feat(ts): scenario revision flow — refine with feedback (AC-441)

[jayscambler]

Summary

Adds scenario revision flow to TypeScript — users can refine created scenarios with feedback instead of starting over from scratch.

Problem

Python has agent_task_revision.py for iterative spec refinement. TypeScript had nothing — once a scenario was created, the only option was delete and retry. Scenario creation is iterative; the first attempt rarely matches what the user wants.

What's new

ts/src/scenarios/scenario-revision.ts — two levels of revision:

1. Spec revision (reviseSpec)

Takes current spec + user feedback → LLM produces updated spec.

const result = await reviseSpec({
  currentSpec: { description: "Basic task", taskPrompt: "Do X" },
  feedback: "Add edge cases and make it harder",
  family: "agent_task",
  provider,
});
// result.revised has the updated spec
// result.changesApplied === true
// result.original preserved for diff/audit

• Works for all families (agent_task, simulation, workflow, etc.)
• Preserves original spec on LLM failure (never loses user's work)
• Includes family-specific context in the revision prompt

2. Output revision (reviseAgentTaskOutput)

Builds revision prompts from judge feedback for the ImprovementLoop.

• Ports Python's build_revision_prompt()
• Highlights weak dimensions below threshold
• Includes task-specific revision instructions

MCP integration

New revise_scenario tool — agents and external tools can trigger revision:

{ "currentSpec": {...}, "feedback": "Add error handling", "family": "simulation" }

Tests

10 new tests covering:

• Prompt building (spec context, family context, judge feedback, weak dimensions)
• Spec revision (successful update, LLM failure recovery, multi-family)
• Output revision (judge feedback, revision instructions, weak dim highlighting)
• Result shape validation

1300 total tests pass, lint clean.

Linear: AC-441

[linear]

AC-441 TS has no scenario revision flow — users must start over on every change