@falai/agent 1.2.8 → 2.0.0

package/docs/guides/flow-control.md

@@ -0,0 +1,409 @@
+---
+title: "Flow control"
+description: "Redirect, complete, abort, or speak verbatim from a tool, hook, or webhook by emitting a directive."
+type: guide
+order: 3
+---
+
+# Flow control
+
+> **Where this is introduced:** [Directives](../concepts/directives.md)
+
+Most steps run, ask the model, write fields, and move on. Some don't.
+A permission tool finds the caller is not eligible. A booking tool
+finishes its work and wants the flow to end without the LLM phrasing a
+confirmation. A webhook decides the next turn should start in a
+different flow than where the user left off. Every one of these is the
+same primitive: emit a [`Directive`](../reference/directive.md).
+
+This guide is a tour of the recipes. Each section is a task and a
+code-block that does it. The shape stays the same across all of them;
+what changes is which fields you set and where the directive comes
+from. The [Directive reference](../reference/directive.md) is the
+canonical contract for every shorthand, object form, and validation
+rule.
+
+## The shape
+
+```typescript
+import type { Directive } from "@falai/agent";
+
+const d: Directive = {
+  goTo: "Booking",                     // position (one max)
+  reply: "Routing you to booking.",    // verbatim utterance
+  dataUpdate: { source: "tool" },      // state write
+};
+```
+
+Every field is optional. State writes (`dataUpdate`,
+`contextUpdate`) and `reply` ride alongside any position field.
+
+## Position fields and precedence
+
+Position answers "where does the conversation go after this turn?"
+
+| Field      | Effect                                                                |
+|------------|-----------------------------------------------------------------------|
+| `goTo`     | Jump to another flow.                                                 |
+| `goToStep` | Jump to a step (within this flow, or — in object form — another flow).|
+| `complete` | Mark the current flow done. Run the flow's completion path.           |
+| `abort`    | End the conversation. Optionally clear the session.                   |
+| `reset`    | Restart the current flow. Optionally clear its declared fields.       |
+
+The fields are **mutually exclusive** — at most one per directive.
+Setting two throws `FlowConfigurationError`. When more than one
+emitter writes a position field on the same turn, the per-turn merge
+picks one winner by precedence:
+
+```
+abort > complete > goTo / goToStep > reset
+```
+
+`abort` always wins — there's no "somewhere else" after the
+conversation has ended. `complete` beats `goTo` — if a follow-up
+jump belongs after completion, put it in `complete.next`. `goTo` and
+`goToStep` share a tier (last emission wins). `reset` is lowest.
+
+State writes and `reply` ride alongside whichever position wins.
+
+## Recipe 1 — Redirect from a tool
+
+Tools have two ways to emit a directive: imperative (`ctx.dispatch`)
+mid-handler, or declarative (`ToolResult.directive`) on return. Both
+land on the same per-turn bus and merge identically.
+
+### Imperative — `ctx.dispatch`
+
+```typescript
+import type { Tool } from "@falai/agent";
+
+const checkEligibility: Tool<{ userId: string }, BookingData, { ok: boolean }> = {
+  id: "check_eligibility",
+  description: "Verify the caller is allowed to book this destination.",
+  isReadOnly: () => true,
+  async handler(ctx) {
+    const ok = await isEligible(ctx.context.userId, ctx.data.destination);
+    if (!ok) {
+      ctx.dispatch({
+        goTo: "Denial",
+        reply: "Sorry — you're not eligible to book that destination.",
+        dataUpdate: { denialReason: "ineligible" },
+      });
+      return { ok: false };
+    }
+    return { ok: true };
+  },
+};
+```
+
+Multiple `dispatch` calls in one handler are allowed — they
+concatenate alongside emissions from other tools and hooks before the
+merge runs.
+
+### Declarative — `ToolResult.directive`
+
+```typescript
+async handler(ctx) {
+  const ok = await isEligible(ctx.context.userId, ctx.data.destination);
+  if (!ok) {
+    return {
+      data: { ok: false },
+      directive: {
+        goTo: "Denial",
+        reply: "Sorry — you're not eligible to book that destination.",
+      },
+    };
+  }
+  return { data: { ok: true } };
+}
+```
+
+Reach for **imperative** when the handler still has work after the
+decision. Reach for **declarative** when there's a single return
+point.
+
+## Recipe 2 — Complete with a chained next
+
+A booking tool reserved the room. The flow is done, and the next
+thing the agent should do is open a feedback flow. `complete` accepts
+an object form whose `next` field is **another directive** applied
+immediately after the flow's completion path runs:
+
+```typescript
+import type { Tool, Directive } from "@falai/agent";
+
+const bookHotel: Tool<unknown, BookingData, { id: string }> = {
+  id: "book_hotel",
+  description: "Reserve the hotel for the collected fields.",
+  async handler(ctx, args) {
+    const id = await reserve(args);
+    const directive: Directive = {
+      complete: {
+        reason: "reservation confirmed",
+        next: { goTo: "Feedback", reply: "Booked. Mind a quick survey?" },
+      },
+      dataUpdate: { bookingId: id },
+    };
+    return { data: { id }, directive };
+  },
+};
+```
+
+What runs, in order: the tool's directive lands on the bus, the merge
+picks `complete`, the flow's `hooks.onComplete` runs, then
+`complete.next` is applied — `goTo: "Feedback"` redirects with the
+verbatim reply as that turn's assistant message.
+
+`complete.next` is one level deep on purpose — chains do not nest.
+For the simple case, the shorthand `complete: true` is the right
+call:
+
+```typescript
+return { data, directive: { complete: true, dataUpdate: { bookingId: id } } };
+```
+
+## Recipe 3 — Abort on a permission failure
+
+`abort` ends the conversation. Use it when there's no flow to
+redirect *to*:
+
+```typescript
+import type { Tool, Directive } from "@falai/agent";
+
+const verifyAccess: Tool = {
+  id: "verify_access",
+  isReadOnly: () => true,
+  async handler(ctx) {
+    const allowed = await acl.check(ctx.context.userId);
+    if (!allowed) {
+      const directive: Directive = {
+        abort: { reason: "caller is not on the allow-list", clearSession: true },
+      };
+      return { data: { allowed: false }, directive };
+    }
+    return { data: { allowed: true } };
+  },
+};
+```
+
+Two things to know:
+
+- **`abort` cannot co-exist with `reply`.** Aborted conversations
+  don't deliver replies. To say something on the way out, use
+  `complete` plus `reply` instead.
+- **`clearSession: true`** purges the session at the next persistence
+  write. Without it, the aborted session sticks around for traces.
+
+## Recipe 4 — Reply verbatim from a finalize hook
+
+Some utterances should be exact: confirmations, bridges, refusals,
+boilerplate at flow boundaries. `reply: string` skips the LLM and
+emits the literal text — no templating, no model rephrasing.
+
+```typescript
+import type { Directive } from "@falai/agent";
+
+const step = {
+  id: "confirm_handoff",
+  prompt: "Confirm the handoff if the queue is clear.",
+  hooks: {
+    finalize: ({ data }): Directive | void => {
+      if (data.handoffReady) {
+        return { reply: "Connecting you with a specialist now." };
+      }
+    },
+  },
+};
+```
+
+The turn ends with `stoppedReason: "reply"` and the literal string as
+`response.message`. State writes and a position field can ride
+alongside:
+
+```typescript
+finalize: ({ data }) => {
+  if (data.bookingId) {
+    return {
+      reply: `Booked. Confirmation: ${data.bookingId}.`,
+      dataUpdate: { confirmedAt: new Date().toISOString() },
+      complete: true,
+    };
+  }
+}
+```
+
+That single directive does three things at once. They're orthogonal
+payloads — `reply`, `dataUpdate`, and the position field don't
+compete for the same slot.
+
+To skip the LLM **before** it runs (rather than from a finalize hook
+*after*), use a [`PreDirective`](../reference/pre-directive.md) from
+a prepare hook with `halt: true` — see below.
+
+## PreDirective extras (pre-LLM only)
+
+Pre-LLM hooks (`flow.hooks.onEnter`, `step.hooks.onEnter`,
+`step.hooks.prepare`) return a [`PreDirective`](../reference/pre-directive.md) —
+the `Directive` variant with three extra fields that only make sense
+before this turn's LLM call.
+
+```typescript
+interface PreDirective extends Directive {
+  appendPrompt?: string[];
+  injectTools?: Tool[];
+  halt?: boolean;
+}
+```
+
+Lifetime is one turn. None of the three fields persist on
+`session.pendingDirective` — they're stripped before the write.
+Returning a PreDirective from a post-LLM hook drops these fields with
+a debug log.
+
+### `appendPrompt` — nudge the system prompt
+
+```typescript
+const flow = {
+  title: "Booking",
+  hooks: {
+    onEnter: (ctx) => {
+      if (ctx.context.user.tier === "vip") {
+        return { appendPrompt: ["Caller is a VIP — confirm preferences first."] };
+      }
+    },
+  },
+};
+```
+
+Each string is appended to the system prompt for this turn only.
+Multiple emitters' arrays concatenate in emission order; duplicates
+are preserved.
+
+### `injectTools` — one-shot tool surface
+
+```typescript
+const step = {
+  id: "verify",
+  prompt: "Verify the caller before proceeding.",
+  hooks: {
+    prepare: async (ctx) => {
+      if (!ctx.data.verified) return { injectTools: [lookupAccount] };
+    },
+  },
+};
+```
+
+Tools listed here are added for this turn only. Multiple emitters'
+arrays concatenate, then dedupe by `Tool.id` (last definition wins).
+
+### `halt` — skip the LLM call
+
+```typescript
+prepare: async (ctx) => {
+  if (ctx.data.alreadyVerified) {
+    return { halt: true, reply: "Already verified. How can I help?" };
+  }
+}
+```
+
+When any pre-phase emitter sets `halt: true`, the LLM call is
+skipped. With `reply`, the turn ends `stoppedReason: "reply"`. Without
+`reply`, the turn ends `stoppedReason: "halt"` and an empty body.
+Multiple emitters merge by logical-OR.
+
+## Recipe 5 — Dispatch from outside a turn
+
+Tools and hooks emit directives onto the **per-turn** bus. Sometimes
+the redirect comes from outside any turn — a webhook fires, a
+scheduled job notices an idle session, an external system marks a
+caller upgraded. There's no turn running, so there's no bus.
+
+`Agent.dispatch(target, session)` writes a `pendingDirective` onto
+the session without invoking a turn. The directive is consumed at the
+**start of the next** `respond` call — before routing, before
+pre-extraction, before any phase runs.
+
+```typescript
+// Webhook handler: redirect a session from outside a turn.
+import type { Directive } from "@falai/agent";
+
+app.post("/webhook/account-upgraded", async (req, res) => {
+  const session = await agent.session.getOrCreate(req.body.sessionId);
+  await agent.dispatch(
+    { goTo: "VipFlow", reply: "You've been upgraded. Let's start fresh." },
+    session
+  );
+  res.sendStatus(204);
+});
+```
+
+Two forms:
+
+```typescript
+// String shorthand — desugars to { goTo: "Feedback" }
+await agent.dispatch("Feedback", session);
+
+// Full directive
+await agent.dispatch({ goTo: "Billing", reply: "Transferring you now." }, session);
+```
+
+The call validates the directive (`flow.validate`), confirms any
+`goTo`-named flow exists (throws `FlowConfigurationError` if not),
+strips PreDirective-only fields, writes `pendingDirective` onto the
+session, and persists if an adapter is configured.
+
+`pendingDirective` is **single-shot** — consumed exactly once and
+cleared. Calling `dispatch` again before the next turn overwrites the
+previous one (last-wins). To merge with a pending directive instead
+of overwriting, use `flow.merge`:
+
+```typescript
+import { flow } from "@falai/agent";
+
+const merged = flow.merge(
+  session.pendingDirective ?? {},
+  { dataUpdate: { tier: "vip" } }
+);
+await agent.dispatch(merged, session);
+```
+
+## Picking the right tool
+
+Where to emit:
+
+| You're in a... | Type | Use |
+|----------------|------|-----|
+| Tool handler, mid-flight | `Directive` | `ctx.dispatch(d)` |
+| Tool handler, on return | `Directive` | `return { data, directive: d }` |
+| `prepare` / `onEnter` hook | `PreDirective` | `return d` |
+| `finalize` / `onComplete` hook | `Directive` | `return d` |
+| Branch `then` target | `Directive` | `then: d` (see [Branching](./branching.md)) |
+| Outside a turn (webhook, job) | `Directive` | `await agent.dispatch(d, session)` |
+
+Which position field:
+
+- Inside the same flow → `goToStep`.
+- Another flow → `goTo` (or `goToStep` with `flow:` set).
+- Work is done → `complete` (with `complete.next` for follow-ups).
+- No path forward → `abort` (`clearSession: true` if reuse is unsafe).
+- Start the flow over → `reset`.
+
+Code or model speaking:
+
+- Verbatim → set `reply`.
+- From the model → leave `reply` unset; let the LLM call run.
+
+## See also
+
+- [Directive reference](../reference/directive.md) — every field,
+  every shorthand, every validation rule.
+- [PreDirective reference](../reference/pre-directive.md) —
+  pre-LLM-only fields and merge rules.
+- [Directives concept](../concepts/directives.md) — the mental model
+  and the inheritance chain.
+- [Branching](./branching.md) — when the redirect is source-local
+  rather than dynamic.
+- [Turn pipeline](../concepts/pipeline.md) — when and where directives
+  apply within a turn.
+
+**Next:** [Instructions](./instructions.md)

package/docs/guides/instructions.md

@@ -0,0 +1,210 @@
+---
+title: "Instructions"
+description: "Shape how the agent talks with must / never / should at agent, flow, and step scope, and read back exactly which ones fired."
+type: guide
+order: 4
+---
+
+# Instructions
+
+Use `Instruction` when the agent should always confirm dates, never quote prices it has not looked up, or — only inside the booking flow — try to compare two options before committing. One primitive covers all three. Severity comes from `kind`. Reach comes from where you declare the instruction. The framework reports back which instructions actually rendered on each turn, so the behavior surface is observable, not guessed.
+
+This guide assumes a `createAgent` scaffold. If you need one, start from [Your first agent](../start/02-first-agent.md).
+
+## Instruction shape
+
+An `Instruction` is a plain object with one required field — `prompt` — and a handful of optional ones. The full contract lives in the [Instruction reference](../reference/instruction.md); the working subset for this guide is:
+
+```typescript
+import type { Instruction } from "@falai/agent";
+
+const ins: Instruction = {
+  kind: "must",                                     // 'must' | 'never' | 'should' (default 'should')
+  when: "User asks about pricing",                  // optional AI-evaluated activation
+  if: (ctx) => ctx.context.tier === "pro",          // optional code-evaluated activation
+  prompt: "Quote only rates fetched from the live API this turn.",
+};
+```
+
+`prompt` is the behavioral text. `kind` declares severity. `when` and `if` gate the instruction by activation, with `if` running first and `when` only evaluated if `if` passes. Everything else (`id`, `enabled`, `tags`, `metadata`) is bookkeeping you can ignore until you need it.
+
+## The three `kind` values
+
+`kind` answers *how strict is this?* Three values, no inheritance, no ordering:
+
+| `kind` | Use it for | Example |
+|--------|------------|---------|
+| `must` | Absolute do. The agent is required to follow it whenever rendered. | `"Validate dates are in the future before booking."` |
+| `never` | Absolute don't. The agent is forbidden from doing it. | `"Promise rates you have not looked up."` |
+| `should` | Conditional nudge. Default kind. The agent should try, but it's not a hard line. | `"Offer to compare two options before committing."` |
+
+Default is `should`. Reach for `must` or `never` only when the behavior is non-negotiable — they are louder in the prompt and less forgiving in tone. If three flows each need the same `must`, declare it once at the agent scope; if it only matters for one flow, declare it there.
+
+## Agent vs flow vs step scope
+
+The same shape attaches at three positions:
+
+```typescript
+createAgent({
+  instructions: [/* agent scope — every turn, every flow */],
+  flows: [
+    {
+      title: "Booking",
+      instructions: [/* flow scope — only when 'Booking' is active */],
+      steps: [
+        {
+          id: "payment",
+          instructions: [/* step scope — only when 'payment' is the current step */],
+        },
+      ],
+    },
+  ],
+});
+```
+
+Reach narrows as you nest. Agent-scope instructions render on every turn for every flow. Flow-scope instructions render only while that flow is active. Step-scope instructions render only while that step is current. Conditions on `when` and `if` apply on top — an agent-scope instruction with `when: "User is angry"` still requires that activation to fire, just like a step-scope one.
+
+The composer renders the resolved set under a single `## Instructions` heading and tags each line with a scope caption so the model can read both *what* and *where from* in one pass:
+
+| Scope | Caption | When it appears |
+|-------|---------|-----------------|
+| Agent | `[Always]` | Every turn. |
+| Flow | `[In: <FlowTitle>]` | While the named flow is active. |
+| Step | `[Step: <stepId>]` | While the named step is current. |
+
+## Rendering format
+
+Each active instruction lands in the prompt as a single bullet:
+
+```
+- [<kind>] [<scope>] <prompt>
+```
+
+A composed block looks like this:
+
+```
+## Instructions
+
+- [must] [Always] Validate dates are in the future before booking.
+- [never] [Always] Promise rates you have not looked up.
+- [should] [In: Booking] Offer to compare two options before committing.
+- [must] [Step: payment] If the card is declined, never retry without confirmation.
+```
+
+The format is fixed. The kind prefix is always present (defaulting to `[should]`), the scope caption is always present, and the prompt text follows verbatim. There is nothing to configure — what you write in `prompt` is what the model reads.
+
+## `appliedInstructions` on the response
+
+Every `respond()` call returns an `appliedInstructions` array listing exactly which instructions were rendered into that turn's prompt. The set is deterministic — it comes from the prompt composer, not the model — so you can use it for observability, audits, and tests:
+
+```typescript
+const response = await agent.respond("I want to book a room.");
+
+for (const a of response.appliedInstructions ?? []) {
+  console.log(`${a.scope}${a.scopeRef ? `:${a.scopeRef}` : ""} → ${a.id}`);
+}
+// global → ins_validate_dates
+// global → ins_no_unquoted_prices
+// flow:Booking → ins_offer_two_options
+```
+
+Each `AppliedInstruction` carries the firing instruction's `id`, the originating `scope`, and a `scopeRef` that is the `flowTitle` for flow scope, the `stepId` for step scope, and `undefined` for agent scope. If you want stable ids in the report, set `id` on the instructions you care about — otherwise the framework auto-generates them.
+
+The same array lands on the final chunk of `respondStream`:
+
+```typescript
+for await (const chunk of agent.respondStream("I want to book a room.")) {
+  if (chunk.done) {
+    console.log("rendered:", chunk.appliedInstructions);
+  }
+}
+```
+
+`appliedInstructions` is empty on intermediate chunks and populated only when `done: true`.
+
+## Recipe: agent-level absolutes
+
+Two house rules every flow must respect:
+
+```typescript
+const agent = createAgent({
+  name: "BookingBot",
+  provider: new GeminiProvider({ apiKey: process.env.GEMINI_API_KEY! }),
+  instructions: [
+    { id: "ins_validate_dates", kind: "must",  prompt: "Validate dates are in the future before booking." },
+    { id: "ins_no_unquoted",    kind: "never", prompt: "Promise rates you have not looked up." },
+  ],
+  flows: [/* ... */],
+});
+```
+
+These render with caption `[Always]` on every turn, regardless of which flow is active.
+
+## Recipe: flow-level conditional
+
+A flow-scope nudge that only fires when the user is comparison shopping:
+
+```typescript
+const booking = {
+  title: "Booking",
+  instructions: [
+    {
+      id: "ins_offer_two_options",
+      kind: "should",
+      when: "User is comparing options or asking which is better",
+      prompt: "Offer to compare two options side by side before recommending one.",
+    },
+  ],
+  steps: [/* ... */],
+};
+```
+
+While `Booking` is active, the line renders with caption `[In: Booking]` only on turns where the AI condition resolves true. On other turns the entry stays out of the prompt and out of `appliedInstructions`.
+
+## Recipe: step-level reminder
+
+A `must` that scopes to a specific step, gated by a code condition:
+
+```typescript
+const paymentStep = {
+  id: "payment",
+  prompt: "Take payment.",
+  instructions: [
+    {
+      id: "ins_no_retry_on_decline",
+      kind: "must",
+      if: (ctx) => ctx.data.lastChargeStatus === "declined",
+      prompt: "If the card is declined, never retry without explicit user confirmation.",
+    },
+  ],
+};
+```
+
+Renders with caption `[Step: payment]` only while `payment` is the current step *and* the most recent charge was declined. Move off the step or change `lastChargeStatus`, and it drops out cleanly.
+
+## Recipe: reading `appliedInstructions` in a test
+
+Because the set is deterministic, you can assert against it directly:
+
+```typescript
+import { test, expect } from "bun:test";
+
+test("payment step renders the no-retry rule when the card was declined", async () => {
+  const response = await agent.respond("Try again.", {
+    sessionId: "s_1",
+    initialContext: {},
+    initialData: { lastChargeStatus: "declined" },
+  });
+
+  const ids = response.appliedInstructions?.map(a => a.id) ?? [];
+  expect(ids).toContain("ins_no_retry_on_decline");
+});
+```
+
+No fixtures, no LLM mocks — the rendered set is computable from configuration and condition state.
+
+## Where this fits
+
+`Instruction` shapes how the agent *talks*. To shape what it *does* — redirect, complete, abort — return a [Directive](../reference/directive.md) from a tool or hook (covered in [Flow control](./flow-control.md)). The two surfaces compose: an instruction nudges the model to confirm dates, a tool's directive completes the flow once the booking is written.
+
+**Next:** [Persistence](./persistence.md)