merge: upstream a5b56b7 (#444 system prompt, #443 locale tests, #435 cron week, #434 deferred rebuild)

README.md

@@ -178,6 +178,28 @@ Section 3 of [harness integration](docs/harness-integration.md) has the full
 endpoint reference with captured request/response examples, the error
 contract, and a minimal client in JavaScript and Python.
 
+### Lightweight conversations
+
+For simple conversation or general knowledge questions, start a new Session with
+`--mode lightweight`:
+
+```bash
+kcode --mode lightweight "Explain how DNS caching works."
+kcode exec --mode lightweight "Summarize the CAP theorem."
+```
+
+Lightweight mode sends a small conversational system prompt and no tool schemas. It omits workspace
+instructions, Skills, memory blocks, MCP schemas, and the full environment block from provider
+context, so it cannot inspect or change local files. Start a new standard Session for coding or tool use with
+`kcode --mode standard` or `kcode exec --mode standard`; omitting `--mode` is identical to standard
+mode.
+
+The mode is fixed when a root Session is created. `--mode lightweight` therefore cannot be combined
+with `--session` or `--continue`. Child/sub-agent Sessions do not inherit it, `/compact` continues to
+use the standard compaction context, and ACP and desktop Sessions remain standard. Reopening a
+lightweight Session keeps its mode and shows `Lightweight` in the TUI status line; opening a standard
+Session from a lightweight launch remains standard and shows no indicator.
+
 ### Continue your work
 
 ```bash

docs/demo.md

@@ -6,15 +6,13 @@
 
 ## A new feature, then a new requirement
 
-The starter is a small, predesigned browser pet that blinks and says hello. The first request asks the agent to add a Pomodoro timer. The follow-up changes pause to an 800 ms hold and adds a completion celebration. The finished example runs locally without hardware or frontend dependencies.
+The starter is a small, predesigned browser pet that blinks and says hello. The first request asks MiniMax Code to add a Pomodoro timer. The follow-up changes pause to an 800 ms hold and adds a completion celebration. The finished example runs locally without hardware or frontend dependencies.
 
 The video pairs actual browser recordings with clearly labeled prompt excerpts, generated source, and excerpts of real headless output. The interface shown in the final demonstration includes subsequent maintainer design work. It is an edited demonstration, not a recording of the interactive TUI. See the [full prompts](../examples/pocket-pet) and try the same changes in your own copy.
 
-The source in this repository names **Kinetick Code**. The video and still are the upstream recording and can still show **MiniMax Code** in the page chrome. That caption is the recording, not the identity of this fork.
-
 ## Recording and verification
 
-- Recorded upstream on September 28–29, 2026 (Asia/Shanghai), using installed MiniMax Code **0.5.8** with an existing **BYOK** configuration. This is a CLI workflow demonstration, not a model evaluation. Kinetick Code keeps its own version (**0.6.8**).
+- Recorded on September 28–29, 2026 (Asia/Shanghai), using the installed MiniMax Code **0.5.8**, with the existing **BYOK** configuration. This is a CLI workflow demonstration, not a MiniMax-model evaluation.
 - The two real agent turns took approximately **102 seconds** and **154 seconds**. Model waiting time is omitted from the 36-second edit; it is not a speed benchmark. Browser footage plays at its captured speed.
 - The first run added the timer and passed 6 tests. The follow-up added hold-to-pause and the celebration and passed all 13 tests. Both runs explicitly said browser behavior was unverified; the maintainer then checked it in Chrome.
 - Browser checks cover start, short-click behavior, pointer hold/pause/resume, keyboard hold/repeat/release/resume, the default 25-minute setting, canceled holds, completion, reduced-motion behavior, and a 390 px viewport without horizontal overflow.
@@ -23,8 +21,37 @@ The source in this repository names **Kinetick Code**. The video and still are t
 - Only synthetic project content appears. Private run logs, account state, provider aliases, local paths and session identifiers are excluded. Captions and layout are editorial; the cursor highlight follows actual pointer events. Browser footage is cropped for readability and its interaction segment plays continuously at 1×. No tool results were invented.
 - Video assembly uses HyperFrames with locally staged browser footage. The video is silent so it works in muted README and social contexts. The revised cut uses a warm page and a cobalt device, with larger interaction shots and visible action labels.
 
-To edit the starter, install Kinetick Code and configure `kcode` with your own model, or with a MiniMax account that has available credits. Calls may incur charges. Running the finished example requires only Node.js and a browser.
+To edit the starter, install and configure MCode with a MiniMax account with available credits or your own compatible model API. Calls may incur charges. Running the finished example requires only Node.js and a browser.
+
+---
+
+## Earlier demo: a small code repair
+
+![Real MiniMax Code terminal replay: request, failing tests, a code fix, and passing tests](assets/tui-demo.gif)
+
+[Play or download the MP4](assets/tui-demo.mp4) · [View the full-size still](assets/tui-demo.png) · [Example source](../examples/clamp)
+
+## What was recorded
+
+The interactive TUI built from this repository ran in a temporary directory containing only `clamp.mjs` and `clamp.test.mjs`, using an existing MiniMax Token Plan login and `MiniMax-M3`. The model actually read the files, ran `node --test`, edited the function, and reran the tests. Initially two failed and one passed; after the fix, all three passed. The test file was unchanged.
+
+The recorded task:
+
+> Read clamp.mjs and clamp.test.mjs. Run node --test to reproduce the failure, fix clamp without changing the tests, then run the tests again. Reply briefly in English.
+
+Full access was used only for this isolated synthetic project. Use `/permission` to choose an appropriate mode for your own work.
+
+## Recording and editing
+
+- Recorded on 2026-09-11 from source commit `fe49bbd73d4bb2df6e5873801af12695378c0b62`, TUI version 0.3.11.
+- Actual ANSI output was captured through a PTY and replayed at 110 columns × 36 rows. Application text, colors, and layout are retained; the outer window title and provenance caption are recording decoration.
+- The still shows the same session after `Ctrl+O` expanded tool details and the viewport was scrolled upward, including the actual diff and test output.
+- The 20-second edit keeps task order, shortens waits and thinking, and holds on the final diff and tests. **It is not real-time playback or a performance benchmark.** No model output, tool calls, or test results were fabricated.
+- `@xterm/headless` interpreted terminal cells, which were rendered as SVG / PNG. FFmpeg encoded the GIF and silent H.264 MP4. A border keeps the dark terminal frame visible on both light and dark GitHub backgrounds.
+- No private project, account page, or key was recorded, and no user test image was used. Raw recordings and runtime logs remain local and are not committed as release assets.
+
+This demonstrates one small code repair, not acceptance of every project, provider, or tool. See the [verification records](verification.md) for broader evidence.
 
-## Code repair exercise
+## Brand assets
 
-[`examples/clamp`](../examples/clamp) is the small failing-test exercise. Copy it out of the repository before asking Kinetick Code to fix it. The README screenshot is a Kinetick Code session. Older terminal recordings stay in `docs/assets` as historical captures.
+The README's light and dark wordmarks use the existing TUI welcome logo, with blue and cyan from its theme. `assets/social-preview.png` is a local 1280 × 640 sharing card for maintainers to configure as the GitHub Social Preview at release time. Creating these assets does not publish the repository; nothing was uploaded to a third-party media host.

docs/examples.md

@@ -156,6 +156,40 @@ text-only again, remove `image` from `modalities.input` and remove or set
 declare image input support. Restart KCode after editing configuration and select
 the configured model with `/model`, or use `exec --model` for a single run.
 
+Images stay in the conversation history, so a long session can carry more images
+than a provider accepts in one request. KCode sends only the 20 most recent images
+per request and replaces older ones with a short text note naming the original
+file; the stored history keeps every image. If your provider allows fewer images
+per request, set the limit on the model (Mistral's API is limited to 8
+automatically):
+
+```yaml
+custom_provider:
+  my-vision-provider:
+    models:
+      my-vision-model:
+        capabilities:
+          support_image: true
+          max_images_per_request: 8
+```
+
+Images stay in the conversation history, so a long session can carry more images
+than a provider accepts in one request. MCode sends only the 20 most recent images
+per request and replaces older ones with a short text note naming the original
+file; the stored history keeps every image. If your provider allows fewer images
+per request, set the limit on the model (Mistral's API is limited to 8
+automatically):
+
+```yaml
+custom_provider:
+  my-vision-provider:
+    models:
+      my-vision-model:
+        capabilities:
+          support_image: true
+          max_images_per_request: 8
+```
+
 After signing in to MiniMax, try a task that explicitly requires search:
 
 > Use web_search to find the official Node.js test runner documentation. Summarize how to run tests and include the source URL. If the tool is unavailable, say so.