@lazyingart/agintiflow 0.20.211 → 0.20.213

package/docs/supervision-campaign-ledger.md

@@ -75,3 +75,58 @@ The incident established two general contracts:
   canvas/download filename derived from the task title and source purpose.
   Internal collision identifiers are short suffixes, never the leading or only
   visible filename information.
+
+### QA repair continuity and evidence intent
+
+`qa-incident-metrics-001` passed after two reusable runtime fixes. A normal,
+underspecified QA prompt led the DeepSeek-backed agent to reproduce and diagnose
+compound-duration parsing, percentile interpolation/mutation, and deterministic
+summary-order defects. The first partial patch advanced the mutation revision,
+but the runtime then forgot the retained failing test and prematurely reduced
+the tool surface to test-only mode. Source-next restored the failed-test repair
+state until a fresh current-revision test passed, allowing the same durable
+session to finish the coherent patch, add regressions, run 15 tests at 100%
+statement coverage, clean debris, and commit `667891f`.
+
+Independent `pytest` and the hidden `qa_incident_metrics_contract.py` checker
+both passed. The run also exposed an evidence-intent false positive: `figure
+out` and `clean up generated test debris` were interpreted as a request for a
+canvas artifact. Evidence inference now excludes those non-production phrases
+while retaining the artifact gate for real generated figures. Exact session
+evidence remains in
+`~/.agintiflow/sessions/aginti-qa-incident-metrics-001/events.jsonl` and the
+machine ledger records the run as `passed_after_fix`.
+
+### GitHub maintenance hidden acceptance and goal-scoped evidence
+
+`github-safe-maintenance-012` started from a normal maintenance prompt rather
+than a checker-shaped instruction. The agent repaired and verified the target,
+but its first completion omitted the exact `docs/maintenance-handoff.md` file
+required by independent acceptance. The same retained session inspected the
+external failure, created the missing handoff, committed and pushed target
+commit `36fa6c0`, and left `main` clean and synchronized. The hidden
+`github_maintenance_contract.py` checker then passed.
+
+Supervision of the repair exposed four runtime defects that could affect other
+profiles:
+
+- A test command wrapped as `command; echo "EXIT:$?"` could have shell exit zero
+  even when the real command failed. Explicit final `EXIT`, `STATUS`, or
+  `RESULT` probes are now parsed; a missing or nonzero marker is failing
+  evidence.
+- A genuinely new continuation could inherit completed artifact, SCS,
+  project-verification, and repair state from the prior goal. New goals now
+  clear only goal-scoped execution evidence while preserving conversation,
+  durable goal history, and goal-keyed research memory.
+- An acceptance sentence listing screenshots, PDFs, reports, or app launches
+  "as appropriate" could force irrelevant visual work. Optional evidence
+  examples no longer become mandatory categories.
+- Merely naming a read-only checker such as `contract.py` could force a file
+  artifact. File evidence now requires both mutation intent and a workspace
+  file/source target; a virtual canvas artifact remains artifact evidence only.
+
+The patched source resumed `aginti-github-maintenance-001`, invoked the hidden
+checker exactly once, performed no file, canvas, commit, or push side effect,
+and completed with a clean repository-state check. The full npm suite and the
+focused dynamic-budget, SCS/model-role, and web-canvas regressions pass for
+AgInTiFlow `0.20.213`.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lazyingart/agintiflow",
-  "version": "0.20.211",
+  "version": "0.20.213",
   "type": "module",
   "description": "AgInTiFlow is a project-aware agent workspace for hybrid wet-dry R&D, hardware-aware intelligence, software automation, and industrial workflows.",
   "license": "Apache-2.0",

package/scripts/smoke-cli-chat.js

@@ -74,6 +74,12 @@ function cellWidth(value = "") {
   return [...String(value || "").replace(/\x1b\[[0-9;?]*[ -/]*[@-~]/g, "")].reduce((sum, char) => sum + charCellWidth(char), 0);
 }
 
+function stripTerminalControls(value = "") {
+  return String(value || "")
+    .replace(/\x1b\][\s\S]*?(?:\x07|\x1b\\)/g, "")
+    .replace(/\x1b\[[0-9;?]*[ -/]*[@-~]/g, "");
+}
+
 function runChat(inputText) {
   return runCli(
     ["chat", "--provider", "mock", "--routing", "manual", "--profile", "code", "--headless", "--sandbox-mode", "host"],
@@ -596,7 +602,7 @@ try {
     animated: false,
     webAppUrl: "http://127.0.0.1:3210",
   }).join("\n");
-  if (!launchHeaderWithWeb.includes("webapp: http://127.0.0.1:3210")) {
+  if (!stripTerminalControls(launchHeaderWithWeb).includes("webapp: http://127.0.0.1:3210")) {
     throw new Error("launch header did not render the active webapp URL in the tagline row");
   }
   const launchHeaderWithWebError = buildLaunchHeaderLines({
@@ -749,17 +755,21 @@ try {
     throw new Error("terminal prompt layout did not render live input queue and cwd footer");
   }
   const hintPromptLayout = buildPromptLayout("/mo", 3, 90, 24, { suggestions: ["/models", "/model"], suggestionIndex: 1 });
-  const hintText = hintPromptLayout.renderedRows
-    .map((line) => line.replace(/\x1b\[[0-9;?]*[ -/]*[@-~]/g, ""))
-    .join("\n");
-  if (!hintText.includes("  user> /mo") || !hintText.includes("hint    /models  >/model")) {
-    throw new Error("terminal prompt layout did not align user and hint text columns");
+  const hintRaw = hintPromptLayout.renderedRows.join("\n");
+  const hintText = stripTerminalControls(hintRaw);
+  const selectedHintRendered =
+    hintText.includes("hint    /models  >/model") ||
+    (hintText.includes("hint    /models  /model") && hintRaw.includes("\x1b[1m/model\x1b[0m"));
+  if (!hintText.includes("  user> /mo") || !selectedHintRendered) {
+    throw new Error(`terminal prompt layout did not align user and hint text columns: ${JSON.stringify(hintText)}`);
   }
   const exactHintLayout = buildPromptLayout("/model", 6, 90, 24);
-  const exactHintText = exactHintLayout.renderedRows
-    .map((line) => line.replace(/\x1b\[[0-9;?]*[ -/]*[@-~]/g, ""))
-    .join("\n");
-  if (!exactHintText.includes("hint    >/model") || exactHintText.includes("/models")) {
+  const exactHintRaw = exactHintLayout.renderedRows.join("\n");
+  const exactHintText = stripTerminalControls(exactHintRaw);
+  const exactHintSelected =
+    exactHintText.includes("hint    >/model") ||
+    (exactHintText.includes("hint    /model") && exactHintRaw.includes("\x1b[1m/model\x1b[0m"));
+  if (!exactHintSelected || exactHintText.includes("/models")) {
     throw new Error("exact slash commands should not show broader prefix matches");
   }
   if (classifyEscapeAction({ active: false }) !== "noop") {

package/scripts/smoke-dynamic-step-budget.js

@@ -26,6 +26,7 @@ import {
   recordProjectVerificationOutcome,
   recordExactOutputProgress,
   recordStaticDiscoveryProgress,
+  resetGoalScopedRuntimeState,
   resetStaticDiscoveryAfterContextLoss,
   rememberCompletedDeepResearch,
   repeatedNoProgressToolBlock,
@@ -37,6 +38,7 @@ import {
 import {
   augmentScsTaskContractWithProjectVerification,
   buildScsEvidenceLedger,
+  deriveScsTaskContract,
   evaluateScsEvidence,
 } from "../src/scs-evidence.js";
 import {
@@ -72,6 +74,128 @@ function toolMessage(payload) {
 
 try {
   assert(normalizeDynamicStepsMode("off") === "off", "dynamic mode off did not normalize");
+  const failedExitProbeState = { meta: {} };
+  const failedExitProbeResult = {
+    toolName: "run_command",
+    ok: true,
+    exitCode: 0,
+    args: { command: 'python -m unittest discover -s tests; echo "EXIT:$?"' },
+    stdout: "external contract did not pass\nEXIT:1\n",
+    stderr: "",
+  };
+  recordProjectVerificationOutcome(failedExitProbeState, failedExitProbeResult, {
+    commandCwd: workspace,
+    taskProfile: "qa",
+  });
+  assert(
+    failedExitProbeResult.projectTest?.passed === false &&
+      failedExitProbeResult.projectTest?.explicitExitStatus === 1,
+    "a wrapped nonzero test status was recorded as passing"
+  );
+  const missingExitProbeState = { meta: {} };
+  const missingExitProbeResult = {
+    toolName: "run_command",
+    ok: true,
+    exitCode: 0,
+    args: { command: 'python -m unittest discover -s tests; printf "EXIT=%s\\n" "$?"' },
+    stdout: "external contract output ended before the status marker\n",
+    stderr: "",
+  };
+  recordProjectVerificationOutcome(missingExitProbeState, missingExitProbeResult, {
+    commandCwd: workspace,
+    taskProfile: "qa",
+  });
+  assert(
+    missingExitProbeResult.projectTest?.passed === false &&
+      missingExitProbeResult.projectTest?.explicitExitStatus === null,
+    "a missing wrapped test status was recorded as passing"
+  );
+  const passingExitProbeState = { meta: {} };
+  const passingExitProbeResult = {
+    toolName: "run_command",
+    ok: true,
+    exitCode: 0,
+    args: { command: 'python -m unittest discover -s tests; echo "EXIT=$?"' },
+    stdout: "contract checks passed\nEXIT=0\n",
+    stderr: "",
+  };
+  recordProjectVerificationOutcome(passingExitProbeState, passingExitProbeResult, {
+    commandCwd: workspace,
+    taskProfile: "qa",
+  });
+  assert(
+    passingExitProbeResult.projectTest?.passed === true &&
+      passingExitProbeResult.projectTest?.explicitExitStatus === 0,
+    "a wrapped zero test status was not accepted"
+  );
+
+  const newGoalState = {
+    meta: {
+      artifactProgress: { complete: true },
+      completionEvidenceRepair: { attempts: 1 },
+      dataProjectWorkflow: { ready: true },
+      durableEvidenceCategories: ["file", "visual"],
+      durableGitActions: ["commit"],
+      durableGitEvidence: [{ action: "commit", goalRevision: 1 }],
+      failedTestRecoveryPacket: { content: "old failure" },
+      goalContract: { revision: 2 },
+      projectVerification: { mutationRevision: 4 },
+      scs: { taskContract: { exactOutputPaths: ["old-output.md"] } },
+      completedDeepResearch: [{ goalKey: "retained-other-goal" }],
+    },
+  };
+  const removedGoalState = resetGoalScopedRuntimeState(newGoalState);
+  assert(removedGoalState.includes("artifactProgress"), "new goal did not clear stale artifact progress");
+  assert(!newGoalState.meta.projectVerification, "new goal retained stale project verification");
+  assert(!newGoalState.meta.scs, "new goal retained the previous SCS task contract");
+  assert(!newGoalState.meta.durableEvidenceCategories, "new goal inherited completed evidence categories");
+  assert(newGoalState.meta.goalContract?.revision === 2, "new goal reset its durable goal contract");
+  assert(newGoalState.meta.completedDeepResearch?.length === 1, "new goal discarded goal-keyed research cache");
+
+  const optionalVisualContract = deriveScsTaskContract({
+    goal: "Repair the repository, verify it, commit, and push the intentional work.",
+    taskProfile: "github",
+    acceptanceCriteria: [
+      "Do not rely only on chat summaries; verify files, commands, screenshots, PDFs, reports, or app launches as appropriate.",
+    ],
+  });
+  assert(
+    !optionalVisualContract.requiredEvidence.some((item) => item.category === "visual"),
+    "an optional evidence example forced irrelevant visual validation"
+  );
+  const explicitVisualContract = deriveScsTaskContract({
+    goal: "Capture and inspect a screenshot of the repaired interface.",
+    taskProfile: "website",
+  });
+  assert(
+    explicitVisualContract.requiredEvidence.some((item) => item.category === "visual"),
+    "an explicit screenshot request lost visual validation"
+  );
+  const readOnlyCheckerContract = deriveScsTaskContract({
+    goal:
+      "Re-run /tmp/acceptance/github_maintenance_contract.py once, verify the repository, and do not edit, commit, or push anything.",
+    taskProfile: "github",
+  });
+  assert(
+    !readOnlyCheckerContract.requiredEvidence.some((item) => item.category === "file"),
+    "a read-only checker path was mistaken for a requested file change"
+  );
+  const canvasOnlyContract = deriveScsTaskContract({
+    goal: "Create a canvas artifact preview for this smoke test.",
+  });
+  assert(
+    canvasOnlyContract.requiredEvidence.some((item) => item.category === "artifact") &&
+      !canvasOnlyContract.requiredEvidence.some((item) => item.category === "file"),
+    "a virtual canvas artifact was mistaken for a workspace-file mutation"
+  );
+  const sourceRepairContract = deriveScsTaskContract({
+    goal: "Fix src/runtime.py and verify the focused tests.",
+    taskProfile: "python",
+  });
+  assert(
+    sourceRepairContract.requiredEvidence.some((item) => item.category === "file"),
+    "an explicit source repair lost its file-change evidence gate"
+  );
   assert(
     completionEvidenceNeedsCommand({ missingProjectCommands: ["python analysis.py"] }),
     "a pending canonical command did not reopen command execution"

package/scripts/smoke-model-roles.js

@@ -986,6 +986,14 @@ const generatedReviewContract = deriveScsTaskContract({
   ].join("\n"),
   taskProfile: "review",
 });
+const qaCleanupContract = deriveScsTaskContract({
+  goal: "Inspect the project, figure out what is actually wrong, fix the failing tests, clean up generated test debris, and commit only the intentional fix.",
+  taskProfile: "qa",
+});
+const generatedFigureContract = deriveScsTaskContract({
+  goal: "Generate a publication-ready figure that compares the two repair strategies.",
+  taskProfile: "design",
+});
 assert(
   !explainCodeContract.requiresExternalEvidence,
   "code profile alone should not force external evidence for a pure explanation"
@@ -1037,6 +1045,16 @@ assert(
     !generatedReviewContract.requiredEvidence.some((item) => item.category === "artifact"),
   "review profile should not turn review-format boilerplate into file/artifact production requirements"
 );
+assert(
+  qaCleanupContract.requiredEvidence.some((item) => item.category === "command") &&
+    qaCleanupContract.requiredEvidence.some((item) => item.category === "git") &&
+    !qaCleanupContract.requiredEvidence.some((item) => item.category === "artifact"),
+  "cleaning generated test debris should not invent an unrelated artifact-delivery requirement"
+);
+assert(
+  generatedFigureContract.requiredEvidence.some((item) => item.category === "artifact"),
+  "tightening QA intent phrases removed the real generated-figure artifact gate"
+);
 const fileOnlyLedger = buildScsEvidenceLedger({
   context: { events: [{ type: "file.changed", data: { path: "src/app.js" } }] },
 });

package/scripts/smoke-progressive-tool-selection.js

@@ -725,6 +725,47 @@ const constrainedInstructionWrite = failedTestRepairWithRequiredInstruction.find
 assertStrict.deepEqual(constrainedInstructionWrite.function.parameters.properties.path.enum, ["AGINTI.md"]);
 assertStrict.deepEqual(constrainedInstructionWrite.function.parameters.properties.mode.enum, ["create"]);
 
+const continuedFailedTestRepairRuntime = nextStepRuntimeConfig(
+  { provider: "localllm", taskProfile: "qa" },
+  {
+    meta: {
+      projectVerification: {
+        mutationRevision: 1,
+        discoveredTests: ["python -m pytest -q"],
+        requiredOutputs: [],
+        testRuns: [
+          {
+            command: "python -m pytest -q",
+            mutationRevision: 0,
+            passed: false,
+            failureSignature: "baseline-failure",
+          },
+        ],
+      },
+    },
+    messages: [],
+  }
+);
+assertStrict.equal(
+  continuedFailedTestRepairRuntime.testFailureRepairActive,
+  true,
+  "a partial repair forgot the retained failed test after the mutation revision advanced"
+);
+assertStrict.equal(
+  continuedFailedTestRepairRuntime.testFailureCommand,
+  "python -m pytest -q",
+  "a partial repair lost the retained failing test command"
+);
+sameNames(
+  selectProgressiveTools(allTools, {
+    config: continuedFailedTestRepairRuntime,
+    goal: "Continue the coherent repair, then retest.",
+    profile: "qa",
+  }),
+  ["read_file", "search_files", "apply_patch", "run_command", "finish"],
+  "a partial repair was forced into test-only mode before the coherent patch was complete"
+);
+
 const pendingTestTools = selectProgressiveTools(allTools, {
   config: {
     provider: "localllm",

package/src/agent-runner.js

@@ -2423,6 +2423,28 @@ async function finishWithDirectAnswer({ config, state, store, observers, session
   };
 }
 
+export function resetGoalScopedRuntimeState(state = {}) {
+  state.meta = state.meta || {};
+  const keys = [
+    "artifactProgress",
+    "completionEvidenceRepair",
+    "dataProjectWorkflow",
+    "durableEvidenceCategories",
+    "durableGitActions",
+    "durableGitEvidence",
+    "failedTestRecoveryPacket",
+    "projectVerification",
+    "scs",
+  ];
+  const removed = [];
+  for (const key of keys) {
+    if (!(key in state.meta)) continue;
+    delete state.meta[key];
+    removed.push(key);
+  }
+  return removed;
+}
+
 async function applyContinuationPrompt(state, config, observers) {
   if (!config.resume || !config.goal) return null;
 
@@ -2439,6 +2461,9 @@ async function applyContinuationPrompt(state, config, observers) {
   state.meta = state.meta || {};
   const preserveTaskBoundary = preservesCurrentTaskBoundary(state, config.goal);
   const goalUpdate = updateGoalContract(state, config.goal, { preserveTaskBoundary });
+  if (!preserveTaskBoundary) {
+    resetGoalScopedRuntimeState(state);
+  }
   if (
     preserveTaskBoundary &&
     continuationAddsConcreteRequirement(config.goal) &&
@@ -2961,6 +2986,19 @@ function commandReportsTestFailure(result = {}) {
   );
 }
 
+function explicitExitProbeStatus(command = "", result = {}) {
+  const normalizedCommand = normalizeProjectCommand(command);
+  const hasExitProbe = /(?:^|;)\s*(?:echo|printf)\b[^;&|]*(?:EXIT|STATUS|RESULT)[^;&|]*\$\?[^;&|]*$/i.test(
+    normalizedCommand
+  );
+  if (!hasExitProbe) return { present: false, status: null };
+
+  const output = `${String(result.stdout || "")}\n${String(result.stderr || "")}`;
+  const matches = [...output.matchAll(/(?:^|\n)\s*(?:EXIT|STATUS|RESULT)(?:_CODE)?\s*[:=]\s*(-?\d+)\s*(?=\n|$)/gim)];
+  if (!matches.length) return { present: true, status: null };
+  return { present: true, status: Number(matches.at(-1)[1]) };
+}
+
 function actionableTestWarnings(result = {}) {
   const output = redactSensitiveText(`${String(result.stderr || "")}\n${String(result.stdout || "")}`);
   const warnings = [];
@@ -3133,17 +3171,24 @@ export function recordProjectVerificationOutcome(state = {}, toolResult = {}, co
 
   if (toolName === "run_command") {
     const command = normalizeProjectCommand(toolResult.args?.command || "");
+    const exitProbe = explicitExitProbeStatus(command, toolResult);
     const run = {
       command,
       at: now,
-      ok: toolResult.ok !== false && Number(toolResult.exitCode ?? 0) === 0,
+      ok:
+        toolResult.ok !== false &&
+        Number(toolResult.exitCode ?? 0) === 0 &&
+        (!exitProbe.present || exitProbe.status === 0),
       mutationRevision: verification.mutationRevision,
+      ...(exitProbe.present ? { explicitExitStatus: exitProbe.status } : {}),
     };
     verification.commandRuns = [...verification.commandRuns, run].slice(-40);
     toolResult.projectMutationRevision = verification.mutationRevision;
     if (isSubstantiveTestCommand(command)) {
       const zeroTests = commandReportsZeroTests(toolResult);
-      const reportedFailure = commandReportsTestFailure(toolResult);
+      const reportedFailure =
+        commandReportsTestFailure(toolResult) ||
+        (exitProbe.present && exitProbe.status !== 0);
       const qualityWarnings = actionableTestWarnings(toolResult);
       const passed = run.ok && !zeroTests && !reportedFailure && qualityWarnings.length === 0;
       const failedEvidence = passed ? {} : compactFailedTestEvidence(toolResult, config);
@@ -4239,13 +4284,23 @@ export function nextStepRuntimeConfig(config = {}, state = {}) {
   };
   const verification = state.meta?.projectVerification || {};
   const mutationRevision = Number(verification.mutationRevision || 0);
-  const latestCurrentTest = [...(verification.testRuns || [])]
+  const testRuns = Array.isArray(verification.testRuns) ? verification.testRuns : [];
+  const latestCurrentTest = [...testRuns]
     .reverse()
     .find((run) => Number(run.mutationRevision || 0) === mutationRevision);
-  if (latestCurrentTest && latestCurrentTest.passed !== true) {
+  const latestRecordedTest = [...testRuns]
+    .reverse()
+    .find((run) => String(run?.command || "").trim());
+  const retainedFailedTest =
+    latestCurrentTest && latestCurrentTest.passed !== true
+      ? latestCurrentTest
+      : !latestCurrentTest && latestRecordedTest?.passed === false
+        ? latestRecordedTest
+        : null;
+  if (retainedFailedTest) {
     runtimeConfig.testFailureRepairActive = true;
-    runtimeConfig.testFailureCommand = String(latestCurrentTest.command || "");
-    runtimeConfig.testFailureSignature = String(latestCurrentTest.failureSignature || "");
+    runtimeConfig.testFailureCommand = String(retainedFailedTest.command || "");
+    runtimeConfig.testFailureSignature = String(retainedFailedTest.failureSignature || "");
     const completedOutputs = new Set(
       (state.meta?.artifactProgress?.completed || [])
         .map((item) => String(item || "").replace(/\\/g, "/").replace(/^\.\//, ""))
@@ -4256,10 +4311,7 @@ export function nextStepRuntimeConfig(config = {}, state = {}) {
       .filter((item) => /(?:^|\/)(?:AGINTI|AGENTS)\.md$/i.test(item))
       .slice(0, 8);
   } else if (!latestCurrentTest && mutationRevision > 0 && (verification.discoveredTests || []).length) {
-    const retainedTestCommand = [...(verification.testRuns || [])]
-      .reverse()
-      .map((run) => String(run.command || ""))
-      .find(Boolean);
+    const retainedTestCommand = String(latestRecordedTest?.command || "");
     if (retainedTestCommand) {
       runtimeConfig.testVerificationPending = true;
       runtimeConfig.testVerificationCommand = retainedTestCommand;

package/src/scs-evidence.js

@@ -487,6 +487,31 @@ function codeProfileRequiresCommand(goal = "") {
   return substantiveCodeWork && !simpleDocumentWrite;
 }
 
+function goalRequestsWorkspaceMutation(goal = "") {
+  const text = normalizedText(goal);
+  return (
+    /\b(?:append|build|convert|copy|create|delete|edit|fix|generate|implement|modify|move|patch|refactor|remove|rename|repair|replace|rewrite|save|update|write)\b/.test(
+      text
+    ) ||
+    /创建|写入|编辑|修复|实现|修改|更新|生成|保存|复制|移动|转换|删除|重命名|替换|追加/.test(text)
+  );
+}
+
+function goalRequestsFileMutation(goal = "") {
+  const text = normalizedText(goal);
+  if (!goalRequestsWorkspaceMutation(text)) return false;
+  return (
+    /\b(?:code|codebase|document(?:ation)?|file|notes?|path|readme|repo(?:sitory)?|script|source|workspace)\b/.test(
+      text
+    ) ||
+    /(?:^|[\s`'"(])(?:\.{0,2}\/|\/)?[a-z0-9_.-]+(?:\/[a-z0-9_.{}-]+)+/i.test(text) ||
+    /\.(?:c|cc|cpp|cs|css|csv|go|h|hpp|html?|java|js|jsx|json|kt|md|mjs|php|py|rb|rs|sh|swift|tex|ts|tsx|txt|ya?ml)\b/i.test(
+      text
+    ) ||
+    /文件|文档|代码|代码库|仓库|脚本|源码|路径|工作区|说明书|笔记/.test(text)
+  );
+}
+
 function profileRequirementsForGoal(taskProfile = "", goal = "") {
   const profile = String(taskProfile || "").toLowerCase();
   const defaults = PROFILE_REQUIREMENTS[profile] || [];
@@ -510,8 +535,14 @@ function profileRequirementsForGoal(taskProfile = "", goal = "") {
     "aaps",
   ]);
   if (!codeLikeProfiles.has(profile)) return defaults;
-  if (codeProfileRequiresCommand(goal)) return defaults;
-  return defaults.filter((category) => category !== "command");
+  let requirements = [...defaults];
+  if (!goalRequestsWorkspaceMutation(goal)) {
+    requirements = requirements.filter((category) => category !== "file");
+  }
+  if (!codeProfileRequiresCommand(goal)) {
+    requirements = requirements.filter((category) => category !== "command");
+  }
+  return requirements;
 }
 
 function isReadOnlyReadinessTask(goal = "") {
@@ -544,18 +575,22 @@ function requiresSourceGrounding(goal = "") {
 function inferRequirementCategories(goal = "", taskProfile = "", acceptanceCriteria = []) {
   const positiveGoal = stripForbiddenLanguage(goal);
   const text = normalizedText(positiveGoal);
+  const artifactSignalText = text
+    .replace(
+      /\b(?:clean(?:\s+up)?|remove|delete|clear|purge)\b[^.\n;]{0,120}\b(?:generated|temporary|stale|test)?\s*(?:test\s+)?(?:debris|caches?|byproducts?)\b/gi,
+      ""
+    )
+    .replace(/\bfigure\s+out\b/gi, "");
+  const mandatoryEvidenceText = artifactSignalText.replace(
+    /[^.\n]{0,240}\b(?:as appropriate|if appropriate|when useful|where applicable)\b/gi,
+    " "
+  );
   const profile = String(taskProfile || "").toLowerCase();
   const categories = new Set(
     goalRequiresEvidence(positiveGoal, "") ? profileRequirementsForGoal(taskProfile, positiveGoal) : []
   );
 
-  if (
-    textHas(
-      text,
-      /\b(file|path|workspace|edit|patch|fix|repair|refactor|convert|copy|move|remove|delete|source|script|code)\b|\.(?:md|txt|js|jsx|ts|tsx|mjs|cjs|json|ya?ml|py|tex|html|css|svg|csv)\b|\b(?:markdown|json|yaml|html|css|tex|latex)\s+file\b|\bfile\s+(?:as|in)\s+(?:markdown|json|yaml|html|css|tex|latex)\b/
-    ) ||
-    /文件|写入文件|编辑|修复|转换|复制|移动|删除|脚本|代码/.test(text)
-  ) {
+  if (goalRequestsFileMutation(positiveGoal)) {
     categories.add("file");
   }
   const directCommandSignal =
@@ -571,13 +606,13 @@ function inferRequirementCategories(goal = "", taskProfile = "", acceptanceCrite
   if (directCommandSignal || (validationSignal && codeProfileRequiresCommand(positiveGoal))) {
     categories.add("command");
   }
-  if (textHas(text, /\b(artifact|canvas|pdf|image|video|screenshot|cover|plot|chart|figure|docx|archive|copy to|export|generated|generate|draft)\b/) || /输出|产物|图片|视频|截图|封面|生成/.test(text)) {
+  if (textHas(mandatoryEvidenceText, /\b(artifact|canvas|pdf|image|video|screenshot|cover|plot|chart|figure|docx|archive|copy to|export|generated|generate|draft)\b/) || /输出|产物|图片|视频|截图|封面|生成/.test(mandatoryEvidenceText)) {
     categories.add("artifact");
   }
-  if (textHas(text, /\b(browser|chrome|chromium|cdp|devtools|playwright|selenium|web[- ]?ui|website|page|tab|composer|click|type|upload|attach|submit|form)\b/) || /浏览器|网页|页面|上传|提交|附件|资产库/.test(text)) {
+  if (textHas(mandatoryEvidenceText, /\b(browser|chrome|chromium|cdp|devtools|playwright|selenium|web[- ]?ui|website|page|tab|composer|click|type|upload|attach|submit|form)\b/) || /浏览器|网页|页面|上传|提交|附件|资产库/.test(mandatoryEvidenceText)) {
     categories.add("browser");
   }
-  if (textHas(text, /\b(screenshot|visible|visual|see|inspect image|open image|read_image|thumbnail)\b/) || /截图|可见|缩略图/.test(text)) {
+  if (textHas(mandatoryEvidenceText, /\b(screenshot|visible|visual|see|inspect image|open image|read_image|thumbnail)\b/) || /截图|可见|缩略图/.test(mandatoryEvidenceText)) {
     categories.add("visual");
   }
   if (