@zhixuan92/multi-model-agent-core 0.1.1 → 0.3.0

package/dist/runners/codex-runner.js

@@ -1,9 +1,16 @@
 import OpenAI from 'openai';
 import { z } from 'zod';
+import { createHash } from 'node:crypto';
 import { getCodexAuth } from '../auth/codex-oauth.js';
-import { withTimeout } from '../types.js';
+import { withTimeout, computeCostUSD, computeSavedCostUSD, } from '../types.js';
 import { FileTracker } from '../tools/tracker.js';
 import { createToolImplementations } from '../tools/definitions.js';
+import { TextScratchpad } from '../tools/scratchpad.js';
+import { buildSystemPrompt, buildBudgetHint, buildReGroundingMessage, buildBudgetPressureNudge, RE_GROUNDING_INTERVAL_TURNS, } from './prevention.js';
+import { validateCompletion, validateCoverage, buildRePrompt, sameDegenerateOutput, resolveInputTokenSoftLimit, checkWatchdogThreshold, logWatchdogEvent, trimProgressTrace, } from './supervision.js';
+import { injectionTypeFor } from './injection-type.js';
+import { classifyError } from './error-classification.js';
+import { findModelProfile } from '../routing/model-profiles.js';
 // CODEX_DEBUG=1 causes the runner to log raw HTTP request/response bodies to
 // stderr. Those bodies routinely include the user's prompt, file contents,
 // tool arguments, and other sensitive data — fine for local debugging,
@@ -16,6 +23,11 @@ if (process.env.CODEX_DEBUG === '1') {
         'bodies (including prompts and file contents) will be logged to stderr. ' +
         'Disable in any environment where logs may be retained or shared.');
 }
+/**
+ * Hard cap on supervision re-prompts before we give up and salvage. Three is
+ * the value chosen in the spec (A.2.2); mirrors openai-runner and claude-runner.
+ */
+const MAX_SUPERVISION_RETRIES = 3;
 export function createCodexClient(capture) {
     const debug = process.env.CODEX_DEBUG === '1';
     // A custom fetch that tees error-response bodies into `capture`.
@@ -146,7 +158,37 @@ export async function runCodex(prompt, options, providerConfig, defaults) {
     const sandboxPolicy = options.sandboxPolicy ?? providerConfig.sandboxPolicy ?? 'cwd-only';
     const effort = options.effort ?? providerConfig.effort;
     const abortController = new AbortController();
-    const tracker = new FileTracker();
+    // --- Progress event emission (Task 11) ----------------------------------
+    //
+    // `onProgress` is already wrapped in `safeSink` by the orchestrator
+    // (Task 8), so any throw from the consumer callback is swallowed
+    // upstream and cannot corrupt this loop. We do not need to wrap it
+    // again here.
+    const onProgress = options.onProgress;
+    const shouldCaptureTrace = options.includeProgressTrace ?? false;
+    const traceBuffer = [];
+    const emit = (event) => {
+        if (shouldCaptureTrace)
+            traceBuffer.push(event);
+        if (onProgress)
+            onProgress(event);
+    };
+    // Accumulated state (hoisted so the timeout callback can read partial
+    // progress, AND so the FileTracker callback closure — constructed below
+    // — can read the running turn count at firing time).
+    //
+    // Turn attribution for tool calls: in codex-runner, tool calls fire in
+    // the tool-execution loop AFTER the model's stream for that turn has
+    // completed but BEFORE the next iteration of `while` starts. The `turns`
+    // variable already reflects the current turn at that point (it was
+    // incremented at the top of the iteration), so the callback can read it
+    // directly — no +1 offset.
+    let inputTokens = 0;
+    let outputTokens = 0;
+    let turns = 0;
+    const tracker = new FileTracker((summary) => {
+        emit({ kind: 'tool_call', turn: turns, toolSummary: summary });
+    });
     const toolImpls = createToolImplementations(tracker, cwd, sandboxPolicy, abortController.signal);
     const codexTools = toolMode === 'full' ? buildCodexTools(toolImpls, sandboxPolicy) : [];
     const toolsByName = new Map(codexTools.map(t => [t.name, t]));
@@ -167,29 +209,111 @@ export async function runCodex(prompt, options, providerConfig, defaults) {
         ? configuredHostedTools.map(t => ({ type: t }))
         : [];
     const allTools = [...responsesTools, ...hostedTools];
-    // Accumulated state (hoisted so the timeout callback can read partial progress)
-    let inputTokens = 0;
-    let outputTokens = 0;
-    let turns = 0;
+    // --- Prevention layer: system prompt + budget hint ---
+    //
+    // buildSystemPrompt() is deliberately static and parameter-free (same
+    // decision as openai-runner and claude-runner: Task 1 review rejected
+    // provider/maxTurns options). The budget hint is prepended to the user
+    // prompt so the model sees it as part of its task brief, while the system
+    // prompt is threaded through the Responses API `instructions` field.
+    const systemPrompt = buildSystemPrompt();
+    const budgetHint = buildBudgetHint({ maxTurns });
+    const promptWithBudgetHint = `${budgetHint}\n\n${prompt}`;
+    // --- onInitialRequest (Task 12) ----------------------------------------
+    //
+    // Fire once per attempt with the canonical orchestrator-side initial
+    // brief: `${systemPrompt}\n\n${promptWithBudgetHint}`. This is NOT the
+    // literal request body the OpenAI Responses API transmits — codex
+    // sends the systemPrompt via the Responses API `instructions` field
+    // and the user prompt as a structured `input` message array. We hash
+    // the canonical form instead so the hash is cross-runner stable:
+    // openai-runner and claude-runner compute the same hash from the same
+    // canonical string. See `AttemptRecord.initialPromptHash` in types.ts
+    // for the full wire-level caveat.
+    if (options.onInitialRequest) {
+        const canonicalInitialBrief = `${systemPrompt}\n\n${promptWithBudgetHint}`;
+        try {
+            options.onInitialRequest({
+                lengthChars: canonicalInitialBrief.length,
+                sha256: createHash('sha256').update(canonicalInitialBrief).digest('hex'),
+            });
+        }
+        catch {
+            // Swallow — a broken callback must not affect dispatch.
+        }
+    }
+    // --- Scratchpad: buffers every text emission the codex backend streams
+    // through our loop. Every termination path (ok / incomplete / max_turns /
+    // error / timeout / force_salvage) salvages `scratchpad.latest()` when
+    // the final message is empty or degenerate. ---
+    const scratchpad = new TextScratchpad();
+    // --- Watchdog: resolve the input-token soft limit once per run ---
+    const profile = findModelProfile(providerConfig.model);
+    const softLimit = resolveInputTokenSoftLimit(providerConfig, profile);
+    // --- Task timing + parent model (Task 9) --------------------------------
+    const taskStartMs = Date.now();
+    const parentModel = options.parentModel;
     const run = async () => {
         const capture = {};
         const client = createCodexClient(capture);
         const input = [
             // eslint-disable-next-line @typescript-eslint/no-explicit-any
-            { role: 'user', content: prompt },
+            { role: 'user', content: promptWithBudgetHint },
         ];
         let output = '';
+        // --- Abort-path investigation (plan Step 2) ---------------------------
+        //
+        // The 2026-04-10 Fate dispatch captured an error "Request was aborted |
+        // last response status: completed". The "completed" suffix was
+        // misleading: it was captured from a PREVIOUS successful turn, not the
+        // failed one. Mechanism:
+        //
+        //   1. Turn N's stream emits `response.completed` with status
+        //      `'completed'`. We update `lastResponseStatus = 'completed'`.
+        //   2. Turn N+1 starts; `client.responses.create(...)` opens a new
+        //      stream, but the abort signal fires before any
+        //      `response.completed` event is received.
+        //   3. The thrown error is caught below. The catch branch reads
+        //      `lastResponseStatus` — which is STILL `'completed'` from turn N
+        //      — and appends it as "last response status: completed", making
+        //      the error look like it originated from a successful response.
+        //
+        // Fix: track which turn the status was captured on. If the status was
+        // NOT captured on the current (failed) turn, drop the suffix. That way
+        // we never emit a status that belongs to a different, already-
+        // concluded request. Users saw the misleading suffix and wasted time
+        // debugging a phantom "the request completed but was aborted" condition
+        // that doesn't exist.
         let lastResponseStatus = null;
+        let lastResponseStatusTurn = null;
+        // --- Supervision / watchdog bookkeeping ---
+        let supervisionRetries = 0;
+        // Initialised to `null` (NOT ''): on the first turn there is no
+        // previous degenerate output to compare against, so the same-output
+        // early-out must be skipped. Initialising to '' would cause
+        // sameDegenerateOutput('', '') to fire on a first-turn empty output
+        // and break the loop before any retries run. See openai-runner
+        // regression #5.
+        let lastDegenerateOutput = null;
+        // High-watermark guard for the watchdog warning nudge — fire at most
+        // once per distinct input-token level. Mirrors openai-runner and
+        // claude-runner.
+        let lastWarnedInputTokens = -1;
         try {
             while (turns < maxTurns) {
                 turns++;
+                // Emit turn_start AFTER incrementing so `turn` matches the 1-indexed
+                // turn number we use everywhere else in this runner (the scratchpad
+                // append, watchdog logs, error diagnostics, result.turns).
+                emit({ kind: 'turn_start', turn: turns, provider: 'codex' });
                 // Codex backend requires streaming. The Codex backend's
                 // `response.completed` event does NOT populate `response.output` —
                 // we must accumulate content from individual stream events.
-                // `instructions` is required (mirrors gumi-agent's proven shape).
+                // `instructions` carries the prevention-layer system prompt; the
+                // per-run budget hint is already prepended to the first user input.
                 const stream = await client.responses.create({
                     model: providerConfig.model,
-                    instructions: prompt,
+                    instructions: systemPrompt,
                     input,
                     stream: true,
                     store: false,
@@ -241,8 +365,10 @@ export async function runCodex(prompt, options, providerConfig, defaults) {
                             inputTokens += r.usage.input_tokens ?? 0;
                             outputTokens += r.usage.output_tokens ?? 0;
                         }
-                        if (r?.status)
+                        if (r?.status) {
                             lastResponseStatus = r.status;
+                            lastResponseStatusTurn = turns;
+                        }
                     }
                 }
                 if (process.env.CODEX_DEBUG === '1') {
@@ -256,6 +382,20 @@ export async function runCodex(prompt, options, providerConfig, defaults) {
                 if (!sawCompleted) {
                     throw new Error('Codex stream ended without a response.completed event');
                 }
+                // Buffer this turn's text into the scratchpad BEFORE any exit so
+                // every termination path (including supervision exhaustion and
+                // force_salvage) can salvage it. Codex does not emit <think> tags
+                // by default, so there is no stripping step here.
+                if (textThisTurn) {
+                    scratchpad.append(turns, textThisTurn);
+                    emit({
+                        kind: 'text_emission',
+                        turn: turns,
+                        chars: textThisTurn.length,
+                        preview: textThisTurn.slice(0, 200),
+                    });
+                    output = textThisTurn;
+                }
                 // Replay only function_call items into the next turn's input.
                 //
                 // We send `store: false` to the Responses API, which means the server
@@ -287,26 +427,175 @@ export async function runCodex(prompt, options, providerConfig, defaults) {
                         });
                     }
                 }
-                if (textThisTurn) {
-                    output = textThisTurn;
+                // --- Watchdog checks after tokens are updated -------------------
+                const watchdogStatus = checkWatchdogThreshold(inputTokens, softLimit);
+                if (watchdogStatus !== 'ok') {
+                    logWatchdogEvent(watchdogStatus, {
+                        provider: 'codex',
+                        model: providerConfig.model,
+                        turn: turns,
+                        inputTokens,
+                        softLimit,
+                        scratchpadChars: scratchpad.toString().length,
+                    });
                 }
-                else if (toolCalls.length === 0) {
-                    output = `[codex returned no text — items streamed: ${itemTypesSeen.join(', ') || '(none)'}]`;
+                if (watchdogStatus === 'force_salvage') {
+                    // `watchdog_force_salvage` is not an injected message — no
+                    // re-prompt is sent — but observers still want to see exactly
+                    // why the run is being killed. Emit with contentLengthChars: 0
+                    // to reflect the "nothing was injected, we just terminated"
+                    // semantics (mirrors openai-runner and claude-runner).
+                    emit({
+                        kind: 'injection',
+                        injectionType: 'watchdog_force_salvage',
+                        turn: turns,
+                        contentLengthChars: 0,
+                    });
+                    const salvaged = buildCodexForceSalvageResult({
+                        tracker,
+                        scratchpad,
+                        providerConfig,
+                        inputTokens,
+                        outputTokens,
+                        turns,
+                        softLimit,
+                        durationMs: Date.now() - taskStartMs,
+                        parentModel,
+                        traceBuffer: shouldCaptureTrace ? traceBuffer : undefined,
+                    });
+                    emit({ kind: 'done', status: salvaged.status });
+                    return salvaged;
+                }
+                // Warning-band nudge: fire at most once per distinct input-token
+                // high-watermark. Pushed as a user message so the next turn of
+                // the codex loop addresses the budget-pressure prompt. We use
+                // the shared prevention helper (NOT an inline string) so every
+                // runner emits byte-identical wording.
+                if (watchdogStatus === 'warning' && inputTokens > lastWarnedInputTokens) {
+                    lastWarnedInputTokens = inputTokens;
+                    const warning = buildBudgetPressureNudge({ inputTokens, softLimit });
+                    input.push({
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                        role: 'user',
+                        content: warning,
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                    });
+                    emit({
+                        kind: 'injection',
+                        injectionType: 'watchdog_warning',
+                        turn: turns,
+                        contentLengthChars: warning.length,
+                    });
                 }
-                // If the model made no tool calls, it's done
+                // --- Periodic re-grounding inside the loop ---------------------
+                if (turns > 0 && turns % RE_GROUNDING_INTERVAL_TURNS === 0) {
+                    const reground = buildReGroundingMessage({
+                        originalPromptExcerpt: prompt,
+                        currentTurn: turns,
+                        maxTurns,
+                        toolCallsSoFar: tracker.getToolCalls().length,
+                        filesReadSoFar: tracker.getReads().length,
+                    });
+                    input.push({
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                        role: 'user',
+                        content: reground,
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                    });
+                    emit({
+                        kind: 'injection',
+                        injectionType: 'reground',
+                        turn: turns,
+                        contentLengthChars: reground.length,
+                    });
+                }
+                // --- turn_complete: one event per while-iteration. Fires after the
+                // watchdog + re-grounding checks have run (so cumulative token
+                // counts and any injection events are already on the wire) and
+                // BEFORE the supervision branching / tool-execution loop. Every
+                // continue/return in the branches below happens AFTER this event,
+                // so the sequence "turn_start ... text_emission ... turn_complete"
+                // is guaranteed per iteration.
+                emit({
+                    kind: 'turn_complete',
+                    turn: turns,
+                    cumulativeInputTokens: inputTokens,
+                    cumulativeOutputTokens: outputTokens,
+                });
+                // If the model made no tool calls, the turn ended with either a
+                // final answer or a degenerate emission. Wrap in the supervision
+                // state machine: valid text is an immediate ok-exit; degenerate
+                // either re-prompts (and continues the loop) or — if the retry
+                // budget is spent / same-output early-out fires — exits as
+                // incomplete with scratchpad salvage.
                 if (toolCalls.length === 0) {
-                    return {
-                        output,
-                        status: 'ok',
-                        usage: {
+                    const stripped = textThisTurn; // codex does not emit <think> tags
+                    const validation = validateCompletion(stripped);
+                    // NEW: coverage check — only when syntactic validation passes
+                    if (validation.valid && options.expectedCoverage) {
+                        const coverageValidation = validateCoverage(stripped, options.expectedCoverage);
+                        if (!coverageValidation.valid) {
+                            validation.valid = false;
+                            validation.kind = coverageValidation.kind;
+                            validation.reason = coverageValidation.reason;
+                        }
+                    }
+                    if (validation.valid) {
+                        const ok = buildCodexOkResult({
+                            tracker,
+                            scratchpad,
+                            providerConfig,
                             inputTokens,
                             outputTokens,
-                            totalTokens: inputTokens + outputTokens,
-                            costUSD: null,
-                        },
-                        turns,
-                        files: tracker.getFiles(),
-                    };
+                            turns,
+                            output: stripped,
+                            durationMs: Date.now() - taskStartMs,
+                            parentModel,
+                            traceBuffer: shouldCaptureTrace ? traceBuffer : undefined,
+                        });
+                        emit({ kind: 'done', status: ok.status });
+                        return ok;
+                    }
+                    // Same-output early-out: only compare when we have a previous
+                    // degenerate output. First-turn degeneracy must still get
+                    // retries — see openai-runner regression #5.
+                    if ((lastDegenerateOutput !== null &&
+                        sameDegenerateOutput(stripped, lastDegenerateOutput)) ||
+                        supervisionRetries >= MAX_SUPERVISION_RETRIES) {
+                        const exhausted = buildCodexIncompleteResult({
+                            tracker,
+                            scratchpad,
+                            providerConfig,
+                            inputTokens,
+                            outputTokens,
+                            turns,
+                            reason: `supervision loop exhausted after ${supervisionRetries} re-prompts (last kind: ${validation.kind ?? 'unknown'})`,
+                            durationMs: Date.now() - taskStartMs,
+                            parentModel,
+                            traceBuffer: shouldCaptureTrace ? traceBuffer : undefined,
+                        });
+                        emit({ kind: 'done', status: exhausted.status });
+                        return exhausted;
+                    }
+                    // Inject the re-prompt as the next user input and continue
+                    // the loop. The next turn of the codex backend will respond
+                    // to the re-prompt directly.
+                    lastDegenerateOutput = stripped;
+                    supervisionRetries++;
+                    const rePrompt = buildRePrompt(validation);
+                    input.push({
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                        role: 'user',
+                        content: rePrompt,
+                        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                    });
+                    emit({
+                        kind: 'injection',
+                        injectionType: injectionTypeFor(validation.kind),
+                        turn: turns,
+                        contentLengthChars: rePrompt.length,
+                    });
+                    continue;
                 }
                 // Execute tool calls and feed outputs back
                 for (const call of toolCalls) {
@@ -331,19 +620,23 @@ export async function runCodex(prompt, options, providerConfig, defaults) {
                     });
                 }
             }
-            // Max turns exhausted
-            return {
-                output: output || `Agent exceeded max turns (${maxTurns}).`,
-                status: 'max_turns',
-                usage: {
-                    inputTokens,
-                    outputTokens,
-                    totalTokens: inputTokens + outputTokens,
-                    costUSD: null,
-                },
+            // Max turns exhausted — salvage any buffered text.
+            const maxTurnsResult = buildCodexMaxTurnsResult({
+                tracker,
+                scratchpad,
+                providerConfig,
+                inputTokens,
+                outputTokens,
                 turns,
-                files: tracker.getFiles(),
-            };
+                maxTurns,
+                lastOutput: output,
+                reason: `hand-rolled loop exited after completing ${turns} of ${maxTurns} user-declared turns without producing a clean final answer`,
+                durationMs: Date.now() - taskStartMs,
+                parentModel,
+                traceBuffer: shouldCaptureTrace ? traceBuffer : undefined,
+            });
+            emit({ kind: 'done', status: maxTurnsResult.status });
+            return maxTurnsResult;
         }
         catch (err) {
             // OpenAI SDK's APIError carries status/body/headers — surface them
@@ -374,30 +667,230 @@ export async function runCodex(prompt, options, providerConfig, defaults) {
             }
             if (e?.requestID)
                 pieces.push(`req_id=${e.requestID}`);
-            if (lastResponseStatus)
+            // Only include `last response status` when it was captured on the
+            // CURRENT (failing) turn — otherwise it belongs to a previous,
+            // separate request and appending it is actively misleading. See the
+            // abort-path investigation comment at the top of `run()`.
+            if (lastResponseStatus && lastResponseStatusTurn === turns) {
                 pieces.push(`last response status: ${lastResponseStatus}`);
+            }
+            else if (lastResponseStatus && lastResponseStatusTurn !== turns) {
+                pieces.push(`note: a previous request (turn ${lastResponseStatusTurn}) completed with status ` +
+                    `"${lastResponseStatus}" — it is unrelated to this failure`);
+            }
             const detailed = pieces.join(' | ') || String(err);
+            // Classify the thrown error into a finer-grained RunStatus. Task 7
+            // introduces api_aborted / api_error / network_error alongside the
+            // catch-all 'error' status. The turn-scoped `lastResponseStatus`
+            // disambiguation above is ORTHOGONAL to this classification: the
+            // `detailed` message is still the rich operator-facing diagnostic,
+            // and `classifyError` only decides which RunStatus bucket the
+            // failure lands in.
+            const { status } = classifyError(err);
+            // Salvage: if the scratchpad has buffered text from earlier turns,
+            // return it as the output. Pre-Task-5 behavior returned only the
+            // error string, losing 30k+ tokens of work on abort.
+            emit({ kind: 'done', status });
+            const hasSalvage = !scratchpad.isEmpty();
+            const costUSD = computeCostUSD(inputTokens, outputTokens, providerConfig);
+            const savedCostUSD = computeSavedCostUSD(costUSD, inputTokens, outputTokens, parentModel);
             return {
-                output: `Sub-agent error: ${detailed}`,
-                status: 'error',
+                output: hasSalvage ? scratchpad.latest() : `Sub-agent error: ${detailed}`,
+                status,
                 usage: {
                     inputTokens,
                     outputTokens,
                     totalTokens: inputTokens + outputTokens,
-                    costUSD: null,
+                    costUSD,
+                    savedCostUSD,
                 },
                 turns,
-                files: tracker.getFiles(),
+                filesRead: tracker.getReads(),
+                directoriesListed: tracker.getDirectoriesListed(),
+                filesWritten: tracker.getWrites(),
+                toolCalls: tracker.getToolCalls(),
+                outputIsDiagnostic: !hasSalvage,
+                escalationLog: [],
                 error: detailed,
+                durationMs: Date.now() - taskStartMs,
+                ...(shouldCaptureTrace && { progressTrace: trimProgressTrace(traceBuffer) }),
             };
         }
     };
-    return withTimeout(run(), timeoutMs, () => ({
-        output: `Agent timed out after ${timeoutMs}ms.`,
-        status: 'timeout',
-        files: tracker.getFiles(),
-        usage: { inputTokens, outputTokens, totalTokens: inputTokens + outputTokens, costUSD: null },
+    return withTimeout(run(), timeoutMs, () => {
+        emit({ kind: 'done', status: 'timeout' });
+        const hasSalvage = !scratchpad.isEmpty();
+        const costUSD = computeCostUSD(inputTokens, outputTokens, providerConfig);
+        const savedCostUSD = computeSavedCostUSD(costUSD, inputTokens, outputTokens, parentModel);
+        return {
+            // Preserve any text the scratchpad buffered before the timeout fired.
+            // Partial usage is read from the running accumulators hoisted above —
+            // hardcoded zeros would discard every token counted on partial turns.
+            output: hasSalvage ? scratchpad.latest() : `Agent timed out after ${timeoutMs}ms.`,
+            status: 'timeout',
+            filesRead: tracker.getReads(),
+            directoriesListed: tracker.getDirectoriesListed(),
+            filesWritten: tracker.getWrites(),
+            toolCalls: tracker.getToolCalls(),
+            usage: {
+                inputTokens,
+                outputTokens,
+                totalTokens: inputTokens + outputTokens,
+                costUSD,
+                savedCostUSD,
+            },
+            turns,
+            outputIsDiagnostic: !hasSalvage,
+            escalationLog: [],
+            durationMs: Date.now() - taskStartMs,
+            ...(shouldCaptureTrace && { progressTrace: trimProgressTrace(traceBuffer) }),
+        };
+    }, abortController);
+}
+function buildCodexOkResult(args) {
+    const { tracker, providerConfig, inputTokens, outputTokens, turns, output, durationMs, parentModel, traceBuffer } = args;
+    const costUSD = computeCostUSD(inputTokens, outputTokens, providerConfig);
+    const savedCostUSD = computeSavedCostUSD(costUSD, inputTokens, outputTokens, parentModel);
+    return {
+        output,
+        status: 'ok',
+        usage: {
+            inputTokens,
+            outputTokens,
+            totalTokens: inputTokens + outputTokens,
+            costUSD,
+            savedCostUSD,
+        },
+        turns,
+        filesRead: tracker.getReads(),
+        directoriesListed: tracker.getDirectoriesListed(),
+        filesWritten: tracker.getWrites(),
+        toolCalls: tracker.getToolCalls(),
+        // `ok` always carries a real model answer — never a diagnostic.
+        outputIsDiagnostic: false,
+        escalationLog: [],
+        durationMs,
+        ...(traceBuffer && { progressTrace: trimProgressTrace(traceBuffer) }),
+    };
+}
+/**
+ * Supervision-exhausted path: retry cap hit or same-output early-out. Prefer
+ * scratchpad salvage; fall back to the incomplete diagnostic.
+ */
+function buildCodexIncompleteResult(args) {
+    const { tracker, scratchpad, providerConfig, inputTokens, outputTokens, turns, reason, durationMs, parentModel, traceBuffer } = args;
+    const filesRead = tracker.getReads();
+    const filesWritten = tracker.getWrites();
+    const costUSD = computeCostUSD(inputTokens, outputTokens, providerConfig);
+    const savedCostUSD = computeSavedCostUSD(costUSD, inputTokens, outputTokens, parentModel);
+    const hasSalvage = !scratchpad.isEmpty();
+    return {
+        output: hasSalvage
+            ? scratchpad.latest()
+            : buildCodexIncompleteDiagnostic({
+                turns,
+                inputTokens,
+                outputTokens,
+                filesRead,
+                filesWritten,
+            }),
+        status: 'incomplete',
+        usage: {
+            inputTokens,
+            outputTokens,
+            totalTokens: inputTokens + outputTokens,
+            costUSD,
+            savedCostUSD,
+        },
         turns,
-    }), abortController);
+        filesRead,
+        directoriesListed: tracker.getDirectoriesListed(),
+        filesWritten,
+        toolCalls: tracker.getToolCalls(),
+        outputIsDiagnostic: !hasSalvage,
+        escalationLog: [],
+        error: reason,
+        durationMs,
+        ...(traceBuffer && { progressTrace: trimProgressTrace(traceBuffer) }),
+    };
+}
+function buildCodexForceSalvageResult(args) {
+    const { tracker, scratchpad, providerConfig, inputTokens, outputTokens, turns, softLimit, durationMs, parentModel, traceBuffer } = args;
+    const costUSD = computeCostUSD(inputTokens, outputTokens, providerConfig);
+    const savedCostUSD = computeSavedCostUSD(costUSD, inputTokens, outputTokens, parentModel);
+    const hasSalvage = !scratchpad.isEmpty();
+    return {
+        output: hasSalvage
+            ? scratchpad.latest()
+            : `[codex sub-agent forcibly terminated at ${inputTokens} input tokens (soft limit ${softLimit}). No usable text was buffered.]`,
+        status: 'incomplete',
+        usage: {
+            inputTokens,
+            outputTokens,
+            totalTokens: inputTokens + outputTokens,
+            costUSD,
+            savedCostUSD,
+        },
+        turns,
+        filesRead: tracker.getReads(),
+        directoriesListed: tracker.getDirectoriesListed(),
+        filesWritten: tracker.getWrites(),
+        toolCalls: tracker.getToolCalls(),
+        outputIsDiagnostic: !hasSalvage,
+        escalationLog: [],
+        durationMs,
+        ...(traceBuffer && { progressTrace: trimProgressTrace(traceBuffer) }),
+    };
+}
+function buildCodexMaxTurnsResult(args) {
+    const { tracker, scratchpad, providerConfig, inputTokens, outputTokens, turns, maxTurns, lastOutput, reason, durationMs, parentModel, traceBuffer } = args;
+    const hasSalvage = !scratchpad.isEmpty();
+    // Note: `lastOutput` here is the model's final text for the max-turns
+    // boundary — real model content, not a diagnostic template. Only the
+    // `Agent exceeded max turns…` fallback (empty scratchpad AND empty
+    // lastOutput) is a diagnostic.
+    const output = hasSalvage
+        ? scratchpad.latest()
+        : (lastOutput || `Agent exceeded max turns (${maxTurns}).`);
+    const outputIsDiagnostic = !hasSalvage && !lastOutput;
+    const costUSD = computeCostUSD(inputTokens, outputTokens, providerConfig);
+    const savedCostUSD = computeSavedCostUSD(costUSD, inputTokens, outputTokens, parentModel);
+    return {
+        output,
+        status: 'max_turns',
+        usage: {
+            inputTokens,
+            outputTokens,
+            totalTokens: inputTokens + outputTokens,
+            costUSD,
+            savedCostUSD,
+        },
+        turns,
+        filesRead: tracker.getReads(),
+        directoriesListed: tracker.getDirectoriesListed(),
+        filesWritten: tracker.getWrites(),
+        toolCalls: tracker.getToolCalls(),
+        outputIsDiagnostic,
+        escalationLog: [],
+        error: reason,
+        durationMs,
+        ...(traceBuffer && { progressTrace: trimProgressTrace(traceBuffer) }),
+    };
+}
+function buildCodexIncompleteDiagnostic(opts) {
+    return [
+        '[codex sub-agent terminated without producing a final answer]',
+        '',
+        'The model emitted no tool calls and no usable text on its final turn, and',
+        'supervision re-prompts did not recover a valid response.',
+        '',
+        `Turns used:    ${opts.turns}`,
+        `Input tokens:  ${opts.inputTokens}`,
+        `Output tokens: ${opts.outputTokens}`,
+        `Files read:    ${opts.filesRead.length}`,
+        `Files written: ${opts.filesWritten.length}`,
+        '',
+        'Recommended action: re-dispatch with a tighter brief, or escalate provider tier.',
+    ].join('\n');
 }
 //# sourceMappingURL=codex-runner.js.map