feat(athena): add session handoff with Question tool for Atlas/Prometheus routing

After Athena synthesizes council findings, presents user with Question tool
TUI to choose: Atlas (fix now), Prometheus (create plan), or no action.
On selection, session_handoff tool stores intent + calls updateSessionAgent(),
then agent-handoff hook fires on session.idle to switch the main session's
active agent via promptAsync with synthesis context.
This commit is contained in:
ismeth
2026-02-13 12:57:16 +01:00
committed by YeonGyu-Kim
parent fea732a6d2
commit 7a71d4fb4f
16 changed files with 282 additions and 12 deletions

View File

@@ -80,26 +80,36 @@ Step 4: After collecting ALL council member responses via background_output, syn
- Solo findings are potential false positives — flag the risk explicitly
- Add your own assessment and rationale to each finding
Step 5: Present synthesized findings to the user grouped by agreement level (unanimous first, then majority, minority, solo). End with action options: "fix now" (Atlas) or "create plan" (Prometheus).
Step 5: Present synthesized findings to the user grouped by agreement level (unanimous first, then majority, minority, solo). Then use the Question tool to ask which action to take:
Step 6: Wait for explicit user confirmation before delegating. NEVER delegate without confirmation.
- Direct fixes → delegate to Atlas using the task tool (background is fine — Atlas executes autonomously)
- Planning → do NOT spawn Prometheus as a background task. Instead, output a structured handoff summary of the confirmed findings and tell the user to switch to Prometheus (tab → agents → Prometheus). Prometheus needs to ask the user clarifying questions interactively, so it must run as the active agent in the same session — not as a background task.
Question({
questions: [{
question: "How should we proceed with these findings?",
header: "Action",
options: [
{ label: "Fix now (Atlas)", description: "Hand off to Atlas for direct implementation" },
{ label: "Create plan (Prometheus)", description: "Hand off to Prometheus for planning and phased execution" },
{ label: "No action", description: "Review only — no delegation" }
],
multiple: false
}]
})
## Prometheus Handoff Format
When the user confirms planning, output:
1. A clear summary of confirmed findings for Prometheus to work with
2. The original question for context
3. Tell the user: "Switch to Prometheus to start planning. It will see this conversation and can ask you questions."
Step 6: After the user selects an action:
- **"Fix now (Atlas)"** → Call session_handoff with agent="atlas" and context containing the confirmed findings summary, the original question, and instruction to implement the fixes.
- **"Create plan (Prometheus)"** → Call session_handoff with agent="prometheus" and context containing the confirmed findings summary, the original question, and instruction to create a phased plan.
- **"No action"** → Acknowledge and end. Do not delegate.
The session_handoff tool switches the active agent. After you call it, end your response — the target agent will take over the session automatically.
## Constraints
- Use the Question tool for member selection BEFORE calling athena_council (unless user pre-specified).
- Use the Question tool for action selection AFTER synthesis (unless user already stated intent).
- After athena_council, use background_output for each returned task ID before synthesizing.
- Do NOT write or edit files directly.
- Do NOT delegate without explicit user confirmation.
- Do NOT delegate without explicit user confirmation via Question tool.
- Do NOT ignore solo finding false-positive warnings.
- Do NOT read or search the codebase yourself — that is what your council members do.
- Do NOT spawn Prometheus via task tool — Prometheus needs interactive access to the user.`
- Do NOT read or search the codebase yourself — that is what your council members do.`
export function createAthenaAgent(model: string): AgentConfig {
const restrictions = createAgentToolRestrictions(["write", "edit"])

View File

@@ -49,6 +49,7 @@ export const HookNameSchema = z.enum([
"write-existing-file-guard",
"anthropic-effort",
"hashline-read-enhancer",
"agent-handoff",
])
export type HookName = z.infer<typeof HookNameSchema>

View File

@@ -0,0 +1,2 @@
export { setPendingHandoff, consumePendingHandoff, _resetForTesting } from "./state"
export type { PendingHandoff } from "./state"

View File

@@ -0,0 +1,50 @@
import { describe, test, expect, beforeEach } from "bun:test"
import { setPendingHandoff, consumePendingHandoff, _resetForTesting } from "./state"
describe("agent-handoff state", () => {
beforeEach(() => {
_resetForTesting()
})
//#given a pending handoff is set
//#when consumePendingHandoff is called
//#then it returns the handoff and removes it
test("should store and consume a pending handoff", () => {
setPendingHandoff("session-1", "atlas", "Fix these findings")
const handoff = consumePendingHandoff("session-1")
expect(handoff).toEqual({ agent: "atlas", context: "Fix these findings" })
expect(consumePendingHandoff("session-1")).toBeUndefined()
})
//#given no pending handoff exists
//#when consumePendingHandoff is called
//#then it returns undefined
test("should return undefined when no handoff is pending", () => {
expect(consumePendingHandoff("session-1")).toBeUndefined()
})
//#given a pending handoff is set
//#when a new handoff is set for the same session
//#then the latest handoff wins
test("should overwrite previous handoff for same session", () => {
setPendingHandoff("session-1", "atlas", "Fix A")
setPendingHandoff("session-1", "prometheus", "Plan B")
const handoff = consumePendingHandoff("session-1")
expect(handoff).toEqual({ agent: "prometheus", context: "Plan B" })
})
//#given handoffs for different sessions
//#when consumed separately
//#then each session gets its own handoff
test("should isolate handoffs by session", () => {
setPendingHandoff("session-1", "atlas", "Fix A")
setPendingHandoff("session-2", "prometheus", "Plan B")
expect(consumePendingHandoff("session-1")).toEqual({ agent: "atlas", context: "Fix A" })
expect(consumePendingHandoff("session-2")).toEqual({ agent: "prometheus", context: "Plan B" })
})
})

View File

@@ -0,0 +1,23 @@
export interface PendingHandoff {
agent: string
context: string
}
const pendingHandoffs = new Map<string, PendingHandoff>()
export function setPendingHandoff(sessionID: string, agent: string, context: string): void {
pendingHandoffs.set(sessionID, { agent, context })
}
export function consumePendingHandoff(sessionID: string): PendingHandoff | undefined {
const handoff = pendingHandoffs.get(sessionID)
if (handoff) {
pendingHandoffs.delete(sessionID)
}
return handoff
}
/** @internal For testing only */
export function _resetForTesting(): void {
pendingHandoffs.clear()
}

View File

@@ -0,0 +1,37 @@
import type { PluginInput } from "@opencode-ai/plugin"
import { consumePendingHandoff } from "../../features/agent-handoff"
import { log } from "../../shared/logger"
const HOOK_NAME = "agent-handoff" as const
export function createAgentHandoffHook(ctx: PluginInput) {
return {
event: async (input: { event: { type: string; properties?: Record<string, unknown> } }): Promise<void> => {
if (input.event.type !== "session.idle") return
const props = input.event.properties as Record<string, unknown> | undefined
const sessionID = props?.sessionID as string | undefined
if (!sessionID) return
const handoff = consumePendingHandoff(sessionID)
if (!handoff) return
log(`[${HOOK_NAME}] Executing handoff to ${handoff.agent}`, { sessionID })
try {
await ctx.client.session.promptAsync({
path: { id: sessionID },
body: {
agent: handoff.agent,
parts: [{ type: "text", text: handoff.context }],
},
query: { directory: ctx.directory },
})
log(`[${HOOK_NAME}] Handoff to ${handoff.agent} complete`, { sessionID })
} catch (err) {
log(`[${HOOK_NAME}] Handoff failed`, { sessionID, error: String(err) })
}
},
}
}

View File

@@ -0,0 +1 @@
export { createAgentHandoffHook } from "./hook"

View File

@@ -50,3 +50,5 @@ export { createRuntimeFallbackHook, type RuntimeFallbackHook, type RuntimeFallba
export { createWriteExistingFileGuardHook } from "./write-existing-file-guard";
export { createHashlineReadEnhancerHook } from "./hashline-read-enhancer";
export { createJsonErrorRecoveryHook, JSON_ERROR_TOOL_EXCLUDE_LIST, JSON_ERROR_PATTERNS, JSON_ERROR_REMINDER } from "./json-error-recovery";
export { createHashlineEditDiffEnhancerHook } from "./hashline-edit-diff-enhancer";
export { createAgentHandoffHook } from "./agent-handoff";

View File

@@ -156,6 +156,7 @@ export function createEventHandler(args: {
await Promise.resolve(hooks.compactionTodoPreserver?.event?.(input));
await Promise.resolve(hooks.writeExistingFileGuard?.event?.(input));
await Promise.resolve(hooks.atlasHook?.handler?.(input));
await Promise.resolve(hooks.agentHandoffHook?.event?.(input));
};
const recentSyntheticIdles = new Map<string, number>();

View File

@@ -9,6 +9,7 @@ import {
createCompactionContextInjector,
createCompactionTodoPreserverHook,
createAtlasHook,
createAgentHandoffHook,
} from "../../hooks"
import { safeCreateHook } from "../../shared/safe-create-hook"
import { createUnstableAgentBabysitter } from "../unstable-agent-babysitter"
@@ -21,6 +22,7 @@ export type ContinuationHooks = {
unstableAgentBabysitter: ReturnType<typeof createUnstableAgentBabysitter> | null
backgroundNotificationHook: ReturnType<typeof createBackgroundNotificationHook> | null
atlasHook: ReturnType<typeof createAtlasHook> | null
agentHandoffHook: ReturnType<typeof createAgentHandoffHook> | null
}
type SessionRecovery = {
@@ -111,6 +113,10 @@ export function createContinuationHooks(args: {
}))
: null
const agentHandoffHook = isHookEnabled("agent-handoff")
? safeHook("agent-handoff", () => createAgentHandoffHook(ctx))
: null
return {
stopContinuationGuard,
compactionContextInjector,
@@ -119,5 +125,6 @@ export function createContinuationHooks(args: {
unstableAgentBabysitter,
backgroundNotificationHook,
atlasHook,
agentHandoffHook,
}
}

View File

@@ -11,6 +11,7 @@ import {
createBackgroundTools,
createCallOmoAgent,
createAthenaCouncilTool,
createSessionHandoffTool,
createLookAt,
createSkillMcpTool,
createSkillTool,
@@ -133,6 +134,7 @@ export function createToolRegistry(args: {
...backgroundTools,
call_omo_agent: callOmoAgent,
athena_council: athenaCouncilTool,
session_handoff: createSessionHandoffTool(),
...(lookAt ? { look_at: lookAt } : {}),
task: delegateTask,
skill_mcp: skillMcpTool,

View File

@@ -38,6 +38,7 @@ export { createCallOmoAgent } from "./call-omo-agent"
export { createAthenaCouncilTool } from "./athena-council"
export { createLookAt } from "./look-at"
export { createDelegateTask } from "./delegate-task"
export { createSessionHandoffTool } from "./session-handoff"
export {
createTaskCreateTool,
createTaskGetTool,

View File

@@ -0,0 +1 @@
export { createSessionHandoffTool } from "./tools"

View File

@@ -0,0 +1,91 @@
import { describe, test, expect, beforeEach } from "bun:test"
import { createSessionHandoffTool } from "./tools"
import { consumePendingHandoff, _resetForTesting as resetHandoff } from "../../features/agent-handoff"
import { getSessionAgent, _resetForTesting as resetSession } from "../../features/claude-code-session-state"
describe("session_handoff tool", () => {
const sessionID = "test-session-123"
const messageID = "msg-456"
const agent = "athena"
const toolContext = {
sessionID,
messageID,
agent,
abort: new AbortController().signal,
}
beforeEach(() => {
resetHandoff()
resetSession()
})
//#given valid atlas handoff args
//#when execute is called
//#then it stores pending handoff and updates session agent
test("should queue handoff to atlas", async () => {
const tool = createSessionHandoffTool()
const result = await tool.execute(
{ agent: "atlas", context: "Fix the auth bug based on council findings" },
toolContext
)
expect(result).toContain("atlas")
expect(result).toContain("Handoff queued")
const handoff = consumePendingHandoff(sessionID)
expect(handoff).toEqual({
agent: "atlas",
context: "Fix the auth bug based on council findings",
})
expect(getSessionAgent(sessionID)).toBe("atlas")
})
//#given valid prometheus handoff args
//#when execute is called
//#then it stores pending handoff for prometheus
test("should queue handoff to prometheus", async () => {
const tool = createSessionHandoffTool()
const result = await tool.execute(
{ agent: "Prometheus", context: "Create a plan for the refactoring" },
toolContext
)
expect(result).toContain("prometheus")
expect(result).toContain("Handoff queued")
const handoff = consumePendingHandoff(sessionID)
expect(handoff?.agent).toBe("prometheus")
})
//#given an invalid agent name
//#when execute is called
//#then it returns an error
test("should reject invalid agent names", async () => {
const tool = createSessionHandoffTool()
const result = await tool.execute(
{ agent: "librarian", context: "Some context" },
toolContext
)
expect(result).toContain("Invalid handoff target")
expect(result).toContain("librarian")
expect(consumePendingHandoff(sessionID)).toBeUndefined()
})
//#given agent name with different casing
//#when execute is called
//#then it normalizes to lowercase
test("should handle case-insensitive agent names", async () => {
const tool = createSessionHandoffTool()
await tool.execute(
{ agent: "ATLAS", context: "Fix things" },
toolContext
)
const handoff = consumePendingHandoff(sessionID)
expect(handoff?.agent).toBe("atlas")
expect(getSessionAgent(sessionID)).toBe("atlas")
})
})

View File

@@ -0,0 +1,37 @@
import { tool, type ToolDefinition } from "@opencode-ai/plugin"
import { setPendingHandoff } from "../../features/agent-handoff"
import { updateSessionAgent } from "../../features/claude-code-session-state"
import type { SessionHandoffArgs } from "./types"
const DESCRIPTION =
"Switch the active session agent. After calling this tool, the session will transition to the specified agent " +
"with the provided context as its starting prompt. Use this to hand off work to another agent " +
"(e.g., Atlas for fixes, Prometheus for planning). The handoff executes when the current agent's turn completes."
const ALLOWED_AGENTS = new Set(["atlas", "prometheus", "sisyphus", "hephaestus"])
export function createSessionHandoffTool(): ToolDefinition {
return tool({
description: DESCRIPTION,
args: {
agent: tool.schema
.string()
.describe("Target agent name to hand off to (e.g., 'atlas', 'prometheus')"),
context: tool.schema
.string()
.describe("Context message for the target agent — include confirmed findings, the original question, and what action to take"),
},
async execute(args: SessionHandoffArgs, toolContext) {
const agentName = args.agent.toLowerCase()
if (!ALLOWED_AGENTS.has(agentName)) {
return `Invalid handoff target: "${args.agent}". Allowed agents: ${[...ALLOWED_AGENTS].join(", ")}`
}
updateSessionAgent(toolContext.sessionID, agentName)
setPendingHandoff(toolContext.sessionID, agentName, args.context)
return `Handoff queued. Session will switch to ${agentName} when your turn completes.`
},
})
}

View File

@@ -0,0 +1,4 @@
export interface SessionHandoffArgs {
agent: string
context: string
}