@clearlysh/beehaven 0.8.36 → 0.8.38

package/dist/relay.js

@@ -518,7 +518,16 @@ export class Relay {
                     expiredNoRefresh = true;
             }
         }
-        catch { }
+        catch {
+            // A damaged cached ID token has no decodable exp, but a valid refresh session can still
+            // recover it. Do not strand that account in an unauthenticated reconnect loop. Opaque
+            // self-hosted tokens without a refresh session retain their existing dial behavior.
+            if (this.config.refreshToken) {
+                const refreshed = await this.refreshIdToken();
+                if (refreshed && this.config)
+                    token = this.config.token;
+            }
+        }
         // If the token is already expired AND refresh failed, dialing would just
         // get rejected — spinning the reconnect loop forever with a dead
         // credential (the classic "daemon stuck reconnecting" state). Back off
@@ -777,10 +786,14 @@ export class Relay {
             this.heartbeatInterval = null;
             this.wsPingInterval = null;
             this.wsWakeDetector = null;
-            // Reject pending requests
-            for (const [, pending] of this.pendingRequests) {
+            // Reject pending requests.
+            // ⚠ CAN2-175 — SAY THE OUTCOME IS UNKNOWN, as the timeout does. A dropped socket loses the REPLY,
+            // not the request: a canvas batch carries on and lands on the workspace. A bare "WebSocket closed"
+            // read as "nothing happened", the caller resent at once, and could not tell which state was real.
+            for (const [requestId, pending] of this.pendingRequests) {
                 clearTimeout(pending.timer);
-                pending.reject(new Error('WebSocket closed'));
+                pending.reject(new Error(`WebSocket closed (code ${code ?? '?'}) before the reply (action=${pending.action ?? '?'}, requestId=${requestId}); outcome unknown: `
+                    + 'the request may still be running on the workspace or may already have landed. Read its state before resending.'));
             }
             this.pendingRequests.clear();
             this.backoff.closed(openedAt);
@@ -1098,12 +1111,12 @@ export class Relay {
                 return reject(new Error('WebSocket not connected'));
             }
             const requestId = Math.random().toString(36).slice(2) + Date.now().toString(36);
+            const action = msg.type === 'agent-invoke' ? msg.action : msg.type;
             const timer = setTimeout(() => {
                 this.pendingRequests.delete(requestId);
-                const action = msg.type === 'agent-invoke' ? msg.action : msg.type;
                 reject(new Error(`Request timeout after ${timeout}ms (action=${action}, requestId=${requestId}); outcome unknown`));
             }, timeout);
-            this.pendingRequests.set(requestId, { resolve, reject, timer });
+            this.pendingRequests.set(requestId, { resolve, reject, timer, action });
             // Send both rid (legacy) and requestId (workspace actions) for compatibility
             if (!this.wsSend({ ...msg, rid: requestId, requestId })) {
                 clearTimeout(timer);

package/dist/utf8-stream.js

@@ -0,0 +1,23 @@
+/**
+ * utf8-stream — decode a socket's bytes ONCE PER CHARACTER, not once per chunk (CA-63).
+ *
+ * ⚠⚠ WHY THIS EXISTS (staging, 2026-10-23). A 402-step `beehaven call canvas-act @build.json` (90,714
+ * bytes, the file holding the correct C2 B7) stored ONE text node's middle dot as "��" — the
+ * same node, the same character, on every resend — while dozens of identical dots elsewhere in the batch
+ * were stored intact. Every IPC reader did `buffer += chunk.toString()`. A local socket hands the reader
+ * the stream in pieces (8 KB on macOS, 64 KB on Linux), so any multi-byte character that straddles a
+ * piece boundary was decoded as two halves, each half an invalid sequence, each a U+FFFD. Which character
+ * lands on a boundary depends only on the bytes before it, so the same payload broke the same character
+ * every time. The CLI → daemon direction corrupted what was STORED; the daemon → CLI direction corrupted
+ * what a reader was SHOWN (a long scene read could carry U+FFFD the canvas does not have).
+ *
+ * A `StringDecoder` holds an incomplete trailing sequence until its remaining bytes arrive, so the text is
+ * exactly what one decode of the whole stream would give. One decoder per stream (it carries state), and
+ * never shared between two sockets.
+ */
+import { StringDecoder } from 'node:string_decoder';
+/** A stateful UTF-8 decoder for ONE byte stream. A string chunk (a socket with an encoding set) passes through. */
+export function utf8Stream() {
+    const decoder = new StringDecoder('utf8');
+    return (chunk) => (typeof chunk === 'string' ? chunk : decoder.write(chunk));
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@clearlysh/beehaven",
-  "version": "0.8.36",
+  "version": "0.8.38",
   "description": "Beehaven CLI — command your Clearly workspace, agents, and canvas from the terminal",
   "private": false,
   "type": "module",

package/skills/clearly-canvas/SKILL.md

@@ -32,7 +32,7 @@ For a new composition or substantial redesign, use the workflow below.
 
 Keep the requested scope. A landing concept with a logo needs its page layouts and a usable
 logo; add application boards or a full guideline system only when the brief calls for them.
-Read only relevant references. Save large tool responses once and print compact summaries.
+Read only relevant references. Save large tool responses once (`beehaven call … --out file.json`) and print compact summaries.
 
 ## The studio bar — what "done" means, every time
 
@@ -46,7 +46,8 @@ A composition is done when every requested artboard meets all five conditions:
    remains: text colliding or covered, contrast under WCAG (4.5:1 body, 3:1 display),
    type below the artboard's legible floor, missing media, text overflowing its box,
    placeholder copy. Fix it. For an intentional finding, use
-   `accept:[{check,ids:[node IDs],reason}]`. Accepted findings stay visible and return `look`.
+   `accept:[{check,ids:[node IDs],reason}]`; one rule may name every layer with the same
+   deliberate finding. Accepted findings stay visible and return `look`.
    Legacy `ignore` can suppress multiple checks and is not an unconditional pass.
 2. **Every `look` finding was looked at**: type-size and palette sprawl, near-miss
    alignment, copy against the trim, text over imagery. Act on it or say why not.
@@ -117,8 +118,9 @@ Read `clearly-agent` before your first workspace call. Pin the intended workspac
 sign in once. Do not switch a user's explicit workspace to `home`.
 
 Your surface: the shell. The `beehaven` examples in this skill are literal commands to
-run; they are not shapes to translate into an MCP canvas-action tool. MCP artifact tools
-can read and edit artifact records, while canvas-specific actions use the CLI.
+run; they are not shapes to translate into an MCP canvas-action tool. The Clearly connector
+has its own canvas tools for agents without a shell (batched per kind of step, see
+authoring.md); in a shell the CLI's mixed `canvas-act` batch takes fewer calls.
 
 ```sh
 beehaven env
@@ -156,7 +158,7 @@ builds on the last and can be reviewed alone. Build ONE commit per canvas-act ba
 `canvas.review` of the artboard and carries
 `"ticket":{"id":"<key>","status":"done","resolution":"<what changed>"}` — it lands only if that
 review is clean or look; `result.ticket` says whether it did, and the ticket logs the canvas
-revision and the layers the commit changed. Auto-layout, gradients, shadows, drawing and placeholders:
+revision, the layers the commit changed and a link to its frame. Auto-layout, gradients, shadows, drawing and placeholders:
 see "Technique" in [authoring.md](authoring.md).
 
 `composition-board-get {compositionId}` reads the plan back. `design-start` with a project, ticket or
@@ -279,7 +281,7 @@ a detail view catches collisions, wrapping, cropping and legibility.
 The bundled helper saves pixels directly to disk so base64 does not fill your context:
 
 ```sh
-node <skill-directory>/review.mjs <composition-id> <frame-id-or-unique-name> /tmp/canvas-review/01
+node <skill-directory>/review.mjs <composition-id> <frame-id-or-unique-name> canvas-review/01
 # Open the returned PNG using your image-viewing tool; read the JSON findings.
 ```
 
@@ -300,9 +302,15 @@ The helper prints the verdict and every finding. Work until the verdict is `clea
 use a real frame ID with `--perceive` for a preview-only check and run the text measurement
 and audit separately. A failed or missing render means visual verification is incomplete.
 
+A deliberate finding (say why in the handoff): write the findings you are keeping to a file and
+pass it to the same helper. They stay visible in the report and return `look`.
+
+```json
+[{"check":"text-over-image","ids":["<text-id>","<image-id>"],"reason":"Intentional type over this photograph; rendered contrast was inspected."}]
+```
+
 ```sh
-# A deliberate finding — say why in the handoff:
-beehaven call canvas-act '{"compositionId":"<id>","action":"canvas.review","args":{"ids":["<frame>"],"accept":[{"check":"text-over-image","ids":["<text-id>","<image-id>"],"reason":"Intentional type over this photograph; rendered contrast was inspected."}]}}'
+node <skill-directory>/review.mjs <composition-id> <frame> canvas-review/02 --accept accepted.json
 ```
 
 Asset storage is a reference-aware pipeline: an SVG may have `svgRef` instead of inline markup.

package/skills/clearly-canvas/authoring.md

@@ -38,15 +38,34 @@ add('canvas.update-nodes', { ids: ['opening-title', 'opening-deck'], patch: { la
 add('canvas.auto-layout', { id: 'hero-copy', direction: 'v', gap: 24, sizingH: 'fixed', sizingV: 'hug' });
 add('canvas.auto-layout', { id: '01-opening', direction: 'v', gap: 64, padding: grid.margin, sizingH: 'fixed', sizingV: 'hug' });
 // Continue with the brief's central visual, supporting evidence and footer.
-writeFileSync('/tmp/canvas-build.json', JSON.stringify({ compositionId, batch }));
+writeFileSync('canvas-build.json', JSON.stringify({ compositionId, batch }));
 ```
 
 ```sh
 beehaven call composition-create '{"title":"Project — concept","projectId":"<project-id>"}'
 node build.mjs <returned-id>
-beehaven call canvas-act @/tmp/canvas-build.json > /tmp/canvas-build-result.json
+beehaven call canvas-act @canvas-build.json --brief
 ```
 
+`--brief` prints a short receipt instead of the full JSON:
+- step counts by kind;
+- layers added, changed and removed;
+- every failed step with its error;
+- rollbacks and skips;
+- lint to fix and review verdicts.
+
+The exit code stays the same. Use it for routine batches instead of piping the JSON through a
+summariser of your own. Drop it when you need a step's full value.
+
+Keep builders, payloads and results in your **working directory**:
+- Write files with your file-writing tool, not shell heredocs or `>` redirects into `/tmp`.
+- Run each `beehaven` command on its own rather than chained with `&&`, `;`, `| tee` or `| head`.
+- Save a large result with `beehaven call <action> '…' --out scene.json` (add `--brief` for the
+  receipt), not `> file`: the whole JSON goes to the file and only a line reaches the console.
+
+Agent hosts that gate shell commands (Claude Code, Codex) refuse heredocs containing JSON and
+compound pipelines outright. One recorded run lost four tool calls that way (MCP2-20).
+
 Check `ok`, each `results[].ok`, `count`, `skipped` and `rolledBack`. A failed headless
 batch may roll back its scene writes. Re-read the scene before deciding what to retry;
 do not blindly resend only skipped operations. Mixed live/headless actions may have
@@ -155,6 +174,7 @@ The same rules the in-app assistant works to (canvas-catalog returns them as `te
 - COMPUTED GEOMETRY: shapes from math (radial petals, spirals, grids) are computed and drawn in ONE call: canvas-act { compositionId, script } runs JavaScript with geo.petal/circle/arc/star/polygon/spiral/smooth/rotate/radial/grid/bounds (degrees, 0 = 3 o'clock, clockwise) and a read-only scene list, then applies the { action, args } steps it returns. Never copy numbers from one tool result into another call. A new top-level artboard goes beside or below what scene already holds (or leave out both x and y), never at 0,0 over an existing one.
 - SHEETS: several concepts or variants go on canvas.sheet { name, title, subtitle, cards or count, cols, cardW, cardH }: a live wrap grid of card frames that own their captions, so rows never overlap. Draw into a card with parentId = its id.
 - MEASURE: answer spacing, overlap and alignment questions with canvas.measure { frame } or { ids } and quote its numbers, never from memory or a preview.
+- REFERENCES: canvas.link {id:<exact frame/item id>} returns a stable URL and Markdown for tickets, documents, comments and inbox refs. Use the returned link instead of a screenshot or a guessed canvas URL. It preserves the parent canvas permissions; rename/move retains the layer id.
 - PLACEHOLDERS: until real imagery exists, draw a native placeholder at the final size in a muted brand tone, never a flat accent block: a rect with the slot radius plus a simple motif (an ellipse sun, a vector.create mountain), grouped with canvas.group and named "<section> image placeholder", with any label in its own text node.
 - SECTIONS: one frame per page section, named exactly as its ticket says, inside the page artboard. Build a section in one canvas-act batch that ends with canvas.review of the artboard and carries ticket: { id: "<its key>", status: "done", resolution: "<what changed>" }. The ticket moves only when that review is clean or look; result.ticket says whether it did. Then start the next section.
 
@@ -215,15 +235,13 @@ the backing shows through as dark wedges at the corners.
 
 ### Images from MCP
 
-The artifact MCP tools do not upload local binary files or dispatch `canvas-act`.
-For an image already reachable by an absolute URL, read the composition with
-`clearly_read {type:"composition",target:<id>}`; append an image node
-`{id,type:"image",src,x,y,w,h,fit:"contain",parentId?}` to the returned complete node
-array; then call `clearly_replace {type:"composition",target:<id>,content:<JSON array>,
-expectedRev:<canvasNodesRev>}`. Preserve every existing node. A stale revision is a
-merge prompt, never a reason to drop `expectedRev`. `clearly_edit` can change an
-existing image's `src`, `fit` or geometry, but cannot create a node. For a local file,
-use the CLI upload path first; an MCP-only client needs a URL it can already access.
+A connector-only agent places an image already reachable by an absolute URL (or a
+`data:` URI) as an `image` node in `clearly_canvas_add_nodes`:
+`{id,type:"image",name,src,x,y,w,h,fit:"contain",parentId?}`. Use PNG or JPEG: the headless
+review cannot draw WebP yet and does not say so (CAN2-188), so a WebP image looks blank in the
+review PNG while the editor shows it. `clearly_canvas_update_nodes`
+changes an existing image's `src`, `fit` or geometry. For a local file in a shell, use the
+CLI upload path first.
 
 Current headless rendering can resolve images, including workspace media. A blank image
 can indicate fetch, decoding or budget failure; check its source and inspect a live view.
@@ -236,8 +254,8 @@ design. Verify the returned artifact in a browser at realistic viewport sizes.
 ## Review and scoped perception
 
 ```sh
-beehaven call canvas-act '{"compositionId":"<id>","action":"canvas.review","args":{"ids":["<frame-id>"]}}' > /tmp/review.json
-beehaven call canvas-perceive '{"compositionId":"<id>","includePixels":true,"fitNodeIds":["<frame-id>"],"pixelScale":1}' > /tmp/perceive.json
+beehaven call canvas-act '{"compositionId":"<id>","action":"canvas.review","args":{"ids":["<frame-id>"]}}' --brief
+beehaven call canvas-perceive '{"compositionId":"<id>","includePixels":true,"fitNodeIds":["<frame-id>"],"pixelScale":1}' > perceive.json
 ```
 
 Use the bundled `review.mjs` to decode the PNG and keep base64 out of the conversation.
@@ -268,7 +286,12 @@ MCP's artifact tools are split by operation:
   node array — pass `expectedRev` from the read.
 - `clearly_edit` patches existing nodes with `changes:{ids,patch}`.
 
-Canvas action dispatch and targeted pixels use the CLI. Do not invent
-`clearly_canvas_act`/`clearly_canvas_perceive` tools. If only MCP is available, use the
-artifact schemas for supported edits and state that pixel review remains unverified.
-Do not claim the design was visually checked from scene JSON alone.
+The connector also has native canvas tools for agents without a shell, batched per kind of
+step rather than mixed (MCP2-22): `clearly_canvas_add_nodes` builds a section per call (up to
+300 frames, shapes, text, pen paths and groups of its own new layers),
+`clearly_canvas_update_nodes {edits:[{ids,patch}]}` revises many layers at once,
+`clearly_canvas_measure_text {texts:[…]}` measures several lines, and
+`clearly_canvas_review` returns the saved artboard as an MCP image to inspect. There is no
+`clearly_canvas_act`: a mixed batch is the CLI's `canvas-act`. In a shell, use the CLI; it
+is fewer calls for the same scene. Do not claim the design was visually checked from scene
+JSON alone.

package/skills/clearly-canvas/review.mjs

@@ -123,14 +123,35 @@ export async function saveReviewPixels(report, prefix, fetchImage = fetch) {
   return saveReview({ ...report, preview: { ...report.preview, dataUrl: `data:image/png;base64,${bytes.toString('base64')}` } }, prefix);
 }
 
+/**
+ * `--accept <findings.json>`: findings you have looked at and are keeping, each
+ * `{ check, ids: [node IDs], reason }`. They stay visible in the report and return `look`.
+ * Without this, a deliberate overlap (steam rising out of a bowl) could only be accepted by
+ * hand-writing the canvas.review call — agents did exactly that (MCP2-20).
+ */
+export function takeAccept(argv) {
+  const i = argv.indexOf('--accept');
+  if (i < 0) return { argv, accept: undefined };
+  const file = argv[i + 1];
+  if (!file || file.startsWith('--')) throw new Error('--accept needs a JSON file: [{"check":"text-over-image","ids":["…"],"reason":"…"}]');
+  const accept = JSON.parse(readFileSync(file, 'utf8'));
+  if (!Array.isArray(accept) || !accept.every(a => a && typeof a.check === 'string' && Array.isArray(a.ids) && a.ids.length && typeof a.reason === 'string' && a.reason.trim())) {
+    throw new Error('--accept file must be an array of { check, ids: [node IDs], reason } with a non-empty reason.');
+  }
+  return { argv: argv.filter((_, j) => j !== i && j !== i + 1), accept };
+}
+
 export async function main(argv = process.argv.slice(2)) {
+  const taken = takeAccept(argv);
+  argv = taken.argv;
+  const accept = taken.accept ? { accept: taken.accept } : {};
   if (argv.includes('--many')) {
     const args = argv.filter(a => a !== '--many');
-    if (args.length !== 3) throw new Error('Usage: node review.mjs --many <compositionId> <artboard-ids.json> <output-directory>');
+    if (args.length !== 3) throw new Error('Usage: node review.mjs --many <compositionId> <artboard-ids.json> <output-directory> [--accept findings.json]');
     const [compositionId, manifest, directory] = args;
     const ids = JSON.parse(readFileSync(manifest, 'utf8'));
     if (!Array.isArray(ids) || !ids.length) throw new Error('The manifest must be a non-empty artboard ID/name array.');
-    const response = rpc('canvas-act', { compositionId, action: 'canvas.review-many', args: { ids, preview: 'files' } });
+    const response = rpc('canvas-act', { compositionId, action: 'canvas.review-many', args: { ids, preview: 'files', ...accept } });
     const reports = response.results?.[0]?.value?.reviews;
     if (!Array.isArray(reports) || reports.length !== ids.length) throw new Error('Incomplete review-many result; do not claim all artboards passed.');
     const saved = [];
@@ -140,7 +161,7 @@ export async function main(argv = process.argv.slice(2)) {
   }
   const legacy = argv.includes('--perceive');
   const args = argv.filter((a) => a !== '--perceive');
-  if (args.length !== 3) throw new Error('Usage: node review.mjs <compositionId> <frame-id-or-name> <output-prefix> [--perceive]\n--perceive: older workers, preview only; pass a frame ID.');
+  if (args.length !== 3) throw new Error('Usage: node review.mjs <compositionId> <frame-id-or-name> <output-prefix> [--accept findings.json] [--perceive]\n--perceive: older workers, preview only; pass a frame ID.');
   const [compositionId, id, prefix] = args;
   let report;
   if (legacy) {
@@ -150,7 +171,7 @@ export async function main(argv = process.argv.slice(2)) {
     report = { preview: response.room?.pixels, visualReviewRequired: true,
       limitations: ['Preview-only compatibility mode. Text metrics and scoped audit were not run.'] };
   } else {
-    const response = rpc('canvas-act', { compositionId, action: 'canvas.review', args: { ids: [id], pixelScale: 1 } });
+    const response = rpc('canvas-act', { compositionId, action: 'canvas.review', args: { ids: [id], pixelScale: 1, ...accept } });
     report = response.results?.[0]?.value;
     if (!report?.scope) throw new Error('canvas.review unavailable. Check canvas-catalog; use --perceive for an explicitly preview-only check on an older worker.');
   }

package/skills/index.json

@@ -17,7 +17,7 @@
       "project-plans.md",
       "review.mjs"
     ],
-    "sha": "e23e25bd46b02d9c"
+    "sha": "4b478740d8ef22c2"
   },
   {
     "name": "agent-home",