mindweave 3.2.5 → 3.2.7

package/README.md

@@ -1,113 +1,113 @@
-<h1 align="center">Mindweave</h1>
-
-<p align="center">
-  A coding agent for your terminal.<br>
-  You bring your own key. It runs on your machine. Nothing goes anywhere else.
-</p>
-
-<p align="center">
-  <a href="LICENSE">Apache 2.0</a> &nbsp;•&nbsp;
-  <a href="https://mindweavedev.netlify.app/">Website</a> &nbsp;•&nbsp;
-  <a href="https://github.com/mindweave-cli/mwcode">Desktop app</a> &nbsp;•&nbsp;
-  <a href="https://x.com/mindweavecli">X</a>
-</p>
-
----
-
-Mindweave works inside your repository: it reads, searches, edits, runs commands and
-checks its own work. There is no backend and no account. Your code and your key only
-ever go to the model provider you picked.
-
-Prompts are kept thin on purpose, so the context goes on your code rather than on
-scaffolding.
-
-Prefer a window to a terminal? [mwcode](https://github.com/mindweave-cli/mwcode) is the
-desktop app, built on the same engine, for Windows, macOS and Linux.
-
-## Install
-
-Needs **Node.js 24+**. The npm package is tested on **Windows** and **Linux**. It has not
-been run on a real Mac yet, so on macOS the desktop app is the safe choice
-([details](KNOWN-ISSUES.md)).
-
-```bash
-npm install -g mindweave
-cd your-project
-mindweave          # or mw
-```
-
-It asks for an API key on first launch and keeps it in `~/.mindweave/.env`.
-
-From source:
-
-```bash
-git clone https://github.com/mindweave-cli/mindweave
-cd mindweave
-npm install
-npm run build
-npm link
-```
-
-## What it does
-
-- **16 providers, 63 models** — DeepSeek, Anthropic, OpenAI, Gemini, xAI, Mistral, Groq, Cerebras,
-  Qwen, Kimi, GLM, Meta, MiniMax, Tencent, OpenRouter, and local models through Ollama
-  with no key at all. [PROVIDERS.md](src/drivers/PROVIDERS.md)
-- **Real tools** — file reads and edits, search, a shell with background jobs,
-  sub-agents, web search, and a tool that tests your app by using it.
-- **Code intelligence** — tree-sitter and language servers index the repo in the
-  background, at no token cost, and every edit comes back with that file's errors.
-- **Goals that run on their own** — `/marathon` takes a goal and keeps going until it is
-  verified done, blocked, or out of budget.
-- **Memory** — project notes in MINDWEAVE.md, compaction that keeps what matters, and
-  earlier sessions it can look back at.
-- **Safety nets** — `/undo`, `/rewind`, approval modes, per-project rules for paths
-  and commands it must not touch, and a project's own servers and hooks never run unasked.
-- **MCP servers** — connect external tools with `/mcp`. [docs/MCP.md](docs/MCP.md)
-
-`/help` lists every command. `shift+tab` switches modes. Type while it works and your
-message queues.
-
-## Your own rules and hooks
-
-Both live in your state folder (`~/.mindweave`, or `projects/<name>` inside it for one
-project), never in the repository, so a project you clone cannot add them.
-
-**`command-rules.md`** sets what Sentinel asks about, one line per rule, matched on the start
-of the command:
-
-```
-allow npm test
-prompt git push
-forbid rm -rf :: use the trash folder
-```
-
-`forbid` refuses the command in every mode and tells the agent your reason. `prompt` always
-asks. `allow` stops the question for that command, but never for one that writes a file or
-that Mindweave cannot read with certainty.
-
-**`hooks.json`** runs your own commands at set moments. Each gets one JSON object on its
-input, and exit code 2 stops the action and tells the agent why:
-
-```json
-{ "hooks": {
-  "PreToolUse": [ { "matcher": "run_command|edit", "command": "node check.js" } ],
-  "Stop": [ { "command": "npm test --silent" } ]
-} }
-```
-
-The moments are `PreToolUse`, `PostToolUse`, `UserPromptSubmit`, `Stop` and `SessionStart`.
-A hook is stopped after 60 seconds unless it sets `timeoutMs`.
-
-## More
-
-[CHANGELOG](CHANGELOG.md) · [Known issues](KNOWN-ISSUES.md) ·
-[Philosophy](PHILOSOPHY.md) · [Contributing](CONTRIBUTING.md) · [Security](SECURITY.md)
-
-Contributions are welcome, model drivers especially. Bug reports are the most useful
-thing you can send ([what to include](CONTRIBUTING.md#reporting-a-bug)). Questions:
-zallinimann@gmail.com
-
-## License
-
-[Apache License 2.0](LICENSE)
+<h1 align="center">Mindweave</h1>
+
+<p align="center">
+  A coding agent for your terminal.<br>
+  You bring your own key. It runs on your machine. Nothing goes anywhere else.
+</p>
+
+<p align="center">
+  <a href="LICENSE">Apache 2.0</a> &nbsp;•&nbsp;
+  <a href="https://mindweavedev.netlify.app/">Website</a> &nbsp;•&nbsp;
+  <a href="https://github.com/mindweave-cli/mwcode">Desktop app</a> &nbsp;•&nbsp;
+  <a href="https://x.com/mindweavecli">X</a>
+</p>
+
+---
+
+Mindweave works inside your repository: it reads, searches, edits, runs commands and
+checks its own work. There is no backend and no account. Your code and your key only
+ever go to the model provider you picked.
+
+Prompts are kept thin on purpose, so the context goes on your code rather than on
+scaffolding.
+
+Prefer a window to a terminal? [mwcode](https://github.com/mindweave-cli/mwcode) is the
+desktop app, built on the same engine, for Windows, macOS and Linux.
+
+## Install
+
+Needs **Node.js 24+**. Every change is tested on **Windows**, **macOS** (Apple silicon) and
+**Linux**. Driving a native desktop app and taking window screenshots are Windows-only for
+now, and the rest of the limits are in [KNOWN-ISSUES.md](KNOWN-ISSUES.md).
+
+```bash
+npm install -g mindweave
+cd your-project
+mindweave          # or mw
+```
+
+It asks for an API key on first launch and keeps it in `~/.mindweave/.env`.
+
+From source:
+
+```bash
+git clone https://github.com/mindweave-cli/mindweave
+cd mindweave
+npm install
+npm run build
+npm link
+```
+
+## What it does
+
+- **16 providers, 63 models** — DeepSeek, Anthropic, OpenAI, Gemini, xAI, Mistral, Groq, Cerebras,
+  Qwen, Kimi, GLM, Meta, MiniMax, Tencent, OpenRouter, and local models through Ollama
+  with no key at all. [PROVIDERS.md](src/drivers/PROVIDERS.md)
+- **Real tools** — file reads and edits, search, a shell with background jobs,
+  sub-agents, web search, and a tool that tests your app by using it.
+- **Code intelligence** — tree-sitter and language servers index the repo in the
+  background, at no token cost, and every edit comes back with that file's errors.
+- **Goals that run on their own** — `/marathon` takes a goal and keeps going until it is
+  verified done, blocked, or out of budget.
+- **Memory** — project notes in MINDWEAVE.md, compaction that keeps what matters, and
+  earlier sessions it can look back at.
+- **Safety nets** — `/undo`, `/rewind`, approval modes, per-project rules for paths
+  and commands it must not touch, and a project's own servers and hooks never run unasked.
+- **MCP servers** — connect external tools with `/mcp`. [docs/MCP.md](docs/MCP.md)
+
+`/help` lists every command. `shift+tab` switches modes. Type while it works and your
+message queues.
+
+## Your own rules and hooks
+
+Both live in your state folder (`~/.mindweave`, or `projects/<name>` inside it for one
+project), never in the repository, so a project you clone cannot add them.
+
+**`command-rules.md`** sets what Sentinel asks about, one line per rule, matched on the start
+of the command:
+
+```
+allow npm test
+prompt git push
+forbid rm -rf :: use the trash folder
+```
+
+`forbid` refuses the command in every mode and tells the agent your reason. `prompt` always
+asks. `allow` stops the question for that command, but never for one that writes a file or
+that Mindweave cannot read with certainty.
+
+**`hooks.json`** runs your own commands at set moments. Each gets one JSON object on its
+input, and exit code 2 stops the action and tells the agent why:
+
+```json
+{ "hooks": {
+  "PreToolUse": [ { "matcher": "run_command|edit", "command": "node check.js" } ],
+  "Stop": [ { "command": "npm test --silent" } ]
+} }
+```
+
+The moments are `PreToolUse`, `PostToolUse`, `UserPromptSubmit`, `Stop` and `SessionStart`.
+A hook is stopped after 60 seconds unless it sets `timeoutMs`.
+
+## More
+
+[CHANGELOG](CHANGELOG.md) · [Known issues](KNOWN-ISSUES.md) ·
+[Philosophy](PHILOSOPHY.md) · [Contributing](CONTRIBUTING.md) · [Security](SECURITY.md)
+
+Contributions are welcome, model drivers especially. Bug reports are the most useful
+thing you can send ([what to include](CONTRIBUTING.md#reporting-a-bug)). Questions:
+zallinimann@gmail.com
+
+## License
+
+[Apache License 2.0](LICENSE)

package/dist/cli/App.js

@@ -76,7 +76,7 @@ import { isTight } from "./blockSpacing.js";
 import { parseScreenArg, screenChoices, screenNotice, startupMode } from "./screenMode.js";
 import { applyScreenMode } from "./screenShell.js";
 import { saveScreenMode } from "./screenStore.js";
-import { needsMeasure, pruneHeights } from "./blockHeights.js";
+import { FULL_TAIL_BLOCKS, needsMeasure, pruneHeights, seedHeights } from "./blockHeights.js";
 import { expandableAt, hitBlock, hitItem } from "./expandHits.js";
 import { shellNote } from "./shellNotes.js";
 import { readFileSync, statSync } from "node:fs";
@@ -3761,15 +3761,27 @@ export function App({ resumeSessionId, initialScreen }) {
     // bottom — so in practice this prefix is everything but the last block or two.
     let known = 0;
     const knownHeights = [];
-    while (known < rendered.length) {
-        const block = rendered[known];
-        const entry = blockHeights.current.get(block.id);
-        // The identity check IS the invalidation: a block that changed is a new object, so
-        // its recorded height belongs to the version before the change.
-        if (entry === undefined || entry.block !== block)
-            break;
-        knownHeights.push(entry.height);
-        known++;
+    const extendKnown = () => {
+        while (known < rendered.length) {
+            const block = rendered[known];
+            const entry = blockHeights.current.get(block.id);
+            // The identity check IS the invalidation: a block that changed is a new object, so
+            // its recorded height belongs to the version before the change.
+            if (entry === undefined || entry.block !== block)
+                break;
+            knownHeights.push(entry.height);
+            known++;
+        }
+    };
+    extendKnown();
+    // A long conversation reopened has no heights at all, and laying every block out in full just to
+    // learn them cost about a millisecond each: 28 seconds for a very long session, with nothing on the
+    // screen until it was done. Only the newest few hundred are laid out now (they are the ones in view);
+    // the older ones take an ESTIMATE that is marked as one, exactly as after a resize, and each is
+    // measured for real the first time it comes near the window. See seedHeights.
+    if (rendered.length - known > FULL_TAIL_BLOCKS) {
+        seedHeights(blockHeights.current, rendered, known, width);
+        extendKnown();
     }
     const win = virtualWindow(knownHeights, -chatOffset, chatRows);
     if (perfEnabled()) {