…del generations

- thinking.py now normalizes model IDs to their claude- core, so Bedrock
  (anthropic.claude-..., us.anthropic....-v1:0), LiteLLM
  (anthropic/claude-...), and Vertex (claude-...@Date) IDs classify like
  bare IDs instead of falling into the budget_tokens path (400 on >=4.7
  generation models).
- Inverted classification: adaptive thinking is the default; only the
  frozen legacy budget_tokens families (2.x/3.x, Haiku 4.5, Sonnet 4/4.5,
  Opus 4/4.1/4.5) keep the integer budget, so unknown future Claude
  models work by default.
- New accepts_sampling_params() / supports_disabled_thinking() helpers;
  AndroidAgent act defaults now use make_non_thinking_settings(), which
  drops temperature on models that reject sampling params (Opus 4.7+,
  Sonnet 5, Fable 5) and omits thinking "disabled" on always-on models
  (Fable 5 / Mythos 5).
- EffortLevel gains "xhigh" (supported from the Opus 4.7 generation on).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(models): thinking config for gateway model IDs + new model generations

OpenAI-style usage counts prompt_tokens_details.cached_tokens INSIDE
prompt_tokens, while UsageParam consumers (conversation statistics
callback, HTML reporting) treat the fields as disjoint, Anthropic-style —
so cached tokens were counted and billed twice for OpenAI-route models
(observed ~5x cost overstatement on cache-heavy Gemini agentic runs).
Subtract the cached subset at the response boundary so UsageParam means
the same thing on every route. Guarded so absent/malformed details pass
through unchanged.

fix(openai): report cached tokens disjoint from input_tokens

When the AskUI controller returns a 0×0 bitmap (e.g. because the
display is minimised or unavailable), PIL cannot encode the resulting
image to PNG. This caused the SimpleHtmlReporter to crash and be
permanently disabled by ReporterErrorHandler, losing the entire HTML
report for the run.

Two-layer fix:
- askui_controller.py: `_check_bitmap_dimensions()` raises
  `AskUiControllerError` immediately after a zero-size bitmap is
  received, before any PIL image is constructed or passed downstream.
- reporting.py: `normalize_to_pil_images()` now filters out any
  zero-size PIL images as defence-in-depth, so no reporter can crash
  on an empty image regardless of its origin.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: reject zero-size screen captures before they reach the reporter

bump version to 0.43.0

Reading a remote file or directory that does not exist raised a
DesktopAgentOsError that inherited from BaseException, so the tool-calling
loop's `except Exception` handler never caught it. The error propagated all
the way up and crashed the run instead of being surfaced to the agent as a
tool error result, as recoverable tool failures are everywhere else.

Introduce two error types that map onto the existing fatal/recoverable
handling in the tool-calling loop:

- DesktopAgentOsException (recoverable): a plain Exception for failures the
  agent can work around (path not found, undecodable file contents). The
  loop's generic `except Exception` catches it and returns it to the agent.
- DesktopAgentOsError (unfixable): now inherits from AutomationError for
  protocol violations (unexpected response type, missing error and response).
  The loop's existing `except (AgentError, AutomationError): raise` re-raises
  it, terminating the run cleanly instead of crashing on an uncaught
  BaseException.

Controller operation failures (res.error) and undecodable payloads now raise
the recoverable type; contract violations keep the fatal type. No changes to
the generic tool-calling loop are required.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…models

Models of the Sonnet 5 generation onward default the thinking `display`
setting to "omitted": the API returns thinking blocks whose text is empty
while the full thinking tokens are still billed. Any consumer relying on
make_thinking_settings() therefore silently loses all visible reasoning in
reports and logs on the newest models.

Set display: "summarized" explicitly in the adaptive branch. Billing is
identical for both display modes, so this only restores visibility.
Verified end-to-end against Vertex claude-sonnet-5 (empty before, 3k chars
of summarized thinking after).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Make the tool description explicit that the tool lists only regular files
and never includes folders/subdirectories, so the agent does not try to use
it to discover or navigate directories.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…iption

docs: clarify get_file_names_tool returns files only, not folders

fix(models): request visible thinking summaries on adaptive-thinking models

…ble-exception

fix: split Desktop Agent OS errors into recoverable vs unfixable