@akanjs/cli 3.0.0-alpha.49 → 3.0.0-alpha.50

package/.build-stamp

@@ -1 +1 @@
-461d177fea7bea8b38847fed3e7046a101f7ad110bf3e5e3e577f6fb1ce9531f
+fc931b5c84401e9bf2bda97ced4e93b6df47eb3293b40b306161d09542b821d7

package/guidelines/agentRule/agentRule.instruction.md

@@ -385,6 +385,15 @@ apps and libs never import it directly (`no-import-external-library`) — everyt
   `highlight` scrolls its target into view and flashes it **once the scroll lands**, since a smooth scroll across a
   long page outlasts the flash; it is the one built-in that exists for the *user's* benefit, because showing where
   a control is beats writing directions to it.
+- **Every tool's return value is bounded at 20,000 characters before it enters the transcript** (`ToolOutput.limit`,
+  clipped with a note that gives the real size and tells the model to read a narrower part). `readScreen` caps
+  itself; nothing else did, and a store key is sized for a screen rather than for a model's window — one
+  `readState` of a list whose rows carry inlined bytes is megabytes, which the loop then posts on this turn *and on
+  every turn after it*, so the window fills after four messages. Compaction cannot save that: it summarizes what is
+  above the cut and a result this large arrives below it. Clipped rather than dropped, because a model handed a
+  value it cannot see the end of asks a narrower question, where one handed nothing answers from the field names.
+  The cap is a safety rail, not the design: a tool should return what answers the question — an id, a count, the
+  three fields the agent asked about — and `AgentContext` inlines only small primitives for the same reason.
 - **A slow tool reports its own progress with `AgentProgress.report(message, { done, total })`** from wherever the
   work is — a store action, an upload loop, an adapter — reached through a module slot rather than a parameter, and
   a no-op when nobody is rendering it. The chat shows it on that call's row until the row resolves. It is the

package/guidelines/conventions/conventions.instruction.md

@@ -531,6 +531,12 @@ package; apps and libs never import it directly (`no-import-external-library`)
   nothing — normalize with the control's `transform` prop, and multi-write with a `_postSet<Field>` store method.
 - **Reading is per key, not per store.** `st.use.x({ agent: false })` keeps a subscribed key off the surface;
   a key no component reads is unreadable. A value with populated `hidden`/`secret` fields is refused at read.
+- **Return what answers the question, not the record.** A tool's value is capped at 20,000 characters before it
+  enters the transcript and clipped with a note the model reads (`ToolOutput.limit`), because a store key is sized
+  for a screen and not for a model's window: one `readState` of a list whose rows carry inlined bytes is megabytes,
+  and from there it rides *every* later turn. Each tool row in the chat carries its own estimated token cost and
+  opens onto the value the model was handed — that row is where a conversation that filled up in four messages
+  explains itself.
 - **Route guidance is `<Agent.Guide instructions="..." />`** rendered from a `_layout.tsx` or page — the render
   tree is the cascade. It is a component, not a `pageConfig` field, and `*.abstract.md` is never served to agents.
 - The framework publishes five built-ins on every store surface — `navigate`, `goBack`, `readScreen`,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@akanjs/cli",
-  "version": "3.0.0-alpha.49",
+  "version": "3.0.0-alpha.50",
   "sourceType": "module",
   "type": "module",
   "publishConfig": {
@@ -34,7 +34,7 @@
     "@langchain/openai": "^1.4.6",
     "@tailwindcss/node": "^4.3.0",
     "@trapezedev/project": "^7.1.4",
-    "akanjs": "3.0.0-alpha.49",
+    "akanjs": "3.0.0-alpha.50",
     "chalk": "^5.6.2",
     "commander": "^14.0.3",
     "dayjs": "^1.11.20",