zsp/everything-claude-code
1
0
Fork 0
mirror of https://github.com/affaan-m/everything-claude-code.git synced 2026-03-30 13:43:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(harness-audit): make scoring deterministic with scripted rubric (#524)

Browse Source
This commit is contained in:
Chandan Semwal
2026-03-17 02:04:30 +05:30
committed by GitHub
parent ebd8c8c6fa
commit 0ff1b594d0
5 changed files with 660 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
commands/harness-audit.md
View File

@@ -1,6 +1,6 @@
# Harness Audit Command
Audit the current repository's agent harness setup and return a prioritized scorecard.
Run a deterministic repository harness audit and return a prioritized scorecard.
## Usage
@@ -9,9 +9,19 @@ Audit the current repository's agent harness setup and return a prioritized scor
- `scope` (optional): `repo` (default), `hooks`, `skills`, `commands`, `agents`
- `--format`: output style (`text` default, `json` for automation)
## What to Evaluate
## Deterministic Engine
Score each category from `0` to `10`:
Always run:
```bash
node scripts/harness-audit.js <scope> --format <text|json>
```
This script is the source of truth for scoring and checks. Do not invent additional dimensions or ad-hoc points.
Rubric version: `2026-03-16`.
The script computes 7 fixed categories (`0-10` normalized each):
1. Tool Coverage
2. Context Efficiency
@@ -21,34 +31,37 @@ Score each category from `0` to `10`:
6. Security Guardrails
7. Cost Efficiency
Scores are derived from explicit file/rule checks and are reproducible for the same commit.
## Output Contract
Return:
1. `overall_score` out of 70
1. `overall_score` out of `max_score` (70 for `repo`; smaller for scoped audits)
2. Category scores and concrete findings
3. Top 3 actions with exact file paths
4. Suggested ECC skills to apply next
3. Failed checks with exact file paths
4. Top 3 actions from the deterministic output (`top_actions`)
5. Suggested ECC skills to apply next
## Checklist
- Inspect `hooks/hooks.json`, `scripts/hooks/`, and hook tests.
- Inspect `skills/`, command coverage, and agent coverage.
- Verify cross-harness parity for `.cursor/`, `.opencode/`, `.codex/`.
- Flag broken or stale references.
- Use script output directly; do not rescore manually.
- If `--format json` is requested, return the script JSON unchanged.
- If text is requested, summarize failing checks and top actions.
- Include exact file paths from `checks[]` and `top_actions[]`.
## Example Result
```text
Harness Audit (repo): 52/70
- Quality Gates: 9/10
- Eval Coverage: 6/10
- Cost Efficiency: 4/10
Harness Audit (repo): 66/70
- Tool Coverage: 10/10 (10/10 pts)
- Context Efficiency: 9/10 (9/10 pts)
- Quality Gates: 10/10 (10/10 pts)
Top 3 Actions:
1) Add cost tracking hook in scripts/hooks/cost-tracker.js
2) Add pass@k docs and templates in skills/eval-harness/SKILL.md
3) Add command parity for /harness-audit in .opencode/commands/
1) [Security Guardrails] Add prompt/tool preflight security guards in hooks/hooks.json. (hooks/hooks.json)
2) [Tool Coverage] Sync commands/harness-audit.md and .opencode/commands/harness-audit.md. (.opencode/commands/harness-audit.md)
3) [Eval Coverage] Increase automated test coverage across scripts/hooks/lib. (tests/)
```
## Arguments