { method: string; path: string; body?: unknown } }) => {
  const { event } = input;
  if (event.method !== "POST" || event.path !== "/hooks/pr") {
    return { status: 404, body: { error: "not found" } };
  }
  const triage = await chidori.prompt(
    `Triage this pull request event:\n${JSON.stringify(event.body)}`,
    { type: "final" },
  );
  return { status: 200, body: { triage } };
});

Branch on input.event early, as above: every request runs the whole agent, so return a cheap 4xx for non-events before any model call. A run that pauses (on input(), a signal, or a policy approval) is stored as a real session and answered 202 with the session view, so the caller can resume or signal it later — a webhook can open a long-lived, durable workflow. Deep doc: Running Modes (which also covers outbound requests from a handler — examples/agents/webhook.ts).

Give the model tools, keep the loop

prompt() with tools runs the whole provider tool-use loop for you. Hand-roll with context().respond() / toolResult() only when you need per-step control:

const answer = await chidori.prompt(input.question, {
  tools: [searchNotes, wikiSearch], // defineTool handles and/or registry names
  maxTurns: 6,
});

Every model turn and tool invocation is journaled, so the loop shows up in chidori trace — and replays for $0. Examples: examples/agents/tool_use.ts (built-in loop), examples/agents/worker.ts (author-driven loop).

Fan out drafts, concurrently

const [a, b, c] = await chidori.util.parallel(
  [
    () => chidori.prompt("Draft: crisp and technical", { type: "draft" }),
    () => chidori.prompt("Draft: warm and story-led", { type: "draft" }),
    () => chidori.prompt("Draft: contrarian angle", { type: "draft" }),
  ],
  { concurrency: 3 },
);
const winner = await chidori.prompt(
  `Pick the strongest draft and say why:\n\nA:\n${a}\n\nB:\n${b}\n\nC:\n${c}`,
  { type: "final" },
);

util.parallel is in-VM control flow — only the prompts inside are durable calls. When the strategies differ (different code, not just different prompts), use chidori.branch instead: each variant runs its own module from the current state, and the whole fan-out replays as one recorded call. Example: examples/branching/.

Multiplayer review

A run can wait on named signals from several parties — humans or other agents — with a durable mailbox absorbing whatever arrives early:

// Wait for either reviewer; the result's `name` says who fired.
const first = await chidori.signal(["design-review", "security-review"], {
  timeoutMs: 24 * 60 * 60 * 1000,
});
if (first.timedOut) return { escalated: true };

Deliver with POST /sessions/{id}/signal — the server resolves a matching pause, pushes to a live streaming run, or queues into the mailbox. Example: examples/multiplayer-review/, deep doc: Signals.

A service that sleeps between events

A detached agent owns its own run, journal, and mailbox, and outlives the run that spawned it. Waiting costs nothing: signal() hibernates with no thread and no VM, and alarm() is the durable "wake me every N hours" timer:

// services/inbox-triager.ts
import { chidori, run } from "chidori:agent";

run(async () => {
  for (;;) {
    const msg = await chidori.signal(["email", "shutdown"], {
      timeoutMs: 6 * 60 * 60 * 1000, // or use chidori.alarm for pure maintenance ticks
    });
    if (msg.timedOut) { await runMaintenance(); continue; }
    if (msg.name === "shutdown") return { stopped: true };
    await triage(msg.payload);
  }
});

Spawn it once with chidori.agents.spawn(..., { name: "inbox-triager" }); send to it from any run or via POST /agents/detached/{name}/send. The fleet re-arms at boot, so a server restart loses nothing. Deep doc: Detached Agents.

Don't re-pay expensive compute

Replay re-executes your TypeScript. That's usually free — but a genuinely expensive pure computation (parsing a big corpus, building an index) would re-run on every resume. Journal its value once:

const index = await chidori.step("build-index", () => buildIndex(corpus));

Resume returns the recorded value without re-running the callback. The callback must be pure, synchronous compute — host effects inside a step throw. Deep doc: Value Checkpoints, example: examples/agents/value_checkpoint.ts.

Checkpoint-as-test

A recorded run is a complete, deterministic specification of your agent's behavior. Commit one and assert against it in CI:

chidori run agent.ts --input question="smoke test"   # record once
git add .chidori/runs/<run_id>                        # commit the recording
chidori verify agent.ts <run_id>                      # in CI: exit 0 = no drift

verify replays with no provider configured and a deny-all policy — if the agent's prompts, tool calls, or control flow drift from the recording, it fails. A full integration test that costs $0 and runs in milliseconds. Observing with Tael builds its golden regression cases on the same mechanism.

Keep long conversations inside the window

Compaction is explicit — it changes what the model sees, so it's an author decision, and it's a recorded prompt call, so it replays deterministically:

const chat = chidori.conversation({
  system: PERSONA,
  compact: { budgetTokens: 8000 }, // no-op until the tail exceeds budget
});

Or on a raw context: ctx = await ctx.compact({ budgetTokens: 8000 }). Deep doc: Context Management.