@pushpalsdev/cli 1.0.86 → 1.0.93

package/dist/pushpals-cli.js

@@ -1071,7 +1071,7 @@ function loadPushPalsConfig(options = {}) {
           const parsed = Number.parseFloat(String(firstNonEmpty(process.env.REMOTEBUDDY_AUTONOMY_ALERT_AUTONOMY_FAILURE_RATE_THRESHOLD, asString(remoteAutonomyNode.alert_autonomy_failure_rate_threshold, "0.45"), "0.45")));
           return Number.isFinite(parsed) ? parsed : 0.45;
         })())),
-        allowReadAnywhere: parseBoolEnv("REMOTEBUDDY_AUTONOMY_ALLOW_READ_ANYWHERE") ?? asBoolean(remoteAutonomyNode.allow_read_anywhere, false),
+        allowReadAnywhere: parseBoolEnv("REMOTEBUDDY_AUTONOMY_ALLOW_READ_ANYWHERE") ?? asBoolean(remoteAutonomyNode.allow_read_anywhere, true),
         prFeedbackCommentRows: Math.max(1, Math.min(200, asInt(parseIntEnv("REMOTEBUDDY_AUTONOMY_PR_FEEDBACK_COMMENT_ROWS") ?? remoteAutonomyNode.pr_feedback_comment_rows, 16))),
         prFeedbackCommentChars: Math.max(32, Math.min(20000, asInt(parseIntEnv("REMOTEBUDDY_AUTONOMY_PR_FEEDBACK_COMMENT_CHARS") ?? remoteAutonomyNode.pr_feedback_comment_chars, 600))),
         prFeedbackSummaryChars: Math.max(32, Math.min(20000, asInt(parseIntEnv("REMOTEBUDDY_AUTONOMY_PR_FEEDBACK_SUMMARY_CHARS") ?? remoteAutonomyNode.pr_feedback_summary_chars, 600))),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pushpalsdev/cli",
-  "version": "1.0.86",
+  "version": "1.0.93",
   "description": "PushPals terminal CLI for LocalBuddy -> RemoteBuddy orchestration",
   "license": "MIT",
   "repository": {
@@ -12,7 +12,7 @@
     "url": "https://github.com/PushPalsDev/pushpals/issues"
   },
   "bin": {
-    "pushpals": "./bin/pushpals.cjs"
+    "pushpals": "bin/pushpals.cjs"
   },
   "files": [
     "bin",

package/runtime/prompts/remotebuddy/autonomy_ideation_system_prompt.md

@@ -56,9 +56,10 @@ Constraints:
 - `objective_type` is a governance lane, not a fixed feature catalog. Feature ideas are free-form and should be expressed in `title`, `problem_statement`, and `feature_hypotheses`.
 - `feature_hypotheses` may contain any suitable product/engineering features; keep each item concise and actionable.
 - target_paths must be literal repo-relative paths.
-- write_globs must be repo-relative globs.
+- write_globs must be repo-relative globs used as starting-point/relevance hints, not hard write boundaries.
 - Choose target_paths that own the behavior being improved, not thin route wrappers, re-export files, or shell components, unless the requested change is explicitly at that wrapper boundary.
 - For UI/game/product-surface objectives, prefer files that render or compute the relevant state directly; use wrapper files only for navigation, mounting, or screen-level chrome work.
+- Workers have repo-wide sandbox write access and may expand from these hints to the behavior-owning files; the review agent will judge whether the final diff stays relevant.
 - do not invent evidence ids.
 - If all signals are low/noisy, it is valid to return zero candidates.
 - Treat a low `sig_queue_health` value as maintenance-window evidence for safe proactive work, not only incident response.

package/runtime/prompts/remotebuddy/autonomy_planning_system_prompt.md

@@ -1,5 +1,5 @@
 Write objective instruction text for a worker.
 Return strict JSON:
 { "instruction": "..." }
-Keep it concise, executable, and scoped to target_paths and write_globs only.
+Keep it concise and executable. Treat target_paths and write_globs as starting-point/relevance hints, not hard write boundaries; the worker may edit other behavior-owning repo files when needed and the review agent will judge relevance.
 If you mention commands, use Bun/Bunx command forms (`bun ...`, `bunx ...`), never npm/npx/pnpm/yarn forms.

package/runtime/prompts/remotebuddy/remotebuddy_system_prompt.md

@@ -11,7 +11,7 @@ Repository boundary policy:
 
 - Treat `{{repo_root}}` as the only allowed repository scope.
 - Never plan edits or checks outside this repository root.
-- Prefer explicit repo-relative targets; use broad `"."` scope only when the user explicitly requests whole-repo work.
+- Prefer explicit repo-relative targets as review/relevance hints. WorkerPals have repo-wide sandbox write access, so do not over-constrain write scope unless the user explicitly asks for a hard path limit.
 
 Intent taxonomy (choose the single best fit):
 
@@ -44,7 +44,7 @@ Execution policy:
 - Scope policy (for `requires_worker=true`):
   - `scope.read_anywhere` should default to `true` (do not set `false` unless user explicitly requested restrictive reading)
   - `scope.write_allowed` should default to `true`
-  - `scope.write_globs` should be included only when you need to constrain edits
+  - `scope.write_globs` should be included as starting-point/relevance hints, not as hard write boundaries
   - `scope.forbidden_globs` should be included only when specific paths must be blocked
   - `scope.max_files_to_edit` should be included only when a cap is needed
 

package/runtime/prompts/workerpals/miniswe_completion_requirement.md

@@ -1 +1 @@
-Completion requirement: handle all requested edits across all explicit target paths before setting done=true.
+Completion requirement: solve the requested task before setting done=true. Target paths are relevance hints; edit other behavior-owning files when needed and explain why.

package/runtime/prompts/workerpals/miniswe_explicit_targets_block.md

@@ -1,2 +1,2 @@
-Explicit target paths:
+Target path hints:
 {{targets_block}}

package/runtime/prompts/workerpals/openai_codex_task_execute_system_prompt.md

@@ -11,7 +11,10 @@ Non-negotiable runtime invariants:
 
 Execution rules:
 
-- Keep edits minimal, correct, and scoped to the requested task.
+- Keep edits minimal, correct, and relevant to the requested task.
+- You have repo-wide read/write access inside an isolated WorkerPal sandbox. Target paths and write globs are starting-point/relevance hints, not hard write boundaries.
+- If the hinted file is a thin wrapper or the behavior lives elsewhere, edit the behavior-owning file(s) needed to solve the task and explain the scope expansion in your final response.
+- Avoid irrelevant sprawl; the review agent will judge whether changed files are necessary for the requested outcome.
 - Read relevant files before editing, then run focused validation.
 - Use direct commands without shell wrappers. Prefer plain commands like `git diff -- path`, `git add <path>`, `git status --porcelain`, and `pwd`.
 - Do not wrap commands in `/bin/bash -lc`, `sh -lc`, `cmd /c`, or `powershell -Command`, and avoid pipelines, `awk`, heredocs, or multi-command shell snippets unless they are truly unavoidable.

package/runtime/prompts/workerpals/openhands_minimal_system_prompt.j2

@@ -1,10 +1,12 @@
 You are PushPals WorkerPal running inside OpenHands.
 
-Operate with strict scope:
+Operate with strict relevance:
 - Focus only on the current user task.
 - If the task is a question, answer directly and do not edit files.
 - If code/file changes are requested, implement them end-to-end.
 - Reuse existing repository conventions and avoid unrelated refactors/docs.
+- Treat target paths and write globs as starting-point/relevance hints, not hard write boundaries.
+- If the behavior lives outside the hinted files, edit the owning files and explain the scope expansion.
 
 Execution loop:
 - Prefer concrete execution over planning chatter.

package/runtime/prompts/workerpals/openhands_task_execute_system_prompt.md

@@ -3,7 +3,8 @@ You are PushPals WorkerPal running inside OpenHands.
 Execution rules:
 
 - Focus only on the task below.
-- Keep changes minimal, correct, and scoped to the request.
+- Keep changes minimal, correct, and relevant to the request.
+- Target paths and write globs are starting-point/relevance hints, not hard write boundaries; if the behavior lives elsewhere, edit the owning files and explain why.
 - Read relevant files before editing.
 - Reuse existing project conventions and tooling.
 - If the task is a question/explanation, answer directly and do not edit files.

package/runtime/prompts/workerpals/workerpals_system_prompt.md

@@ -6,7 +6,7 @@ Your mission:
 
 - Take the user (or RemoteBuddy) request and fully execute it end-to-end.
 - You are responsible for breaking the work down into concrete subtasks, completing them, validating, reviewing your own changes, and preparing a high-quality commit message when the work is ready.
-- You have full read/write access within the assigned repository only; do not access files outside that repository.
+- You have full read/write access within the assigned repository sandbox only; do not access files outside that repository. Target paths and write globs are starting-point/relevance hints, not hard write boundaries.
 
 Mindset:
 
@@ -64,7 +64,7 @@ Execution workflow (you MUST follow this):
      - Performance: no unnecessary work on UI thread, no extra network calls, no large bundles
      - Cross-platform: iOS/Android/Web differences guarded appropriately
      - Security: no secret leakage, safe networking defaults, no unsafe shell usage
-   - Make any final polish edits that improve clarity without changing scope.
+   - Make any final polish edits that improve clarity without drifting from the task.
 
 7. Prepare to commit (when appropriate)
    - When the work is ready, produce a detailed commit message (do NOT actually commit unless your system explicitly allows it).

package/runtime/sandbox/.pushpals-remotebuddy-fallback.js

@@ -86,11 +86,13 @@ class CommunicationManager {
   sessionId;
   from;
   authToken;
+  fetchImpl;
   constructor(opts) {
     this.serverUrl = opts.serverUrl;
     this.sessionId = opts.sessionId;
     this.from = opts.from;
     this.authToken = opts.authToken ?? null;
+    this.fetchImpl = opts.fetchImpl ?? fetch;
   }
   headers() {
     const headers = { "Content-Type": "application/json" };
@@ -132,7 +134,7 @@ class CommunicationManager {
         body.turnId = meta.turnId;
       if (meta.parentId)
         body.parentId = meta.parentId;
-      const response = await fetch(this.commandUrl(sessionId), {
+      const response = await this.fetchImpl(this.commandUrl(sessionId), {
         method: "POST",
         headers: this.headers(),
         body: JSON.stringify(body)
@@ -588,7 +590,8 @@ function validateScopeInvariants(componentArea, targetPathsInput, writeGlobsInpu
   const scopeSeeds = collectScopeSeedPaths(targetPathsInput, writeGlobsInput);
   const normalizedComponentArea = normalizeAutonomyComponentArea(componentArea) ?? deriveAutonomyComponentArea(targetPathsInput, writeGlobsInput);
   const allowMultipleComponentRoots = options?.allowMultipleComponentRoots === true;
-  if (!normalizedComponentArea && scopeSeeds.length > 1 && !allowMultipleComponentRoots) {
+  const hintsOnly = options?.hintsOnly === true;
+  if (!hintsOnly && !normalizedComponentArea && scopeSeeds.length > 1 && !allowMultipleComponentRoots) {
     errors.push(`scope spans multiple component roots: ${scopeSeeds.slice(0, 6).join(", ")}`);
   }
   const rootPrefix = normalizedComponentArea ? componentRootPrefix(normalizedComponentArea) : "";
@@ -600,7 +603,7 @@ function validateScopeInvariants(componentArea, targetPathsInput, writeGlobsInpu
       errors.push(`invalid target_path: ${String(raw ?? "")}`);
       continue;
     }
-    if (rootPrefix && !underRoot(normalized, rootPrefix)) {
+    if (!hintsOnly && rootPrefix && !underRoot(normalized, rootPrefix)) {
       errors.push(`target_path outside component root: ${normalized}`);
       continue;
     }
@@ -621,20 +624,20 @@ function validateScopeInvariants(componentArea, targetPathsInput, writeGlobsInpu
       errors.push(`invalid write_glob: ${String(raw ?? "")}`);
       continue;
     }
-    if (hasForbiddenBroadGlob(normalized)) {
+    if (!hintsOnly && hasForbiddenBroadGlob(normalized)) {
       errors.push(`forbidden broad write_glob: ${normalized}`);
       continue;
     }
     const prefix = literalPrefix(normalized);
-    if (!prefix) {
+    if (!hintsOnly && !prefix) {
       errors.push(`write_glob literal prefix cannot be empty: ${normalized}`);
       continue;
     }
-    if (rootPrefix && !underRoot(prefix, rootPrefix)) {
+    if (!hintsOnly && rootPrefix && !underRoot(prefix, rootPrefix)) {
       errors.push(`write_glob outside component root: ${normalized}`);
       continue;
     }
-    if (!normalizedTargetPaths.some((targetPath) => targetPath === prefix || targetPath.startsWith(`${prefix}/`))) {
+    if (!hintsOnly && !normalizedTargetPaths.some((targetPath) => targetPath === prefix || targetPath.startsWith(`${prefix}/`))) {
       errors.push(`write_glob prefix does not align with target_paths: ${normalized}`);
       continue;
     }
@@ -647,14 +650,14 @@ function validateScopeInvariants(componentArea, targetPathsInput, writeGlobsInpu
   if ((options?.requireWriteGlobs ?? true) && normalizedWriteGlobs.length === 0) {
     errors.push("write_globs must be provided and non-empty");
   }
-  if (normalizedTargetPaths.length > 0 && normalizedWriteGlobs.length > 0) {
+  if (!hintsOnly && normalizedTargetPaths.length > 0 && normalizedWriteGlobs.length > 0) {
     for (const targetPath of normalizedTargetPaths) {
       const covered = normalizedWriteGlobs.some((glob) => matchesGlob(targetPath, glob));
       if (!covered)
         errors.push(`target_path not covered by write_globs: ${targetPath}`);
     }
   }
-  if (!normalizedComponentArea && !allowMultipleComponentRoots) {
+  if (!hintsOnly && !normalizedComponentArea && !allowMultipleComponentRoots) {
     errors.push("component_area could not be derived from scope");
   }
   const breadth = classifyGlobBreadth(normalizedWriteGlobs);
@@ -1357,7 +1360,7 @@ function loadPushPalsConfig(options = {}) {
           const parsed = Number.parseFloat(String(firstNonEmpty(process.env.REMOTEBUDDY_AUTONOMY_ALERT_AUTONOMY_FAILURE_RATE_THRESHOLD, asString(remoteAutonomyNode.alert_autonomy_failure_rate_threshold, "0.45"), "0.45")));
           return Number.isFinite(parsed) ? parsed : 0.45;
         })())),
-        allowReadAnywhere: parseBoolEnv("REMOTEBUDDY_AUTONOMY_ALLOW_READ_ANYWHERE") ?? asBoolean(remoteAutonomyNode.allow_read_anywhere, false),
+        allowReadAnywhere: parseBoolEnv("REMOTEBUDDY_AUTONOMY_ALLOW_READ_ANYWHERE") ?? asBoolean(remoteAutonomyNode.allow_read_anywhere, true),
         prFeedbackCommentRows: Math.max(1, Math.min(200, asInt(parseIntEnv("REMOTEBUDDY_AUTONOMY_PR_FEEDBACK_COMMENT_ROWS") ?? remoteAutonomyNode.pr_feedback_comment_rows, 16))),
         prFeedbackCommentChars: Math.max(32, Math.min(20000, asInt(parseIntEnv("REMOTEBUDDY_AUTONOMY_PR_FEEDBACK_COMMENT_CHARS") ?? remoteAutonomyNode.pr_feedback_comment_chars, 600))),
         prFeedbackSummaryChars: Math.max(32, Math.min(20000, asInt(parseIntEnv("REMOTEBUDDY_AUTONOMY_PR_FEEDBACK_SUMMARY_CHARS") ?? remoteAutonomyNode.pr_feedback_summary_chars, 600))),
@@ -4257,11 +4260,6 @@ var POLICY = {
   }
 };
 var RISK_ORDER = { low: 0, medium: 1, high: 2 };
-var BREADTH_ORDER = {
-  narrow: 0,
-  medium: 1,
-  broad: 2
-};
 var IDEATION_SYSTEM_PROMPT = loadPromptTemplate("remotebuddy/autonomy_ideation_system_prompt.md").trim();
 var SCORING_SYSTEM_PROMPT = loadPromptTemplate("remotebuddy/autonomy_scoring_system_prompt.md").trim();
 var PLANNING_SYSTEM_PROMPT = loadPromptTemplate("remotebuddy/autonomy_planning_system_prompt.md").trim();
@@ -4996,7 +4994,8 @@ function chooseRepoObjectiveTargetProfile(profiles, objective) {
 function adaptCandidateShapeToRepo(params) {
   const shape = params.shape;
   const scopeValidation = validateScopeInvariants(shape.component_area, shape.target_paths, shape.write_globs, {
-    requireWriteGlobs: true
+    requireWriteGlobs: true,
+    hintsOnly: true
   });
   const pathsExist = params.repoRoot && scopeValidation.ok ? findMissingRepoTargetPaths(params.repoRoot, scopeValidation.normalizedTargetPaths).length === 0 : scopeValidation.ok;
   if (scopeValidation.ok && pathsExist) {
@@ -5782,7 +5781,7 @@ ${pattern.tags.join(" ")}`.toLowerCase();
   const targetPaths = asStringArray2(metadataShape.target_paths ?? metadataShape.targetPaths ?? metadata.target_paths);
   const writeGlobs = asStringArray2(metadataShape.write_globs ?? metadataShape.writeGlobs ?? metadata.write_globs);
   const validationIdeas = asStringArray2(metadataShape.expected_validation ?? metadataShape.expectedValidation ?? metadata.expected_validation ?? pattern.validationIdeas);
-  const scopeCheck = validateScopeInvariants(componentArea, targetPaths.length > 0 ? targetPaths : defaults.target_paths, writeGlobs.length > 0 ? writeGlobs : defaults.write_globs, { requireWriteGlobs: true });
+  const scopeCheck = validateScopeInvariants(componentArea, targetPaths.length > 0 ? targetPaths : defaults.target_paths, writeGlobs.length > 0 ? writeGlobs : defaults.write_globs, { requireWriteGlobs: true, hintsOnly: true });
   return adaptCandidateShapeToRepo({
     shape: {
       objective_type: objectiveType,
@@ -6217,7 +6216,7 @@ function buildRepoVisionFallbackCandidates(params) {
       component_area: componentArea,
       target_paths: targetPaths,
       scope: {
-        read_anywhere: false,
+        read_anywhere: true,
         write_globs: writeGlobs
       },
       risk_level: "low",
@@ -6271,7 +6270,7 @@ function buildEngineFallbackCandidates(params) {
       component_area: candidateShape.component_area,
       target_paths: candidateShape.target_paths,
       scope: {
-        read_anywhere: false,
+        read_anywhere: true,
         write_globs: candidateShape.write_globs
       },
       risk_level: candidateShape.risk_level,
@@ -7367,15 +7366,11 @@ ${JSON.stringify(input.messages ?? [])}`),
             recordDropReason(`${source}_risk_exceeds_policy`);
             continue;
           }
-          const scopeValidation = validateScopeInvariants(candidate.component_area, candidate.target_paths, candidate.scope.write_globs, { requireWriteGlobs: true });
+          const scopeValidation = validateScopeInvariants(candidate.component_area, candidate.target_paths, candidate.scope.write_globs, { requireWriteGlobs: true, hintsOnly: true });
           if (!scopeValidation.ok) {
             recordDropReason(`${source}_scope_validation_failed`);
             continue;
           }
-          if (BREADTH_ORDER[scopeValidation.breadth] > BREADTH_ORDER[policy.maxBreadth]) {
-            recordDropReason(`${source}_scope_breadth_exceeds_policy`);
-            continue;
-          }
           if (candidate.scope.read_anywhere && !this.cfg.allowReadAnywhere) {
             recordDropReason(`${source}_read_anywhere_not_allowed`);
             continue;
@@ -8266,12 +8261,13 @@ function buildExecutionGuidance(plan, targetPaths, requiredValidationSteps = [])
   const lines = [];
   const targets = normalizePathHints(targetPaths.length > 0 ? targetPaths : plan.scope.write_globs ?? []);
   if (targets.length > 0) {
-    lines.push("Target paths:");
+    lines.push("Target paths / starting points:");
     for (const path of targets)
       lines.push(`- ${path}`);
     lines.push("Path handling:");
     lines.push("- Treat all target paths as repo-relative to the current working directory.");
     lines.push("- Do not prepend a leading slash to target paths.");
+    lines.push("- These paths are relevance hints, not hard write boundaries; edit the behavior-owning files needed for the task and explain any expansion.");
   }
   lines.push("Scope:");
   lines.push(`- read_anywhere: ${plan.scope.read_anywhere ? "true" : "false"}`);
@@ -8280,7 +8276,7 @@ function buildExecutionGuidance(plan, targetPaths, requiredValidationSteps = [])
     lines.push(`- max_files_to_edit: ${plan.scope.max_files_to_edit}`);
   }
   if (Array.isArray(plan.scope.write_globs) && plan.scope.write_globs.length > 0) {
-    lines.push("Write globs:");
+    lines.push("Write intent hints:");
     for (const glob of plan.scope.write_globs)
       lines.push(`- ${glob}`);
   }
@@ -8556,6 +8552,7 @@ class RemoteBuddyOrchestrator {
   server;
   sessionId;
   authToken;
+  fetchImpl;
   repo;
   jobsDbPath;
   workerOnlineTtlMs;
@@ -8624,6 +8621,7 @@ class RemoteBuddyOrchestrator {
     this.server = opts.server;
     this.sessionId = opts.sessionId;
     this.authToken = opts.authToken;
+    this.fetchImpl = opts.fetchImpl ?? fetch;
     this.brain = opts.brain;
     this.idempotency = opts.idempotency;
     this.persistentMemory = opts.persistentMemory;
@@ -8687,7 +8685,8 @@ class RemoteBuddyOrchestrator {
       serverUrl: this.server,
       sessionId: this.sessionId,
       authToken: this.authToken,
-      from: `agent:${this.agentId}`
+      from: `agent:${this.agentId}`,
+      fetchImpl: this.fetchImpl
     });
     this.autonomousEngine = new RemoteBuddyAutonomousEngine({
       server: this.server,
@@ -8749,7 +8748,7 @@ class RemoteBuddyOrchestrator {
   async ensureSessionWithRetry(sessionId = this.sessionId, maxRetries = 20, baseDelayMs = 500, maxDelayMs = 5000) {
     for (let attempt = 1;attempt <= maxRetries && !this.disposed; attempt++) {
       try {
-        const res = await fetch(`${this.server}/sessions`, {
+        const res = await this.fetchImpl(`${this.server}/sessions`, {
           method: "POST",
           headers: this.authHeaders(),
           body: JSON.stringify({ sessionId })
@@ -8796,7 +8795,7 @@ class RemoteBuddyOrchestrator {
   }
   async fetchJobLogs(jobId, limit = 80) {
     try {
-      const res = await fetch(`${this.server}/jobs/${jobId}/logs?limit=${Math.max(1, Math.min(500, limit))}`, {
+      const res = await this.fetchImpl(`${this.server}/jobs/${jobId}/logs?limit=${Math.max(1, Math.min(500, limit))}`, {
         method: "GET",
         headers: this.authHeaders()
       });
@@ -8812,7 +8811,7 @@ class RemoteBuddyOrchestrator {
   }
   async fetchJobToolRuns(jobId, limit = 20) {
     try {
-      const res = await fetch(`${this.server}/jobs/${jobId}/tool-runs?limit=${Math.max(1, Math.min(100, limit))}`, {
+      const res = await this.fetchImpl(`${this.server}/jobs/${jobId}/tool-runs?limit=${Math.max(1, Math.min(100, limit))}`, {
         method: "GET",
         headers: this.authHeaders()
       });
@@ -8868,7 +8867,7 @@ class RemoteBuddyOrchestrator {
     query.set("feedbackLimit", "3");
     const suffix = query.toString();
     try {
-      const res = await fetch(`${this.server}/autonomy/insights${suffix ? `?${suffix}` : ""}`, {
+      const res = await this.fetchImpl(`${this.server}/autonomy/insights${suffix ? `?${suffix}` : ""}`, {
         method: "GET",
         headers: this.authHeaders()
       });
@@ -9169,7 +9168,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
         payload.dedupeKey = dedupeKey;
       if (targetWorkerId)
         payload.targetWorkerId = targetWorkerId;
-      const res = await fetch(`${this.server}/jobs/enqueue`, {
+      const res = await this.fetchImpl(`${this.server}/jobs/enqueue`, {
         method: "POST",
         headers: this.authHeaders(),
         body: JSON.stringify(payload)
@@ -9439,7 +9438,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
   }
   async fetchWorkers() {
     try {
-      const res = await fetch(`${this.server}/workers?ttlMs=${this.workerOnlineTtlMs}`, {
+      const res = await this.fetchImpl(`${this.server}/workers?ttlMs=${this.workerOnlineTtlMs}`, {
         method: "GET",
         headers: this.authHeaders()
       });
@@ -9453,7 +9452,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
   }
   async fetchWorkerAutoscaleSnapshot() {
     try {
-      const res = await fetch(`${this.server}/workers/autoscale?ttlMs=${this.workerOnlineTtlMs}`, {
+      const res = await this.fetchImpl(`${this.server}/workers/autoscale?ttlMs=${this.workerOnlineTtlMs}`, {
         method: "GET",
         headers: this.authHeaders()
       });
@@ -9715,7 +9714,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
     const prompt = String(request.prompt ?? "").trim();
     if (!prompt) {
       console.warn(`[RemoteBuddy] Request ${requestId} missing prompt; marking failed`);
-      await fetch(`${this.server}/requests/${requestId}/fail`, {
+      await this.fetchImpl(`${this.server}/requests/${requestId}/fail`, {
         method: "POST",
         headers: this.authHeaders(),
         body: JSON.stringify({ message: "Request missing prompt" })
@@ -9749,7 +9748,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
           plan.job_kind = "task.execute";
           plan.lane = "worker";
         }
-        plan.scope.read_anywhere = false;
+        plan.scope.read_anywhere = true;
         plan.scope.write_allowed = true;
         plan.scope.write_globs = [...autonomyMetadata.writeGlobs];
       }
@@ -9774,7 +9773,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
         if (scopeCoverage.addedGlobs.length > 0) {
           console.warn(`[RemoteBuddy] Planner write_globs did not cover target paths. Added scope globs: ${scopeCoverage.addedGlobs.join(", ")}`);
         }
-        if (forceWorker) {
+        if (forceWorker && !autonomyMetadata) {
           const concreteTargetCount = targetPaths.filter((entry) => entry && entry !== ".").length;
           if (concreteTargetCount > 0) {
             const currentMax = Number.isFinite(Number(plan.scope.max_files_to_edit)) && Number(plan.scope.max_files_to_edit) > 0 ? Math.floor(Number(plan.scope.max_files_to_edit)) : 0;
@@ -9783,9 +9782,6 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
             }
           }
         }
-        if (autonomyMetadata && (!plan.scope.write_globs || plan.scope.write_globs.length === 0)) {
-          throw new Error("Autonomy-origin request requires non-empty planning.scope.write_globs before task dispatch.");
-        }
         if (plan.acceptance_criteria.length === 0) {
           plan.acceptance_criteria = ["Produce a correct and helpful result for the user request."];
         }
@@ -9844,7 +9840,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
             await this.assistantMessage(requestSessionId, "Should I have a WorkerPal implement this? Reply to confirm and I'll enqueue the work, or clarify what you'd like focused on.", { turnId, correlationId: requestId, from: eventFrom });
           }
         }
-        await fetch(`${this.server}/requests/${requestId}/complete`, {
+        await this.fetchImpl(`${this.server}/requests/${requestId}/complete`, {
           method: "POST",
           headers: this.authHeaders(),
           body: JSON.stringify({
@@ -9875,7 +9871,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
             correlationId: requestId,
             from: eventFrom
           });
-          await fetch(`${this.server}/requests/${requestId}/fail`, {
+          await this.fetchImpl(`${this.server}/requests/${requestId}/fail`, {
             method: "POST",
             headers: this.authHeaders(),
             body: JSON.stringify({
@@ -10010,7 +10006,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
         await this.assistantMessage(requestSessionId, "I could not queue this WorkerPal task. No task was started.", { turnId, correlationId: requestId, from: eventFrom });
         this.rememberPersistentMemory("job_enqueue_failed", `enqueue_failed lane=${lane} intent=${plan.intent}`, requestId, requestSessionId);
       }
-      await fetch(`${this.server}/requests/${requestId}/complete`, {
+      await this.fetchImpl(`${this.server}/requests/${requestId}/complete`, {
         method: "POST",
         headers: this.authHeaders(),
         body: JSON.stringify({
@@ -10041,7 +10037,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
         correlationId: requestId,
         from: eventFrom
       });
-      await fetch(`${this.server}/requests/${requestId}/fail`, {
+      await this.fetchImpl(`${this.server}/requests/${requestId}/fail`, {
         method: "POST",
         headers: this.authHeaders(),
         body: JSON.stringify({
@@ -10056,7 +10052,7 @@ Please reply with the missing details and I will enqueue a follow-up request.` :
     while (!this.disposed) {
       try {
         await this.maybeAutoscaleWorkers();
-        const res = await fetch(`${this.server}/requests/claim`, {
+        const res = await this.fetchImpl(`${this.server}/requests/claim`, {
           method: "POST",
           headers: this.authHeaders(),
           body: JSON.stringify({ agentId: this.agentId })

package/runtime/sandbox/apps/workerpals/src/backends/miniswe/miniswe_executor.py

@@ -465,25 +465,9 @@ def _extract_write_globs_from_payload(payload: Optional[Dict[str, Any]]) -> List
 
 
 def _assert_write_allowed(repo: str, path: str, write_globs: Optional[List[str]]) -> None:
-    if not write_globs:
-        return
-    normalized = _normalize_concrete_repo_path(repo, path)
-    if not normalized:
-        raise RuntimeError(f"Invalid write path for scope enforcement: {path!r}")
-    for glob in write_globs:
-        pattern = str(glob or "").strip()
-        if not pattern:
-            continue
-        if any(ch in pattern for ch in "*?[]"):
-            if fnmatch.fnmatchcase(normalized, pattern):
-                return
-            continue
-        if normalized == pattern or normalized.startswith(pattern + "/"):
-            return
-    raise RuntimeError(
-        "Scope violation: attempted write outside writeGlobs. "
-        f"path={normalized!r} write_globs={write_globs!r}"
-    )
+    # WorkerPal jobs run in isolated sandboxes. Scope hints are used for review
+    # relevance, not per-write filesystem enforcement.
+    return
 
 
 def _read_text_file(repo: str, path: str, max_chars: int = 60000) -> str:
@@ -1587,11 +1571,6 @@ def _broker_run(
     if expected_targets and changed_paths:
         changed_set = {str(p).strip().replace("\\", "/") for p in changed_paths}
         expected_set = {str(p).strip().replace("\\", "/") for p in expected_targets}
-        strict_target_match = bool(
-            explicit_target_set
-            and not any(t in {".", "/"} for t in explicit_target_set)
-            and not any(any(ch in t for ch in "*?[]") for t in explicit_target_set)
-        )
         matched = any(
             _target_hint_matches_changed_path(expected, changed)
             for expected in expected_set
@@ -1602,14 +1581,6 @@ def _broker_run(
                 "Expected one of target paths to change, but observed different files. "
                 f"expected={sorted(expected_set)} observed={sorted(changed_set)}"
             )
-            if strict_target_match:
-                return {
-                    "ok": False,
-                    "summary": "tool broker failed: changed files do not match explicit target paths",
-                    "stdout": stdout + "\n\nChanged files:\n" + "\n".join(f"- {p}" for p in changed_paths),
-                    "stderr": msg,
-                    "exitCode": 3,
-                }
             stdout += "\n\nTarget-path mismatch (heuristic, non-fatal):\n" + msg
     if edits_made and not shell_validation_ran:
         stdout += (
@@ -1786,10 +1757,10 @@ def _run_miniswe_task(
     agent = None
     agent_messages: List[Dict[str, Any]] = []
     broker_enabled = _tool_broker_enabled(base_url)
-    prefer_broker_for_scoped_writes = bool(explicit_write_globs)
+    prefer_broker_for_scoped_writes = False
     ran_primary_broker = False
     if prefer_broker_for_scoped_writes and broker_enabled:
-        log.info("Using tool broker shim for strict per-write scope enforcement.")
+        log.info("Using tool broker shim for task execution.")
         broker_result = _run_broker_with_recovery()
         if not bool(broker_result.get("ok")):
             return {

package/runtime/sandbox/apps/workerpals/src/backends/openhands/openhands_executor.py

@@ -352,9 +352,10 @@ def _build_path_handling_message(target_paths: List[str], repo: str) -> str:
         "- Prefer the repo-relative paths for shell commands.\n"
         "- If FileEditor rejects a repo-relative path, retry with the matching absolute path.\n"
         "- Do not run broad filesystem scans when concrete target paths are listed.\n"
-        "Concrete target paths (repo-relative):\n"
+        "- These paths are starting points, not hard write boundaries; edit other behavior-owning files when needed and explain why.\n"
+        "Target path hints (repo-relative):\n"
         f"{listed_rel}\n"
-        "Concrete target paths (absolute):\n"
+        "Target path hints (absolute):\n"
         f"{listed_abs}"
     )
 

package/runtime/sandbox/apps/workerpals/src/execute_job.ts

@@ -19,7 +19,6 @@ import {
   normalizeTargetPath,
   requirementsForValidationCommand,
   sanitizeSourceControlIdentityField,
-  validateScopeInvariants,
   type AutonomyComponentArea,
   type SourceControlCommitIdentity,
   type ToolRequirement,
@@ -2503,12 +2502,21 @@ function buildStageTargets(kind: string, params?: Record<string, unknown>): stri
   }
 }
 
-function buildStageCommand(kind: string, params?: Record<string, unknown>): string[] | null {
+export function buildStageCommand(kind: string, params?: Record<string, unknown>): string[] | null {
+  if (kind === "task.execute") {
+    return [
+      "add",
+      "-A",
+      "--",
+      ".",
+      ":(exclude)workspace/**",
+      ":(exclude)outputs/**",
+      ":(exclude).codex",
+      ":(exclude).codex/**",
+    ];
+  }
   const targets = buildStageTargets(kind, params);
   if (targets.length === 0) {
-    if (kind === "task.execute") {
-      return ["add", "-A", "--", ".", ":(exclude)workspace/**", ":(exclude)outputs/**"];
-    }
     return null;
   }
   return ["add", "-A", "--", ...targets];
@@ -3949,13 +3957,10 @@ function taskExecuteOrigin(params: Record<string, unknown>): "autonomy" | "user"
   return "user";
 }
 
-function collectWriteScopeIssuesFromChangedPaths(
+export function collectWriteScopeIssuesFromChangedPaths(
   changedPaths: string[],
   planning: TaskExecutePlanning,
 ): string[] {
-  const writeGlobs = toStringArray(planning.scope.writeGlobs ?? []);
-  if (writeGlobs.length === 0) return [];
-
   const normalizedChangedPaths = changedPaths
     .map((entry) => normalizeStagePath(entry))
     .filter((entry): entry is string => Boolean(entry) && entry !== ".");
@@ -3963,12 +3968,6 @@ function collectWriteScopeIssuesFromChangedPaths(
 
   const forbidden = toStringArray(planning.scope.forbiddenGlobs ?? []);
   const issues: string[] = [];
-  const outOfScope = normalizedChangedPaths.filter(
-    (path) => !writeGlobs.some((glob) => matchesGlob(path, glob)),
-  );
-  if (outOfScope.length > 0) {
-    issues.push(`modified paths outside writeGlobs: ${outOfScope.join(", ")}`);
-  }
   const forbiddenTouched = normalizedChangedPaths.filter((path) =>
     forbidden.some((glob) => matchesGlob(path, glob)),
   );
@@ -4105,41 +4104,17 @@ function validateTaskExecutePlanning(
       reviewAgentAllowsMultiRootScope(options?.reviewAgentResolutionType);
     if (origin === "autonomy") {
       const declaredComponentArea = asAutonomyComponentArea(options?.autonomyComponentArea);
-      const inferredComponentArea = allowMultiRootAutonomyScope
-        ? null
-        : deriveAutonomyComponentArea(normalizedTargetPaths, normalizedWriteGlobs);
-      const componentArea = allowMultiRootAutonomyScope
-        ? declaredComponentArea
-        : declaredComponentArea ?? inferredComponentArea;
-      if (!allowMultiRootAutonomyScope && !componentArea) {
-        return {
-          ok: false,
-          message:
-            "task.execute planning.targetPaths must resolve to a repo-relative componentArea",
-        };
-      }
-      if (
-        !allowMultiRootAutonomyScope &&
-        declaredComponentArea &&
-        inferredComponentArea &&
-        declaredComponentArea !== inferredComponentArea
-      ) {
-        return {
-          ok: false,
-          message: "task.execute planning.targetPaths do not match autonomy componentArea",
-        };
-      }
-      const validatedScope = validateScopeInvariants(
-        componentArea,
-        normalizedTargetPaths,
-        normalizedWriteGlobs,
-        { requireWriteGlobs: false, allowMultipleComponentRoots: allowMultiRootAutonomyScope },
-      );
-      if (!validatedScope.ok) {
-        return {
-          ok: false,
-          message: `task.execute scope invariants failed: ${validatedScope.errors.join("; ")}`,
-        };
+      if (!allowMultiRootAutonomyScope && declaredComponentArea) {
+        const inferredComponentArea = deriveAutonomyComponentArea(
+          normalizedTargetPaths,
+          normalizedWriteGlobs,
+        );
+        if (inferredComponentArea && declaredComponentArea !== inferredComponentArea) {
+          return {
+            ok: false,
+            message: "task.execute planning.targetPaths do not match autonomy componentArea",
+          };
+        }
       }
     } else if (normalizedWriteGlobs.length > 0) {
       const uncoveredPaths = normalizedTargetPaths.filter(

package/runtime/sandbox/apps/workerpals/src/workerpals_main.ts

@@ -65,6 +65,8 @@ type WorkerJobResult = JobResult & {
 
 const DEFAULT_LLM_MODEL = "local-model";
 const CODEX_UNAVAILABLE_WORKER_EXIT_CODE = 86;
+const CODEX_UNAVAILABLE_DOCKER_SHUTDOWN_GRACE_MS = 5_000;
+const CODEX_UNAVAILABLE_WORKER_FORCE_EXIT_MS = 4_000;
 const CONFIG = loadPushPalsConfig();
 const LOG = new Logger("WorkerPals");
 
@@ -360,6 +362,36 @@ function shouldRecycleWorkerForCodexUnavailableFailure(
   ].some((needle) => text.includes(needle));
 }
 
+async function shutdownDockerExecutorBeforeCodexRecycle(
+  dockerExecutor: DockerExecutor | null,
+): Promise<void> {
+  if (!dockerExecutor) return;
+
+  let timeout: ReturnType<typeof setTimeout> | null = null;
+  let timedOut = false;
+  try {
+    await Promise.race([
+      dockerExecutor.shutdown(),
+      new Promise<void>((resolvePromise) => {
+        timeout = setTimeout(() => {
+          timedOut = true;
+          resolvePromise();
+        }, CODEX_UNAVAILABLE_DOCKER_SHUTDOWN_GRACE_MS);
+      }),
+    ]);
+  } catch (err) {
+    console.error(`[WorkerPals] Docker shutdown cleanup failed: ${String(err)}`);
+  } finally {
+    if (timeout) clearTimeout(timeout);
+  }
+
+  if (timedOut) {
+    console.warn(
+      `[WorkerPals] Docker shutdown cleanup exceeded ${CODEX_UNAVAILABLE_DOCKER_SHUTDOWN_GRACE_MS}ms; exiting worker for Codex recycle anyway.`,
+    );
+  }
+}
+
 function parseArgs(): {
   server: string;
   pollMs: number;
@@ -1667,21 +1699,46 @@ async function workerLoop(
             }
           } finally {
             clearInterval(busyHeartbeat);
-            if (
-              !recycleWorkerAfterJob &&
-              job.sessionId &&
-              result?.cooldownMs &&
-              result.cooldownMs > 0
-            ) {
-              await transport.queueSessionCommand(job.sessionId, {
-                type: "assistant_message",
-                payload: {
-                  text: `WorkerPal is cooling down for ${formatDurationMs(result.cooldownMs)} after transient infrastructure failures.`,
+            if (recycleWorkerAfterJob) {
+              runtimeState.shutdownRequested = true;
+              const forceExitTimer = setTimeout(() => {
+                console.warn(
+                  `[WorkerPals] Forcing worker recycle ${CODEX_UNAVAILABLE_WORKER_FORCE_EXIT_MS}ms after Codex backend failure.`,
+                );
+                process.exit(CODEX_UNAVAILABLE_WORKER_EXIT_CODE);
+              }, CODEX_UNAVAILABLE_WORKER_FORCE_EXIT_MS);
+              try {
+                await maybeHeartbeat("offline", null, true);
+                if (directWorktreePath) {
+                  await removeIsolatedWorktree(opts.repo, directWorktreePath).catch((err) => {
+                    console.error(
+                      `[WorkerPals] Failed to remove isolated worktree before Codex recycle: ${String(
+                        err,
+                      )}`,
+                    );
+                  });
+                  directWorktreePath = null;
+                }
+                await shutdownDockerExecutorBeforeCodexRecycle(dockerExecutor);
+              } finally {
+                clearTimeout(forceExitTimer);
+                process.exit(CODEX_UNAVAILABLE_WORKER_EXIT_CODE);
+              }
+            }
+            if (job.sessionId && result?.cooldownMs && result.cooldownMs > 0) {
+              await transport.queueSessionCommand(
+                job.sessionId,
+                {
+                  type: "assistant_message",
+                  payload: {
+                    text: `WorkerPal is cooling down for ${formatDurationMs(result.cooldownMs)} after transient infrastructure failures.`,
+                  },
+                  from: `worker:${opts.workerId}`,
                 },
-                from: `worker:${opts.workerId}`,
-              }, { priority: "high" });
+                { priority: "high" },
+              );
             }
-            if (!recycleWorkerAfterJob && result?.cooldownMs && result.cooldownMs > 0) {
+            if (result?.cooldownMs && result.cooldownMs > 0) {
               const cooldownMs = Math.max(0, Math.floor(result.cooldownMs));
               console.warn(
                 `[WorkerPals] Entering cooldown for ${formatDurationMs(cooldownMs)} after retry exhaustion.`,
@@ -1697,18 +1754,6 @@ async function workerLoop(
                 console.error(`[WorkerPals] Failed to remove isolated worktree: ${String(err)}`);
               });
             }
-            if (recycleWorkerAfterJob) {
-              runtimeState.shutdownRequested = true;
-              await maybeHeartbeat("offline", null, true);
-              if (dockerExecutor) {
-                try {
-                  await dockerExecutor.shutdown();
-                } catch (err) {
-                  console.error(`[WorkerPals] Docker shutdown cleanup failed: ${String(err)}`);
-                }
-              }
-              process.exit(CODEX_UNAVAILABLE_WORKER_EXIT_CODE);
-            }
           }
         }
       }

package/runtime/sandbox/packages/shared/src/autonomy_policy.ts

@@ -319,7 +319,11 @@ export function validateScopeInvariants(
   componentArea: AutonomyComponentArea | null | undefined,
   targetPathsInput: unknown[],
   writeGlobsInput: unknown[],
-  options?: { requireWriteGlobs?: boolean; allowMultipleComponentRoots?: boolean },
+  options?: {
+    requireWriteGlobs?: boolean;
+    allowMultipleComponentRoots?: boolean;
+    hintsOnly?: boolean;
+  },
 ): ScopeValidationResult {
   const errors: string[] = [];
   const scopeSeeds = collectScopeSeedPaths(targetPathsInput, writeGlobsInput);
@@ -327,7 +331,8 @@ export function validateScopeInvariants(
     normalizeAutonomyComponentArea(componentArea) ??
     deriveAutonomyComponentArea(targetPathsInput, writeGlobsInput);
   const allowMultipleComponentRoots = options?.allowMultipleComponentRoots === true;
-  if (!normalizedComponentArea && scopeSeeds.length > 1 && !allowMultipleComponentRoots) {
+  const hintsOnly = options?.hintsOnly === true;
+  if (!hintsOnly && !normalizedComponentArea && scopeSeeds.length > 1 && !allowMultipleComponentRoots) {
     errors.push(
       `scope spans multiple component roots: ${scopeSeeds.slice(0, 6).join(", ")}`,
     );
@@ -341,7 +346,7 @@ export function validateScopeInvariants(
       errors.push(`invalid target_path: ${String(raw ?? "")}`);
       continue;
     }
-    if (rootPrefix && !underRoot(normalized, rootPrefix)) {
+    if (!hintsOnly && rootPrefix && !underRoot(normalized, rootPrefix)) {
       errors.push(`target_path outside component root: ${normalized}`);
       continue;
     }
@@ -362,20 +367,21 @@ export function validateScopeInvariants(
       errors.push(`invalid write_glob: ${String(raw ?? "")}`);
       continue;
     }
-    if (hasForbiddenBroadGlob(normalized)) {
+    if (!hintsOnly && hasForbiddenBroadGlob(normalized)) {
       errors.push(`forbidden broad write_glob: ${normalized}`);
       continue;
     }
     const prefix = literalPrefix(normalized);
-    if (!prefix) {
+    if (!hintsOnly && !prefix) {
       errors.push(`write_glob literal prefix cannot be empty: ${normalized}`);
       continue;
     }
-    if (rootPrefix && !underRoot(prefix, rootPrefix)) {
+    if (!hintsOnly && rootPrefix && !underRoot(prefix, rootPrefix)) {
       errors.push(`write_glob outside component root: ${normalized}`);
       continue;
     }
     if (
+      !hintsOnly &&
       !normalizedTargetPaths.some(
         (targetPath) => targetPath === prefix || targetPath.startsWith(`${prefix}/`),
       )
@@ -393,13 +399,13 @@ export function validateScopeInvariants(
     errors.push("write_globs must be provided and non-empty");
   }
 
-  if (normalizedTargetPaths.length > 0 && normalizedWriteGlobs.length > 0) {
+  if (!hintsOnly && normalizedTargetPaths.length > 0 && normalizedWriteGlobs.length > 0) {
     for (const targetPath of normalizedTargetPaths) {
       const covered = normalizedWriteGlobs.some((glob) => matchesGlob(targetPath, glob));
       if (!covered) errors.push(`target_path not covered by write_globs: ${targetPath}`);
     }
   }
-  if (!normalizedComponentArea && !allowMultipleComponentRoots) {
+  if (!hintsOnly && !normalizedComponentArea && !allowMultipleComponentRoots) {
     errors.push("component_area could not be derived from scope");
   }
 

package/runtime/sandbox/packages/shared/src/communication.ts

@@ -53,6 +53,7 @@ export interface CommunicationManagerOptions {
   sessionId: string;
   from: string;
   authToken?: string | null;
+  fetchImpl?: typeof fetch;
 }
 
 export class CommunicationManager {
@@ -60,12 +61,14 @@ export class CommunicationManager {
   private readonly sessionId: string;
   private readonly from: string;
   private readonly authToken: string | null;
+  private readonly fetchImpl: typeof fetch;
 
   constructor(opts: CommunicationManagerOptions) {
     this.serverUrl = opts.serverUrl;
     this.sessionId = opts.sessionId;
     this.from = opts.from;
     this.authToken = opts.authToken ?? null;
+    this.fetchImpl = opts.fetchImpl ?? fetch;
   }
 
   private headers(): Record<string, string> {
@@ -121,7 +124,7 @@ export class CommunicationManager {
       if (meta.turnId) body.turnId = meta.turnId;
       if (meta.parentId) body.parentId = meta.parentId;
 
-      const response = await fetch(this.commandUrl(sessionId), {
+      const response = await this.fetchImpl(this.commandUrl(sessionId), {
         method: "POST",
         headers: this.headers(),
         body: JSON.stringify(body),

package/runtime/sandbox/packages/shared/src/config.ts

@@ -1929,7 +1929,7 @@ export function loadPushPalsConfig(options: LoadOptions = {}): PushPalsConfig {
         ),
         allowReadAnywhere:
           parseBoolEnv("REMOTEBUDDY_AUTONOMY_ALLOW_READ_ANYWHERE") ??
-          asBoolean(remoteAutonomyNode.allow_read_anywhere, false),
+          asBoolean(remoteAutonomyNode.allow_read_anywhere, true),
         prFeedbackCommentRows: Math.max(
           1,
           Math.min(

package/runtime/sandbox/prompts/workerpals/miniswe_completion_requirement.md

@@ -1 +1 @@
-Completion requirement: handle all requested edits across all explicit target paths before setting done=true.
+Completion requirement: solve the requested task before setting done=true. Target paths are relevance hints; edit other behavior-owning files when needed and explain why.

package/runtime/sandbox/prompts/workerpals/miniswe_explicit_targets_block.md

@@ -1,2 +1,2 @@
-Explicit target paths:
+Target path hints:
 {{targets_block}}

package/runtime/sandbox/prompts/workerpals/openai_codex_task_execute_system_prompt.md

@@ -11,7 +11,10 @@ Non-negotiable runtime invariants:
 
 Execution rules:
 
-- Keep edits minimal, correct, and scoped to the requested task.
+- Keep edits minimal, correct, and relevant to the requested task.
+- You have repo-wide read/write access inside an isolated WorkerPal sandbox. Target paths and write globs are starting-point/relevance hints, not hard write boundaries.
+- If the hinted file is a thin wrapper or the behavior lives elsewhere, edit the behavior-owning file(s) needed to solve the task and explain the scope expansion in your final response.
+- Avoid irrelevant sprawl; the review agent will judge whether changed files are necessary for the requested outcome.
 - Read relevant files before editing, then run focused validation.
 - Use direct commands without shell wrappers. Prefer plain commands like `git diff -- path`, `git add <path>`, `git status --porcelain`, and `pwd`.
 - Do not wrap commands in `/bin/bash -lc`, `sh -lc`, `cmd /c`, or `powershell -Command`, and avoid pipelines, `awk`, heredocs, or multi-command shell snippets unless they are truly unavoidable.

package/runtime/sandbox/prompts/workerpals/openhands_minimal_system_prompt.j2

@@ -1,10 +1,12 @@
 You are PushPals WorkerPal running inside OpenHands.
 
-Operate with strict scope:
+Operate with strict relevance:
 - Focus only on the current user task.
 - If the task is a question, answer directly and do not edit files.
 - If code/file changes are requested, implement them end-to-end.
 - Reuse existing repository conventions and avoid unrelated refactors/docs.
+- Treat target paths and write globs as starting-point/relevance hints, not hard write boundaries.
+- If the behavior lives outside the hinted files, edit the owning files and explain the scope expansion.
 
 Execution loop:
 - Prefer concrete execution over planning chatter.

package/runtime/sandbox/prompts/workerpals/openhands_task_execute_system_prompt.md

@@ -3,7 +3,8 @@ You are PushPals WorkerPal running inside OpenHands.
 Execution rules:
 
 - Focus only on the task below.
-- Keep changes minimal, correct, and scoped to the request.
+- Keep changes minimal, correct, and relevant to the request.
+- Target paths and write globs are starting-point/relevance hints, not hard write boundaries; if the behavior lives elsewhere, edit the owning files and explain why.
 - Read relevant files before editing.
 - Reuse existing project conventions and tooling.
 - If the task is a question/explanation, answer directly and do not edit files.

package/runtime/sandbox/prompts/workerpals/workerpals_system_prompt.md

@@ -6,7 +6,7 @@ Your mission:
 
 - Take the user (or RemoteBuddy) request and fully execute it end-to-end.
 - You are responsible for breaking the work down into concrete subtasks, completing them, validating, reviewing your own changes, and preparing a high-quality commit message when the work is ready.
-- You have full read/write access within the assigned repository only; do not access files outside that repository.
+- You have full read/write access within the assigned repository sandbox only; do not access files outside that repository. Target paths and write globs are starting-point/relevance hints, not hard write boundaries.
 
 Mindset:
 
@@ -64,7 +64,7 @@ Execution workflow (you MUST follow this):
      - Performance: no unnecessary work on UI thread, no extra network calls, no large bundles
      - Cross-platform: iOS/Android/Web differences guarded appropriately
      - Security: no secret leakage, safe networking defaults, no unsafe shell usage
-   - Make any final polish edits that improve clarity without changing scope.
+   - Make any final polish edits that improve clarity without drifting from the task.
 
 7. Prepare to commit (when appropriate)
    - When the work is ready, produce a detailed commit message (do NOT actually commit unless your system explicitly allows it).