[Muizzkolapo]

Batch the 62 unreleased changelog entries into 3.0.0 and bump the version in pyproject.toml and agent_actions/__version__.py. Major, because the batch holds 9 Breaking Change entries.

• CHANGELOG.md and .changes/3.0.0.md hold every entry, as changie batch 3.0.0 and changie merge wrote them. One stale entry from the v2.0.0 release ("Release v2.0.0 — batch changie entries, bump version") is dropped.
• uv.lock takes the anthropic bound that pyproject.toml already declares (Update anthropic requirement from <2,>=0.30.0 to >=0.30.0,<3 #1599 updated only pyproject.toml).

Release notes (for the GitHub release after merge)

v3.0.0 has 9 breaking changes, 4 features and 39 bug fixes since v2.0.0. The full list is in CHANGELOG.md.

Breaking changes

• A config edit no longer re-runs work. Answers already stored stand, and each run names the edited action. agac run --rerun <action> answers an action and its readers again. A changed staged input re-answers only the records it changed.
• More refusals at load and preflight. A missing schema file; the dead keys constraints, enable_caching, max_execution_time and where_clause; skip_condition and skip_if (use a guard with on_false: skip).
• Schema dispatch. schema: dispatch_task('fn') declares schema_options:, and the function returns one of those names.
• FILE tools. A record the action cannot be handed fails the action, with every missing field named. A roll-up row carries only the upstream fields its inputs share.

New features

• endpoints: declares OpenAI-compatible providers; select one with model_vendor: <name>.
• request_timeout bounds each online LLM request; record_concurrency sends up to 32 records of an online action at once.
• A guard can compare two fields: extract.score < extract.limit.

Main fixes

• One run per workflow: a second agac run, agac retry, retry --filter, agac clean or reset_workflow.py --full exits 1 while a run is in progress.
• schema: dispatch_task works online and in batch, resolved once per record.
• Retry waits as long as Retry-After asks. Schema bounds reach the provider and the mock. Each repair attempt has its own prompt trace.
• The run output shows guard skips, failing verdicts and a final action that holds no record. Report tables print whole on a narrow terminal.
• The docs site builds again.