Skip to content

feat(v0.2.7): harness v3 + tool professionalization + CLI experience#54

Open
Linux2010 wants to merge 1 commit into
mainfrom
v0.2.7
Open

feat(v0.2.7): harness v3 + tool professionalization + CLI experience#54
Linux2010 wants to merge 1 commit into
mainfrom
v0.2.7

Conversation

@Linux2010

Copy link
Copy Markdown
Owner

openhorse v0.2.7

Based on docs/targets/first-class-coding-agent-vision.md and docs/targets/ui-runtime-boundary.md.

Changes

Round Feature Tests
1 LLM-Driven Compact Summary (already existed)
2 Evidence Ranking with includedCount learning signal 7 new
3 Intent clarification type 16 new
4 Checkpoint & undo for file edits 11 new
5 Tool result structured serialization (v1 envelope) 7 new
6 Enhanced Doctor (artifacts, cache, warn-dedup) existing

Key additions

  • Evidence learning: Frequently-included evidence gets boosted in ranking
  • Clarification intent: "什么意思", "Can you explain" → does NOT replace contract
  • File checkpoints: Auto-snapshot before edit_file/write_file, 7-day TTL, /checkpoint restore <turnId>
  • Tool envelope: ToolResult.schemaVersion + serializeToolResult/parseToolResultEnvelope
  • Doctor: 3 new checks (artifact size, prompt cache status, suppressed warnings)

Stats

  • 66 suites, 740 tests (1 skipped)
  • 14 files changed, 884 insertions

🤖 Generated with Claude Code

Round 2: Evidence Ranking with learning signal
- EvidenceRecord.includedCount field tracks prompt inclusion frequency
- rankEvidence boosts frequently-included evidence (capped at 10x)
- bumpIncludedEvidence called after each prompt assembly
- /harness explain shows included counts

Round 3: Intent classification enhancement
- New 'clarification' intent kind for explanation questions
- 6 clarification patterns (CN/EN)
- clarification does NOT replace active instruction

Round 4: Checkpoint & undo
- src/core/checkpoint.ts: file snapshots before edits
- createCheckpoint/restoreCheckpoint/listCheckpoints/cleanupCheckpoints
- 7-day TTL auto-cleanup, stored in ~/.openhorse/.../_checkpoints/

Round 5: Tool result structured serialization
- ToolResult.schemaVersion field (v1)
- serializeToolResult/parseToolResultEnvelope helpers
- Backward compatible: legacy JSON and non-JSON auto-handled

Round 6: Enhanced Doctor diagnostics
- Artifact directory size and file count check
- Prompt cache configuration status check
- Warning dedup statistics (suppressed duplicates)

Tests: evidence-ranking (7), intent-classification (16), checkpoint (11),
tool-serializer (7) — 41 new tests total

Co-Authored-By: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant