Files
deepseek-harness/packages/core/tools/src/code-mode.ts
Tianyi Cui 77f0f3e632 Merge branch 'codex/code-mode-typed-results' into codex/code-mode-complete-result-card
# Conflicts:
#	.agents/notes/implemented/architecture/2026-07-20-canonical-tool-output-contract.i18n.yaml
#	.agents/notes/implemented/feature/2026-07-20-code-mode-typed-tool-returns.i18n.yaml
2026-07-23 01:38:52 +08:00

380 lines
17 KiB
TypeScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
/**
* Code Mode `run_code` transport. Programs call the registry's agent-visible
* tools through nested, sequential executions; each sub-dispatch is logged for
* reconstruction, while only the outer curated result enters model history.
* @module @deepseek-ai/dsh-tools/src/code-mode
*/
import { parse } from 'node:path'
import { CallId, HarnessError } from '@deepseek-ai/dsh-llm'
import type { ContentBlock } from '@deepseek-ai/dsh-llm'
import type { CodeBindingFunction, CodeRunResult, CodeRuntime } from '@deepseek-ai/dsh-code-runtime'
import { snapshotJsonValue } from '@deepseek-ai/dsh-session'
import type { JsonValue } from '@deepseek-ai/dsh-session'
import { defineTool } from './schema.ts'
import type { ToolDefinition, ToolRegistry } from './index.ts'
declare module '@deepseek-ai/dsh-session' {
interface SessionEventMap {
/**
* One bridged sub-dispatch from a `run_code` program: the parent
* `run_code` call id, the deterministic sub-call id
* (`<parent>:code:<n>`), the tool `name` with its JSON-normalized
* `arguments` — the exact value dispatched, normalized BEFORE dispatch,
* so this append can never fail on payload shape — whether the sub-call
* errored, and a bounded `resultSummary` of its model-facing text. Before
* bounding, occurrences of a non-root session workspace path are
* normalized to `.` so host-specific absolute path lengths cannot change
* the summary.
* Log-only: `deriveMessages()` ignores it, so sub-calls never re-enter
* model context; persistence and UIs get every call. Appended inside the
* parent `run_code`'s execution (the bridge drains its queue before
* returning), so the turn-enclosure invariant holds by construction.
*/
'tool/code-dispatch': { parentCallId: CallId; subCallId: CallId; name: string; arguments: unknown; isError: boolean; resultSummary: string }
}
}
/** The model-facing name of the Code Mode tool. */
export const RUN_CODE_NAME = 'run_code'
/** The `tools:sdk` section order: inside the 100199 tool-guidance band, after per-tool guidance sections. */
export const SDK_SECTION_ORDER = 150
/**
* Thrown by `run_code` when the program run itself failed — a program
* exception, a budget expiry, an abort, or substrate death. Extends
* {@link HarnessError} (`code: 'CODE_RUN_FAILED'`); the registry's execution
* pipeline converts it into a structured `isError` result whose text carries
* the failure kind plus the captured logs, so the model can self-correct.
*/
export class CodeRunFailedError extends HarnessError {
constructor(message: string) {
super(message, 'CODE_RUN_FAILED')
this.name = 'CodeRunFailedError'
}
}
/**
* Cap for a `tool/code-dispatch` event's `resultSummary`. A log-ergonomics
* constant, not config: the full result already flows to the program; the
* summary exists so log readers see what a sub-call returned at a glance.
*/
const SUMMARY_MAX_CHARS = 200
/** Join Native content for the bounded durable sub-dispatch summary; non-text blocks become diagnostic placeholders. */
function textOf(content: ContentBlock[]): string {
return content
.map((block) => {
switch (block.type) {
case 'text': return block.text
// ContentBlockMap is merge-extensible — future block kinds land here
// deliberately (no assertNever on merge-extensible unions).
default: return `[${block.type} content]`
}
})
.join('\n')
}
/** Normalize workspace paths, then bound a sub-call's model-facing text for its durable log summary. */
function summarize(text: string, cwd: string | undefined): string {
const stableText = cwd === undefined || cwd === parse(cwd).root
? text
: text.replaceAll(cwd, '.')
return stableText.length > SUMMARY_MAX_CHARS ? `${stableText.slice(0, SUMMARY_MAX_CHARS)}` : stableText
}
/**
* Snapshot one binding call's argument as lossless JSON, then snapshot that
* detached value again so dispatch and logging stay independent without
* reintroducing structured-clone's platform-specific nesting limit.
*/
function jsonNormalizeArgs(value: unknown): { dispatched: unknown; logged: unknown } {
let snapshot: JsonValue | undefined
try {
snapshot = snapshotJsonValue(value) as JsonValue | undefined
} catch (error: unknown) {
throw new Error(`tool arguments must be lossless JSON: ${error instanceof Error ? error.message : String(error)}`)
}
if (snapshot === undefined) {
throw new Error('tool arguments must be lossless JSON (call the tool with an arguments object, e.g. `{}`)')
}
const logged = snapshotJsonValue(snapshot)
/* v8 ignore next -- snapshot is already a detached lossless JSON value. */
if (logged === undefined) {
throw new Error('tool arguments could not be detached for durable logging')
}
return { dispatched: snapshot, logged }
}
/** Two-space JSON presentation, matching the existing shallow `run_code` text contract. */
const JSON_INDENT = ' '
/**
* ECMAScript caps `JSON.stringify`'s `space` string at ten characters. The
* renderer also caps TOTAL indentation there, compacting deeper subtrees, so
* formatted output remains linear in the canonical JSON size.
*/
const MAX_JSON_INDENT_CHARS = 10
/** A pending fragment in the iterative JSON presentation traversal. */
type JsonRenderTask =
| { kind: 'text'; text: string }
| { kind: 'value'; value: JsonValue; depth: number; compact: boolean }
/** Render one non-string JSON root without recursive traversal or unbounded indentation growth. */
function renderJsonValue(value: Exclude<JsonValue, string>): string {
const chunks: string[] = []
const tasks: JsonRenderTask[] = [{ kind: 'value', value, depth: 0, compact: false }]
for (let task = tasks.pop(); task !== undefined; task = tasks.pop()) {
if (task.kind === 'text') {
chunks.push(task.text)
continue
}
const current = task.value
if (current === null || typeof current === 'boolean' || typeof current === 'number') {
chunks.push(String(current))
continue
}
if (typeof current === 'string') {
chunks.push(JSON.stringify(current))
continue
}
const compact = task.compact || (task.depth + 1) * JSON_INDENT.length > MAX_JSON_INDENT_CHARS
const childDepth = task.depth + 1
if (Array.isArray(current)) {
chunks.push('[')
if (current.length === 0) {
chunks.push(']')
continue
}
tasks.push({ kind: 'text', text: compact ? ']' : `\n${JSON_INDENT.repeat(task.depth)}]` })
for (let index = current.length - 1; index >= 0; index--) {
const item = current[index]
/* v8 ignore next -- canonical JsonValue arrays are dense. */
if (item === undefined) throw new Error('cannot render a sparse JSON array')
tasks.push({ kind: 'value', value: item, depth: childDepth, compact })
tasks.push({
kind: 'text',
text: compact
? index === 0 ? '' : ','
: `${index === 0 ? '\n' : ',\n'}${JSON_INDENT.repeat(childDepth)}`,
})
}
continue
}
const keys = Object.keys(current)
chunks.push('{')
if (keys.length === 0) {
chunks.push('}')
continue
}
tasks.push({ kind: 'text', text: compact ? '}' : `\n${JSON_INDENT.repeat(task.depth)}}` })
for (let index = keys.length - 1; index >= 0; index--) {
const key = keys[index]
/* v8 ignore next -- the loop is bounded by the captured key count. */
if (key === undefined) throw new Error('cannot render a missing JSON object key')
const item = current[key]
/* v8 ignore next -- canonical JsonValue records contain no undefined properties. */
if (item === undefined) throw new Error('cannot render an undefined JSON object property')
tasks.push({ kind: 'value', value: item, depth: childDepth, compact })
tasks.push({
kind: 'text',
text: compact
? `${index === 0 ? '' : ','}${JSON.stringify(key)}:`
: `${index === 0 ? '\n' : ',\n'}${JSON_INDENT.repeat(childDepth)}${JSON.stringify(key)}: `,
})
}
}
return chunks.join('')
}
/** Render one present program completion value for the model-facing result text. */
function renderValue(value: JsonValue): string {
return typeof value === 'string' ? value : renderJsonValue(value)
}
/** Canonical value returned by the outer Code Mode transport. */
type RunCodeOutput = { logs: string[]; result?: JsonValue }
/**
* Build the `run_code` {@link ToolDefinition}: one required `code` parameter,
* executed through the dispatch bridge described above. The
* registry reserves it as presentation infrastructure under non-native modes,
* outside the filterable global/scoped capability layers.
* @param registry - the owning registry (sub-calls go through its `execute`,
* bindings cover its registered tools).
* @param requireRuntime - resolves `ctx.codeRuntime` or throws the loud
* misconfiguration error (shared with the registry's assembly-time checks).
* @returns the registry-ready definition.
*/
export function createRunCodeTool(registry: ToolRegistry, requireRuntime: () => CodeRuntime): ToolDefinition {
return defineTool({
name: RUN_CODE_NAME,
description:
'Execute a TypeScript program against the available tools. Write the BODY of an '
+ 'async function (erasable syntax only; top-level `await` and `return` work) and '
+ 'call tools as `await tools.name(args)` per the declarations in the system prompt. '
+ 'Only what you print or return comes back — curate it.',
parameters: {
code: { type: 'string', required: true, description: 'The program: the body of an async TypeScript function.' },
},
output: {
schema: {
type: 'object',
additionalProperties: false,
properties: {
logs: { type: 'array', required: true, items: { type: 'string' } },
result: { type: 'json' },
},
},
render: (_args, value) => {
const rendered = value.result === undefined ? '' : renderValue(value.result)
const parts = [value.logs.join('\n'), rendered].filter(part => part.length > 0)
return [{ type: 'text', text: parts.length > 0 ? parts.join('\n') : '(run_code completed with no output)' }]
},
},
async execute(args, exec): Promise<RunCodeOutput> {
const runtime = requireRuntime()
// The run-scoped abort: follows the outer signal in, and fires when the
// run settles for ANY reason, so an in-flight sub-dispatch is aborted
// (its executor kills on this signal) instead of orphaned, and
// queued-unstarted dispatches are abandoned.
const runController = new AbortController()
const onOuterAbort = (): void => { runController.abort(exec.signal.reason) }
exec.signal.addEventListener('abort', onOuterAbort, { once: true })
let dispatches = 0
// The per-run serialization queue: every binding call chains onto the tail, so even
// `Promise.all` executes the underlying tool calls one at a time in submission order (the
// tool contract carries no concurrency-safety metadata yet).
let queue: Promise<void> = Promise.resolve()
const enqueue = <T>(task: () => Promise<T>): Promise<T> => {
const turn = queue.then(() => {
if (runController.signal.aborted) {
throw new Error(`run_code run is over (${String(runController.signal.reason)}); tool call abandoned`)
}
return task()
})
queue = turn.then(() => undefined, () => undefined)
return turn
}
// Read through a call, not a bare property: the abort state genuinely
// changes across awaits, and a direct `.aborted` re-check after one
// would be narrowed away by control flow analysis.
const runOver = (): boolean => runController.signal.aborted
const binding = (name: string): CodeBindingFunction => async (rawArgs: unknown): Promise<JsonValue> => {
if (runOver()) {
throw new Error(`run_code run is over (${String(runController.signal.reason)}); ${name} not dispatched`)
}
const normalized = jsonNormalizeArgs(rawArgs)
const outcome = await enqueue(async () => {
const n = ++dispatches
const subCallId = CallId(`${String(exec.callId)}:code:${n}`)
const result = await registry.execute({
callId: subCallId,
name,
arguments: normalized.dispatched,
...exec.agent ? { agent: exec.agent } : {},
parent: exec.token,
signal: runController.signal,
})
for (const context of result.additionalContexts ?? []) {
exec.deferContext(context)
}
const text = textOf(result.content)
exec.agent?.session.append('tool/code-dispatch', {
parentCallId: exec.callId,
subCallId,
name,
// The SIBLING parse of the dispatched value: byte-identical JSON,
// but a separate object — a tool mutating its args cannot desync
// this record from what it actually received.
arguments: normalized.logged,
isError: result.isError,
resultSummary: summarize(text, exec.agent.session.header.cwd),
})
return result.isError
? { isError: true as const, message: result.error.message }
: { isError: false as const, value: result.value }
})
// A budget expiry or outer cancel that lands while this call was in
// flight already aborted the dispatch; stop the program now rather
// than hand it a result from a run that is over.
if (runOver()) {
throw new Error(`run_code run is over (${String(runController.signal.reason)}); ${name} result discarded`)
}
// The worker turns a binding rejection into ToolCallError and adds
// only the binding name. Native content and internal error metadata
// stay outside the program-facing failure contract.
if (outcome.isError) throw new Error(outcome.message)
return outcome.value
}
// Null-prototype + defineProperty, mirroring the worker-side namespace
// build: a registered tool named `__proto__` must become an ordinary
// own key (a plain-object assignment would hit the prototype setter,
// silently dropping the binding), and the runtime host resolves
// binding names as own properties only.
const functions: Record<string, CodeBindingFunction> = Object.create(null) as Record<string, CodeBindingFunction>
// Enumerate the CALLING AGENT's visible set (scoped tools join,
// restricted globals vanish) — the same view the SDK section declared,
// so a program can bind exactly what its prompt promised; sub-dispatch
// re-resolves per call through the same view (exec.agent threads down).
for (const schema of registry.schemas(exec.agent)) {
if (schema.name === RUN_CODE_NAME) continue
Object.defineProperty(functions, schema.name, { enumerable: true, value: binding(schema.name) })
}
try {
let result: CodeRunResult
try {
result = await runtime.run({
program: args.code,
bindings: [{
global: 'tools',
functions,
errorClass: { name: 'ToolCallError', memberNameProperty: 'toolName' },
}],
signal: runController.signal,
})
} finally {
// Abort sub-dispatches and drain the folded queue before closing the turn.
// Binding failures remain observable through their individual promises.
runController.abort('run_code settled')
await queue
}
if (result.error) {
const logsText = result.logs.length > 0 ? `\nCaptured output:\n${result.logs.join('\n')}` : ''
throw new CodeRunFailedError(`code run failed (${result.error.kind}): ${result.error.message}${logsText}`)
}
return {
logs: result.logs,
...result.value !== undefined ? { result: result.value } : {},
}
} finally {
exec.signal.removeEventListener('abort', onOuterAbort)
}
},
// ACP execute cards use the program as their visible title.
presentCall: args => ({
card: 'generic',
title: args.code,
kind: 'execute',
rawInput: args.code,
}),
// Title omitted on the result: an update replaces only the fields it
// carries, so the pending card's program title persists through
// completion. The durable final content already includes logs plus the
// return value, failure, or post-policy spill preview.
presentResult: (_args, result) => ({ card: 'generic', content: result.content }),
})
}