Migrate fast/ module (fast_plan_tasks) to AgentField native .harness()

[santoshkumarradha]

Part of epic #21

Summary

Migrate fast_plan_tasks in swe_af/fast/planner.py from custom AgentAI(...).run() to AgentField's native router.harness() (or potentially .ai() since it uses only 3 turns and no explicit tools).

AgentField now natively supports .harness() and .ai() as first-class primitives — this replaces the custom abstraction.

No dependencies — can be worked independently. Good first issue (single agent, single file).

---

Agent to migrate

Agent	File	Current tools	Role
fast_plan_tasks	swe_af/fast/planner.py	(none specified)	Single-pass task decomposition for FastBuild mode

Note: This agent uses AgentAIConfig with max_turns=3 and no explicit allowed_tools, so it gets defaults. Check whether it actually needs tool access — if not, .ai() may be more appropriate.

Current pattern

from swe_af.agent_ai import AgentAI, AgentAIConfig

ai = AgentAI(
    AgentAIConfig(
        provider=ai_provider,
        model=pm_model,
        cwd=repo_path,
        max_turns=3,
        permission_mode=permission_mode or None,
    )
)
response = await ai.run(
    task_prompt,
    system_prompt=FAST_PLANNER_SYSTEM_PROMPT,
    output_schema=FastPlanResult,
)

Target pattern (option A — .harness())

result = await fast_router.harness(
    prompt=task_prompt,
    schema=FastPlanResult,
    provider=provider,
    model=pm_model,
    max_turns=3,
    permission_mode=permission_mode or None,
    system_prompt=FAST_PLANNER_SYSTEM_PROMPT,
    cwd=repo_path,
)

Target pattern (option B — .ai() if no tools needed)

result = await fast_router.ai(
    task_prompt,
    system=FAST_PLANNER_SYSTEM_PROMPT,
    schema=FastPlanResult,
    model=pm_model,
)

Decision: .harness() vs .ai()

• If the planner needs to read repository files to plan → .harness()
• If it only uses the goal text and context passed in the prompt → .ai()
• Check the prompt template in swe_af/fast/prompts.py to determine

Files to modify

• swe_af/fast/planner.py — fast_plan_tasks
• Remove from swe_af.agent_ai import AgentAI, AgentAIConfig

Acceptance criteria

• fast_plan_tasks uses fast_router.harness() or fast_router.ai() instead of AgentAI.run()
• No imports from swe_af.agent_ai remain in swe_af/fast/planner.py
• Output schema FastPlanResult preserved exactly
• Fallback plan (_fallback_plan) preserved on failure
• max_tasks truncation logic preserved
• Provider mapping: "claude" → "claude-code" (if using .harness())

Tests needed

• Unit test: returns FastPlanResult dict on success (mock router method)
• Unit test: returns fallback plan with fallback_used=True on LLM failure
• Unit test: returns fallback plan when parsed response is None
• Unit test: truncates to max_tasks when LLM returns too many tasks
• Verify fast/executor.py and fast/verifier.py are unaffected (they use app.call(), not AgentAI directly)

Notes

• executor.py and verifier.py in the fast/ module use app.call() to dispatch reasoners — they don't use AgentAI directly and are NOT affected by this migration
• Only planner.py has the direct AgentAI import
• Default model is haiku — cheap, appropriate for fast planning

[Parth-Vasave]

Hey would like to work on this if its available, can you assign me?

[santoshkumarradha]

Already migrated on main branch. swe_af/fast/planner.py already uses fast_router.harness() (lines 99-111). Tests in tests/fast/test_planner.py and tests/test_malformed_responses.py already mock router.harness. No code changes needed — closing as complete.