By allowing fetch to send results directly into jq we can prevent a lot
of tokens from entering the context, and compress several tool calls.

Descriptions should describe when a skill applies, not how it does its
job. Leaving language like "using go doc", "the jq command-line
processor", etc. in the trigger encourages the model to guess its way
through without reading SKILL.md.

Models frequently skipped viewing a skill's SKILL.md and inferred its
behavior from the description alone; making the load mandatory in
<critical_rules> prevents LLMs from taking that shortcut.

For example: Apple Terminal.

Do not run these two new tests in parallel.

💘 Generated with Crush

Assisted-by: Claude Opus 4.7 via Crush <crush@charm.land>

Also, two local skills now exist for Crush development:

* builtin-skills, for adding new builtin skills
* shell-builtins, for adding new shell builtins

- Display loaded/errored skills in sidebar and landing page
- Use pubsub event to carry skill states directly to UI, avoiding
  global mutable state
- Subscribe to skills discovery events via broker, consistent with
  MCP/LSP patterns
- Rename State to DiscoveryState for clarity

It takes a lot of time to re-record these VCRs for so many models.
Sometimes they fail intermittently, and you need to manually re-record
again and again until it works.

Let's keep a single provider and models which should work well enough for
us.

When a session is interrupted mid-tool-call (Ctrl-C, network timeout),
the assistant message contains a tool_use block but no corresponding
tool_result is ever recorded. On session resume, the Anthropic API
rejects the conversation with invalid_request_error, permanently
locking the session.

The existing filterOrphanedToolResults handles the inverse case
(tool_result without a matching tool_use). This commit adds the
symmetric fix: during preparePrompt, detect tool_use IDs that have
no matching tool_result and inject a synthetic error tool_result so
the API accepts the history and the session can continue.

Co-authored-by: Christian Rocha <christian@rocha.is>
Co-authored-by: Andrey Nering <andreynering@users.noreply.github.com>

See: https://github.com/charmbracelet/ultraviolet/pull/109

This confuses dumber models, that might respond in Portuguese even when
the user prompt was in English.

This is more common for small models. We already made this fix on the
title generation prompt, we're now just porting to the coder prompt.

* Closes #2627
* Follow-up #2497

fix(agent): validate tool calls + strip tags from titles

This fixes a case where tool results could not have a matching tool
call, resulting in corrupted sessions.

There's also an interleaved fix for trimming rogue "</think>" tags from
session titles while we're at it.