Implement CompletionCheck trait and standard implementations

[sbeardsley]

Context

CompletionCheck is referenced in two loop strategies in the Harness (#3):

• LoopStrategy::Ralph { completion_check: Box<dyn CompletionCheck> } — the external check that determines when a Ralph continuation loop is actually done
• LoopStrategy::SelfVerifying { verifier, evaluator_harness } — uses the termination policy's CompletionCheck to determine when the build phase should stop and hand off to the evaluator

Currently a stub in the harness module tagged // SPEC: full trait lives in this issue. The Ralph and SelfVerifying strategies both return HaltReason::StrategyNotYetImplemented until this trait is implemented.

Trait Definition

// Returns None if the task is complete, Some(reason) if not done yet.
// reason is injected into the next turn's context — tells the agent
// what it still needs to do. This is what prevents premature victory.
trait CompletionCheck {
  async fn check(state: &SessionStateSnapshot) -> Option<String>

  // Human-readable description of what this check evaluates.
  // Injected into agent context at session start so it understands
  // what "done" means for this task.
  fn description() -> String
}

Standard Implementations

FeatureListCheck

Reads feature_list.json from the workspace. Returns Some with the list of incomplete features if any have passes: false. Returns None when all features pass.

FeatureListCheck {
  path: PathBuf,   // default: "feature_list.json"
}

TestSuiteCheck

Runs the test suite and returns Some(failure_summary) if any tests fail. Returns None when the full suite passes.

TestSuiteCheck {
  command: String,          // e.g. "npm test", "cargo test", "pytest"
  working_dir: PathBuf,
  timeout: Duration,
  sandbox: Arc<dyn SandboxProvider>,
}

QuestionAnsweredCheck

LLM-as-judge: evaluates whether the agent's final response actually answered the user's question. Used for RAG and conversational agents.

QuestionAnsweredCheck {
  judge_model: ModelConfig,
  original_question: String,
  rubric: Option<String>,    // custom evaluation criteria
}

SqlResultCheck

Validates that the SQL result set is non-empty and structurally correct (column names match expectation). Used for NL-to-SQL agents.

SqlResultCheck {
  expected_columns: Option<Vec<String>>,
  min_rows: Option<usize>,
}

AlwaysComplete

Returns None immediately — task is always considered done when the model claims it is. Used for simple single-turn tasks where the model's self-assessment is sufficient.

AlwaysComplete

Relationship to TerminationPolicy

CompletionCheck is injected into TerminationPolicy and called only when agent_claims_done: true. The TerminationPolicy evaluates budget limits first (unconditionally), then sensor results, then calls CompletionCheck. This is the mechanism that prevents premature victory — the agent claims done, the check says "not yet, here's what's missing", and the harness injects that reason into the next turn.

Checklist

• Rust: CompletionCheck trait + all standard implementations
• TypeScript: CompletionCheck trait + all standard implementations
• Python: CompletionCheck trait + all standard implementations
• Go: CompletionCheck trait + all standard implementations
• Unit tests: each implementation returns correct None / Some(reason) for its domain
• Harness (Implement Harness runtime loop #3) stubs replaced with real CompletionCheck in Ralph and TerminationPolicy
• TerminationPolicy (Implement TerminationPolicy #13) updated to call CompletionCheck correctly
• Fixture: fixtures/completion_checks/feature_list_complete.jsonl

Related Issues

• Implement Harness runtime loop #3 Harness runtime loop (Ralph and SelfVerifying strategies depend on this)
• Implement TerminationPolicy #13 TerminationPolicy (calls CompletionCheck when agent claims done)
• Implement MetricEvaluator trait and HillClimbing loop strategy #23 MetricEvaluator (sibling — HillClimbing's equivalent of CompletionCheck)