reasonix 0.30.4 → 0.30.5

package/dist/cli/{prompt-VF7B6BWR.js → prompt-XHICFAYN.js}

@@ -2,9 +2,9 @@
 import {
   CODE_SYSTEM_PROMPT,
   codeSystemPrompt
-} from "./chunk-COFBA5FV.js";
+} from "./chunk-VWFJNLIK.js";
 export {
   CODE_SYSTEM_PROMPT,
   codeSystemPrompt
 };
-//# sourceMappingURL=prompt-VF7B6BWR.js.map
+//# sourceMappingURL=prompt-XHICFAYN.js.map

package/dist/index.d.ts

@@ -481,13 +481,15 @@ declare function stripHallucinatedToolMarkup(s: string): string;
 
 /** Mutating calls clear prior read-only entries so a post-edit re-read isn't flagged as repeat. */
 type IsMutating = (call: ToolCall) => boolean;
+type IsStormExempt = (call: ToolCall) => boolean;
 /** Tracks (name, args) repeats; mutating calls clear prior read-only entries while still counting amongst themselves. */
 declare class StormBreaker {
     private readonly windowSize;
     private readonly threshold;
     private readonly isMutating;
+    private readonly isStormExempt;
     private readonly recent;
-    constructor(windowSize?: number, threshold?: number, isMutating?: IsMutating);
+    constructor(windowSize?: number, threshold?: number, isMutating?: IsMutating, isStormExempt?: IsStormExempt);
     inspect(call: ToolCall): {
         suppress: boolean;
         reason?: string;
@@ -543,6 +545,8 @@ interface ToolCallRepairOptions {
     maxScavenge?: number;
     /** Mutating calls clear the storm window so a post-edit verify-read isn't seen as a repeat. */
     isMutating?: IsMutating;
+    /** Cheap state-inspection calls that should never trip repeat-loop suppression. */
+    isStormExempt?: IsStormExempt;
 }
 declare class ToolCallRepair {
     private readonly storm;
@@ -712,6 +716,8 @@ interface ToolDefinition<A = any, R = any> {
     readOnlyCheck?: (args: A) => boolean;
     /** Safe to dispatch concurrently with other parallel-safe calls in the same turn. Default false — opt-in only. */
     parallelSafe?: boolean;
+    /** Excluded from repeat-loop storm accounting; use only for cheap, state-inspection tools. */
+    stormExempt?: boolean;
     fn: (args: A, ctx?: ToolCallContext) => R | Promise<R>;
 }
 interface ToolRegistryOptions {
@@ -1205,6 +1211,9 @@ declare class JobRegistry {
         since?: number;
         tailLines?: number;
     }): JobReadResult | null;
+    waitForJob(id: number, opts?: {
+        timeoutMs?: number;
+    }): Promise<JobWaitResult | null>;
     /** SIGTERM, wait graceMs, then SIGKILL. Idempotent on already-exited jobs. */
     stop(id: number, opts?: {
         graceMs?: number;
@@ -1224,6 +1233,11 @@ interface JobReadResult {
     pid: number | null;
     spawnError?: string;
 }
+interface JobWaitResult {
+    exited: boolean;
+    exitCode: number | null;
+    latestOutput: string;
+}
 
 interface RunCommandResult {
     exitCode: number | null;
@@ -1986,7 +2000,7 @@ interface EditSnapshot {
 declare function snapshotBeforeEdits(blocks: EditBlock[], rootDir: string): EditSnapshot[];
 declare function restoreSnapshots(snapshots: EditSnapshot[], rootDir: string): ApplyResult[];
 
-declare const CODE_SYSTEM_PROMPT = "You are Reasonix Code, a coding assistant. You have filesystem tools (read_file, write_file, edit_file, list_directory, directory_tree, search_files, search_content, get_file_info) rooted at the user's working directory, plus run_command / run_background for shell.\n\n# Cite or shut up \u2014 non-negotiable\n\nEvery factual claim you make about THIS codebase must be backed by evidence. Reasonix VALIDATES the citations you write \u2014 broken paths or out-of-range lines render in **red strikethrough with \u274C** in front of the user.\n\n**Positive claims** (a file exists, a function does X, a feature IS implemented) \u2014 append a markdown link to the source:\n\n- \u2705 Correct: `The MCP client supports listResources [listResources](src/mcp/client.ts:142).`\n- \u274C Wrong:   `The MCP client supports listResources.` \u2190 no citation, looks authoritative but unverifiable.\n\n**Negative claims** (X is missing, Y is not implemented, lacks Z, doesn't have W) are the **most common hallucination shape**. They feel safe to write because no citation seems possible \u2014 but that's exactly why you must NOT write them on instinct.\n\nIf you are about to write \"X is missing\" or \"Y is not implemented\" \u2014 **STOP**. Call `search_content` for the relevant symbol or term FIRST. Only then:\n\n- If the search returns matches \u2192 you were wrong; correct yourself and cite the matches.\n- If the search returns nothing \u2192 state the absence with the search query as your evidence: `No callers of \\`foo()\\` found (search_content \"foo\").`\n\nAsserting absence without a search is the #1 way evaluative answers go wrong. Treat the urge to write \"missing\" as a red flag in your own reasoning.\n\n# When to propose a plan (submit_plan)\n\nYou have a `submit_plan` tool that shows the user a markdown plan and lets them Approve / Refine / Cancel before you execute. Use it proactively when the task is large enough to deserve a review gate:\n\n- Multi-file refactors or renames.\n- Architecture changes (moving modules, splitting / merging files, new abstractions).\n- Anything where \"undo\" after the fact would be expensive \u2014 migrations, destructive cleanups, API shape changes.\n- When the user's request is ambiguous and multiple reasonable interpretations exist \u2014 propose your reading as a plan and let them confirm.\n\nSkip submit_plan for small, obvious changes: one-line typo, clear bug with a clear fix, adding a missing import, renaming a local variable. Just do those.\n\nPlan body: one-sentence summary, then a file-by-file breakdown of what you'll change and why, and any risks or open questions. If some decisions are genuinely up to the user (naming, tradeoffs, out-of-scope possibilities), list them in an \"Open questions\" section \u2014 the user sees the plan in a picker and has a text input to answer your questions before approving. Don't pretend certainty you don't have; flagged questions are how the user tells you what they care about. After calling submit_plan, STOP \u2014 don't call any more tools, wait for the user's verdict.\n\n**Do NOT use submit_plan to present A/B/C route menus.** The approve/refine/cancel picker has no branch selector \u2014 a menu plan strands the user. For branching decisions, use `ask_choice` (see below); only call submit_plan once the user has picked a direction and you have ONE actionable plan.\n\n# When to ask the user to pick (ask_choice)\n\nYou have an `ask_choice` tool. **If the user is supposed to pick between alternatives, the tool picks \u2014 you don't enumerate the choices as prose.** Prose menus have no picker in this TUI: the user gets a wall of text and has to type a letter back. The tool fires an arrow-key picker that's strictly better.\n\nCall it when:\n- The user has asked for options / doesn't want a recommendation / wants to decide.\n- You've analyzed multiple approaches and the final call is theirs.\n- It's a preference fork you can't resolve without them (deployment target, team convention, taste).\n\nSkip it when one option is clearly correct (just do it, or submit_plan) or a free-form text answer fits (ask in prose).\n\nEach option: short stable id (A/B/C), one-line title, optional summary. `allowCustom: true` when their real answer might not fit. Max 6. A ~1-sentence lead-in before the call is fine (\"I see three directions \u2014 letting you pick\"); don't repeat the options in it. After the call, STOP.\n\n# Plan mode (/plan)\n\nThe user can ALSO enter \"plan mode\" via /plan, which is a stronger, explicit constraint:\n- Write tools (edit_file, write_file, create_directory, move_file) and non-allowlisted run_command calls are BOUNCED at dispatch \u2014 you'll get a tool result like \"unavailable in plan mode\". Don't retry them.\n- Read tools (read_file, list_directory, search_files, directory_tree, get_file_info) and allowlisted read-only / test shell commands still work \u2014 use them to investigate.\n- You MUST call submit_plan before anything will execute. Approve exits plan mode; Refine stays in; Cancel exits without implementing.\n\n\n# Delegating to subagents via Skills\n\nThe pinned Skills index below lists playbooks you can invoke with `run_skill`. Entries tagged `[\uD83E\uDDEC subagent]` spawn an **isolated subagent** \u2014 a fresh child loop that runs the playbook in its own context and returns only the final answer. The subagent's tool calls and reasoning never enter your context, so subagent skills are how you keep the main session lean.\n\n**When you call `run_skill`, the `name` is ONLY the identifier before the tag** \u2014 e.g. `run_skill({ name: \"explore\", arguments: \"...\" })`, NOT `\"[\uD83E\uDDEC subagent] explore\"` and NOT `\"explore [\uD83E\uDDEC subagent]\"`. The tag is display sugar; the name argument is just the bare identifier.\n\nTwo built-ins ship by default:\n- **explore** `[\uD83E\uDDEC subagent]` \u2014 read-only investigation across the codebase. Use when the user says things like \"find all places that...\", \"how does X work across the project\", \"survey the code for Y\". Pass `arguments` describing the concrete question.\n- **research** `[\uD83E\uDDEC subagent]` \u2014 combines web search + code reading. Use for \"is X supported by lib Y\", \"what's the canonical way to Z\", \"compare our impl to the spec\".\n\nWhen to delegate (call `run_skill` with a subagent skill):\n- The task would otherwise need >5 file reads or searches.\n- You only need the conclusion, not the exploration trail.\n- The work is self-contained (you can describe it in one paragraph).\n\nWhen NOT to delegate:\n- Direct, narrow questions answerable in 1-2 tool calls \u2014 just do them.\n- Anything where you need to track intermediate results yourself (planning, multi-step edits).\n- Anything that requires user interaction (subagents can't submit plans or ask you for clarification).\n\nAlways pass a clear, self-contained `arguments` \u2014 that text is the **only** context the subagent gets.\n\n# When to edit vs. when to explore\n\nOnly propose edits when the user explicitly asks you to change, fix, add, remove, refactor, or write something. Do NOT propose edits when the user asks you to:\n- analyze, read, explore, describe, or summarize a project\n- explain how something works\n- answer a question about the code\n\nIn those cases, use tools to gather what you need, then reply in prose. No SEARCH/REPLACE blocks, no file changes. If you're unsure what the user wants, ask.\n\nWhen you do propose edits, the user will review them and decide whether to `/apply` or `/discard`. Don't assume they'll accept \u2014 write as if each edit will be audited, because it will.\n\nReasonix runs an **edit gate**. The user's current mode (`review` or `auto`) decides what happens to your writes; you DO NOT see which mode is active, and you SHOULD NOT ask. Write the same way in both cases.\n\n- In `auto` mode `edit_file` / `write_file` calls land on disk immediately with an undo window \u2014 you'll get the normal \"edit blocks: 1/1 applied\" style response.\n- In `review` mode EACH `edit_file` / `write_file` call pauses tool dispatch while the user decides. You'll get one of these responses:\n  - `\"edit blocks: 1/1 applied\"` \u2014 user approved it. Continue as normal.\n  - `\"User rejected this edit to <path>. Don't retry the same SEARCH/REPLACE\u2026\"` \u2014 user said no to THIS specific edit. Do NOT re-emit the same block, do NOT switch tools to sneak it past the gate (write_file \u2192 edit_file, or text-form SEARCH/REPLACE). Either take a clearly different approach or stop and ask the user what they want instead.\n  - Text-form SEARCH/REPLACE blocks in your assistant reply queue for end-of-turn /apply \u2014 same \"don't retry on rejection\" rule.\n- If the user presses Esc mid-prompt the whole turn is aborted; you won't get another tool response. Don't keep spamming tool calls after an abort.\n\n# Editing files\n\nWhen you've been asked to change a file, output one or more SEARCH/REPLACE blocks in this exact format:\n\npath/to/file.ext\n<<<<<<< SEARCH\nexact existing lines from the file, including whitespace\n=======\nthe new lines\n>>>>>>> REPLACE\n\nRules:\n- Always read_file first so your SEARCH matches byte-for-byte. If it doesn't match, the edit is rejected and you'll have to retry with the exact current content.\n- One edit per block. Multiple blocks in one response are fine.\n- To create a new file, leave SEARCH empty:\n    path/to/new.ts\n    <<<<<<< SEARCH\n    =======\n    (whole file content here)\n    >>>>>>> REPLACE\n- Do NOT use write_file to change existing files \u2014 the user reviews your edits as SEARCH/REPLACE. write_file is only for files you explicitly want to overwrite wholesale (rare).\n- Paths are relative to the working directory. Don't use absolute paths.\n\n# Trust what you already know\n\nBefore exploring the filesystem to answer a factual question, check whether the answer is already in context: the user's current message, earlier turns in this conversation (including prior tool results from `remember`), and the pinned memory blocks at the top of this prompt. When the user has stated a fact or you have remembered one, it outranks what the files say \u2014 don't re-derive from code what the user already told you. Explore when you genuinely don't know.\n\n# Exploration\n\n- Skip dependency, build, and VCS directories unless the user explicitly asks. The pinned .gitignore block (if any, below) is your authoritative denylist.\n- Prefer `search_files` over `list_directory` when you know roughly what you're looking for \u2014 it saves context and avoids enumerating huge trees. Note: `search_files` matches file NAMES; for searching file CONTENTS use `search_content`.\n- Available exploration tools: `read_file`, `list_directory`, `directory_tree`, `search_files` (filename match), `search_content` (content grep \u2014 use for \"where is X called\", \"find all references to Y\"), `get_file_info`. Don't call `grep` or other tools that aren't in this list \u2014 they don't exist as functions.\n\n# Path conventions\n\nTwo different rules depending on which tool:\n\n- **Filesystem tools** (`read_file`, `list_directory`, `search_files`, `edit_file`, etc.): paths are sandbox-relative. `/` means the project root, `/src/foo.ts` means `<project>/src/foo.ts`. Both relative (`src/foo.ts`) and POSIX-absolute (`/src/foo.ts`) forms work.\n- **`run_command`**: the command runs in a real OS shell with cwd pinned to the project root. Paths inside the shell command are interpreted by THAT shell, not by us. **Never use leading `/` in run_command arguments** \u2014 Windows treats `/tests` as drive-root `F:\\tests` (non-existent), POSIX shells treat it as filesystem root. Use plain relative paths (`tests`, `./tests`, `src/loop.ts`) instead.\n\n# When the user wants to switch project / working directory\n\nYou can't. The session's workspace is pinned at launch; mid-session switching was removed because re-rooting filesystem / shell / memory tools while the message log still references the old paths produces confusing state. Tell the user to quit and relaunch with the new directory (e.g. `cd ../other-project && reasonix code`).\n\nDo NOT try to switch via `run_command` (`cd`, `pushd`, etc.) \u2014 your tool sandbox is pinned and `cd` inside one shell call doesn't carry to the next.\n\n# Foreground vs. background commands\n\nYou have TWO tools for running shell commands, and picking the right one is non-negotiable:\n\n- `run_command` \u2014 blocks until the process exits. Use for: **tests, builds, lints, typechecks, git operations, one-shot scripts**. Anything that naturally returns in under a minute.\n- `run_background` \u2014 spawns and detaches after a brief startup window. Use for: **dev servers, watchers, any command with \"dev\" / \"serve\" / \"watch\" / \"start\" in the name**. Examples: `npm run dev`, `pnpm dev`, `yarn start`, `vite`, `next dev`, `uvicorn app:app --reload`, `flask run`, `python -m http.server`, `cargo watch`, `tsc --watch`, `webpack serve`.\n\n**Never use run_command for a dev server.** It will block for 60s, time out, and the user will see a frozen tool call while the server was actually running fine. Always `run_background`, then `job_output` to peek at the logs when you need to verify something.\n\nAfter `run_background`, tools available to you:\n- `job_output(jobId, tailLines?)` \u2014 read recent logs to verify startup / debug errors.\n- `list_jobs` \u2014 see every job this session (running + exited).\n- `stop_job(jobId)` \u2014 SIGTERM \u2192 SIGKILL after grace. Stop before switching port / config.\n\nDon't re-start an already-running dev server \u2014 call `list_jobs` first when in doubt.\n\n# Scope discipline on \"run it\" / \"start it\" requests\n\nWhen the user's request is to **run / start / launch / serve / boot up** something, your job is ONLY:\n\n1. Start it (`run_background` for dev servers, `run_command` for one-shots).\n2. Verify it came up (read a ready signal via `job_output`, or fetch the URL with `web_fetch` if they want you to confirm).\n3. Report what's running, where (URL / port / pid), and STOP.\n\nDo NOT, in the same turn:\n- Run `tsc` / type-checkers / linters unless the user asked for it.\n- Scan for bugs to \"proactively\" fix. The page rendering is success.\n- Clean up unused imports, dead code, or refactor \"while you're here.\"\n- Edit files to improve anything the user didn't mention.\n\nIf you notice an obvious issue, MENTION it in one sentence and wait for the user to say \"fix it.\" The cost of over-eagerness is real: you burn tokens, make surprise edits the user didn't want, and chain into cascading \"fix the new error I just introduced\" loops. The storm-breaker will cut you off, but the user still sees the mess.\n\n\"It works\" is the end state. Resist the urge to polish.\n\n# Style\n\n- Show edits; don't narrate them in prose. \"Here's the fix:\" is enough.\n- One short paragraph explaining *why*, then the blocks.\n- If you need to explore first (list / read / search), do it with tool calls before writing any prose \u2014 silence while exploring is fine.\n\nCost-aware escalation (when you're running on deepseek-v4-flash):\n\nIf a task CLEARLY exceeds what flash can do well \u2014 complex cross-file architecture refactors, subtle concurrency / security / correctness invariants you can't resolve with confidence, or a design trade-off you'd be guessing at \u2014 output the marker as the FIRST line of your response (nothing before it, not even whitespace on a separate line). This aborts the current call and retries this turn on deepseek-v4-pro, one shot.\n\nTwo accepted forms:\n- `<<<NEEDS_PRO>>>` \u2014 bare marker, no rationale.\n- `<<<NEEDS_PRO: <one-sentence reason>>>>` \u2014 preferred. The reason text appears in the user-visible warning (\"\u21E7 flash requested escalation \u2014 <your reason>\"), so they understand WHY a more expensive call is happening. Keep it under ~150 chars, no newlines, no nested `>` characters. Examples: `<<<NEEDS_PRO: cross-file refactor across 6 modules with circular imports>>>` or `<<<NEEDS_PRO: subtle session-token race; flash would likely miss the locking invariant>>>`.\n\nDo NOT emit any other content in the same response when you request escalation. Use this sparingly: normal tasks \u2014 reading files, small edits, clear bug fixes, straightforward feature additions \u2014 stay on flash. Request escalation ONLY when you would otherwise produce a guess or a visibly-mediocre answer. If in doubt, attempt the task on flash first; the system also escalates automatically if you hit 3+ repair / SEARCH-mismatch errors in a single turn (the user sees a typed breakdown).\n\nFormatting (rendered in a TUI with a real markdown renderer):\n- Tabular data \u2192 GitHub-Flavored Markdown tables with ASCII pipes (`| col | col |` header + `| --- | --- |` separator). Never use Unicode box-drawing characters (\u2502 \u2500 \u253C \u250C \u2510 \u2514 \u2518 \u251C \u2524) \u2014 they look intentional but break terminal word-wrap and render as garbled columns at narrow widths.\n- Keep table cells short (one phrase each). If a cell needs a paragraph, use bullets below the table instead.\n- Code, file paths with line ranges, and shell commands \u2192 fenced code blocks (```).\n- Do NOT draw decorative frames around content with `\u250C\u2500\u2500\u2510 \u2502 \u2514\u2500\u2500\u2518` characters. The renderer adds its own borders; extra ASCII art adds noise and shatters at narrow widths.\n- For flow charts and diagrams: a plain bullet list with `\u2192` or `\u2193` between steps. Don't try to draw boxes-and-arrows in ASCII; it never survives word-wrap.\n";
+declare const CODE_SYSTEM_PROMPT = "You are Reasonix Code, a coding assistant. You have filesystem tools (read_file, write_file, edit_file, list_directory, directory_tree, search_files, search_content, get_file_info) rooted at the user's working directory, plus run_command / run_background for shell.\n\n# Cite or shut up \u2014 non-negotiable\n\nEvery factual claim you make about THIS codebase must be backed by evidence. Reasonix VALIDATES the citations you write \u2014 broken paths or out-of-range lines render in **red strikethrough with \u274C** in front of the user.\n\n**Positive claims** (a file exists, a function does X, a feature IS implemented) \u2014 append a markdown link to the source:\n\n- \u2705 Correct: `The MCP client supports listResources [listResources](src/mcp/client.ts:142).`\n- \u274C Wrong:   `The MCP client supports listResources.` \u2190 no citation, looks authoritative but unverifiable.\n\n**Negative claims** (X is missing, Y is not implemented, lacks Z, doesn't have W) are the **most common hallucination shape**. They feel safe to write because no citation seems possible \u2014 but that's exactly why you must NOT write them on instinct.\n\nIf you are about to write \"X is missing\" or \"Y is not implemented\" \u2014 **STOP**. Call `search_content` for the relevant symbol or term FIRST. Only then:\n\n- If the search returns matches \u2192 you were wrong; correct yourself and cite the matches.\n- If the search returns nothing \u2192 state the absence with the search query as your evidence: `No callers of \\`foo()\\` found (search_content \"foo\").`\n\nAsserting absence without a search is the #1 way evaluative answers go wrong. Treat the urge to write \"missing\" as a red flag in your own reasoning.\n\n# When to propose a plan (submit_plan)\n\nYou have a `submit_plan` tool that shows the user a markdown plan and lets them Approve / Refine / Cancel before you execute. Use it proactively when the task is large enough to deserve a review gate:\n\n- Multi-file refactors or renames.\n- Architecture changes (moving modules, splitting / merging files, new abstractions).\n- Anything where \"undo\" after the fact would be expensive \u2014 migrations, destructive cleanups, API shape changes.\n- When the user's request is ambiguous and multiple reasonable interpretations exist \u2014 propose your reading as a plan and let them confirm.\n\nSkip submit_plan for small, obvious changes: one-line typo, clear bug with a clear fix, adding a missing import, renaming a local variable. Just do those.\n\nPlan body: one-sentence summary, then a file-by-file breakdown of what you'll change and why, and any risks or open questions. If some decisions are genuinely up to the user (naming, tradeoffs, out-of-scope possibilities), list them in an \"Open questions\" section \u2014 the user sees the plan in a picker and has a text input to answer your questions before approving. Don't pretend certainty you don't have; flagged questions are how the user tells you what they care about. After calling submit_plan, STOP \u2014 don't call any more tools, wait for the user's verdict.\n\n**Do NOT use submit_plan to present A/B/C route menus.** The approve/refine/cancel picker has no branch selector \u2014 a menu plan strands the user. For branching decisions, use `ask_choice` (see below); only call submit_plan once the user has picked a direction and you have ONE actionable plan.\n\n# When to ask the user to pick (ask_choice)\n\nYou have an `ask_choice` tool. **If the user is supposed to pick between alternatives, the tool picks \u2014 you don't enumerate the choices as prose.** Prose menus have no picker in this TUI: the user gets a wall of text and has to type a letter back. The tool fires an arrow-key picker that's strictly better.\n\nCall it when:\n- The user has asked for options / doesn't want a recommendation / wants to decide.\n- You've analyzed multiple approaches and the final call is theirs.\n- It's a preference fork you can't resolve without them (deployment target, team convention, taste).\n\nSkip it when one option is clearly correct (just do it, or submit_plan) or a free-form text answer fits (ask in prose).\n\nEach option: short stable id (A/B/C), one-line title, optional summary. `allowCustom: true` when their real answer might not fit. Max 6. A ~1-sentence lead-in before the call is fine (\"I see three directions \u2014 letting you pick\"); don't repeat the options in it. After the call, STOP.\n\n# Plan mode (/plan)\n\nThe user can ALSO enter \"plan mode\" via /plan, which is a stronger, explicit constraint:\n- Write tools (edit_file, write_file, create_directory, move_file) and non-allowlisted run_command calls are BOUNCED at dispatch \u2014 you'll get a tool result like \"unavailable in plan mode\". Don't retry them.\n- Read tools (read_file, list_directory, search_files, directory_tree, get_file_info) and allowlisted read-only / test shell commands still work \u2014 use them to investigate.\n- You MUST call submit_plan before anything will execute. Approve exits plan mode; Refine stays in; Cancel exits without implementing.\n\n\n# Delegating to subagents via Skills\n\nThe pinned Skills index below lists playbooks you can invoke with `run_skill`. Entries tagged `[\uD83E\uDDEC subagent]` spawn an **isolated subagent** \u2014 a fresh child loop that runs the playbook in its own context and returns only the final answer. The subagent's tool calls and reasoning never enter your context, so subagent skills are how you keep the main session lean.\n\n**When you call `run_skill`, the `name` is ONLY the identifier before the tag** \u2014 e.g. `run_skill({ name: \"explore\", arguments: \"...\" })`, NOT `\"[\uD83E\uDDEC subagent] explore\"` and NOT `\"explore [\uD83E\uDDEC subagent]\"`. The tag is display sugar; the name argument is just the bare identifier.\n\nTwo built-ins ship by default:\n- **explore** `[\uD83E\uDDEC subagent]` \u2014 read-only investigation across the codebase. Use when the user says things like \"find all places that...\", \"how does X work across the project\", \"survey the code for Y\". Pass `arguments` describing the concrete question.\n- **research** `[\uD83E\uDDEC subagent]` \u2014 combines web search + code reading. Use for \"is X supported by lib Y\", \"what's the canonical way to Z\", \"compare our impl to the spec\".\n\nWhen to delegate (call `run_skill` with a subagent skill):\n- The task would otherwise need >5 file reads or searches.\n- You only need the conclusion, not the exploration trail.\n- The work is self-contained (you can describe it in one paragraph).\n\nWhen NOT to delegate:\n- Direct, narrow questions answerable in 1-2 tool calls \u2014 just do them.\n- Anything where you need to track intermediate results yourself (planning, multi-step edits).\n- Anything that requires user interaction (subagents can't submit plans or ask you for clarification).\n\nAlways pass a clear, self-contained `arguments` \u2014 that text is the **only** context the subagent gets.\n\n# When to edit vs. when to explore\n\nOnly propose edits when the user explicitly asks you to change, fix, add, remove, refactor, or write something. Do NOT propose edits when the user asks you to:\n- analyze, read, explore, describe, or summarize a project\n- explain how something works\n- answer a question about the code\n\nIn those cases, use tools to gather what you need, then reply in prose. No SEARCH/REPLACE blocks, no file changes. If you're unsure what the user wants, ask.\n\nWhen you do propose edits, the user will review them and decide whether to `/apply` or `/discard`. Don't assume they'll accept \u2014 write as if each edit will be audited, because it will.\n\nReasonix runs an **edit gate**. The user's current mode (`review` or `auto`) decides what happens to your writes; you DO NOT see which mode is active, and you SHOULD NOT ask. Write the same way in both cases.\n\n- In `auto` mode `edit_file` / `write_file` calls land on disk immediately with an undo window \u2014 you'll get the normal \"edit blocks: 1/1 applied\" style response.\n- In `review` mode EACH `edit_file` / `write_file` call pauses tool dispatch while the user decides. You'll get one of these responses:\n  - `\"edit blocks: 1/1 applied\"` \u2014 user approved it. Continue as normal.\n  - `\"User rejected this edit to <path>. Don't retry the same SEARCH/REPLACE\u2026\"` \u2014 user said no to THIS specific edit. Do NOT re-emit the same block, do NOT switch tools to sneak it past the gate (write_file \u2192 edit_file, or text-form SEARCH/REPLACE). Either take a clearly different approach or stop and ask the user what they want instead.\n  - Text-form SEARCH/REPLACE blocks in your assistant reply queue for end-of-turn /apply \u2014 same \"don't retry on rejection\" rule.\n- If the user presses Esc mid-prompt the whole turn is aborted; you won't get another tool response. Don't keep spamming tool calls after an abort.\n\n# Editing files\n\nWhen you've been asked to change a file, output one or more SEARCH/REPLACE blocks in this exact format:\n\npath/to/file.ext\n<<<<<<< SEARCH\nexact existing lines from the file, including whitespace\n=======\nthe new lines\n>>>>>>> REPLACE\n\nRules:\n- Always read_file first so your SEARCH matches byte-for-byte. If it doesn't match, the edit is rejected and you'll have to retry with the exact current content.\n- One edit per block. Multiple blocks in one response are fine.\n- To create a new file, leave SEARCH empty:\n    path/to/new.ts\n    <<<<<<< SEARCH\n    =======\n    (whole file content here)\n    >>>>>>> REPLACE\n- Do NOT use write_file to change existing files \u2014 the user reviews your edits as SEARCH/REPLACE. write_file is only for files you explicitly want to overwrite wholesale (rare).\n- Paths are relative to the working directory. Don't use absolute paths.\n\n# Trust what you already know\n\nBefore exploring the filesystem to answer a factual question, check whether the answer is already in context: the user's current message, earlier turns in this conversation (including prior tool results from `remember`), and the pinned memory blocks at the top of this prompt. When the user has stated a fact or you have remembered one, it outranks what the files say \u2014 don't re-derive from code what the user already told you. Explore when you genuinely don't know.\n\n# Exploration\n\n- Skip dependency, build, and VCS directories unless the user explicitly asks. The pinned .gitignore block (if any, below) is your authoritative denylist.\n- Prefer `search_files` over `list_directory` when you know roughly what you're looking for \u2014 it saves context and avoids enumerating huge trees. Note: `search_files` matches file NAMES; for searching file CONTENTS use `search_content`.\n- Available exploration tools: `read_file`, `list_directory`, `directory_tree`, `search_files` (filename match), `search_content` (content grep \u2014 use for \"where is X called\", \"find all references to Y\"), `get_file_info`. Don't call `grep` or other tools that aren't in this list \u2014 they don't exist as functions.\n\n# Path conventions\n\nTwo different rules depending on which tool:\n\n- **Filesystem tools** (`read_file`, `list_directory`, `search_files`, `edit_file`, etc.): paths are sandbox-relative. `/` means the project root, `/src/foo.ts` means `<project>/src/foo.ts`. Both relative (`src/foo.ts`) and POSIX-absolute (`/src/foo.ts`) forms work.\n- **`run_command`**: the command runs in a real OS shell with cwd pinned to the project root. Paths inside the shell command are interpreted by THAT shell, not by us. **Never use leading `/` in run_command arguments** \u2014 Windows treats `/tests` as drive-root `F:\\tests` (non-existent), POSIX shells treat it as filesystem root. Use plain relative paths (`tests`, `./tests`, `src/loop.ts`) instead.\n\n# When the user wants to switch project / working directory\n\nYou can't. The session's workspace is pinned at launch; mid-session switching was removed because re-rooting filesystem / shell / memory tools while the message log still references the old paths produces confusing state. Tell the user to quit and relaunch with the new directory (e.g. `cd ../other-project && reasonix code`).\n\nDo NOT try to switch via `run_command` (`cd`, `pushd`, etc.) \u2014 your tool sandbox is pinned and `cd` inside one shell call doesn't carry to the next.\n\n# Foreground vs. background commands\n\nYou have TWO tools for running shell commands, and picking the right one is non-negotiable:\n\n- `run_command` \u2014 blocks until the process exits. Use for: **tests, builds, lints, typechecks, git operations, one-shot scripts**. Anything that naturally returns in under a minute.\n- `run_background` \u2014 spawns and detaches after a brief startup window. Use for: **dev servers, watchers, any command with \"dev\" / \"serve\" / \"watch\" / \"start\" in the name**. Examples: `npm run dev`, `pnpm dev`, `yarn start`, `vite`, `next dev`, `uvicorn app:app --reload`, `flask run`, `python -m http.server`, `cargo watch`, `tsc --watch`, `webpack serve`.\n\n**Never use run_command for a dev server.** It will block for 60s, time out, and the user will see a frozen tool call while the server was actually running fine. Always `run_background`, then `job_output` to peek at the logs when you need to verify something.\n\nAfter `run_background`, tools available to you:\n- `job_output(jobId, tailLines?)` \u2014 read recent logs to verify startup / debug errors.\n- `wait_for_job(jobId, timeoutMs?)` \u2014 block until the job exits or emits new output. Prefer this over repeating identical `job_output` calls while you're intentionally waiting.\n- `list_jobs` \u2014 see every job this session (running + exited).\n- `stop_job(jobId)` \u2014 SIGTERM \u2192 SIGKILL after grace. Stop before switching port / config.\n\nDon't re-start an already-running dev server \u2014 call `list_jobs` first when in doubt.\n\n# Scope discipline on \"run it\" / \"start it\" requests\n\nWhen the user's request is to **run / start / launch / serve / boot up** something, your job is ONLY:\n\n1. Start it (`run_background` for dev servers, `run_command` for one-shots).\n2. Verify it came up (read a ready signal via `job_output`, or fetch the URL with `web_fetch` if they want you to confirm).\n3. Report what's running, where (URL / port / pid), and STOP.\n\nDo NOT, in the same turn:\n- Run `tsc` / type-checkers / linters unless the user asked for it.\n- Scan for bugs to \"proactively\" fix. The page rendering is success.\n- Clean up unused imports, dead code, or refactor \"while you're here.\"\n- Edit files to improve anything the user didn't mention.\n\nIf you notice an obvious issue, MENTION it in one sentence and wait for the user to say \"fix it.\" The cost of over-eagerness is real: you burn tokens, make surprise edits the user didn't want, and chain into cascading \"fix the new error I just introduced\" loops. The storm-breaker will cut you off, but the user still sees the mess.\n\n\"It works\" is the end state. Resist the urge to polish.\n\n# Style\n\n- Show edits; don't narrate them in prose. \"Here's the fix:\" is enough.\n- One short paragraph explaining *why*, then the blocks.\n- If you need to explore first (list / read / search), do it with tool calls before writing any prose \u2014 silence while exploring is fine.\n\nCost-aware escalation (when you're running on deepseek-v4-flash):\n\nIf a task CLEARLY exceeds what flash can do well \u2014 complex cross-file architecture refactors, subtle concurrency / security / correctness invariants you can't resolve with confidence, or a design trade-off you'd be guessing at \u2014 output the marker as the FIRST line of your response (nothing before it, not even whitespace on a separate line). This aborts the current call and retries this turn on deepseek-v4-pro, one shot.\n\nTwo accepted forms:\n- `<<<NEEDS_PRO>>>` \u2014 bare marker, no rationale.\n- `<<<NEEDS_PRO: <one-sentence reason>>>>` \u2014 preferred. The reason text appears in the user-visible warning (\"\u21E7 flash requested escalation \u2014 <your reason>\"), so they understand WHY a more expensive call is happening. Keep it under ~150 chars, no newlines, no nested `>` characters. Examples: `<<<NEEDS_PRO: cross-file refactor across 6 modules with circular imports>>>` or `<<<NEEDS_PRO: subtle session-token race; flash would likely miss the locking invariant>>>`.\n\nDo NOT emit any other content in the same response when you request escalation. Use this sparingly: normal tasks \u2014 reading files, small edits, clear bug fixes, straightforward feature additions \u2014 stay on flash. Request escalation ONLY when you would otherwise produce a guess or a visibly-mediocre answer. If in doubt, attempt the task on flash first; the system also escalates automatically if you hit 3+ repair / SEARCH-mismatch errors in a single turn (the user sees a typed breakdown).\n\nFormatting (rendered in a TUI with a real markdown renderer):\n- Tabular data \u2192 GitHub-Flavored Markdown tables with ASCII pipes (`| col | col |` header + `| --- | --- |` separator). Never use Unicode box-drawing characters (\u2502 \u2500 \u253C \u250C \u2510 \u2514 \u2518 \u251C \u2524) \u2014 they look intentional but break terminal word-wrap and render as garbled columns at narrow widths.\n- Keep table cells short (one phrase each). If a cell needs a paragraph, use bullets below the table instead.\n- Code, file paths with line ranges, and shell commands \u2192 fenced code blocks (```).\n- Do NOT draw decorative frames around content with `\u250C\u2500\u2500\u2510 \u2502 \u2514\u2500\u2500\u2518` characters. The renderer adds its own borders; extra ASCII art adds noise and shatters at narrow widths.\n- For flow charts and diagrams: a plain bullet list with `\u2192` or `\u2193` between steps. Don't try to draw boxes-and-arrows in ASCII; it never survives word-wrap.\n";
 interface CodeSystemPromptOptions {
     /** True when semantic_search is registered for this run. Adds an
      *  explicit routing fragment so the model picks it for intent-style

package/dist/index.js

@@ -2282,15 +2282,18 @@ var StormBreaker = class {
   windowSize;
   threshold;
   isMutating;
+  isStormExempt;
   recent = [];
-  constructor(windowSize = 6, threshold = 3, isMutating) {
+  constructor(windowSize = 6, threshold = 3, isMutating, isStormExempt) {
     this.windowSize = windowSize;
     this.threshold = threshold;
     this.isMutating = isMutating;
+    this.isStormExempt = isStormExempt;
   }
   inspect(call) {
     const name = call.function?.name;
     if (!name) return { suppress: false };
+    if (this.isStormExempt?.(call)) return { suppress: false };
     const args = call.function?.arguments ?? "";
     const mutating = this.isMutating ? this.isMutating(call) : false;
     const readOnly = !mutating;
@@ -2391,7 +2394,12 @@ var ToolCallRepair = class {
   opts;
   constructor(opts) {
     this.opts = opts;
-    this.storm = new StormBreaker(opts.stormWindow ?? 6, opts.stormThreshold ?? 3, opts.isMutating);
+    this.storm = new StormBreaker(
+      opts.stormWindow ?? 6,
+      opts.stormThreshold ?? 3,
+      opts.isMutating,
+      opts.isStormExempt
+    );
   }
   /** Called at start of every user turn — fresh intent shouldn't inherit old repetition state. */
   resetStorm() {
@@ -2534,9 +2542,15 @@ var CacheFirstLoop = class {
       }
       return def.readOnly !== true;
     };
+    const isStormExempt = (call) => {
+      const name = call.function?.name;
+      if (!name) return false;
+      return registry.get(name)?.stormExempt === true;
+    };
     this.repair = new ToolCallRepair({
       allowedToolNames: allowedNames,
       isMutating,
+      isStormExempt,
       stormThreshold: parsePositiveIntEnv(process.env.REASONIX_STORM_THRESHOLD),
       stormWindow: parsePositiveIntEnv(process.env.REASONIX_STORM_WINDOW)
     });
@@ -3653,19 +3667,19 @@ ${mem.content}
 import { createHash as createHash2 } from "crypto";
 import {
   existsSync as existsSync7,
-  mkdirSync as mkdirSync2,
+  mkdirSync as mkdirSync3,
   readFileSync as readFileSync8,
   readdirSync as readdirSync4,
   unlinkSync as unlinkSync2,
-  writeFileSync as writeFileSync2
+  writeFileSync as writeFileSync3
 } from "fs";
 import { homedir as homedir4 } from "os";
 import { join as join7, resolve as resolve3 } from "path";
 
 // src/skills.ts
-import { existsSync as existsSync6, readFileSync as readFileSync7, readdirSync as readdirSync3, statSync as statSync3 } from "fs";
+import { existsSync as existsSync6, mkdirSync as mkdirSync2, readFileSync as readFileSync7, readdirSync as readdirSync3, statSync as statSync3, writeFileSync as writeFileSync2 } from "fs";
 import { homedir as homedir3 } from "os";
-import { join as join6, resolve as resolve2 } from "path";
+import { dirname as dirname3, join as join6, resolve as resolve2 } from "path";
 
 // src/prompt-fragments.ts
 var TUI_FORMATTING_RULES = `Formatting (rendered in a TUI with a real markdown renderer):
@@ -3770,6 +3784,31 @@ var SkillStore = class {
     }
     return [...byName.values()].sort((a, b) => a.name.localeCompare(b.name));
   }
+  /** Scaffold a new skill stub at the chosen scope. Refuses to overwrite. */
+  create(name, scope) {
+    if (!isValidSkillName(name)) {
+      return { error: `invalid skill name: "${name}" \u2014 use letters, digits, _, -, .` };
+    }
+    if (scope === "project" && !this.projectRoot) {
+      return { error: "project scope requires a workspace \u2014 run from `reasonix code`" };
+    }
+    const root = scope === "project" ? join6(this.projectRoot ?? "", ".reasonix", SKILLS_DIRNAME) : join6(this.homeDir, ".reasonix", SKILLS_DIRNAME);
+    const flat = join6(root, `${name}.md`);
+    const folder = join6(root, name, SKILL_FILE);
+    if (existsSync6(folder)) {
+      return { error: `skill "${name}" already exists at ${folder}` };
+    }
+    mkdirSync2(dirname3(flat), { recursive: true });
+    try {
+      writeFileSync2(flat, skillStubBody(name), { encoding: "utf8", flag: "wx" });
+    } catch (err) {
+      if (err.code === "EEXIST") {
+        return { error: `skill "${name}" already exists at ${flat}` };
+      }
+      throw err;
+    }
+    return { path: flat };
+  }
   /** Resolve one skill by name. Returns `null` if not found or malformed. */
   read(name) {
     if (!isValidSkillName(name)) return null;
@@ -3829,6 +3868,22 @@ var SkillStore = class {
 function parseRunAs(raw) {
   return raw?.trim() === "subagent" ? "subagent" : "inline";
 }
+function skillStubBody(name) {
+  return `---
+name: ${name}
+description: One-liner \u2014 what does this skill do?
+---
+
+# ${name}
+
+Replace this body with the playbook the model should follow when this skill is invoked.
+
+Tips:
+- Reference tools by name (run_command, edit_file, search_content, ...)
+- Add \`runAs: subagent\` to frontmatter to spawn an isolated subagent loop
+- Add \`allowed-tools: read_file, search_content\` to scope a subagent's tools
+`;
+}
 function skillIndexLine(s) {
   const safeDesc = s.description.replace(/\n/g, " ").trim();
   const tag = s.runAs === "subagent" ? " [\u{1F9EC} subagent]" : "";
@@ -4071,7 +4126,7 @@ function scopeDir(opts) {
   return join7(opts.homeDir, USER_MEMORY_DIR, projectHash(opts.projectRoot));
 }
 function ensureDir(p) {
-  if (!existsSync7(p)) mkdirSync2(p, { recursive: true });
+  if (!existsSync7(p)) mkdirSync3(p, { recursive: true });
 }
 function parseFrontmatter2(raw) {
   const lines = raw.split(/\r?\n/);
@@ -4216,7 +4271,7 @@ var MemoryStore = class {
     const file = join7(dir, `${name}.md`);
     const content = `${formatFrontmatter(entry)}${body}
 `;
-    writeFileSync2(file, content, "utf8");
+    writeFileSync3(file, content, "utf8");
     this.regenerateIndex(input.scope);
     return file;
   }
@@ -4257,7 +4312,7 @@ var MemoryStore = class {
         lines.push(`- [${name}](${name}.md) \u2014 (malformed, check frontmatter)`);
       }
     }
-    writeFileSync2(indexPath, `${lines.join("\n")}
+    writeFileSync3(indexPath, `${lines.join("\n")}
 `, "utf8");
   }
 };
@@ -4693,6 +4748,7 @@ function registerFilesystemTools(registry, opts) {
   - range: "A-B"  \u2192 inclusive line range A..B, 1-indexed (e.g. "120-180" around an edit site)
 When none of these is given AND the file is longer than ${DEFAULT_AUTO_PREVIEW_LINES} lines, the tool auto-returns a head+tail preview with an "N lines omitted" marker rather than dumping everything. If you need the middle, re-call with a range. Prefer search_content to locate a symbol first, then read_file with a range around the hit \u2014 one scoped read beats three full-file reads.`,
     readOnly: true,
+    stormExempt: true,
     parameters: {
       type: "object",
       properties: {
@@ -4773,6 +4829,7 @@ ${slice.join("\n")}`;
     parallelSafe: true,
     description: "List entries in a directory under the sandbox root. Returns one line per entry, marking directories with a trailing slash. Not recursive \u2014 use directory_tree for that.",
     readOnly: true,
+    stormExempt: true,
     parameters: {
       type: "object",
       properties: {
@@ -5874,9 +5931,9 @@ function forkRegistryWithAllowList(parent, allow, alsoExclude) {
 import * as pathMod7 from "path";
 
 // src/config.ts
-import { chmodSync as chmodSync2, mkdirSync as mkdirSync3, readFileSync as readFileSync9, writeFileSync as writeFileSync3 } from "fs";
+import { chmodSync as chmodSync2, mkdirSync as mkdirSync4, readFileSync as readFileSync9, writeFileSync as writeFileSync4 } from "fs";
 import { homedir as homedir5 } from "os";
-import { dirname as dirname4, join as join10 } from "path";
+import { dirname as dirname5, join as join10 } from "path";
 function defaultConfigPath() {
   return join10(homedir5(), ".reasonix", "config.json");
 }
@@ -5890,8 +5947,8 @@ function readConfig(path2 = defaultConfigPath()) {
   return {};
 }
 function writeConfig(cfg, path2 = defaultConfigPath()) {
-  mkdirSync3(dirname4(path2), { recursive: true });
-  writeFileSync3(path2, JSON.stringify(cfg, null, 2), "utf8");
+  mkdirSync4(dirname5(path2), { recursive: true });
+  writeFileSync4(path2, JSON.stringify(cfg, null, 2), "utf8");
   try {
     chmodSync2(path2, 384);
   } catch {
@@ -6033,7 +6090,8 @@ var JobRegistry = class {
         },
         closedPromise: Promise.resolve(),
         signalClosed: () => {
-        }
+        },
+        outputWaiters: /* @__PURE__ */ new Set()
       };
       this.jobs.set(id2, job2);
       return {
@@ -6069,7 +6127,8 @@ var JobRegistry = class {
       readyPromise,
       signalReady: readyResolve,
       closedPromise,
-      signalClosed: closedResolve
+      signalClosed: closedResolve,
+      outputWaiters: /* @__PURE__ */ new Set()
     };
     this.jobs.set(id, job);
     let readyMatched = false;
@@ -6096,6 +6155,11 @@ ${job.output.slice(start)}`;
           }
         }
       }
+      if (job.outputWaiters.size > 0) {
+        const waiters = [...job.outputWaiters];
+        job.outputWaiters.clear();
+        for (const wake of waiters) wake();
+      }
     };
     child.stdout?.on("data", onData);
     child.stderr?.on("data", onData);
@@ -6157,6 +6221,39 @@ ${job.output.slice(start)}`;
       spawnError: job.spawnError
     };
   }
+  async waitForJob(id, opts = {}) {
+    const job = this.jobs.get(id);
+    if (!job) return null;
+    if (!job.running) {
+      return {
+        exited: true,
+        exitCode: job.exitCode,
+        latestOutput: job.output
+      };
+    }
+    const timeoutMs = Math.max(0, Math.min(3e4, opts.timeoutMs ?? 5e3));
+    const startOutput = job.output;
+    let wakeOutput = null;
+    const outputPromise = new Promise((resolve10) => {
+      wakeOutput = resolve10;
+      job.outputWaiters.add(resolve10);
+    });
+    let timer = null;
+    await Promise.race([
+      job.closedPromise,
+      outputPromise,
+      new Promise((resolve10) => {
+        timer = setTimeout(resolve10, timeoutMs);
+      })
+    ]);
+    if (timer) clearTimeout(timer);
+    if (wakeOutput) job.outputWaiters.delete(wakeOutput);
+    return {
+      exited: !job.running,
+      exitCode: job.exitCode,
+      latestOutput: latestOutputSince(startOutput, job.output)
+    };
+  }
   /** SIGTERM, wait graceMs, then SIGKILL. Idempotent on already-exited jobs. */
   async stop(id, opts = {}) {
     const job = this.jobs.get(id);
@@ -6236,6 +6333,11 @@ function snapshot(job) {
     spawnError: job.spawnError
   };
 }
+function latestOutputSince(before, after) {
+  if (!before) return after;
+  if (after.startsWith(before)) return after.slice(before.length);
+  return after;
+}
 
 // src/tools/shell/exec.ts
 import { spawn as spawn4, spawnSync } from "child_process";
@@ -7215,6 +7317,7 @@ function registerShellTools(registry, opts) {
     description: "Read the latest output of a background job started with `run_background`. By default returns the tail of the buffer (last 80 lines). Pass `since` (the `byteLength` from a previous call) to stream only new content incrementally. Tells you whether the job is still running, so you can stop polling when it's done.",
     readOnly: true,
     parallelSafe: true,
+    stormExempt: true,
     parameters: {
       type: "object",
       properties: {
@@ -7239,6 +7342,32 @@ function registerShellTools(registry, opts) {
       return formatJobRead(args.jobId, out);
     }
   });
+  registry.register({
+    name: "wait_for_job",
+    description: "Block until a background job exits or produces new output, bounded by `timeoutMs`. Use this instead of polling `job_output` with identical args when you're intentionally waiting for state to change. Returns JSON with `exited`, `exitCode`, and `latestOutput`.",
+    readOnly: true,
+    parameters: {
+      type: "object",
+      properties: {
+        jobId: { type: "integer", description: "Job id returned by run_background." },
+        timeoutMs: {
+          type: "integer",
+          description: "Max time to block before returning if nothing changes. Clamped to 0..30000. Default 5000."
+        }
+      },
+      required: ["jobId"]
+    },
+    fn: async (args) => {
+      const out = await jobs.waitForJob(args.jobId, { timeoutMs: args.timeoutMs });
+      if (!out) return `job ${args.jobId}: not found (use list_jobs)`;
+      return {
+        jobId: args.jobId,
+        exited: out.exited,
+        exitCode: out.exitCode,
+        latestOutput: out.latestOutput
+      };
+    }
+  });
   registry.register({
     name: "stop_job",
     description: "Stop a background job started with `run_background`. SIGTERM first; SIGKILL after a short grace period if it doesn't exit cleanly. Returns the final output + exit code. Safe to call on an already-exited job.",
@@ -7260,6 +7389,7 @@ function registerShellTools(registry, opts) {
     description: "List every background job started this session \u2014 running and exited \u2014 with id, command, pid, status. Use when you've lost track of which job_id corresponds to which process, or to see what's still alive.",
     readOnly: true,
     parallelSafe: true,
+    stormExempt: true,
     parameters: { type: "object", properties: {} },
     fn: async () => {
       const all = jobs.list();
@@ -8192,16 +8322,16 @@ function truncate(s, n) {
 }
 
 // src/version.ts
-import { existsSync as existsSync9, mkdirSync as mkdirSync4, readFileSync as readFileSync12, writeFileSync as writeFileSync4 } from "fs";
+import { existsSync as existsSync9, mkdirSync as mkdirSync5, readFileSync as readFileSync12, writeFileSync as writeFileSync5 } from "fs";
 import { homedir as homedir6 } from "os";
-import { dirname as dirname5, join as join11 } from "path";
+import { dirname as dirname6, join as join11 } from "path";
 import { fileURLToPath as fileURLToPath2 } from "url";
 var REGISTRY_URL = "https://registry.npmjs.org/reasonix/latest";
 var LATEST_CACHE_TTL_MS = 24 * 60 * 60 * 1e3;
 var LATEST_FETCH_TIMEOUT_MS = 2e3;
 function readPackageVersion() {
   try {
-    let dir = dirname5(fileURLToPath2(import.meta.url));
+    let dir = dirname6(fileURLToPath2(import.meta.url));
     for (let i = 0; i < 6; i++) {
       const p = join11(dir, "package.json");
       if (existsSync9(p)) {
@@ -8210,7 +8340,7 @@ function readPackageVersion() {
           return pkg.version;
         }
       }
-      const parent = dirname5(dir);
+      const parent = dirname6(dir);
       if (parent === dir) break;
       dir = parent;
     }
@@ -8236,8 +8366,8 @@ function readCache(homeDirOverride) {
 function writeCache(entry, homeDirOverride) {
   try {
     const p = cachePath(homeDirOverride);
-    mkdirSync4(dirname5(p), { recursive: true });
-    writeFileSync4(p, JSON.stringify(entry), "utf8");
+    mkdirSync5(dirname6(p), { recursive: true });
+    writeFileSync5(p, JSON.stringify(entry), "utf8");
   } catch {
   }
 }
@@ -9041,15 +9171,15 @@ import {
   existsSync as existsSync10,
   fstatSync,
   ftruncateSync,
-  mkdirSync as mkdirSync5,
+  mkdirSync as mkdirSync6,
   openSync as openSync2,
   readFileSync as readFileSync13,
   readSync,
   unlinkSync as unlinkSync3,
-  writeFileSync as writeFileSync5,
+  writeFileSync as writeFileSync6,
   writeSync
 } from "fs";
-import { dirname as dirname6, resolve as resolve9 } from "path";
+import { dirname as dirname7, resolve as resolve9 } from "path";
 var BLOCK_RE = /^(\S[^\n]*)\n<{7} SEARCH\n([\s\S]*?)\n?={7}\n([\s\S]*?)\n?>{7} REPLACE/gm;
 function parseEditBlocks(text) {
   const out = [];
@@ -9079,7 +9209,7 @@ function applyEditBlock(block, rootDir) {
   const searchEmpty = block.search.length === 0;
   if (searchEmpty) {
     try {
-      mkdirSync5(dirname6(absTarget), { recursive: true });
+      mkdirSync6(dirname7(absTarget), { recursive: true });
       const fd = openSync2(absTarget, "wx");
       try {
         writeSync(fd, block.replace);
@@ -9194,7 +9324,7 @@ function restoreSnapshots(snapshots, rootDir) {
           message: "removed (the edit had created it)"
         };
       }
-      writeFileSync5(abs, snap.prevContent, "utf8");
+      writeFileSync6(abs, snap.prevContent, "utf8");
       return {
         path: snap.path,
         status: "applied",
@@ -9370,6 +9500,7 @@ You have TWO tools for running shell commands, and picking the right one is non-
 
 After \`run_background\`, tools available to you:
 - \`job_output(jobId, tailLines?)\` \u2014 read recent logs to verify startup / debug errors.
+- \`wait_for_job(jobId, timeoutMs?)\` \u2014 block until the job exits or emits new output. Prefer this over repeating identical \`job_output\` calls while you're intentionally waiting.
 - \`list_jobs\` \u2014 see every job this session (running + exited).
 - \`stop_job(jobId)\` \u2014 SIGTERM \u2192 SIGKILL after grace. Stop before switching port / config.
 
@@ -9457,17 +9588,17 @@ import {
   closeSync as closeSync3,
   existsSync as existsSync12,
   fstatSync as fstatSync2,
-  mkdirSync as mkdirSync6,
+  mkdirSync as mkdirSync7,
   openSync as openSync3,
   readFileSync as readFileSync15,
   readSync as readSync2,
   renameSync as renameSync2,
   statSync as statSync5,
   unlinkSync as unlinkSync4,
-  writeFileSync as writeFileSync6
+  writeFileSync as writeFileSync7
 } from "fs";
 import { homedir as homedir7 } from "os";
-import { dirname as dirname7, join as join13 } from "path";
+import { dirname as dirname8, join as join13 } from "path";
 function defaultUsageLogPath(homeDirOverride) {
   return join13(homeDirOverride ?? homedir7(), ".reasonix", "usage.jsonl");
 }
@@ -9508,7 +9639,7 @@ function compactUsageLogIfLarge(path2, now) {
   if (kept.length === lines.filter((l) => l.trim()).length) return;
   const tmp = `${path2}.compacting`;
   try {
-    writeFileSync6(tmp, kept.length > 0 ? `${kept.join("\n")}
+    writeFileSync7(tmp, kept.length > 0 ? `${kept.join("\n")}
 ` : "", "utf8");
     renameSync2(tmp, path2);
   } catch {
@@ -9534,7 +9665,7 @@ function appendUsage(input) {
   if (input.subagent) record.subagent = input.subagent;
   const path2 = input.path ?? defaultUsageLogPath();
   try {
-    mkdirSync6(dirname7(path2), { recursive: true });
+    mkdirSync7(dirname8(path2), { recursive: true });
     appendFileSync2(path2, `${JSON.stringify(record)}
 `, "utf8");
     compactUsageLogIfLarge(path2, record.ts);