AGENT SKILL

bump-version

Collect current git changes, infer the next version number, update HARNESS_VERSION in src/harness.py, and prepend a changelog entry to CHANGELOG.md. Triggers on: bump version, version bump, collect changes and version, release version, new version, up version.

License 许可
MIT
最近更新
2026-06-12
Stars
★ 4
Forks
0

01 前言 Front Matter

name
bump-version
description
Collect current git changes, infer the next version number, update HARNESS_VERSION in src/harness.py, and prepend a changelog entry to CHANGELOG.md. Triggers on: bump version, version bump, collect changes and version, release version, new version, up version.

03 SKILL.md 正文

Bump the hyp-harness version based on the current uncommitted changes.

Steps

0. Clean up junk files

Before anything else, remove known artifacts and stray files:

# Build/test artifacts — always safe to delete
rm -f .coverage
rm -rf dashboard/test-results/
rm -rf __pycache__ src/__pycache__ tests/__pycache__

# Scratch files
find . -maxdepth 3 -name 'working*.md' -not -path './runs/*' -delete

# Root-level stray test files (belong in tests/ — if outside, they're scratch)
find . -maxdepth 1 -name 'test_*.py' -delete

Then check for files that need a judgment call — list them and ask the user before deleting:

git status --short | grep '^??'

Known-good untracked files — include in commit without asking:

  • pyproject.toml — pytest config ([tool.pytest.ini_options]), always keep
  • tests/test_*.py, tests/fixtures/ — test suite, always keep
  • dashboard/e2e/, dashboard/playwright.config.js — e2e harness, always keep
  • tools/*.py — harness tooling, always keep
  • .claude/skills/*/SKILL.md — canonical local skills, always keep
  • .agents/skills/*/SKILL.md — symlinks to canonical local skills, always keep

Flag only truly ambiguous patterns (ask user before committing):

  • Any *.md not in runs/, docs/, tests/, .claude/skills/, or .agents/skills/ — might be scratch
  • Any root-level .py outside src/, tests/, tools/ — likely stray

Also check staged deletions — include them in the commit:

git status --short | grep '^D '

These are already-deleted files (e.g. removed quality module files) that need git add -u to stage properly.

1. Collect changes

git diff HEAD --stat
git diff HEAD

Read the diff carefully. Note every file changed and what was added/removed.

2. Determine current version

grep 'HARNESS_VERSION' src/harness.py

The value is the current version (e.g. "v4.3"). The next version is:

  • minor bump (X.Y → X.Y+1): default for any new features, fixes, or refactors
  • major bump (X.Y → X+1.0): only when the architecture changes significantly (new worker backend, new pipeline stage, breaking config format change)

3. Summarize changes into a changelog entry

Write a concise changelog block following the existing style in CHANGELOG.md:

### v<NEXT> — <TODAY_DATE>
- **<Feature/Fix name>** (`<file.py>`): <one sentence description>
  - Sub-bullet for each distinct change in that file if needed
- **<Next feature>** ...

Rules:

  • Group changes by logical feature, not by file
  • Bold the feature name, then file in backticks
  • Keep bullet text factual and specific — what changed and why (1-2 sentences max)
  • Deleted files: note what was removed and why
  • Test files: mention only if they cover a new public contract
  • Use today's date in YYYY-MM-DD format

4. Update files

src/harness.py — update HARNESS_VERSION:

HARNESS_VERSION = "v<NEXT>"

CHANGELOG.md — prepend the new entry immediately after the # Changelog heading (before the existing top entry). Do not modify any existing entries. (CLAUDE.md's ## Changelog section just links here.)

4b. Sync descriptive sections in CLAUDE.md

CLAUDE.md has two kinds of content: append-only changelog, and current-state descriptions that must mirror the actual code. After writing the changelog entry, check each section below whose corresponding source files appear in the diff. For each affected section: read the actual current source, then rewrite only the stale parts.

Changed file(s) CLAUDE.md section to check
src/harness.py, src/postproc.py ### Post-processor pipeline — fixer count + numbered list must match PostProcessor class exactly
src/harness.py ## Architecture (pipeline flow), ### Evaluator-optimizer
src/evaluator.py ### Evaluator-optimizer — rubric names, retry logic, READER_STYLES list
src/synthesis.py ## Architecture synthesis step, any synthesis retry description
src/harness.py env vars ## Env vars (harness tuning) table — add/remove/change defaults
dashboard/src/ ### Dashboard (\dashboard/`)` — tabs, routes, features
src/render*.py ## Architecture render step description
src/harness.py DEFAULT_WRAPPER / RETRY_WRAPPER ## Kimi tool environment notes on FetchURL, steps
src/harness.py progress.log events ## Architecture progress.log API line

How to update each section:

  1. Read the relevant source file(s) to get the ground truth
  2. Read the current CLAUDE.md section
  3. Edit only the lines that are now wrong — preserve surrounding prose style
  4. Do not expand sections beyond what the diff touched — if a section is correct, leave it

Critical: ### Post-processor pipeline fixer list This is the most commonly stale section. The numbered list must exactly match the fixers in PostProcessor (in src/postproc.py or harness.py). Steps:

grep -n '_fix_\|_inject_\|_autonumber_\|_wrap_\|_stream_\|_fence_' src/postproc.py | grep 'def '

Count the fixers, update the header count ("21 safety-net fixers"), and add/remove numbered entries to match.

5. Sync check — find stale version refs

After updating the two canonical files, scan for any other hardcoded version strings that are older than <NEXT>:

grep -rn 'v[0-9]\+\.[0-9]\+' src/ --include='*.py' | grep -v 'HARNESS_VERSION\|\.pyc'

For each hit, classify:

Pattern Action
HARNESS_VERSION = "vX.Y" in any run copy (runs/poc-*/harness.py) Skip — run copies are frozen snapshots, never update
# mirrors harness.py vX.Y or # Prompt wrappers (mirrors …) in src/ Update the comment to v<NEXT> if the surrounding code was changed in this diff; leave it if untouched
Extracted from harness.py vX.Y in module docstring Leave as-is — records extraction point, not current version
Any other hardcoded version in src/*.py, dashboard/, CLAUDE.md intro Update if it's meant to track current version

Dashboard files (dashboard/src/) read HARNESS_VERSION dynamically — no static sync needed.

5b. Self-update this skill

Read .claude/skills/bump-version/SKILL.md (this file) and check if the diff introduces anything that makes its instructions stale. Update the skill if any of these changed:

What changed in diff What to update in SKILL.md
New junk file pattern (e.g. new artifact dir, new scratch file type) Add rm / find -delete line to step 0
New "review before committing" file pattern Add to the Flag list in step 0
New file added to src/ that carries version refs Add row to sync table in step 5
New CLAUDE.md section added Add row to mapping table in step 4b
New env var added to harness.py No change needed — step 4b already covers env vars table
HARNESS_VERSION moved to a different file Update step 2 grep and step 4 edit instruction
Commit message convention changed Update step 8 format string

Edit only the lines that are now wrong. Do not rewrite the whole skill.

grep 'HARNESS_VERSION' src/harness.py
grep -B 1 -A 3 '^### v' CHANGELOG.md | head -8

Confirm the version matches and the new entry is at the top of the changelog.

7. Report

Print a summary:

Bumped v<PREV> → v<NEXT>: <N> files changed — <headline of what changed>
Synced: <list of files updated beyond harness.py + CLAUDE.md, or "none">
Skipped (historical): <list of stale refs intentionally left, or "none">

8. Commit and push

Stage all changes — modified tracked files plus any deletions from step 0:

git add -u                                       # stages all modifications + deletions to tracked files
git add src/ dashboard/ CLAUDE.md CHANGELOG.md   # catches any untracked files in these dirs that should be included
# Do NOT use git add . — avoids pulling in confirmed-junk untracked files

Commit with message format release: vX.Y — <headline>:

git commit -m "release: v<NEXT> — <one-line headline of main change>"

Then push:

git push

Report the commit hash and confirm push succeeded.

Version bump heuristics

Change type Bump
New evaluator/rubric, new retry loop, new pipeline stage minor
Bug fix, fixer added, verify check added, dashboard UI minor
Refactor / cleanup / test only minor
New worker backend, new CLI, breaking problem.yaml format major
Docs / CLAUDE.md only skip (no bump needed)

File locations

File What to update
src/harness.py HARNESS_VERSION = "vX.Y" constant (canonical)
CHANGELOG.md prepend new entry at top
src/*.py inline comments Only if code around them changed in this diff
runs/poc-*/harness.py Never — frozen run copies
dashboard/src/ Never — reads version dynamically at runtime

Skill file locations

Canonical skill: .claude/skills/bump-version/SKILL.md — edit this file. Symlink: .agents/skills/bump-version/SKILL.md → canonical.

Khi step 5b detect skill cần update: chỉ edit .claude/skills/bump-version/SKILL.md.

04 同仓库更多 Skills