opencode/packages/core/src/session/runner/llm.ts

487 lines
23 KiB
TypeScript

export * as SessionRunnerLLM from "./llm"
import { LLMClient, LLMError, LLMEvent, isContextOverflowFailure, type ProviderErrorEvent, type ToolCall } from "@opencode-ai/ai"
import { Cause, Data, Effect, Exit, Fiber, FiberSet, Layer, Option, Pull, Schedule, Semaphore, Stream } from "effect"
import { Database } from "../../database/database"
import { EventV2 } from "../../event"
import { PermissionV2 } from "../../permission"
import { QuestionTool } from "../../tool/question"
import { ToolOutputStore } from "../../tool-output-store"
import { InstructionState } from "../instruction-state"
import { SessionCompaction } from "../compaction"
import { SessionContext } from "../context"
import { SessionEvent } from "../event"
import { SessionPending } from "../pending"
import { SessionModelRequest } from "../model-request"
import { SessionMessage } from "../message"
import { SessionSchema } from "../schema"
import { SessionStore } from "../store"
import { SessionTitle } from "../title"
import { Service } from "./index"
import { createLLMEventPublisher } from "./publish-llm-event"
import { Snapshot } from "../../snapshot"
import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
import { llmClient } from "../../effect/app-node-platform"
import { StepFailedError } from "../error"
import { toSessionError } from "../to-session-error"
import { SessionRunnerRetry } from "./retry"
import { SessionUsage } from "../usage"
/** How one model call ended: settled, awaiting a scheduled retry, or restarted by compaction. */
type CallOutcome = Data.TaggedEnum<{
Completed: { readonly needsContinuation: boolean; readonly step: number }
Retry: { readonly step: number }
Restart: { readonly step: number; readonly recoveredOverflow: boolean }
}>
const CallOutcome = Data.taggedEnum<CallOutcome>()
// Declining an interactive prompt halts the drain instead of becoming model-facing tool output.
const isUserDeclined = (cause: Cause.Cause<unknown>) =>
cause.reasons.some(
(reason) =>
Cause.isDieReason(reason) &&
(reason.defect instanceof PermissionV2.DeclinedError || reason.defect instanceof QuestionTool.CancelledError),
)
/**
* Classifies how the owned tool fibers ended. Interrupts and interactive declines abort
* the step; a defect from a tool implementation becomes a failed tool call the model can
* read; a typed infrastructure failure must fail the assistant and then the drain.
*/
const classifyToolExits = (settled: Exit.Exit<Array<Exit.Exit<void, ToolOutputStore.Error>>, never>) => {
const causes =
settled._tag === "Failure"
? [settled.cause]
: settled.value.flatMap((exit) => (exit._tag === "Failure" ? [exit.cause] : []))
const failure = causes.find((cause) => !Cause.hasInterrupts(cause) && !isUserDeclined(cause))
return {
interrupted: causes.some(Cause.hasInterrupts),
declined: causes.some(isUserDeclined),
failure,
infraError: failure === undefined ? undefined : Option.getOrUndefined(Cause.findErrorOption(failure)),
}
}
const layer = Layer.effect(
Service,
Effect.gen(function* () {
const events = yield* EventV2.Service
const llm = yield* LLMClient.Service
const store = yield* SessionStore.Service
const context = yield* SessionContext.Service
const modelRequests = yield* SessionModelRequest.Service
const snapshots = yield* Snapshot.Service
const db = (yield* Database.Service).db
const compaction = yield* SessionCompaction.Service
const title = yield* SessionTitle.Service
// Title generation is a side effect of the first step; it must not delay step continuation.
// Tracked per process so repeated wakes before the second user message arrives don't
// re-fire a redundant LLM call; `SessionTitle` itself is idempotent based on durable history.
const titleStarted = new Set<SessionSchema.ID>()
const forkTitle = yield* FiberSet.makeRuntime<never, void, never>()
/**
* Drains eligible manual compaction and user input until the Session becomes idle.
* Execution lifecycle is published per busy period by SessionExecution, not here.
*/
const drain = Effect.fn("SessionRunner.drain")(function* (input: {
readonly sessionID: SessionSchema.ID
readonly force: boolean
}) {
if (!input.force && !(yield* SessionPending.has(db, input.sessionID, "any"))) return
yield* settleStaleToolCalls(input.sessionID)
yield* runPendingCompaction(input.sessionID)
if (!input.force && !(yield* SessionPending.has(db, input.sessionID, "input"))) return
do {
yield* runSteps(input.sessionID)
} while (yield* SessionPending.has(db, input.sessionID, "input"))
})
/**
* Runs logical steps until no tool result or newly admitted steer requires another
* model call. Queued inputs remain pending until the current model work reaches idle.
*/
const runSteps = Effect.fn("SessionRunner.runSteps")(function* (sessionID: SessionSchema.ID) {
// Fresh work may promote queued input; later steps absorb steers only.
let promotable: SessionPending.Promotable = "input"
let step = 1
while (true) {
const result = yield* runStep(sessionID, promotable, step)
yield* startTitleOnce(sessionID)
yield* runPendingCompaction(sessionID)
if (!result.needsContinuation && !(yield* SessionPending.has(db, sessionID, "steer"))) return
promotable = "steer"
step = result.step + 1
}
})
/** Completes one logical model step, transparently retrying or rebuilding after compaction. */
const runStep = Effect.fnUntraced(function* (
sessionID: SessionSchema.ID,
promotable: SessionPending.Promotable,
step: number,
) {
// Minting message identity before any attempt lets retries resume the same durable
// message. A compaction restart re-mints: the old message is stranded behind the new
// compaction boundary, so the rebuilt step needs identity inside the new epoch.
let assistantMessageID = SessionMessage.ID.create()
const retry = yield* Schedule.toStepWithSleep(
SessionRunnerRetry.schedule(events, sessionID, () => assistantMessageID),
)
/**
* Consumes one retry allowance: sleeps the scheduled backoff, or publishes
* Step.Failed and fails once attempts are exhausted. The step loop performs
* the retry itself on the next iteration.
*/
const waitForRetry = (failure: SessionRunnerRetry.RetryableFailure) =>
retry(failure).pipe(
Effect.as(CallOutcome.Retry({ step: failure.step })),
Pull.catchDone(() =>
events
.publish(SessionEvent.Step.Failed, {
sessionID,
assistantMessageID,
error: failure.error,
})
.pipe(Effect.andThen(Effect.fail(failure.cause))),
),
)
let currentPromotable: SessionPending.Promotable | undefined = promotable
let currentStep = step
// Overflow recovery is one-shot: a call after recovery must not recover another overflow.
let recoverOverflow = true
while (true) {
const outcome = yield* callModel(
sessionID,
currentPromotable,
currentStep,
recoverOverflow,
assistantMessageID,
).pipe(Effect.catchTag("SessionRunner.RetryableFailure", waitForRetry))
if (outcome._tag === "Completed") return { needsContinuation: outcome.needsContinuation, step: outcome.step }
if (outcome._tag === "Restart") {
if (outcome.recoveredOverflow) recoverOverflow = false
assistantMessageID = SessionMessage.ID.create()
}
// Neither a retry nor a compaction restart re-promotes input.
currentPromotable = undefined
currentStep = outcome.step
}
})
/**
* Prepares and runs at most one model call, executes its local tools, and durably
* settles the step. Compaction may instead request that the logical step restart.
*/
const callModel = Effect.fn("SessionRunner.callModel")(function* (
sessionID: SessionSchema.ID,
promotable: SessionPending.Promotable | undefined,
step: number,
recoverOverflow: boolean,
assistantMessageID: SessionMessage.ID,
) {
const selected = yield* context.select(sessionID)
// Establish what the model knows before admitting what the user said, so
// a blocked first step leaves pending inputs untouched.
yield* InstructionState.prepare(db, events, selected.instructions, selected.session.id)
const promoted = promotable ? yield* SessionPending.promote(db, events, selected.session.id, promotable) : 0
// Promoted input opens a fresh step allowance.
const currentStep = promoted > 0 ? 1 : step
const loaded = yield* context.load(selected)
const { session, agent } = loaded
const resolved = loaded.model
const model = resolved.model
// Make room: history must fit the context window before the call. A pending manual
// compaction owns this instead; the runner executes it between steps.
const compactionInput = { session, messages: loaded.messages, model, cost: resolved.cost }
if (compaction.required(compactionInput) && !(yield* SessionPending.compaction(db, session.id))) {
const compacted = yield* compaction.compact(compactionInput)
if (compacted.status === "completed")
return CallOutcome.Restart({ step: currentStep, recoveredOverflow: false })
return yield* new StepFailedError({ error: compacted.error })
}
const prepared = yield* modelRequests.prepare({
context: loaded,
step: currentStep,
})
// Every local tool call forked here is owned until it reaches one durable settlement.
const toolRuns: Array<{ readonly call: ToolCall; readonly fiber: Fiber.Fiber<void, ToolOutputStore.Error> }> = []
const interruptTools = Effect.suspend(() => Fiber.interruptAll(toolRuns.map((run) => run.fiber)))
let needsContinuation = false
const startSnapshot = yield* snapshots.capture()
const publisher = createLLMEventPublisher(events, {
sessionID: session.id,
agent: agent.id,
// The selected catalog identity, not model.id: route-level ids are provider API
// model ids (for example gpt-5.5-fast resolves to api id gpt-5.5).
model: resolved.ref,
providerMetadataKey: model.route.providerMetadataKey ?? model.provider,
snapshot: startSnapshot,
assistantMessageID,
})
const publication = Semaphore.makeUnsafe(1)
// Durable publishes are serialized so tool fibers and step settlement never interleave
// mid-event.
const serialized = <A, E, R>(effect: Effect.Effect<A, E, R>) => publication.withPermit(effect)
const publish = (event: LLMEvent) => serialized(publisher.publish(event))
const stepUsage = (settlement: NonNullable<ReturnType<typeof publisher.stepSettlement>>) => ({
cost: SessionUsage.calculateCost(resolved.cost, settlement.tokens),
tokens: settlement.tokens,
})
const captureStepEnd = Effect.fnUntraced(function* () {
const snapshot = yield* snapshots.capture()
const files =
startSnapshot && snapshot
? yield* snapshots
.files({ from: startSnapshot, to: snapshot })
.pipe(Effect.catch(() => Effect.succeed(undefined)))
: undefined
return { snapshot, files }
})
const publishStepEnd = (settlement: NonNullable<ReturnType<typeof publisher.stepSettlement>>) =>
Effect.gen(function* () {
const end = yield* captureStepEnd()
yield* serialized(
events.publish(SessionEvent.Step.Ended, {
sessionID: session.id,
assistantMessageID: yield* publisher.startAssistant(),
finish: settlement.finish,
...stepUsage(settlement),
...end,
}),
)
})
// The stream is defined here but runs inside the settlement mask below: publish each
// event durably, fork one fiber per local tool call, and hold back a virgin
// context-overflow provider error so settlement may recover it via compaction.
let overflowFailure: ProviderErrorEvent | undefined
const providerStream = llm.stream(prepared.request).pipe(
Stream.runForEach((event) =>
Effect.gen(function* () {
if (overflowFailure || publisher.hasProviderError()) return
if (LLMEvent.is.providerError(event)) {
if (isContextOverflowFailure(event) && !publisher.hasRetryEvidence()) {
overflowFailure = event
return
}
}
yield* publish(event)
if (LLMEvent.is.toolInputError(event)) {
if (!prepared.stepLimitReached) needsContinuation = true
return
}
if (event.type !== "tool-call" || event.providerExecuted) return
// Unavailable calls fail individually through the same execution seam;
// continuation depends only on remaining Step allowance.
if (!prepared.stepLimitReached) needsContinuation = true
const assistantMessageID = yield* publisher.assistantMessageID(event.id)
toolRuns.push({
call: event,
fiber: yield* Effect.uninterruptibleMask((restore) =>
restore(
prepared.executeTool({
sessionID: session.id,
agent: agent.id,
messageID: assistantMessageID,
call: event,
progress: (update) => serialized(publisher.progress(event.id, update)),
}),
).pipe(
Effect.flatMap((execution) => serialized(publisher.toolExecution(event.id, event.name, execution))),
),
).pipe(Effect.forkScoped),
})
}),
),
Effect.ensuring(serialized(publisher.flush())),
)
// Settle: only the stream itself is interruptible (restore); every line after it is
// protected so a started call always reaches one durable outcome.
return yield* Effect.uninterruptibleMask((restore) =>
Effect.gen(function* () {
const stream = yield* restore(providerStream).pipe(Effect.exit)
const streamFailure = Option.getOrUndefined(Exit.findErrorOption(stream))
// Note: Exit.hasInterrupts is a type guard whose false branch unsoundly narrows
// away non-interrupt failures, so both interrupt checks stay Cause-based.
const streamInterrupted = stream._tag === "Failure" && Cause.hasInterrupts(stream.cause)
// A context overflow before any assistant output is recoverable: compact and
// restart the step instead of surfacing the provider error.
if (
recoverOverflow &&
!publisher.hasRetryEvidence() &&
isContextOverflowFailure(overflowFailure ?? streamFailure) &&
(yield* restore(compaction.compact(compactionInput))).status === "completed"
)
return CallOutcome.Restart({ step: currentStep, recoveredOverflow: true })
// An unrecovered held-back overflow becomes the step's durable provider error. A
// thrown LLM failure records the assistant failure unless a provider error was
// already recorded from the stream. Terminal publication waits for owned tools.
if (overflowFailure) yield* publish(overflowFailure)
const llmFailure = streamFailure instanceof LLMError ? streamFailure : undefined
if (llmFailure && !publisher.hasProviderError()) {
const error = toSessionError(llmFailure)
if (SessionRunnerRetry.isRetryable(llmFailure) && !publisher.hasRetryEvidence()) {
// RetryScheduled and Step.Failed fold onto an existing assistant message, so
// Step.Started must be durable before the failure escapes.
yield* serialized(publisher.startAssistant())
return yield* new SessionRunnerRetry.RetryableFailure({
cause: llmFailure,
error,
step: currentStep,
})
}
yield* serialized(publisher.failAssistant(error))
}
// Provider error events only arrive from the stream, so the flag is final here.
const providerFailed = publisher.hasProviderError()
// Settle every owned tool run: await all exits, not just the first failure,
// before publishing the terminal step event.
if (streamInterrupted) yield* interruptTools
const settled = yield* restore(
Effect.forEach(toolRuns, (run) => Fiber.await(run.fiber), { concurrency: "unbounded" }),
).pipe(Effect.exit)
if (settled._tag === "Failure") yield* interruptTools
const tools = classifyToolExits(settled)
if (tools.declined || streamInterrupted || tools.interrupted) {
yield* serialized(publisher.failUnsettledTools({ type: "aborted", message: "Tool execution interrupted" }))
yield* serialized(publisher.failAssistant({ type: "aborted", message: "Step interrupted" }))
}
if (tools.failure !== undefined) {
const error = toSessionError(tools.infraError ?? Cause.squash(tools.failure))
yield* serialized(publisher.failUnsettledTools(error))
if (tools.infraError !== undefined) yield* serialized(publisher.failAssistant(error))
}
// Fail unresolved calls before the terminal step event. Local calls have joined, so
// these sweeps only close calls that could not produce a truthful settlement.
if (providerFailed)
yield* serialized(publisher.failUnsettledTools({ type: "aborted", message: "Tool execution interrupted" }))
const resultMissing = {
type: "tool.result-missing",
message: "Provider did not return a tool result",
} as const
if (llmFailure && !providerFailed) yield* serialized(publisher.failUnsettledTools(resultMissing, "hosted"))
// A clean stream that still left hosted calls unresolved fails the step itself.
if (stream._tag === "Success" && !providerFailed) {
const hostedResultMissing = yield* serialized(publisher.failUnsettledTools(resultMissing, "hosted"))
if (hostedResultMissing && !publisher.stepSettlement())
yield* serialized(publisher.failAssistant(resultMissing))
}
const stepFailure = publisher.stepFailure()
const stepSettlement = publisher.stepSettlement()
if (stepSettlement && !stepFailure) yield* publishStepEnd(stepSettlement)
if (stepFailure) {
const end = yield* captureStepEnd()
yield* serialized(
publisher.publishStepFailure({
...(stepSettlement ? stepUsage(stepSettlement) : {}),
...end,
}),
)
}
if (stream._tag === "Failure") return yield* Effect.failCause(stream.cause)
if (tools.declined) return yield* Effect.interrupt
if ((tools.interrupted || tools.infraError !== undefined) && tools.failure)
return yield* Effect.failCause(tools.failure)
if (tools.interrupted && settled._tag === "Failure") return yield* Effect.failCause(settled.cause)
if (stepFailure) return yield* new StepFailedError({ error: stepFailure })
return CallOutcome.Completed({ needsContinuation, step: currentStep })
}),
)
}, Effect.scoped)
/** Executes a previously admitted manual compaction request, if one is pending. */
const runPendingCompaction = Effect.fn("SessionRunner.runPendingCompaction")(function* (
sessionID: SessionSchema.ID,
) {
const pending = yield* SessionPending.compaction(db, sessionID)
if (!pending) return
const session = yield* getSession(sessionID)
return yield* Effect.uninterruptibleMask((restore) =>
Effect.gen(function* () {
const compacted = yield* restore(
Effect.gen(function* () {
return yield* compaction.compactManual({
session,
messages: yield* store.context(sessionID),
inputID: pending.id,
})
}),
).pipe(Effect.exit)
if (Exit.isSuccess(compacted)) return
const unsettled = yield* SessionPending.compaction(db, sessionID)
if (unsettled)
yield* events.publish(SessionEvent.Compaction.Failed, {
sessionID,
reason: "manual",
error: Cause.hasInterruptsOnly(compacted.cause)
? { type: "aborted", message: "Compaction cancelled" }
: { type: "compaction.failed", message: Cause.pretty(compacted.cause) },
inputID: unsettled.id,
})
return yield* Effect.failCause(compacted.cause)
}),
)
})
/** Closes stale tool calls left active by an earlier interrupted drain. */
const settleStaleToolCalls = Effect.fn("SessionRunner.settleStaleToolCalls")(function* (
sessionID: SessionSchema.ID,
) {
for (const message of yield* store.context(sessionID)) {
if (message.type !== "assistant") continue
for (const tool of message.content) {
if (tool.type !== "tool" || (tool.state.status !== "streaming" && tool.state.status !== "running")) continue
yield* events.publish(SessionEvent.Tool.Failed, {
sessionID,
assistantMessageID: message.id,
callID: tool.id,
error: { type: "aborted", message: `Tool execution interrupted: ${tool.name}` },
executed: tool.executed === true,
})
}
}
})
/** Fires title generation once per process after the first step makes a user message visible. */
const startTitleOnce = Effect.fnUntraced(function* (sessionID: SessionSchema.ID) {
if (titleStarted.has(sessionID)) return
titleStarted.add(sessionID)
forkTitle(title.generateForFirstPrompt(yield* getSession(sessionID)).pipe(Effect.ignore))
})
const getSession = Effect.fn("SessionRunner.getSession")(function* (sessionID: SessionSchema.ID) {
const session = yield* store.get(sessionID)
if (!session) return yield* Effect.die(new Error(`Session not found: ${sessionID}`))
return session
})
return Service.of({ drain })
}),
)
export const node = makeLocationNode({
service: Service,
layer,
deps: [
EventV2.node,
llmClient,
SessionContext.node,
SessionModelRequest.node,
SessionStore.node,
SessionCompaction.node,
SessionTitle.node,
Snapshot.node,
Database.node,
],
})