fix(ulw-loop): add fallback for Oracle verification session tracking

The verification_session_id was never reliably set because the
prompt-based attempt_id matching in tool-execute-after depends on
metadata.prompt surviving the delegate-task execution chain. When
this fails silently, the loop never detects Oracle's VERIFIED
emission.

Add a fallback: when exact attempt_id matching fails but oracle
agent + verification_pending state match, still set the session ID.
Add diagnostic logging to trace verification flow failures.
Add integration test covering the full verification chain.
This commit is contained in:
YeonGyu-Kim
2026-03-17 16:20:27 +09:00
parent 78dac0642e
commit 4bc7b1d27c
5 changed files with 220 additions and 44 deletions
+33 -5
View File
@@ -48,22 +48,50 @@ export function createToolExecuteAfterHandler(args: {
const prompt = typeof output.metadata?.prompt === "string" ? output.metadata.prompt : undefined
const verificationAttemptId = prompt?.match(VERIFICATION_ATTEMPT_PATTERN)?.[1]?.trim()
const loopState = directory ? readState(directory) : null
if (
const isVerificationContext =
agent === "oracle"
&& sessionId
&& verificationAttemptId
&& directory
&& !!sessionId
&& !!directory
&& loopState?.active === true
&& loopState.ultrawork === true
&& loopState.verification_pending === true
&& loopState.session_id === input.sessionID
log("[tool-execute-after] ULW verification tracking check", {
tool: input.tool,
agent,
parentSessionID: input.sessionID,
oracleSessionID: sessionId,
hasPromptInMetadata: typeof prompt === "string",
extractedVerificationAttemptId: verificationAttemptId,
})
if (
isVerificationContext
&& verificationAttemptId
&& loopState.verification_attempt_id === verificationAttemptId
) {
writeState(directory, {
...loopState,
verification_session_id: sessionId,
})
log("[tool-execute-after] Stored oracle verification session via attempt match", {
parentSessionID: input.sessionID,
oracleSessionID: sessionId,
verificationAttemptId,
})
} else if (isVerificationContext && !verificationAttemptId) {
writeState(directory, {
...loopState,
verification_session_id: sessionId,
})
log("[tool-execute-after] Fallback: stored oracle verification session without attempt match", {
parentSessionID: input.sessionID,
oracleSessionID: sessionId,
hasPromptInMetadata: typeof prompt === "string",
expectedAttemptId: loopState.verification_attempt_id,
extractedAttemptId: verificationAttemptId,
})
}
}
+6
View File
@@ -79,6 +79,12 @@ export function createToolExecuteBeforeHandler(args: {
if (shouldInjectOracleVerification) {
const verificationAttemptId = randomUUID()
log("[tool-execute-before] Injecting ULW oracle verification attempt", {
sessionID: input.sessionID,
callID: input.callID,
verificationAttemptId,
loopSessionID: loopState.session_id,
})
writeState(ctx.directory, {
...loopState,
verification_attempt_id: verificationAttemptId,
+66 -39
View File
@@ -19,6 +19,27 @@ describe("tool.execute.before ultrawork oracle verification", () => {
}
}
function createOracleTaskArgs(prompt: string): Record<string, unknown> {
return {
subagent_type: "oracle",
run_in_background: true,
prompt,
}
}
function createSyncTaskMetadata(
args: Record<string, unknown>,
sessionId: string,
): Record<string, unknown> {
return {
prompt: args.prompt,
agent: "oracle",
run_in_background: args.run_in_background,
sessionId,
sync: true,
}
}
test("#given ulw loop is awaiting verification #when oracle task runs #then oracle prompt is enforced and sync", async () => {
const directory = join(tmpdir(), `tool-before-ulw-${Date.now()}`)
mkdirSync(directory, { recursive: true })
@@ -38,13 +59,7 @@ describe("tool.execute.before ultrawork oracle verification", () => {
ctx: createCtx(directory) as unknown as Parameters<typeof createToolExecuteBeforeHandler>[0]["ctx"],
hooks: {} as Parameters<typeof createToolExecuteBeforeHandler>[0]["hooks"],
})
const output = {
args: {
subagent_type: "oracle",
run_in_background: true,
prompt: "Check it",
} as Record<string, unknown>,
}
const output = { args: createOracleTaskArgs("Check it") }
await handler({ tool: "task", sessionID: "ses-main", callID: "call-1" }, output)
@@ -64,13 +79,7 @@ describe("tool.execute.before ultrawork oracle verification", () => {
ctx: createCtx(directory) as unknown as Parameters<typeof createToolExecuteBeforeHandler>[0]["ctx"],
hooks: {} as Parameters<typeof createToolExecuteBeforeHandler>[0]["hooks"],
})
const output = {
args: {
subagent_type: "oracle",
run_in_background: true,
prompt: "Check it",
} as Record<string, unknown>,
}
const output = { args: createOracleTaskArgs("Check it") }
await handler({ tool: "task", sessionID: "ses-main", callID: "call-1" }, output)
@@ -80,7 +89,7 @@ describe("tool.execute.before ultrawork oracle verification", () => {
rmSync(directory, { recursive: true, force: true })
})
test("#given ulw loop is awaiting verification #when oracle task finishes #then oracle session id is stored", async () => {
test("#given ulw loop is awaiting verification #when oracle sync task metadata is persisted #then oracle session id is stored", async () => {
const directory = join(tmpdir(), `tool-after-ulw-${Date.now()}`)
mkdirSync(directory, { recursive: true })
writeState(directory, {
@@ -99,14 +108,44 @@ describe("tool.execute.before ultrawork oracle verification", () => {
ctx: createCtx(directory) as unknown as Parameters<typeof createToolExecuteBeforeHandler>[0]["ctx"],
hooks: {} as Parameters<typeof createToolExecuteBeforeHandler>[0]["hooks"],
})
const beforeOutput = {
args: {
subagent_type: "oracle",
run_in_background: true,
prompt: "Check it",
} as Record<string, unknown>,
}
const beforeOutput = { args: createOracleTaskArgs("Check it") }
await beforeHandler({ tool: "task", sessionID: "ses-main", callID: "call-1" }, beforeOutput)
const metadataFromSyncTask = createSyncTaskMetadata(beforeOutput.args, "ses-oracle")
const handler = createToolExecuteAfterHandler({
ctx: createCtx(directory) as unknown as Parameters<typeof createToolExecuteAfterHandler>[0]["ctx"],
hooks: {} as Parameters<typeof createToolExecuteAfterHandler>[0]["hooks"],
})
await handler(
{ tool: "task", sessionID: "ses-main", callID: "call-1" },
{
title: "oracle task",
output: "done",
metadata: metadataFromSyncTask,
},
)
expect(readState(directory)?.verification_session_id).toBe("ses-oracle")
clearState(directory)
rmSync(directory, { recursive: true, force: true })
})
test("#given ulw loop is awaiting verification #when oracle metadata prompt is missing #then oracle session fallback is stored", async () => {
const directory = join(tmpdir(), `tool-after-ulw-fallback-${Date.now()}`)
mkdirSync(directory, { recursive: true })
writeState(directory, {
active: true,
iteration: 3,
completion_promise: ULTRAWORK_VERIFICATION_PROMISE,
initial_completion_promise: "DONE",
started_at: new Date().toISOString(),
prompt: "Ship feature",
session_id: "ses-main",
ultrawork: true,
verification_pending: true,
})
const handler = createToolExecuteAfterHandler({
ctx: createCtx(directory) as unknown as Parameters<typeof createToolExecuteAfterHandler>[0]["ctx"],
@@ -120,13 +159,13 @@ describe("tool.execute.before ultrawork oracle verification", () => {
output: "done",
metadata: {
agent: "oracle",
prompt: String(beforeOutput.args.prompt),
sessionId: "ses-oracle",
sessionId: "ses-oracle-fallback",
sync: true,
},
},
)
expect(readState(directory)?.verification_session_id).toBe("ses-oracle")
expect(readState(directory)?.verification_session_id).toBe("ses-oracle-fallback")
clearState(directory)
rmSync(directory, { recursive: true, force: true })
@@ -156,23 +195,11 @@ describe("tool.execute.before ultrawork oracle verification", () => {
hooks: {} as Parameters<typeof createToolExecuteAfterHandler>[0]["hooks"],
})
const firstOutput = {
args: {
subagent_type: "oracle",
run_in_background: true,
prompt: "Check it",
} as Record<string, unknown>,
}
const firstOutput = { args: createOracleTaskArgs("Check it") }
await beforeHandler({ tool: "task", sessionID: "ses-main", callID: "call-1" }, firstOutput)
const firstAttemptId = readState(directory)?.verification_attempt_id
const secondOutput = {
args: {
subagent_type: "oracle",
run_in_background: true,
prompt: "Check it again",
} as Record<string, unknown>,
}
const secondOutput = { args: createOracleTaskArgs("Check it again") }
await beforeHandler({ tool: "task", sessionID: "ses-main", callID: "call-2" }, secondOutput)
const secondAttemptId = readState(directory)?.verification_attempt_id