fix(core): cap done repair loop

[Sun-sunshine06]

Summary

• enforce a three-error-round budget around the agentic done() repair loop
• terminate the pi-agent tool loop when the budget is exhausted while preserving the latest artifact with warnings
• align done/tool prompt copy and add regression coverage for the repair-limit path

Fixes #251

Testing

• pnpm exec biome check packages/core/src/agent.ts packages/core/src/tools/done.ts packages/core/src/agent.test.ts .changeset/done-repair-limit.md
• pnpm --dir packages/core exec vitest run src/agent.test.ts
• pnpm --dir packages/core exec tsc --noEmit
• pnpm typecheck
• pnpm test

[github-actions]

Review mode: initial

Findings

• [Minor] The errorRounds counter resets to 0 on a successful done() call (status "ok"). This means after a successful verification, the agent gets a fresh three-error budget for subsequent done() calls. While not a bug, this may allow the agent to cycle through error rounds and fixes indefinitely if it introduces errors after each fix. Consider whether a per-session (non-resetting) budget is desired, or at least document the design choice in the constant comment. (packages/core/src/agent.ts:389)
• [Nit] formatDoneRepairLimitText includes 'has_errors' as the first line of the text content returned to the model. The previous done() result's details already indicate status: 'has_errors', so repeating it in the text is redundant. Consider removing that line. (packages/core/src/agent.ts:407)

Summary

This PR caps the done() repair loop at three error rounds, preserving the latest artifact and surfacing warnings when the limit is reached. The changes are clean, well-tested, and align the tool prompt copy with the new limit. The two new tests cover both the direct tool behavior and the full agent termination path. Overall, the implementation direction is sound.

Testing

Good coverage: unit tests for the wrap logic and integration-style test using the executeTool mock. Consider adding a test that verifies errorRounds resets after a successful done() call, to document the expected reset behavior.

Open-CoDesign Bot