fix: classify runtime checkpoint rounds

[jolestar]

Summary

• classify provider assistant rounds as agent_response or runtime_checkpoint and persist checkpoint request metadata in transcript/audit evidence
• prevent runtime checkpoint text from becoming terminal result text, brief source text, max-output recovery history, or curated info/verbose assistant progress
• retain checkpoint evidence on debug/raw surfaces, preserve multi-brief and WorkItem completion behavior, and document the semantic boundary
• reinforce the best-effort response-language instruction for all operator-visible assistant prose without adding language detection, retries, or extra model calls

Verification

• cargo fmt --all -- --check
• RUSTFLAGS="-D warnings" cargo check --all-targets
• cargo test
• npm test (web-gui/app)
• npm run typecheck (web-gui/app)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
holon	Ready	Preview, Comment	Jul 10, 2026 3:34pm

[jolestar]

Review: LGTM ✅

Well-structured and thorough change that correctly classifies runtime checkpoint rounds as private continuity state.

Summary of review:

1. Semantic classification (execution.rs, types.rs): New AssistantRoundPurpose enum cleanly distinguishes agent_response vs runtime_checkpoint rounds. The classification correctly gates last_assistant_round_id, last_assistant_message, and max-output recovery history — preventing checkpoint text from leaking into terminal results or briefs.

2. Operator event visibility (operator_event.rs): Checkpoint rounds are properly demoted to Trace visibility with explicit labeling.

3. Presentation filtering (presentation.rs, session-reducer-core.ts): Both Rust and web-GUI reducers correctly exclude checkpoint rounds from curated timelines while keeping them on debug/raw surfaces.

4. Tests: Comprehensive coverage across unit, integration, and web-GUI tests. Test budget calculations improved from hardcoded values to dynamic computation.

5. RFC documentation: All three relevant RFCs updated.

No blocking issues found. Waiting for CI to pass.

[holonbot]

Review: Approve ✅

Well-structured change that correctly introduces a semantic distinction between ordinary agent response rounds and runtime-private checkpoint rounds, then enforces that distinction consistently across every operator-visible projection layer.

Correctness: Round purpose determination is correct — snapshotting pending_checkpoint_metadata before response processing ensures classification reflects runtime intent. The last_assistant_round_id guard correctly prevents checkpoint rounds from being recorded as terminal text sources. Max-output recovery correctly excludes checkpoint text from truncated history. The = None initialization for last_assistant_round_id is a latent bug fix.

Test coverage: Comprehensive — unit tests for operator events, presentation reducer, web-GUI reducer; integration test for full turn execution with checkpoint metadata verification. Test budgets refactored from hardcoded to dynamically computed.

Documentation: Three RFCs updated with clear semantic boundary docs.

Nit (non-blocking): The same five checkpoint metadata fields (round_purpose, visibility, checkpoint_request_id, checkpoint_mode, checkpoint_requested_at_round) are repeated across four serde_json::json! blocks in execution.rs. A small helper could reduce this repetition surface.

No correctness, security, or safety issues found.

[holonbot]

Holon Run Report

• Result: completed
• Target: fix: classify runtime checkpoint rounds #2152
• Workflow run: https://github.com/holon-run/holon/actions/runs/29104166137
• Model: dashscope-token-plan/qwen3.7-max
• Requested model: dashscope-token-plan/qwen3.7-max
• Fallback active: false
• Token usage: input 13603, output 5860, total 19463
• Provider cache: read 208800, created 34574, cacheable 243374, hit rate 85.79%
• Model rounds: 9
• Tool calls: 15