Phase 2: Procedure replay engine -- skip reasoning for proven workflows

[AlexMikhalev]

Parent Epic

Part of #692 (Operational Skill Store)

Summary

When learn query finds a matching CapturedProcedure with confidence > 0.8, offer to replay it instead of re-planning from scratch. This is the core token-saving mechanism -- proven workflows skip LLM reasoning entirely.

What Changes

New replay subcommand in terraphim_agent

terraphim-agent learn replay "deploy terraphim-llm-proxy config"

Behaviour:

1. Match query against stored procedures (Aho-Corasick + tag matching)
2. Display matched procedure with steps, confidence, and replay count
3. Dry-run by default -- show steps, require confirmation
4. Execute step-by-step with validation:
   • Check exit code against expected_exit_code
   • Optionally match output against expected_output_pattern
   • Respect per-step timeout_secs
5. On divergence (unexpected exit code or output), halt and report
6. Update procedure's success_count or failure_count after completion
7. Recalculate confidence score

Integration point for ADF agents

Expose replay as a library function so terraphim_orchestrator agents can call it programmatically:

pub async fn replay_procedure(
    procedure_id: &str,
    dry_run: bool,
    on_step: impl Fn(&ProcedureStep, StepResult) -> ReplayDecision,
) -> Result<ReplayOutcome, ReplayError>;

Affected Crates

• terraphim_agent (new replay subcommand + library API)

Dependencies

• Phase 1 (Phase 1: Success capture -- extend CapturedLearning with procedural memory #693) -- procedures must exist to replay

Acceptance Criteria

• learn replay matches and displays procedures
• Dry-run mode shows steps without executing
• Step-by-step execution with exit code validation
• Halts on divergence with clear error message
• Updates confidence score after each replay
• Library API available for programmatic replay
• Integration tests with mock commands