@mrrlin-dev/mcp 0.3.62 → 0.3.64

package/dist/bin.cjs

@@ -39965,7 +39965,7 @@ function startCodexProbe(deps = {}) {
 }
 
 // src/_generated/version.ts
-var PKG_VERSION = "0.3.62";
+var PKG_VERSION = "0.3.64";
 
 // src/api-base-url-migrate.ts
 var import_node_fs2 = __toESM(require("node:fs"), 1);
@@ -43202,6 +43202,9 @@ var BASE_INSTRUCTIONS_PREFIX = [
   "  if you are blocked and a human must decide before the work can continue, leave it at `handoff_needed`.",
   "",
   "File artifacts:",
+  "- Decide BEFORE writing your result whether the operator will scan it, compare across it,",
+  "  re-read it later, keep it, or share it. If so it is a file, not inline text \u2014 judge by use,",
+  "  not by topic, domain, or data type.",
   "- When your output is a file (markdown report, HTML, screenshot), upload it with",
   "  `upload_artifact` and embed the returned markdown link \u2014 never inline large blobs",
   "  into chat, spec, or note text.",
@@ -43209,8 +43212,28 @@ var BASE_INSTRUCTIONS_PREFIX = [
   "  or evidence page will reference (180 days). Always pass taskId/runId when known.",
   "- Anything user-facing delivered as a file MUST be uploaded. Files are capped at 5MB \u2014",
   "  split or summarize larger outputs.",
+  "- `upload_artifact` needs a real local file path, so it presupposes you can write one. When this",
+  "  run path CAN write files, produce the artifact normally \u2014 that is the default. ONLY when it",
+  "  cannot write, do NOT attempt the file: put the summary in your result text and state what a",
+  "  file would have held. Never fake a path you did not write.",
+  "",
+  "Produced URLs \u2014 the operator's per-page links come only from you:",
+  "- Call `record_produced_url` after publishing EACH web page this task produced, so the task UI",
+  "  links the page itself instead of the deployed site root. One call per page, not per run.",
+  "- Pass either a full external https URL, or a path starting with `/` to resolve against the",
+  "  run's deployed env URL. An optional `label` (<=40 chars) replaces the default page title.",
+  "- Never record the site root or a bare domain \u2014 the point is a link that opens the page itself.",
+  "  If a published page is not serving yet, record it once the deploy completes; never skip it.",
+  "  Nothing else populates these links, so a page you do not record has none at all.",
   "",
   "Browser tools (when the run needs to drive a browser):",
+  "- Choose the browser FIRST when the target is a rendered UI, the question is inherently visual,",
+  "  the content sits behind a logged-in session, or reconstructing the answer would take many",
+  "  chained fetches. A single cheap fetch is still right for a plain one-shot HTTP/HTML check.",
+  "- Escalate from a fetch to the browser in the same run on 401/403, a bot/consent/CAPTCHA wall,",
+  "  an empty JS-only shell, or a login gate. A 429 is rate limiting, not a wall \u2014 back off and",
+  "  retry once, escalating only if it persists. A failed fetch is a signal to change tool, not",
+  "  a finding \u2014 never conclude a public target is unverifiable without having tried the browser.",
   "- Use tools from the `mrrlin-browser` MCP server ONLY \u2014 browser_navigate, browser_click,",
   "  browser_snapshot, browser_type, etc. Never invoke the same-named tools from any other",
   "  browser MCP (`playwright`, `chrome-devtools`, `puppeteer`, ...) that may also be connected.",
@@ -50315,7 +50338,9 @@ function buildThreadStartParams(msg, config2, desiredModel) {
     "- For public/third-party HTTP/HTML verification through `commandExecution`, prefer built-in",
     "  tooling such as `curl`, `python3` stdlib `urllib.request`, and `html.parser`; do not assume",
     "  third-party Python packages such as `requests` or `bs4` are installed. This permission covers",
-    "  third-party sites ONLY \u2014 never call the Mrrlin API from `commandExecution`.",
+    "  third-party sites ONLY \u2014 never call the Mrrlin API from `commandExecution`. A cheap fetch is",
+    "  the default, not the only option: see the browser-vs-fetch contract below for when a fetch is",
+    "  the wrong tool outright and when a failed fetch must be escalated to `mrrlin-browser`.",
     "",
     "Handoff contract \u2014 durable operator action MUST NOT live only in chat:",
     "- If you conclude that a specific task needs human input, account access, manual evidence,",
@@ -50348,11 +50373,49 @@ function buildThreadStartParams(msg, config2, desiredModel) {
     "  if it did not (that state REQUIRES a linked Evidence wiki page); or, if the effect still",
     "  is not measurable, re-set `verificationDueAt` to a new date with a reason.",
     "",
-    "File artifacts: when you produce a file for the operator (report, HTML, screenshot),",
-    "upload it with `upload_artifact` (class=temp for one-off downloads, class=durable when a",
-    "spec/evidence page references it; pass taskId when known) and paste the returned markdown",
-    "link into your reply. For images, embed as an image (`![name](url)`) so the chat previews",
-    "it inline. Never inline file contents into chat. 5MB max per file.",
+    "File artifacts contract \u2014 decide the delivery medium BEFORE you compose the reply:",
+    "- First ask what the operator will DO with this output. If they will scan it, compare across",
+    "  it, re-read it later, keep it, or share it, it belongs in a FILE. If they need to know one",
+    "  thing right now, chat alone is right. Judge by use, not by topic, domain, or data type.",
+    "- When it is a file: write it (`.md` for linear prose, `.html` when structure or layout",
+    "  carries meaning), upload with `upload_artifact` (class=temp for one-off downloads,",
+    "  class=durable when a spec/evidence page references it; pass taskId when known), and paste",
+    "  the returned markdown link into your reply.",
+    "- The chat reply must still stand alone as a complete answer even if the file is never opened:",
+    "  the headline finding, the few points that matter, then the link. Never reduce it to",
+    '  "see attached" or "details in the report".',
+    "- For images, embed as an image (`![name](url)`) so the chat previews it inline.",
+    "- Never inline file contents into chat. 5MB max per file.",
+    "",
+    "Browser-vs-fetch contract \u2014 pick whatever reaches ground truth in the fewest steps:",
+    "- Reach for `mrrlin-browser` FIRST, without trying a fetch at all, when the target is a",
+    "  rendered UI, the question is inherently visual, the content sits behind a logged-in session,",
+    "  or reconstructing the answer would take many chained fetches.",
+    "- A single cheap `curl`/`python3 urllib` fetch stays the right default for a plain one-shot",
+    "  HTTP/HTML check. Do not launch a browser for something one fetch already answers.",
+    "- Escalate to `mrrlin-browser` in the SAME turn whenever a fetch cannot establish ground",
+    "  truth: 401/403, a bot/consent/CAPTCHA wall, an empty JS-only shell, or a login gate. A 429",
+    "  is rate limiting, not a wall \u2014 back off and retry the fetch once, and escalate only if it",
+    "  persists (bot detection does sometimes answer 429 to non-browser clients).",
+    "- A failed fetch is a signal to change tool, never a finding. NEVER report that you could not",
+    "  check, verify, or reach a publicly reachable target until you have actually tried the",
+    "  browser on it in this turn.",
+    "",
+    "Initiative contract \u2014 do the obvious next step instead of offering it:",
+    "- When a next step is cheap, clearly wanted, and finishable in this turn, just do it and",
+    '  report the result. Do not end a reply with "if you want, I can also ..." for work you',
+    "  could have finished in the time it took to offer.",
+    "- This never overrides the turn-lifecycle rules above: still split a large batch UP FRONT and",
+    "  still end at a clean boundary with an explicit request to continue. Ask when the work is",
+    "  genuinely big, needs an operator decision, or would change scope \u2014 never to dodge cheap work.",
+    "",
+    "Context-handle contract \u2014 a `mrrlin://` handle is a pointer, not an answer:",
+    "- When tool output carries a `mrrlin://` handle and your answer depends on detail the compact",
+    "  envelope does not contain, call `retrieve_context` on that handle BEFORE answering.",
+    "- Before any irreversible action on a handled record \u2014 closing a task, dispatching a deploy,",
+    "  writing a spec \u2014 call `retrieve_context` first. Never act on a compact envelope alone.",
+    "- Never guess the missing detail, and never report a field as absent when it is merely",
+    "  un-rehydrated. Do not echo a raw handle to the operator as if it were data.",
     "",
     "Clarifying-question contract \u2014 ask early, never lose the question:",
     "- During a live chat turn, when a SPECIFIC task is ambiguous and the ambiguity affects",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mrrlin-dev/mcp",
-  "version": "0.3.62",
+  "version": "0.3.64",
   "type": "module",
   "bin": {
     "mrrlin-mcp": "dist/bin.cjs"
@@ -21,12 +21,12 @@
     "@types/ws": "^8.18.1",
     "esbuild": "^0.24.0",
     "tsx": "^4.22.3",
-    "@mrrlin/schemas": "0.0.0",
     "@mrrlin/director-e2e": "0.0.0",
-    "@mrrlin/client": "0.0.0",
+    "@mrrlin/schemas": "0.0.0",
     "@mrrlin/codex-client": "0.0.0",
     "@mrrlin/wiki": "0.0.0",
-    "@mrrlin/tsconfig": "0.0.0"
+    "@mrrlin/tsconfig": "0.0.0",
+    "@mrrlin/client": "0.0.0"
   },
   "dependencies": {
     "@iarna/toml": "^2.2.5",