[pournasserian]

Closes #2554

What

The pre-flight slice of F-CHK (#2420), CHK-04: an agent runs the site's active checks on a draft without storing it or opening a change set. It comes before the change set run (the epic's fifth slice), which waits on a decision about who the check engine acts as. The runner here is the one that run will use.

• Core (FluentCMS.Core.Checks):
  • ICheckRunner runs every active check whose scope applies to a piece of content, at its active version and in name order. Each check sees the fields its scope names, and calls are charged to the caller. CheckSelection.Deterministic leaves out LLM-judged checks, and the run names the ones it left out (CheckRun.Skipped).
  • CheckResult holds the check, version, type, kind, severity, description, status, findings and model call.
  • CheckStatus is pass, warn (a failing check of warn severity), fail or needs review. CheckResult.Overall is fail if any check fails, then needs review, then warn, then pass.
• Tool check_draft (FluentCMS.Core.Builtin.Tools.CheckTools):
  • Arguments: content_type, fields, optional sources, checks (all or deterministic), and id for a draft of an update.
  • It returns status, results (each with its findings and model call), skipped, and fieldErrors, the errors create_entry would give the draft.
  • With id, a check that asks about changes, such as sources, sees only the fields whose value differs from the entry's latest draft.
  • A draft with field errors fails. An unknown content type or entry is refused, with a hint for the call that finds one.
  • It stores nothing and opens no set.
• MCP-01 coverage: "Run the checks on a draft before proposing it". CONTRIBUTING.md: the runner and the tool.

Why

CHK-04 asks for a pre-flight tool, so an agent fixes what the checks find before it proposes the change, instead of after a set fails. S2 is that loop: three findings, three fixes, a pass.

Choices worth a look:

• Before the change set run. The run on every set (CHK-03) must move a set from checks running to checks failed or awaiting review. AgentPermissions refuses those moves for agents, and a job (Add the job queue with leases, retries and progress #2536) runs as whoever enqueued it, the submitting agent. Initiators today are only users and agents, so the run needs a decision on whom it acts as in the audit log. I have asked; this slice does not depend on the answer.
• Not read-only. The tool stores nothing, but an LLM-judged check spends the agent's model budget, so it does not carry the read-only annotation that clients may approve without asking. checks: "deterministic" makes a free, instant call.
• Field errors fail the draft. A draft the content type refuses cannot be proposed, so it never reads as a pass, whatever the checks say.
• Updates. Without id, every field counts as changed, as for a new entry. With it, a headline edit on a priced page does not ask for a source for the unchanged price. The entry is read through IEntryDirectory, so the agent's scope applies.
• Scope. The content type is read through IContentTypeDirectory, so a type outside the agent's scope is refused, and the result shows no check's settings or examples.
• Not in this PR: the check run on every change set with results stored on the set, the tool that reads a set's results, and checks that read entities (F-ENT).

Tests

Unit:

• CheckToolsConformanceTests (ToolProviderConformanceTests).
• CheckRunner:
  • every active check that applies gives its status and evidence, in name order; a warn check's failure only warns
  • scope by content type and field, and an untyped draft
  • the deterministic selection asks no model and names what it left out
  • without a model the run needs review unless something fails
  • status mapping, and the overall status of none
• MCP-01 coverage row.

Integration over MCP, each on SQLite and PostgreSQL:

• S2: a draft with "best in class", an image without alt text and an unsourced price fails three checks with their findings. The fixed draft with a source passes, and nothing is stored.
• An active LLM-judged check needs review without a model, and the run does too. With deterministic it is listed under skipped and the run passes.
• A draft of an update is checked for what it changes: a headline edit passes and a price change without a source fails.
• A draft the content type refuses lists its field errors and fails.
• Unknown content types, entries, arguments and selections are refused with what to change, and a type outside the agent's scope is refused.
• The built-in tool list includes check_draft.

The full suite ran locally against a postgres:17 container as well as SQLite, and markdownlint passes. No new package and no migration.

Checklist

• Linked to an issue a maintainer has confirmed
• Targets v2
• Tests or docs updated where behaviour changed
• No tool-generated co-author trailers or footers