[SDK] HarnessRunner with retry & schema orchestration for .harness()

[santoshkumarradha]

Overview

Implement the HarnessRunner that orchestrates provider execution with retry logic, schema output handling (4-layer recovery), temp file lifecycle, cost tracking, and JSONL event logging.

Branch: feat/harness-v2
Design doc: docs/design/harness-v2-design.md (Sections 3.2, 5.4, 9)

Scope

Python (sdk/python/agentfield/harness/)

• _runner.py — HarnessRunner class:
  • async run(prompt, schema, options) → orchestrates full execution
  • Config resolution: merge HarnessConfig defaults + per-call overrides
  • Provider dispatch via factory
  • Retry with exponential backoff + jitter for transient errors
  • Schema orchestration: inject prompt suffix → execute → 4-layer recovery
  • Layer 1: Read output file → parse → validate
  • Layer 2: Cosmetic repair → re-validate
  • Layer 3: Follow-up prompt in same session (cap 2 attempts)
  • Layer 4: Full retry (limit 1)
  • Temp file cleanup (always, even on failure)
  • Cost/metrics tracking
• Tests: retry behavior, schema orchestration layers, config resolution, error categories, transient error detection

TypeScript (sdk/typescript/src/harness/)

• runner.ts — HarnessRunner class (mirrors Python)
• Tests: mirror Python coverage

Key Design Decisions

• Transient errors: rate limits, timeouts, 5xx, connection errors → retry
• Non-transient errors: auth, billing, invalid request → fail immediately
• Schema follow-up uses same session (cheap, agent has context)
• No separate schema budget — uses overall max_budget_usd
• Always clean up temp files, even on error paths

Acceptance Criteria

• Retry correctly backs off with jitter
• All 4 schema recovery layers work in sequence
• Follow-up prompt capped at 2 attempts
• Full retry capped at 1 attempt
• Transient vs non-transient error classification correct
• Temp files always cleaned up
• All tests pass

Dependencies

• Depends on: [SDK] Core types & provider interface for .harness() #199 (types), [SDK] Schema handling with universal file-write strategy for .harness() #200 (schema)
• Blocks: All provider issues, agent wiring

[github-actions]

Thanks for picking this up, @santoshkumarradha! 🎉

Here's what to expect:

• We'll check in after 7 days if we haven't heard an update
• After 14 days, we'll send a second reminder
• After 21 days without activity, we'll unassign to let others contribute

Tips for success:

• Drop a comment when you start working on it
• Share any blockers - we're here to help!
• Link your PR with Fixes #201 in the description

Happy coding! 🚀