Answer budget within the context window; enforced reviews of completed todos

Context window
- Requests asked for the model's full max_output, but providers count
  input + output against the window: long chats with large-output models
  (DeepSeek V4.1 Flash, 384k) failed with "input too long" before automatic
  compaction (85 %) kicked in. max_tokens is now what the window leaves
  (estimated input = last usage + messages since), the chat is compacted when
  less than 32k would remain for an answer, and a context-overflow error
  compacts and retries once.
- ModelConfig.max_input (catalogue limit.input, editable in the provider
  settings) lowers the compaction threshold for models with an input limit.

Enforced reviews (modes with refine.rounds > 0, refine.after_todos default on)
- Every todo marked completed is reviewed against the work since the last
  review; with findings it is reopened and the feedback goes to the agent
  (max. rounds per todo). Passed reviews leave a visible note.
- Ending the turn with open todos sends the agent back to work while it makes
  progress (max. 10 reminders in a row without progress); such modes may take
  up to 2,000 steps per turn. Reaching the step limit is reported.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

CHANGELOG.md

@@ -3,6 +3,29 @@
 All notable changes to CodeCraft. Versions follow [Semantic Versioning](https://semver.org/);
 until 1.0, minor versions may contain breaking changes.
 
+## [Unreleased]
+
+### Added
+
+- **Enforced reviews:** when a review is configured for a mode, every todo the agent marks
+  completed is reviewed; with findings the todo is reopened and the agent has to fix them and
+  complete it again (up to the configured rounds per todo). Ending the turn with open todos sends
+  the agent back to work as long as it makes progress – for unattended runs, e.g. overnight
+  (such modes may take up to 2,000 steps per turn). The chat shows passed reviews and these
+  reminders; *Settings → Modes* can switch it off per mode (then only one review at the end of a
+  turn, as before).
+- **Max input per model:** providers that limit the input below the context window (e.g. 272k of
+  400k) are taken into account; fetched from the model catalogue, editable in the provider settings.
+
+### Fixed
+
+- **Context overflow instead of compaction** (e.g. DeepSeek V4.1 Flash with 384k max output):
+  requests asked for the model's full max output, and providers count input + output against the
+  window, so long chats failed before automatic compaction kicked in. The output budget is now what
+  the window leaves, the chat is compacted when no reasonable room for an answer is left, and a
+  "context too long" error compacts and retries once instead of stopping. A turn that hits the
+  step limit now says so.
+
 ## [v202611-00017] – 2026-11-10
 
 ### Added