fix: read the judge's reply in either shape $.model.complete resolves

CHANGELOG.md

@@ -11,6 +11,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - taskcut is released under the MIT License, in place of Apache-2.0.
 
+### Fixed
+
+- **Nothing was compacted on Claude Code 2.1.280.** `$.model.complete` now
+  resolves `{ isAnswered, text, usage }` rather than the reply's text; taskcut
+  read the object as text, every judgement failed, and the failure was logged
+  only to the debug log. The reply is now read in either shape, so 2.1.278
+  keeps working, and a failed call is logged with its reason.
+
 ## [0.8.0] - 2026-09-22
 
 ### Changed

docs/compatibility.md

@@ -11,6 +11,7 @@ declarations it is written against are generated per Claude Code version by
 | `session.start`, `turn.step`, `turn.complete` | hook events | taskcut stops working; `claude plugin validate` reports the unknown event before a session loads it |
 | `$.session.usage`, `$.session.messages`, `$.session.compact`, `$.fs.read`, `$.env.get`, `$.model.complete`, `$.turn.abort`, `$.prompt.submit`, `$.ui.log` | engine calls | same: refused at load, named in the validation output |
 | a hook's budget counting only its own code, not its `$` calls | engine rule | the judgement inside a step would overrun it, and the engine would skip the hook: taskcut would compact only at the end of a turn, and say so in the transcript |
+| what `$.model.complete` resolves to | data shape | `claude plugin validate` does not see it; only `tsc` against regenerated types does. 2.1.278, which taskcut was measured on, resolved the reply's text; 2.1.280 resolves `{ isAnswered, text, usage }`, and 0.8.0 read that object as text, so every judgement failed and nothing was ever compacted. The reply is now taken as `unknown` and read by `readReply`, which knows both shapes and treats anything else as no reply |
 | `session.start`'s `isInteractive` | data shape | taskcut would stop ending turns early; a compaction at the end of a turn still works |
 | `CLAUDE_CODE_ENABLE_FUNCTION_HOOKS` | early-access flag | **nothing**, by design — see below |
 

hooks/judge.ts

@@ -155,6 +155,24 @@ export function judgePrompt(messages: readonly SessionMessage[], step: Step, mem
  * the emphasis a model sometimes puts round it. Anything else, including no
  * answer, is not.
  */
+export type Reply = { text: string } | { reason: string }
+
+/**
+ * The judge's reply, whichever shape `$.model.complete` resolved: the reply's
+ * text up to Claude Code 2.1.278, `{ isAnswered, text }` or
+ * `{ isAnswered: false, reason }` from 2.1.280. Either is read here, where the
+ * engine's answer enters, and nothing past this point knows there were two.
+ */
+export function readReply(answer: unknown): Reply {
+  if (typeof answer === 'string') return { text: answer }
+  if (typeof answer === 'object' && answer !== null && 'isAnswered' in answer) {
+    const { isAnswered, text, reason } = answer as { isAnswered: unknown; text?: unknown; reason?: unknown }
+    if (isAnswered === true && typeof text === 'string') return { text }
+    if (isAnswered === false) return { reason: String(reason) }
+  }
+  return { reason: 'a reply of no shape taskcut knows' }
+}
+
 export function saysDone(answer: string): boolean {
   const lines = answer.trim().split('\n')
   const last = lines[lines.length - 1] ?? ''

hooks/register.ts

@@ -27,7 +27,7 @@ import type { EngineInterface, Register } from 'claude-code'
 
 import { ENV_VAR, INERT, decideActivation, type Activation } from './activation'
 import { readConfig, type Config } from './config'
-import { ANSWER_TOKENS, JUDGE_SYSTEM, judgePrompt, saysDone, type Step } from './judge'
+import { ANSWER_TOKENS, JUDGE_SYSTEM, judgePrompt, readReply, saysDone, type Step } from './judge'
 
 /**
  * Resolved once, at `session.start`. It starts inert so that a session in which
@@ -105,8 +105,15 @@ async function memoryFiles($: EngineInterface): Promise<string[]> {
 async function judge($: EngineInterface, step: Step, config: Config): Promise<boolean> {
   try {
     const prompt = judgePrompt(await $.session.messages(), step, await memoryFiles($))
-    const answer = await $.model.complete({ model: config.model, system: JUDGE_SYSTEM, prompt, maxTokens: ANSWER_TOKENS })
-    return saysDone(answer)
+    // `unknown`: what the call resolves to has changed between releases, and
+    // readReply takes every shape it has had.
+    const answer: unknown = await $.model.complete({ model: config.model, system: JUDGE_SYSTEM, prompt, maxTokens: ANSWER_TOKENS })
+    const reply = readReply(answer)
+    if ('reason' in reply) {
+      $.ui.log(`could not judge the step (${reply.reason})`, { to: 'debug' })
+      return false
+    }
+    return saysDone(reply.text)
   } catch (error) {
     $.ui.log(`could not judge the step (${String(error)})`, { to: 'debug' })
     return false

test/judge.test.ts

@@ -12,6 +12,7 @@ import {
   beforeStep,
   conversationLines,
   judgePrompt,
+  readReply,
   saysDone,
 } from '../hooks/judge.ts'
 
@@ -126,6 +127,24 @@ describe('judgePrompt', () => {
   })
 })
 
+describe('readReply', () => {
+  test('reads the text $.model.complete resolved up to 2.1.278', () => {
+    assert.deepEqual(readReply('It reports task 2 complete.\nDONE'), { text: 'It reports task 2 complete.\nDONE' })
+  })
+
+  test('reads the result it resolves from 2.1.280', () => {
+    const usage = { input_tokens: 1, output_tokens: 1, cache_read_input_tokens: 0, cache_creation_input_tokens: 0 }
+    assert.deepEqual(readReply({ isAnswered: true, text: 'WORKING', usage }), { text: 'WORKING' })
+    assert.deepEqual(readReply({ isAnswered: false, reason: 'api-error', status: 529, error: 'overloaded', usage }), { reason: 'api-error' })
+  })
+
+  test('anything else is no reply, never a verdict', () => {
+    for (const odd of [undefined, null, 42, {}, { isAnswered: true }, { text: 'DONE' }]) {
+      assert.ok('reason' in readReply(odd), JSON.stringify(odd))
+    }
+  })
+})
+
 describe('saysDone', () => {
   test('reads the verdict off the last line', () => {
     assert.equal(saysDone('It reports task 2 complete and starts task 3.\nDONE'), true)