Clickable poster (the mode picker) links to docs/iClaw_video.mp4, a ~50s no-sound walkthrough of inspecting an untrusted shell script in an isolated sandbox.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Use <picture> with prefers-color-scheme so the star-history chart matches the viewer's GitHub theme, instead of a single light-only image.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…oster

Demo now embeds the GitHub user-attachments asset so it plays inline in the README, like other repos. Removes the redundant committed docs/iClaw_video.mp4 (3.2 MB) and poster image now that GitHub hosts the video.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Long scheduled messages (and queued ones — they share the card) rendered at
full height and could take over the whole composer area. Clamp the message
text to 2 lines with a "Show more"/"Show less" toggle that appears only when
the content actually overflows; short messages are unchanged.

Restructure the card into a text block above a full-width footer strip:
schedule time (⏰) + toggle on the left, action buttons pushed to the right.
Overflow is measured client-side (applyPendingRowClamp) across both render
paths — server EJS first paint and JS live render — and both lists.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Pain-led pitch + comparison table vs OpenClaw, Hermes and Claude Code (sandbox safety, project isolation, cost / no lock-in, one-key setup, minimalist UI). Setup difficulty shown as a skull meter.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Work/Secure/Incognito sent image_url blocks to whatever ICLAW_MODEL is set, so a text-only default (deepseek/deepseek-v4-flash) made OpenRouter return "404 No endpoints found that support image input" (hit in chats 314/321). Full Power was unaffected — it forwards to OpenClaw's gateway, whose model is vision-capable.

Mirror OpenClaw's capability gate (iClaw is OpenRouter-only): new agent/model-capabilities.ts reads architecture.input_modalities from OpenRouter /models (cached 1h, stale-on-error, ICLAW_VISION_MODELS allowlist escape hatch). resolveTurnModel() only looks up on image turns and routes a confirmed text-only model to ICLAW_VISION_MODEL (default google/gemini-2.5-flash); text turns are untouched. Wired into both agent loops (loop.ts, secure-runner.ts) via an effectiveModel.

describeApiError is exported from loop.ts + gains an image-404 safety net for the unknown-capability path; secure-runner now uses it too. Tests: test/unit/modelCapabilities.test.ts (8).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Switch the context-compaction (summary) model default from
google/gemini-2.5-flash-lite to tencent/hy3-preview. Still
overridable via ICLAW_SUMMARY_MODEL.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- waveform now advances on a fixed 55ms cadence (was once per animation
  frame), so it scrolls ~3x slower — matching the Flutter app amplitude
  stream (~50ms)
- raise gain (2.4 -> 3.0) and carry the peak between samples so bars track
  loudness (taller when loud)
- hide the floating lock pill once locked; it was overlapping the hint
- locked state: drop the "tap mic to send" text and move Cancel down to the
  toolbar row, centered and level with the send button (wave gets the full row)
- center the quick-tap "hold the mic to record" hint

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The secure-workspace bar keyed its visibility (and Export/Apply/Destroy +
TTL) off `rawChatId` — a const captured once at page load. A chat that
starts as a draft has no id until the first message adopts it; adoptDraftChat
updates messagesEl.dataset.chatId but rawChatId stays empty, so the bar
stayed hidden until a reload re-rendered at /chats/:id.

Read the id live via a secureChatId() helper instead, and refresh the bar on
turn-ended so it also appears after a slow first-turn sandbox start (the fixed
post-send retry timers can miss it).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Model defaults (all overridable via ICLAW_*_MODEL):
- Main agent, web-search, runtime summaries → minimax/minimax-m2.7
- Host chat compaction → tencent/hy3-preview; chat titles now reuse it
- STT / Vision stay on google/gemini-2.5-flash

Remove the unused "Ask" direct-OpenRouter path (replaced by Incognito):
- Drop streamComplete()/StreamCompletionOpts (zero callers) and the
  askModel config field; complete() now requires an explicit model.
- Refresh stale "Ask mode" copy in Settings (tab/section -> "Voice"),
  .env.example, and code comments. The live task-page Ask is untouched.

typecheck + full vitest suite (421 tests) pass; Settings page verified to render.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Secure Mode shipped no web-research tools: web_fetch/web_search are
host-side (a host fetch would bypass the container's --network gate), so
the agent fell back to hand-rolling `curl ... | jq` for any web/API call.

Add Secure variants that keep the sandbox boundary intact:
- web_fetch runs the fetch as curl INSIDE the sandbox (runInSandbox),
  then reuses the existing canonicalize / htmlToText / summarize / cache
  path on the host. The URL is shQuote'd against shell injection; the
  body is capped in-container (head -c) under the docker-exec stdout
  buffer, with the status+content-type marker emitted before the body.
- web_search uses only the OpenRouter web plugin (host->OpenRouter, same
  trust as the chat stream), gated on networkEnabled; no host DDG.

Wire both into the secure loop and steer the secure prompt to prefer
them over raw curl. Work/Incognito are unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Remove the unreachable "Apply changes" feature end-to-end — UI button,
host route /chats/:id/apply-sandbox, runtime /sessions/:id/apply,
secure-export helpers, and its tests. Secure Mode has no folder-ingest
UI, so apply could only ever report "nothing to apply". The dormant
copyFolders ingest plumbing is left in place.

Secure workspace bar:
- TTL is now per-chat (default 7 days); the whole bar stays hidden until
  the sandbox actually holds something.
- Unified "delete" vocabulary: "deletes in 6d" with an inline-clickable
  value that opens the retention menu; Destroy and Save are icon buttons
  that slide out their label on hover.
- "Save to my computer" (export) confirms with a plain-language warning
  that the files leave the sandbox; exported folders are named by local
  date-time instead of an epoch number.

Chat header: Share and Delete collapse to icons with hover-reveal labels
on desktop (mobile was already icon-only).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>