feat(goal): add model-facing goal tools

2026-07-19 19:22:10 +08:00
parent e9940d35cf
commit 0129063ae7
24 changed files with 1388 additions and 1 deletions
@@ -0,0 +1,105 @@
+/** Execution-time authority checks for the model-facing goal tools. */
+
+import type { Context } from 'cordis'
+import type { Agent } from '@deepseek-ai/dsh-agent'
+import type { GoalView } from '@deepseek-ai/dsh-goal'
+import { HarnessError } from '@deepseek-ai/dsh-llm'
+import type { SessionEvent } from '@deepseek-ai/dsh-session'
+import type { ToolRunContext } from '@deepseek-ai/dsh-tools'
+
+type TurnStartEvent = Extract<SessionEvent, { type: 'turn/start' }>
+
+/** Current open turn plus the events accepted after its start boundary. */
+export interface GoalToolExecution {
+  readonly agent: Agent
+  readonly start: TurnStartEvent
+  readonly events: readonly SessionEvent[]
+}
+
+/** Hard authority granted to one state-changing call. */
+export type GoalToolAuthority =
+  | { readonly kind: 'direct-human' }
+  | { readonly kind: 'goal-round'; readonly goal: GoalView }
+
+/** Throw one structured tool-policy failure. */
+function reject(message: string, code = 'GOAL_TOOL_AUTHORITY_REQUIRED'): never {
+  throw new HarnessError(message, code)
+}
+
+/** Locate the open turn enclosing a model tool call. */
+function openTurn(agent: Agent): { start: TurnStartEvent; events: readonly SessionEvent[] } {
+  const events = agent.session.events
+  for (let index = events.length - 1; index >= 0; index -= 1) {
+    const boundary = events[index]
+    if (boundary?.type === 'turn/end') {
+      reject('goal tools require an open model turn', 'GOAL_TOOL_DRIVER_REQUIRED')
+    }
+    if (boundary?.type === 'turn/start') {
+      return { start: boundary, events: events.slice(index + 1) }
+    }
+  }
+  return reject('goal tools require an open model turn', 'GOAL_TOOL_DRIVER_REQUIRED')
+}
+
+/**
+ * Resolve and authenticate the calling agent and its driver boundary.
+ * @param ctx - Context carrying the live agent registry.
+ * @param exec - Tool execution metadata supplied by the registry.
+ * @returns The authenticated agent and its current turn window.
+ */
+export function goalToolExecution(ctx: Context, exec: ToolRunContext): GoalToolExecution {
+  const agent = exec.agent
+  if (agent === undefined) {
+    return reject('goal tools require a calling agent', 'GOAL_TOOL_AGENT_REQUIRED')
+  }
+  if (ctx.agents.get(agent.id) !== agent || agent.status !== 'running'
+    || ctx.agents.currentInitiator() !== agent) {
+    return reject(
+      'goal tools require the exact live calling agent inside its active driver',
+      'GOAL_TOOL_DRIVER_REQUIRED',
+    )
+  }
+  return { agent, ...openTurn(agent) }
+}
+
+/** Whether an accepted human message appears in the current root-agent turn. */
+function hasDirectHumanInput(ctx: Context, execution: GoalToolExecution): boolean {
+  if (!ctx.agents.roots().includes(execution.agent)) return false
+  return execution.events.some(event =>
+    (event.type === 'user/message' || event.type === 'steering/message')
+    && event.data.source.kind === 'user')
+}
+
+/** Whether this turn is the current goal's exact admitted round. */
+function isMatchingGoalRound(execution: GoalToolExecution, goal: GoalView): boolean {
+  return execution.events.some(event => event.type === 'user/message'
+    && event.data.source.kind === 'goal'
+    && event.data.source.goalId === goal.id
+    && event.data.source.revision === goal.revision
+    && event.data.source.round === goal.roundsStarted)
+}
+
+/**
+ * Require authority originating in a human message accepted by a runtime root.
+ * @param ctx - Context carrying the live agent graph.
+ * @param execution - Authenticated current tool execution.
+ */
+export function requireDirectHuman(ctx: Context, execution: GoalToolExecution): void {
+  if (hasDirectHumanInput(ctx, execution)) return
+  reject('this goal operation requires a direct human turn on a top-level agent')
+}
+
+/**
+ * Resolve completion authority from either direct human input or the exact goal round.
+ * @param ctx - Context carrying live agents and goal state.
+ * @param execution - Authenticated current tool execution.
+ * @returns The direct-human or exact-goal-round authority grant.
+ */
+export function completionAuthority(ctx: Context, execution: GoalToolExecution): GoalToolAuthority {
+  if (hasDirectHumanInput(ctx, execution)) return { kind: 'direct-human' }
+  const goal = ctx.goals.get(execution.agent)
+  if (goal !== undefined && isMatchingGoalRound(execution, goal)) {
+    return { kind: 'goal-round', goal }
+  }
+  return reject('complete and blocked require a direct human turn or the current goal round')
+}
@@ -0,0 +1,222 @@
+/**
+ * Model-facing `get_goal`, `create_goal`, and `update_goal` tools over the
+ * persisted same-session goal domain.
+ * @module @deepseek-ai/dsh-tool-goal
+ */
+
+import type { Context } from 'cordis'
+import z from 'schemastery'
+import { GoalId } from '@deepseek-ai/dsh-goal'
+import type { GoalRef, GoalView } from '@deepseek-ai/dsh-goal'
+import { HarnessError } from '@deepseek-ai/dsh-llm'
+import { defineTool } from '@deepseek-ai/dsh-tools'
+import type { GenericCallView } from '@deepseek-ai/dsh-tools'
+import type {} from '@deepseek-ai/dsh-system-prompt'
+import {
+  completionAuthority,
+  goalToolExecution,
+  requireDirectHuman,
+} from './authority.ts'
+
+export const name = 'tool-goal'
+export const inject = ['agents', 'goals', 'tools', 'systemPrompt']
+
+/** Model policy and hard lower bounds for goal-state updates. */
+export interface Config {
+  /** Minimum admitted goal rounds before the model may self-report `blocked`. */
+  blockedAfterConsecutiveRounds?: number
+}
+
+/** Schemastery config for the goal-tool policy. */
+export const Config: z<Config> = z.object({
+  blockedAfterConsecutiveRounds: z.number().step(1).min(1).default(3),
+})
+
+/** Fully materialized tool policy. */
+interface ResolvedConfig {
+  readonly blockedAfterConsecutiveRounds: number
+}
+
+type UpdateAction = 'edit' | 'pause' | 'resume' | 'complete' | 'blocked'
+
+const UPDATE_ACTIONS: UpdateAction[] = ['edit', 'pause', 'resume', 'complete', 'blocked']
+
+const CREATE_DESCRIPTION =
+  'Create one persisted same-session completion goal when the current direct human request '
+  + 'is a long-running objective that should continue across autonomous goal rounds. You may '
+  + 'infer that intent without requiring the user to say "create a goal". Do not use this for '
+  + 'trivial single-turn work. Execution rejects non-human and subagent authority.'
+
+const GET_DESCRIPTION =
+  'Read the current same-session goal, including its exact id/revision, durable phase, admitted '
+  + 'round count, cap, and live process-local activation. Call this before updating a goal.'
+
+/** Render policy guidance with its deployment-selected blocked threshold. */
+function guidance(blockedAfter: number): string {
+  return 'Use goal tools for one long-running completion objective in the current session. '
+    + 'create_goal may infer goal intent from a direct human request in any language; do not '
+    + 'create a goal for routine single-turn work. Call get_goal before update_goal and copy its '
+    + 'exact goal_id and revision. After session resume or fork, an active goal is disarmed: when '
+    + 'a human asks to continue or resume in any wording or language, use update_goal action '
+    + 'resume to rearm it. Mark complete only when the objective is actually achieved. Mark '
+    + `blocked only after the same blocking condition persists for at least ${blockedAfter} `
+    + 'consecutive goal rounds; difficulty, uncertainty, or useful remaining work is not blocked.'
+}
+
+/** Validate config even when apply is called directly outside Loader normalization. */
+function resolveConfig(config: Config): ResolvedConfig {
+  const blockedAfter = config.blockedAfterConsecutiveRounds ?? 3
+  if (!Number.isSafeInteger(blockedAfter) || blockedAfter < 1) {
+    throw new TypeError('blockedAfterConsecutiveRounds must be a positive safe integer')
+  }
+  return { blockedAfterConsecutiveRounds: blockedAfter }
+}
+
+/** Build the exact compare-and-set ref from model arguments. */
+function goalRef(goalId: string, revision: number): GoalRef {
+  if (goalId.length === 0 || goalId !== goalId.trim()
+    || !Number.isSafeInteger(revision) || revision < 1) {
+    throw new HarnessError(
+      'goal_id must be non-empty and revision must be a positive safe integer',
+      'GOAL_TOOL_INVALID_UPDATE',
+    )
+  }
+  return { id: GoalId(goalId), revision }
+}
+
+/** Stable compact model result; activation is an observation, not replay state. */
+function renderGoal(goal: GoalView | undefined): string {
+  if (goal === undefined) return JSON.stringify({ goal: null })
+  return JSON.stringify({
+    goal: {
+      id: goal.id,
+      revision: goal.revision,
+      objective: goal.objective,
+      phase: goal.phase,
+      roundsStarted: goal.roundsStarted,
+      maxGoalRounds: goal.maxGoalRounds,
+    },
+    activation: goal.activation,
+  })
+}
+
+/** Generic, args-only pending presentation shared by the goal tools. */
+function present(title: string, kind: 'read' | 'other', rawInput?: unknown): GenericCallView {
+  return { card: 'generic', title, kind, ...rawInput === undefined ? {} : { rawInput } }
+}
+
+/** Register the three Codex-shaped goal tools and their shared policy section. */
+export function apply(ctx: Context, config: Config): void {
+  const resolved = resolveConfig(config)
+  ctx.systemPrompt.section({
+    name: 'tool:goal',
+    order: 114,
+    text: guidance(resolved.blockedAfterConsecutiveRounds),
+  })
+
+  ctx.tools.register(defineTool({
+    name: 'get_goal',
+    description: GET_DESCRIPTION,
+    parameters: {},
+    execute(_args, exec) {
+      const execution = goalToolExecution(ctx, exec)
+      return Promise.resolve([{
+        type: 'text',
+        text: renderGoal(ctx.goals.get(execution.agent)),
+      }])
+    },
+    presentCall: () => present('Read current goal', 'read'),
+  }))
+
+  ctx.tools.register(defineTool({
+    name: 'create_goal',
+    description: CREATE_DESCRIPTION,
+    parameters: {
+      objective: {
+        type: 'string',
+        required: true,
+        description: 'The concrete completion objective inferred from the direct human request.',
+      },
+      max_goal_rounds: {
+        type: 'number',
+        description: 'Optional positive safe-integer cap; omission uses the goal-domain deployment default.',
+      },
+    },
+    execute(args, exec) {
+      const execution = goalToolExecution(ctx, exec)
+      requireDirectHuman(ctx, execution)
+      const goal = ctx.goals.create(execution.agent, {
+        objective: args.objective,
+        ...args.max_goal_rounds === undefined ? {} : { maxGoalRounds: args.max_goal_rounds },
+      })
+      return Promise.resolve([{ type: 'text', text: renderGoal(goal) }])
+    },
+    presentCall: args => present('Create goal', 'other', args.objective),
+  }))
+
+  ctx.tools.register(defineTool({
+    name: 'update_goal',
+    description: 'Update the exact current goal revision. edit, pause, and resume require a direct '
+      + 'top-level human turn. complete and blocked additionally accept the exact admitted goal '
+      + 'round. blocked is rejected before the configured minimum round count; the model remains '
+      + 'responsible for judging that the same condition persisted across those rounds.',
+    parameters: {
+      goal_id: { type: 'string', required: true, description: 'Exact id returned by get_goal.' },
+      revision: { type: 'number', required: true, description: 'Exact positive revision returned by get_goal.' },
+      action: {
+        type: 'string',
+        required: true,
+        enum: UPDATE_ACTIONS,
+        description: 'edit | pause | resume | complete | blocked',
+      },
+      objective: { type: 'string', description: 'Replacement objective; valid only with action edit.' },
+      max_goal_rounds: { type: 'number', description: 'Replacement cap; valid only with action edit.' },
+    },
+    execute(args, exec) {
+      const execution = goalToolExecution(ctx, exec)
+      const ref = goalRef(args.goal_id, args.revision)
+      const replacements = {
+        ...args.objective === undefined ? {} : { objective: args.objective },
+        ...args.max_goal_rounds === undefined ? {} : { maxGoalRounds: args.max_goal_rounds },
+      }
+      if (args.action === 'edit') {
+        requireDirectHuman(ctx, execution)
+        return Promise.resolve([{
+          type: 'text',
+          text: renderGoal(ctx.goals.edit(execution.agent, ref, replacements)),
+        }])
+      }
+      if (args.objective !== undefined || args.max_goal_rounds !== undefined) {
+        throw new HarnessError(
+          'objective and max_goal_rounds are valid only with action edit',
+          'GOAL_TOOL_INVALID_UPDATE',
+        )
+      }
+      if (args.action === 'pause' || args.action === 'resume') {
+        requireDirectHuman(ctx, execution)
+        const goal = args.action === 'pause'
+          ? ctx.goals.pause(execution.agent, ref)
+          : ctx.goals.resume(execution.agent, ref)
+        return Promise.resolve([{ type: 'text', text: renderGoal(goal) }])
+      }
+      const authority = completionAuthority(ctx, execution)
+      if (args.action === 'blocked' && authority.kind === 'goal-round'
+        && authority.goal.roundsStarted < resolved.blockedAfterConsecutiveRounds) {
+        throw new HarnessError(
+          `blocked requires at least ${resolved.blockedAfterConsecutiveRounds} consecutive goal rounds; `
+          + `current round is ${authority.goal.roundsStarted}`,
+          'GOAL_TOOL_BLOCK_THRESHOLD',
+        )
+      }
+      const goal = args.action === 'complete'
+        ? ctx.goals.complete(execution.agent, ref)
+        : ctx.goals.block(execution.agent, ref)
+      return Promise.resolve([{ type: 'text', text: renderGoal(goal) }])
+    },
+    presentCall: args => present(
+      `${args.action === 'blocked' ? 'Mark' : args.action.charAt(0).toUpperCase() + args.action.slice(1)} goal`,
+      'other',
+      args.objective ?? args.goal_id,
+    ),
+  }))
+}