OpenAI reasoning models (o-series, gpt-5) always reason internally;
omitting the reasoning parameter does not disable reasoning -- it uses
the model's default effort. Those hidden reasoning tokens count against
max_output_tokens, so with a small budget (e.g. title generation) the
model can exhaust all tokens on reasoning and return empty visible text.

When NoThinking is set, explicitly send reasoning_effort: low so the
model spends as few output tokens as possible on reasoning. We use
"low" rather than "minimal" because older models (o3-mini, o1) only
accept low/medium/high.

Also normalize the reasoning field in the VCR request matcher so
cassettes recorded with or without reasoning config still match.

Fixes #2318

The previous budget of 20 max_output_tokens was sufficient for
non-reasoning models, but reasoning models (o-series, gpt-5) include
hidden reasoning tokens in that same budget. Even with low effort,
20 tokens is too tight -- the model can exhaust them on reasoning and
produce no visible title text.

Raise the budget to 256, which gives reasoning models enough room for
low reasoning overhead plus a short title.

Refs #2318

Keep titleMaxTokens at 20 for non-reasoning models. The OpenAI client
now enforces a minimum output-token budget (256) only when NoThinking
is set on a reasoning model, where hidden reasoning tokens would
otherwise starve the visible text.

LLMs occasionally generate structurally invalid JSON for edit_file
arguments when the text payload contains characters that confuse brace
counting (YAML, Dockerfiles, shell scripts). The most common pattern is
an extra closing brace or bracket near the end of the arguments.

Go's json.Unmarshal scanner rejects this before any custom UnmarshalJSON
method runs, so the existing double-serialization fix never fires for
these cases.

Add tryRepairEditFileJSON which iteratively removes the offending
character at the json.SyntaxError offset (up to 3 rounds), handling
extra }, extra ], stray backslash-n, and stray backslash before quotes.

Replace NewHandler with a custom editFileHandler that calls
ParseEditFileArgs (which does repair then unmarshal) before delegating
to handleEditFile. Update all other call sites (ACP filesystem handler,
TUI rendering) to use ParseEditFileArgs as well.

Automated changelog update for release v1.46.0

Assisted-By: Docker Agent

docs: update CHANGELOG.md for v1.46.0

fix: title generation fails with OpenAI reasoning models

Claude Opus 4.7 hides thinking content by default. Add a new

`thinking_display` provider option for Anthropic models accepting

`summarized`, `display`, or `omitted` to control whether thinking

blocks are returned in responses. It applies to both adaptive/

effort-based thinking budgets and explicit token budgets, on both

the standard and Beta Anthropic APIs.

Example:

    models:

      opus-4-7:

        provider: anthropic

        model: claude-opus-4-7

        thinking_budget: adaptive

        provider_opts:

          thinking_display: summarized

While here, consolidate all thinking-related helpers into a dedicated

`pkg/model/provider/anthropic/thinking.go` file and extract two helpers

(`applyThinkingConfig` / `applyBetaThinkingConfig`) so the inline

thinking setup in the two streaming entry points collapses to a single

method call each. `client.go` shrinks from 892 to 723 lines with no

behaviour change.

Updates the JSON schema, provider docs, and the thinking_budget example,

and adds unit tests for the new helpers (apply*, adjustMaxTokensForThinking,

interleavedThinkingEnabled, anthropicThinkingDisplay).

Assisted-By: docker-agent

fix: repair malformed JSON in edit_file tool call arguments

…ig-to-opus-4-7-657e4d3f

Add thinking_display provider_opt for Anthropic models

When a shell command backgrounds a grandchild (e.g. `docker run ... &`),
the grandchild inherits the stdout/stderr pipe fds created by exec.Cmd
for non-*os.File writers. cmd.Wait() then blocks until EOF on those
pipes, which never comes while the grandchild holds them open — making
the tool call hang until the configured timeout (observed wedging an
eval run for minutes after the agent ran a detached `docker run`).

Set cmd.WaitDelay to 500ms in runNativeCommand so that once the direct
shell child has exited, Go force-closes the pipe read ends and Wait()
returns promptly. The grandchild is left untouched — exactly what the
user's `&` intended.

Add two regression tests covering the plain backgrounded-child case
and the detached (setsid) variant where the process-group kill
fallback cannot rescue us.

Assisted-By: docker-agent

shell: fix hang when a tool command backgrounds a child process

Bumps the following direct dependencies:

- charm.land/bubbletea/v2 v2.0.5 -> v2.0.6
- github.com/anthropics/anthropic-sdk-go v1.36.0 -> v1.37.0
- github.com/aws/aws-sdk-go-v2 v1.41.5 -> v1.41.6
- github.com/aws/aws-sdk-go-v2/config v1.32.14 -> v1.32.16
- github.com/aws/aws-sdk-go-v2/credentials v1.19.14 -> v1.19.15
- github.com/aws/aws-sdk-go-v2/service/bedrockruntime v1.50.4 -> v1.50.5
- github.com/aws/aws-sdk-go-v2/service/sts v1.41.10 -> v1.42.0
- github.com/go-git/go-git/v5 v5.17.2 -> v5.18.0
- github.com/openai/openai-go/v3 v3.31.0 -> v3.32.0
- modernc.org/sqlite v1.48.2 -> v1.49.1

Assisted-By: docker-agent

…cies-f372e52e

bump direct Go dependencies