feat(athena): inject intent runtime guidance and action flows

This commit is contained in:
ismeth
2026-03-01 00:18:39 +01:00
committed by YeonGyu-Kim
parent ed2bb1ead7
commit c265200f5f
8 changed files with 825 additions and 12 deletions
@@ -1,3 +1,5 @@
/// <reference types="bun-types" />
import { describe, expect, it } from "bun:test"
import { createAthenaAgent } from "./agent"
@@ -40,9 +42,9 @@ describe("Athena prompt config injection placeholders", () => {
it("#then uses sequential workflow step numbering", () => {
expect(athenaConfig.prompt).toContain("Step 12: Synthesize")
expect(athenaConfig.prompt).toContain("Step 13: Determine the follow-up path")
expect(athenaConfig.prompt).toContain("Step 14: ACTIONABLE findings")
expect(athenaConfig.prompt).toContain("Step 15: INFORMATIONAL findings")
expect(athenaConfig.prompt).toContain("Step 13: Determine follow-up path from council_finalize runtime guidance")
expect(athenaConfig.prompt).toContain("Step 14: Execute the runtime guidance action flow")
expect(athenaConfig.prompt).toContain("Step 15: Fallback behavior if runtime guidance is missing")
})
it("#then omits legacy mixed step labels", () => {
@@ -71,11 +73,10 @@ describe("Athena prompt config injection placeholders", () => {
expect(step2Index).toBeGreaterThan(step1Index)
})
it("#then keeps edge-case disambiguation inside step 3 classification", () => {
it("#then uses primary-objective wording for step 3 classification", () => {
const prompt = athenaConfig.prompt ?? ""
expect(prompt).toContain("Step 3: Classify the question intent.")
expect(prompt).toContain("Classification disambiguation rule:")
expect(prompt).not.toContain("Before selecting a route, apply this interpretation rule")
expect(prompt).toContain("Step 3: Classify the question intent by primary objective.")
expect(prompt).not.toContain("Classification disambiguation rule:")
})
it("#then requires question tool routing in self-answerable path", () => {
@@ -93,7 +94,7 @@ describe("Athena prompt config injection placeholders", () => {
})
it("#then keeps intent classification anchored to step 3 wording", () => {
expect(athenaConfig.prompt).toContain("Step 3: Classify the question intent.")
expect(athenaConfig.prompt).toContain("Step 3: Classify the question intent by primary objective.")
expect(athenaConfig.prompt).toContain("Then proceed to Step 2.")
expect(athenaConfig.prompt).not.toContain("Then classify intent and proceed to Step 2.")
expect(athenaConfig.prompt).not.toContain("Classify intent immediately and proceed to Step 2.")
@@ -102,6 +103,12 @@ describe("Athena prompt config injection placeholders", () => {
it("#then excludes non-interactive mode branch from runtime prompt", () => {
expect(athenaConfig.prompt).not.toContain("Non-interactive mode (Question tool unavailable)")
})
it("#then requires passing intent to council_finalize and honoring injected runtime guidance", () => {
expect(athenaConfig.prompt).toContain("intent=\"{intent from Step 3}\"")
expect(athenaConfig.prompt).toContain("runtime guidance message injected by council_finalize")
expect(athenaConfig.prompt).toContain("<athena_runtime_guidance>")
})
})
})
})
@@ -0,0 +1,562 @@
const VALID_INTENTS = [
"DIAGNOSE",
"AUDIT",
"PLAN",
"EVALUATE",
"EXPLAIN",
"CREATE",
"PERSPECTIVES",
"FREEFORM",
] as const
export type CouncilIntent = (typeof VALID_INTENTS)[number]
const RUNTIME_GUIDANCE_BY_INTENT: Record<CouncilIntent, string> = {
DIAGNOSE: `
<runtime_synthesis_rules>
Use DIAGNOSE synthesis.
- Build: Symptom -> hypotheses -> root cause -> contributing factors -> recommended fix.
- Anchor confidence to agreement level and evidence quality.
</runtime_synthesis_rules>
<runtime_action_paths>
Path type: ACTIONABLE.
1) Treat DIAGNOSE as a single-incident flow.
- Identify the primary root-cause fix candidate as one scoped item.
2) If no actionable fix candidate is identified, ask:
Question({
questions: [{
question: "No actionable findings were identified. What should we do next?",
header: "Next Step",
options: [
{ label: "Ask follow-up", description: "Ask a clarifying question and run another council pass" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
3) If an actionable fix candidate exists, ask action directly (no findings multi-select):
Question({
questions: [{
question: "How should we execute this diagnosis fix?",
header: "Action",
options: [
{ label: "Implement (Hephaestus)", description: "Hand off to Hephaestus for direct implementation" },
{ label: "Implement (Sisyphus)", description: "Hand off to Sisyphus for implementation" },
{ label: "Implement (Sisyphus ultrawork)", description: "Hand off to Sisyphus with ultrawork mode" },
{ label: "No action", description: "Review only - no delegation" }
],
multiple: false
}]
})
4) Execute selected action:
- Implement (Hephaestus) -> switch_agent(agent="hephaestus")
- Implement (Sisyphus) -> switch_agent(agent="sisyphus")
- Implement (Sisyphus ultrawork) -> switch_agent(agent="sisyphus") and prefix the handoff context with "ultrawork "
- No action -> acknowledge and end
</runtime_action_paths>`,
AUDIT: `
<runtime_synthesis_rules>
Use AUDIT synthesis.
- Output numbered findings grouped by confidence: unanimous, majority, minority, solo.
- For each finding: issue, impact, evidence, fix direction.
</runtime_synthesis_rules>
<runtime_action_paths>
Path type: ACTIONABLE.
1) Build actionable findings list from audit synthesis.
2) Ask how the user wants to process findings:
Question({
questions: [{
question: "How would you like to process the findings?",
header: "Processing Mode",
options: [
{ label: "One by one", description: "Review findings individually with per-finding decisions" },
{ label: "By severity/urgency", description: "Select findings by criticality first" },
{ label: "By quorum", description: "Select findings by agreement level" }
],
multiple: false
}]
})
3) Branch by processing mode.
Mode: One by one
- Process findings in batches using ONE Question call per batch.
- Default batch size: 3 findings per batch.
- Hard cap: 5 findings only when findings are short and options are limited.
- For each finding in the batch, provide dynamic options from synthesis (for example A/B/C fix options) plus:
- Skip
- Defer
- Stop review
- Example Question tool call (batch of 3 findings):
Question({
questions: [
{
question: "Finding #10: choose how to proceed.",
header: "#10 Action",
options: [
{ label: "Option A", description: "Apply fix option A from synthesis" },
{ label: "Option B", description: "Apply fix option B from synthesis" },
{ label: "Skip", description: "Do not act on this finding now" },
{ label: "Defer", description: "Keep for later" },
{ label: "Stop review", description: "End one-by-one processing" }
],
multiple: false
},
{
question: "Finding #11: choose how to proceed.",
header: "#11 Action",
options: [
{ label: "Option A", description: "Apply fix option A from synthesis" },
{ label: "Option C", description: "Apply fix option C from synthesis" },
{ label: "Skip", description: "Do not act on this finding now" },
{ label: "Defer", description: "Keep for later" },
{ label: "Stop review", description: "End one-by-one processing" }
],
multiple: false
},
{
question: "Finding #12: choose how to proceed.",
header: "#12 Action",
options: [
{ label: "Option B", description: "Apply fix option B from synthesis" },
{ label: "Skip", description: "Do not act on this finding now" },
{ label: "Defer", description: "Keep for later" },
{ label: "Stop review", description: "End one-by-one processing" }
],
multiple: false
}
]
})
- Keep free-form answers enabled (user may type custom choices like "#10:A, #11:skip"). Parse them.
- Continue batch-by-batch until user stops or all findings are processed.
Mode: By severity/urgency
- Ask multi-select severity buckets using only non-empty groups.
- Typical groups: Critical, High, Medium, Low.
- Example Question tool call:
Question({
questions: [{
question: "Which findings should we act on by severity? You can also type specific finding numbers (e.g. #1, #3, #7).",
header: "Select Findings",
options: [
// Include ONLY severities that actually have findings. Skip empty ones.
// Replace N with the actual count for each category.
{ label: "All Critical (N)", description: "Highest urgency findings" },
{ label: "All High (N)", description: "High-priority findings" },
{ label: "All Medium (N)", description: "Medium-priority findings" },
{ label: "All Low (N)", description: "Lower-priority findings" },
],
multiple: true
}]
})
- Resolve selected buckets into concrete finding IDs.
Mode: By quorum
- Ask multi-select quorum buckets using only non-empty groups.
- Typical groups: Unanimous, Majority, Minority, Solo.
- Example Question tool call:
Question({
questions: [{
question: "Which findings should we act on? You can also type specific finding numbers (e.g. #1, #3, #7).",
header: "Select Findings",
options: [
// Include ONLY categories that actually have findings. Skip empty ones.
// Replace N with the actual count for each category.
{ label: "All Unanimous (N)", description: "Findings agreed on by all members" },
{ label: "All Majority (N)", description: "Findings agreed on by most members" },
{ label: "All Minority (N)", description: "Findings from 2+ members - higher false-positive risk" },
{ label: "All Solo (N)", description: "Single-member findings - potential false positives" },
],
multiple: true
}]
})
- Resolve selected buckets into concrete finding IDs.
4) If actionable findings count is 0 OR user selected no findings, do NOT ask execution action. Ask:
Question({
questions: [{
question: "No findings were selected for action. What should we do next?",
header: "Next Step",
options: [
{ label: "Ask follow-up", description: "Ask a clarifying question and run another council pass" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
5) If selected findings exist, ask what action to take on the selected findings:
Question({
questions: [{
question: "How should we handle the selected 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
}]
})
6) Execute selected action:
- Fix now (Atlas) -> switch_agent(agent="atlas") with ONLY selected findings
- Create plan (Prometheus) -> switch_agent(agent="prometheus") with ONLY selected findings
- No action -> acknowledge and end
</runtime_action_paths>`,
PLAN: `
<runtime_synthesis_rules>
Use PLAN synthesis.
- Consolidate into one execution-ready phased plan.
- For each phase include: goal, tasks, dependencies, risks, effort estimate, and exit criteria.
- If members disagree on sequencing or strategy, preserve alternatives and pick a default recommendation with rationale.
- End with critical path and immediate first step.
</runtime_synthesis_rules>
<runtime_action_paths>
Path type: PLAN_EXECUTION.
1) Ask what to do with this plan:
Question({
questions: [{
question: "What should we do with this plan?",
header: "Plan Next Step",
options: [
{ label: "Execute full plan (Prometheus)", description: "Hand off all phases to Prometheus for execution" },
{ label: "Execute selected phase (Prometheus)", description: "Choose one phase and execute only that phase first" },
{ label: "Write to document", description: "Write the plan under .sisyphus/athena/notes/{council-session-name}" },
{ label: "Ask follow-up", description: "Ask another planning question" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
2) If user chooses Execute selected phase (Prometheus), ask:
Question({
questions: [{
question: "Which phase should we execute first?",
header: "Select Phase",
options: [
// Build from synthesized plan phases (for example: Phase 1, Phase 2, Phase 3).
// Include concise phase goal in each description.
],
multiple: false
}]
})
3) Execute selected action:
- Execute full plan (Prometheus) -> switch_agent(agent="prometheus") with full synthesized plan.
- Execute selected phase (Prometheus) -> switch_agent(agent="prometheus") with only the selected phase plus dependencies.
- Write to document -> write the document directly to ".sisyphus/athena/notes/{council-session-name}" and then report the exact path to the user.
- Ask follow-up -> ask user then restart at Step 3.
- Done -> acknowledge and end.
</runtime_action_paths>`,
EVALUATE: `
<runtime_synthesis_rules>
Use EVALUATE synthesis.
- Compare options against explicit criteria.
- Surface tradeoffs and finish with a primary recommendation plus fallback conditions.
- State confidence and the key uncertainty that could change the recommendation.
</runtime_synthesis_rules>
<runtime_action_paths>
Path type: INFORMATIONAL.
1) Ask what to do with the evaluation:
Question({
questions: [{
question: "What should we do with this evaluation?",
header: "Evaluation Next Step",
options: [
{ label: "Adopt option -> create plan (Prometheus)", description: "Turn a selected option into an execution plan" },
{ label: "Adopt option -> implement now", description: "Implement a selected option immediately" },
{ label: "Write to document", description: "Write under .sisyphus/athena/notes/{council-session-name}" },
{ label: "Ask follow-up", description: "Ask another comparison question" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
2) If user chooses either adopt-option path, ask:
Question({
questions: [{
question: "Which option should we adopt?",
header: "Select Option",
options: [
// Build from synthesized options list (e.g., Option A, Option B, Option C).
],
multiple: false
}]
})
3) If user chooses "Adopt option -> implement now", ask execution agent:
Question({
questions: [{
question: "Which execution agent should implement the selected option?",
header: "Execution Agent",
options: [
{ label: "Hephaestus", description: "Direct implementation with Hephaestus" },
{ label: "Sisyphus", description: "Implementation with Sisyphus" },
{ label: "Sisyphus ultrawork", description: "Implementation with Sisyphus using ultrawork mode" }
],
multiple: false
}]
})
4) Execute selected action:
- Adopt option -> create plan (Prometheus) -> switch_agent(agent="prometheus") with selected option and rationale.
- Adopt option -> implement now + Hephaestus -> switch_agent(agent="hephaestus") with selected option.
- Adopt option -> implement now + Sisyphus -> switch_agent(agent="sisyphus") with selected option.
- Adopt option -> implement now + Sisyphus ultrawork -> switch_agent(agent="sisyphus") and prefix handoff context with "ultrawork ".
- Write to document -> write directly to ".sisyphus/athena/notes/{council-session-name}" and then report the exact path.
- Ask follow-up -> ask user then restart at Step 3.
- Done -> acknowledge and end.
</runtime_action_paths>`,
EXPLAIN: `
<runtime_synthesis_rules>
Use EXPLAIN synthesis.
- Start with thesis.
- Then mechanisms/interactions.
- Then uncertainties and confidence.
- Include a concise "why this matters" section tied to the user question.
</runtime_synthesis_rules>
<runtime_action_paths>
Path type: INFORMATIONAL.
1) Ask what to do with the explanation:
Question({
questions: [{
question: "What should we do with this explanation?",
header: "Explanation Next Step",
options: [
{ label: "Convert to action plan (Prometheus)", description: "Turn insights into a phased plan" },
{ label: "Write to document", description: "Write under .sisyphus/athena/notes/{council-session-name}" },
{ label: "Ask follow-up", description: "Ask another explanatory question" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
2) Execute selected action:
- Convert to action plan (Prometheus) -> switch_agent(agent="prometheus") with synthesized explanation and target outcome.
- Write to document -> write directly to ".sisyphus/athena/notes/{council-session-name}" and then report the exact path.
- Ask follow-up -> ask user then restart at Step 3.
- Done -> acknowledge and end.
</runtime_action_paths>`,
CREATE: `
<runtime_synthesis_rules>
Use CREATE synthesis.
- Preserve creations side-by-side as a gallery.
- Do not collapse into a single merged artifact unless the user asks.
- Assign stable IDs to creations (e.g., C1, C2, C3) so user selections are unambiguous.
</runtime_synthesis_rules>
<runtime_action_paths>
Path type: INFORMATIONAL.
1) Ask which creations to carry forward:
Question({
questions: [{
question: "Which creations should we carry forward? You can also type IDs (e.g. C1, C3).",
header: "Select Creations",
options: [
// Build from synthesized gallery IDs and titles.
],
multiple: true
}]
})
2) If no creations are selected, ask:
Question({
questions: [{
question: "No creations selected. What should we do next?",
header: "Next Step",
options: [
{ label: "Ask follow-up", description: "Ask for clearer selection criteria" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
3) If creations are selected, ask:
Question({
questions: [{
question: "How should we proceed with the selected creations?",
header: "Creation Next Step",
options: [
{ label: "Implement selected creation (Hephaestus)", description: "Direct implementation with Hephaestus" },
{ label: "Implement selected creation (Sisyphus)", description: "Implementation with Sisyphus" },
{ label: "Implement selected creation (Sisyphus ultrawork)", description: "Implementation with Sisyphus using ultrawork mode" },
{ label: "Write selected creation to document", description: "Write under .sisyphus/athena/notes/{council-session-name}" },
{ label: "Ask follow-up", description: "Ask another creation-focused question" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
4) Execute selected action:
- Implement selected creation (Hephaestus) -> switch_agent(agent="hephaestus") with only selected creation(s).
- Implement selected creation (Sisyphus) -> switch_agent(agent="sisyphus") with only selected creation(s).
- Implement selected creation (Sisyphus ultrawork) -> switch_agent(agent="sisyphus") and prefix handoff context with "ultrawork ", including only selected creation(s).
- Write selected creation to document -> write directly to ".sisyphus/athena/notes/{council-session-name}" and then report the exact path.
- Ask follow-up -> ask user then restart at Step 3.
- Done -> acknowledge and end.
</runtime_action_paths>`,
PERSPECTIVES: `
<runtime_synthesis_rules>
Use PERSPECTIVES synthesis.
- Map positions.
- Identify tensions.
- Evaluate evidence strength.
- Take a final stance with conditions.
- Name strongest counter-position and what evidence could overturn the final stance.
</runtime_synthesis_rules>
<runtime_action_paths>
Path type: INFORMATIONAL.
1) Ask what to do with these perspectives:
Question({
questions: [{
question: "What should we do with this perspectives analysis?",
header: "Perspectives Next Step",
options: [
{ label: "Commit to stance -> create plan (Prometheus)", description: "Turn a chosen stance into a phased plan" },
{ label: "Commit to stance -> implement now", description: "Implement based on a chosen stance immediately" },
{ label: "Write to document", description: "Write under .sisyphus/athena/notes/{council-session-name}" },
{ label: "Ask follow-up", description: "Ask another perspective question" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
2) If user chooses a commit-to-stance path, ask:
Question({
questions: [{
question: "Which stance should we commit to?",
header: "Select Stance",
options: [
// Build from synthesized perspective labels and final stance.
],
multiple: false
}]
})
3) If user chooses "Commit to stance -> implement now", ask execution agent:
Question({
questions: [{
question: "Which execution agent should implement this stance?",
header: "Execution Agent",
options: [
{ label: "Hephaestus", description: "Direct implementation with Hephaestus" },
{ label: "Sisyphus", description: "Implementation with Sisyphus" },
{ label: "Sisyphus ultrawork", description: "Implementation with Sisyphus using ultrawork mode" }
],
multiple: false
}]
})
4) Execute selected action:
- Commit to stance -> create plan (Prometheus) -> switch_agent(agent="prometheus") with selected stance and rationale.
- Commit to stance -> implement now + Hephaestus -> switch_agent(agent="hephaestus") with selected stance.
- Commit to stance -> implement now + Sisyphus -> switch_agent(agent="sisyphus") with selected stance.
- Commit to stance -> implement now + Sisyphus ultrawork -> switch_agent(agent="sisyphus") and prefix handoff context with "ultrawork ".
- Write to document -> write directly to ".sisyphus/athena/notes/{council-session-name}" and then report the exact path.
- Ask follow-up -> ask user then restart at Step 3.
- Done -> acknowledge and end.
</runtime_action_paths>`,
FREEFORM: `
<runtime_synthesis_rules>
Use FREEFORM synthesis.
- Preserve meaningful diversity across member responses.
- Avoid forcing rigid structure.
- Produce a clear bottom-line answer plus notable alternatives.
</runtime_synthesis_rules>
<runtime_action_paths>
Path type: INFORMATIONAL.
1) Ask what to do next:
Question({
questions: [{
question: "What should we do with this result?",
header: "Next Step",
options: [
{ label: "Create plan (Prometheus)", description: "Turn the result into a phased execution plan" },
{ label: "Implement now", description: "Implement directly from this result" },
{ label: "Write to document", description: "Write under .sisyphus/athena/notes/{council-session-name}" },
{ label: "Ask follow-up", description: "Ask another question" },
{ label: "Done", description: "No further action needed" }
],
multiple: false
}]
})
2) If user chooses Implement now, ask execution agent:
Question({
questions: [{
question: "Which execution agent should implement this?",
header: "Execution Agent",
options: [
{ label: "Hephaestus", description: "Direct implementation with Hephaestus" },
{ label: "Sisyphus", description: "Implementation with Sisyphus" },
{ label: "Sisyphus ultrawork", description: "Implementation with Sisyphus using ultrawork mode" }
],
multiple: false
}]
})
3) Execute selected action:
- Create plan (Prometheus) -> switch_agent(agent="prometheus") with synthesized result.
- Implement now + Hephaestus -> switch_agent(agent="hephaestus") with synthesized result.
- Implement now + Sisyphus -> switch_agent(agent="sisyphus") with synthesized result.
- Implement now + Sisyphus ultrawork -> switch_agent(agent="sisyphus") and prefix handoff context with "ultrawork ".
- Write to document -> write directly to ".sisyphus/athena/notes/{council-session-name}" and then report the exact path.
- Ask follow-up -> ask user then restart at Step 3.
- Done -> acknowledge and end.
</runtime_action_paths>`,
}
export function getValidCouncilIntents(): readonly CouncilIntent[] {
return VALID_INTENTS
}
export function resolveCouncilIntent(intent?: string): CouncilIntent | null {
if (!intent) return null
const normalized = intent.toUpperCase()
return (VALID_INTENTS as readonly string[]).includes(normalized)
? (normalized as CouncilIntent)
: null
}
export function buildAthenaRuntimeGuidance(intent: CouncilIntent): string {
return [
"<athena_runtime_guidance>",
"source: council_finalize",
`intent: ${intent}`,
RUNTIME_GUIDANCE_BY_INTENT[intent].trim(),
"</athena_runtime_guidance>",
].join("\n\n")
}
+6
View File
@@ -1,3 +1,9 @@
export { createAthenaAgent, ATHENA_PROMPT_METADATA } from "./agent"
export { createCouncilMemberAgent, COUNCIL_MEMBER_PROMPT, COUNCIL_SOLO_ADDENDUM, COUNCIL_DELEGATION_ADDENDUM } from "./council-member-agent"
export { COUNCIL_INTENT_ADDENDUMS } from "./council-intent-addendums"
export {
buildAthenaRuntimeGuidance,
getValidCouncilIntents,
resolveCouncilIntent,
} from "./council-runtime-guidance"
export type { CouncilIntent } from "./council-runtime-guidance"
+2 -1
View File
@@ -32,6 +32,7 @@ import {
createPrepareCouncilPromptTool,
} from "../tools"
import { createCouncilFinalize, createCouncilRead } from "../tools/council-archive"
import { contextCollector } from "../features/context-injector"
import { getMainSessionID } from "../features/claude-code-session-state"
import { filterDisabledTools } from "../shared/disabled-tools"
import { isTaskSystemEnabled, log } from "../shared"
@@ -281,7 +282,7 @@ export function createToolRegistry(args: {
...taskToolsRecord,
...hashlineToolsRecord,
prepare_council_prompt: createPrepareCouncilPromptTool(ctx.directory),
council_finalize: createCouncilFinalize(ctx.directory),
council_finalize: createCouncilFinalize(ctx.directory, { contextCollector }),
council_read: createCouncilRead(ctx.directory),
}
@@ -1,3 +1,5 @@
/// <reference types="bun-types" />
import { describe, expect, it, beforeEach } from "bun:test"
import { mkdtemp, mkdir, writeFile, readFile } from "node:fs/promises"
import { join } from "node:path"
@@ -232,4 +234,165 @@ describe("createCouncilFinalize", () => {
expect(secondContent).toContain("Second response")
})
})
describe("#given intent-based runtime guidance injection", () => {
it("#then registers critical custom context for Athena runtime guidance", async () => {
await writeFile(
join(tmpDir, ".sisyphus", "task-outputs", "bg_intent.md"),
mockTaskOutput("Council: GPT-5", "Plan proposal"),
"utf-8",
)
const calls: Array<{ sessionID: string; content: string; priority?: string; source: string; id: string }> = []
const collector = {
register: (sessionID: string, options: { id: string; source: string; content: string; priority?: string }) => {
calls.push({
sessionID,
id: options.id,
source: options.source,
content: options.content,
priority: options.priority,
})
},
}
const toolDef = createCouncilFinalize(tmpDir, { contextCollector: collector })
const result = await toolDef.execute(
{ task_ids: ["bg_intent"], name: "intent", intent: "PLAN" },
mockCtx,
)
expect(() => JSON.parse(result)).not.toThrow()
expect(calls).toHaveLength(1)
expect(calls[0].sessionID).toBe(mockCtx.sessionID)
expect(calls[0].id).toBe("athena-runtime-guidance")
expect(calls[0].source).toBe("custom")
expect(calls[0].priority).toBe("critical")
expect(calls[0].content).toContain("<athena_runtime_guidance>")
expect(calls[0].content).toContain("intent: PLAN")
expect(calls[0].content).toContain("Execute full plan (Prometheus)")
expect(calls[0].content).toContain("Execute selected phase (Prometheus)")
expect(calls[0].content).toContain(".sisyphus/athena/notes/{council-session-name}")
expect(calls[0].content).not.toContain("Hand off to Atlas to save the plan as .md")
})
it("#then emits diagnose action options for hephaestus and sisyphus", async () => {
await writeFile(
join(tmpDir, ".sisyphus", "task-outputs", "bg_diagnose.md"),
mockTaskOutput("Council: Claude", "Root cause found"),
"utf-8",
)
const calls: Array<{ content: string }> = []
const collector = {
register: (_sessionID: string, options: { content: string }) => {
calls.push({ content: options.content })
},
}
const toolDef = createCouncilFinalize(tmpDir, { contextCollector: collector })
const result = await toolDef.execute(
{ task_ids: ["bg_diagnose"], name: "diagnose", intent: "DIAGNOSE" },
mockCtx,
)
expect(() => JSON.parse(result)).not.toThrow()
expect(calls).toHaveLength(1)
expect(calls[0].content).toContain("Implement (Hephaestus)")
expect(calls[0].content).toContain("Implement (Sisyphus)")
expect(calls[0].content).toContain("Implement (Sisyphus ultrawork)")
expect(calls[0].content).toContain("switch_agent(agent=\"hephaestus\")")
expect(calls[0].content).toContain("switch_agent(agent=\"sisyphus\")")
expect(calls[0].content).toContain("prefix the handoff context with \"ultrawork \"")
expect(calls[0].content).not.toContain("Fix now (Atlas)")
expect(calls[0].content).not.toContain("Create plan (Prometheus)")
})
it("#then emits audit processing mode and batching guidance", async () => {
await writeFile(
join(tmpDir, ".sisyphus", "task-outputs", "bg_audit.md"),
mockTaskOutput("Council: Claude", "Audit findings"),
"utf-8",
)
const calls: Array<{ content: string }> = []
const collector = {
register: (_sessionID: string, options: { content: string }) => {
calls.push({ content: options.content })
},
}
const toolDef = createCouncilFinalize(tmpDir, { contextCollector: collector })
const result = await toolDef.execute(
{ task_ids: ["bg_audit"], name: "audit", intent: "AUDIT" },
mockCtx,
)
expect(() => JSON.parse(result)).not.toThrow()
expect(calls).toHaveLength(1)
expect(calls[0].content).toContain("How would you like to process the findings?")
expect(calls[0].content).toContain("One by one")
expect(calls[0].content).toContain("By severity/urgency")
expect(calls[0].content).toContain("By quorum")
expect(calls[0].content).toContain("Default batch size: 3 findings per batch")
expect(calls[0].content).toContain("Hard cap: 5 findings")
expect(calls[0].content).toContain("Example Question tool call (batch of 3 findings)")
expect(calls[0].content).toContain("Finding #10: choose how to proceed.")
expect(calls[0].content).toContain("#10 Action")
expect(calls[0].content).toContain("Stop review")
expect(calls[0].content).toContain("#10:A, #11:skip")
expect(calls[0].content).toContain("Which findings should we act on by severity?")
expect(calls[0].content).toContain("All Critical (N)")
expect(calls[0].content).toContain("All High (N)")
expect(calls[0].content).toContain("All Medium (N)")
expect(calls[0].content).toContain("All Low (N)")
expect(calls[0].content).toContain("Which findings should we act on? You can also type specific finding numbers")
expect(calls[0].content).toContain("All Unanimous (N)")
expect(calls[0].content).toContain("All Majority (N)")
expect(calls[0].content).toContain("All Minority (N)")
expect(calls[0].content).toContain("All Solo (N)")
expect(calls[0].content).toContain("Fix now (Atlas)")
expect(calls[0].content).toContain("Create plan (Prometheus)")
})
it("#then emits informational write-to-document path without atlas delegation", async () => {
await writeFile(
join(tmpDir, ".sisyphus", "task-outputs", "bg_eval.md"),
mockTaskOutput("Council: Claude", "Option comparison"),
"utf-8",
)
const calls: Array<{ content: string }> = []
const collector = {
register: (_sessionID: string, options: { content: string }) => {
calls.push({ content: options.content })
},
}
const toolDef = createCouncilFinalize(tmpDir, { contextCollector: collector })
const result = await toolDef.execute(
{ task_ids: ["bg_eval"], name: "eval", intent: "EVALUATE" },
mockCtx,
)
expect(() => JSON.parse(result)).not.toThrow()
expect(calls).toHaveLength(1)
expect(calls[0].content).toContain("What should we do with this evaluation?")
expect(calls[0].content).toContain("Adopt option -> create plan (Prometheus)")
expect(calls[0].content).toContain("Adopt option -> implement now")
expect(calls[0].content).toContain(".sisyphus/athena/notes/{council-session-name}")
expect(calls[0].content).not.toContain("Write to document (Atlas)")
})
it("#then rejects invalid intent values", async () => {
const toolDef = createCouncilFinalize(tmpDir)
const result = await toolDef.execute(
{ task_ids: ["bg_none"], name: "invalid-intent", intent: "NOT_A_REAL_INTENT" },
mockCtx,
)
expect(result).toContain("Invalid intent")
expect(result).toContain("NOT_A_REAL_INTENT")
})
})
})
@@ -3,7 +3,13 @@ import { readFile, writeFile, mkdir, rename } from "node:fs/promises"
import { join, isAbsolute, resolve } from "node:path"
import { randomBytes } from "node:crypto"
import { extractCouncilResponse } from "./council-response-extractor"
import {
buildAthenaRuntimeGuidance,
getValidCouncilIntents,
resolveCouncilIntent,
} from "../../agents/athena"
import { log } from "../../shared/logger"
import type { ContextCollector } from "../../features/context-injector"
import type { CouncilFinalizeArgs, CouncilMemberResult, CouncilFinalizeResult } from "./types"
interface MetaMember {
@@ -16,6 +22,12 @@ interface MetaMember {
response_complete: boolean
}
type RegisterContext = Pick<ContextCollector, "register">
type CouncilFinalizeToolContext = {
sessionID?: string
}
function slugify(text: string): string {
return text
.toLowerCase()
@@ -66,19 +78,46 @@ function formatMetaYaml(archiveName: string, createdAt: string, members: MetaMem
return lines.join("\n") + "\n"
}
export function createCouncilFinalize(basePath?: string): ToolDefinition {
export function createCouncilFinalize(
basePath?: string,
options?: { contextCollector?: RegisterContext }
): ToolDefinition {
const collector = options?.contextCollector
return tool({
description:
"Finalize council task outputs: extract COUNCIL_MEMBER_RESPONSE content from raw task output files, write per-member archive files, and create meta.yaml.",
"Finalize council task outputs: extract COUNCIL_MEMBER_RESPONSE content from raw task output files, write per-member archive files, inject intent-specific Athena runtime guidance, and create meta.yaml.",
args: {
task_ids: tool.schema
.array(tool.schema.string())
.describe("Array of background task IDs whose output files should be processed"),
name: tool.schema.string().describe("Council name used in the archive directory name"),
intent: tool.schema
.string()
.optional()
.describe(`Classified question intent used for runtime Athena guidance injection. Valid intents: ${getValidCouncilIntents().join(", ")}`),
question: tool.schema.string().optional().describe("Original user question that triggered the council"),
prompt_file: tool.schema.string().optional().describe("Path to the council prompt temp file (will be moved into the archive)"),
},
async execute(args: CouncilFinalizeArgs) {
async execute(args: CouncilFinalizeArgs, toolContext: CouncilFinalizeToolContext) {
const resolvedIntent = resolveCouncilIntent(args.intent)
if (args.intent && !resolvedIntent) {
return `Invalid intent: "${args.intent}". Valid intents: ${getValidCouncilIntents().join(", ")}.`
}
if (collector && resolvedIntent && toolContext.sessionID) {
collector.register(toolContext.sessionID, {
id: "athena-runtime-guidance",
source: "custom",
priority: "critical",
content: buildAthenaRuntimeGuidance(resolvedIntent),
metadata: {
intent: resolvedIntent,
source: "council_finalize",
},
})
}
const base = basePath ?? process.cwd()
const hexId = randomBytes(2).toString("hex")
const archiveName = `council-${args.name}-${hexId}`
+1
View File
@@ -1,6 +1,7 @@
export interface CouncilFinalizeArgs {
task_ids: string[]
name: string
intent?: string
question?: string
prompt_file?: string
}
+34
View File
@@ -86,6 +86,40 @@ describe("switch_agent tool", () => {
expect(entry?.agent).toBe("prometheus")
})
//#given valid hephaestus switch args
//#when execute is called
//#then it stores pending switch for hephaestus
test("should queue switch to hephaestus", async () => {
const tool = createToolWithMockClient()
const result = await tool.execute(
{ agent: "Hephaestus", context: "Implement the selected diagnosis fix" },
toolContext
)
expect(result).toContain("hephaestus")
expect(result).toContain("switch")
const entry = consumePendingSwitch(sessionID)
expect(entry?.agent).toBe("hephaestus")
})
//#given valid sisyphus switch args
//#when execute is called
//#then it stores pending switch for sisyphus
test("should queue switch to sisyphus", async () => {
const tool = createToolWithMockClient()
const result = await tool.execute(
{ agent: "Sisyphus", context: "Implement the selected diagnosis fix" },
toolContext
)
expect(result).toContain("sisyphus")
expect(result).toContain("switch")
const entry = consumePendingSwitch(sessionID)
expect(entry?.agent).toBe("sisyphus")
})
//#given an invalid agent name
//#when execute is called
//#then it returns an error