Skip to content

CHANGELOG.md

Latest commit

 

History

History
231 lines (174 loc) · 39.8 KB

CHANGELOG.md

File metadata and controls

231 lines (174 loc) · 39.8 KB
Raw

Changelog

All releases follow CalVer — YYYY.0M.MICRO. The marketplace and all plugins ship in lockstep: every release re-tags every plugin with the same version.

2026.07.2 — 2026-07-14

Maintenance release: retires the two multi-CLI wrapper plugins (code-review-external, premortem-multiple) that shelled out to external codex / opencode binaries. The single-perspective premortem plugin is unaffected and stays. This release also folds in django-run-site, added during the 2026.07.1 catalog bump that never shipped as a lockstep release.

Removed

  • code-review-external plugin — removed in full: the wrapper skill plus code-review-codex, code-review-opencode, code-review-claude, and the shared standard-review-prompt.md / target-detection.md / write-directive.md / tmux-runner.md files. For GitHub PR review use the official /code-review:code-review skill; for a local diff, /code-review.
  • premortem-multiple plugin — removed in full: the wrapper skill plus premortem-codex, premortem-opencode, premortem-claude, and the shared standard-premortem-prompt.md / write-directive.md / tmux-runner.md files.
  • Both entries dropped from marketplace.json; all references removed from README.md (plugin tables, the Decision-making and Code-review skill-graph diagrams, the /plugin install lines, the usage bullets, and the code-review-external — argument forms table).

Added

  • django-run-site plugin — reference skill for a local Django dev stack managed by run-site / django-dev-helpers: recognizing a run-site project, locating the .run-site-config sidecar and .dev_helpers_* dotfiles, reading web/Postgres/Redis/SQLite endpoints and ports, the autologin recipe, and the run-site CLI surface. (Introduced in the 2026.07.1 catalog bump; ships in a lockstep release here for the first time.)

Repository

  • Lockstep version bump: the marketplace and all 11 remaining plugins move to 2026.07.2.

2026.05.8 — 2026-05-14

New plugin release: long-file-splitter adds a repo-wide audit for source files that have grown past a configurable line threshold (default 600 LoC), with per-file split proposals that name concrete new modules, the symbols that move into each, and the strategy for preserving the public API. Proposal-first by design — the skill does not auto-refactor; an opt-in guided execution phase performs the moves one file at a time, one commit per logical extraction, with the test suite re-run between commits.

Added — long-file-splitter plugin

  • Phase 1 — Scan. Default extensions cover .py, .pyi, .js/.jsx/.mjs/.cjs, .ts/.tsx, .html/.htm, .vue, .svelte. Discovery uses rg --files + wc -l + awk against a long exclusion list (node_modules, .venv, venv, dist, build, .next, .tox, .pytest_cache, .mypy_cache, .ruff_cache, htmlcov, coverage, vendor, third_party, bower_components, *.min.*, *.bundle.js, lockfiles). Threshold is configurable per invocation.
  • Mega-constant detection (Python). An AST pass surfaces top-level assignments whose literal spans ≥100 lines — catching files that are under the LoC threshold but still problematic because one dict / list / multi-line str dominates the file (e.g., a 400-line module that's 70% one CHOICES dict). Each finding reports the symbol name, line range, span, and share of the file.
  • Generated-file detection. Reads the first ~30 lines of each candidate looking for Generated by Django, DO NOT EDIT, Generated by the protocol buffer compiler, autogenerated by, @generated, AUTOGENERATED. Marks matches SKIP — generated. Django migrations are a tagged subcase: technically generated, but the right move when too long is usually squashmigrations, not a manual split — flagged with that note rather than silently skipped.
  • Phase 2 — Classification into one of 6 categories. A grab-bag utility module (many small unrelated units), B god-class / god-component (one or a few large units), C mega-constant (one literal ≥40% of the file), D many-section HTML template, E mixed (two of the above at once), F SKIP — cohesive (long but justifiably so: tokenizer, state machine, snapshot fixture). The skip path is first-class — the skill is encouraged to refuse a split rather than fabricate one.
  • Phase 3 — Cohesion analysis. Per-category heuristics for deciding what groups with what: name-prefix clusters (parse_*, format_*, slugify_*), import affinity (which functions reach for which third-party libraries), data-type signatures, call-site clustering, lifecycle phases for god-classes (__init__ / load_* / run / cleanup), repetition for HTML (<div class="card"> appearing 14 times → a partial + loop), and an explicit rule that pure data (no callables, no imports, no f-strings) belongs in .json/.yaml/.toml, not Python.
  • Phase 4 — Proposal output. Writes a single markdown report to /tmp/long-file-split-proposal-<ts>.md with a fixed template: punch list at the top, then per-file sections with structural summary, archaeology (why it grew), cohesion analysis, the actual split table (new file / symbols moved / approximate lines / one-sentence rationale), explicit public-API preservation strategy (re-export shim vs. updating call sites vs. both with a deprecation cycle), risks with mitigations, and a suggested commit sequence. Estimates use AST spans rather than guesses.
  • Phase 5 — Opt-in guided execution. Triggered only when the user explicitly asks to execute a specific file's split. Pre-flight gates: clean working tree (refuses if dirty), green test suite (refuses if red), branch creation. Per-row procedure: move symbols → run tests → commit with a focused message → next row. Tests-red anywhere in the sequence aborts and reports rather than attempting to repair in-flow. Final pass runs static checks (ruff, mypy, tsc --noEmit, eslint) but fixes only warnings caused by the split itself.
  • Iron Laws. Proposal before action; no speculative refactors (the layout change must not be mixed with renames, signature changes, or bug fixes noticed in passing); preserve external API (every split picks an explicit strategy); one file per execution session; skip honestly when length is justified.
  • Per-language quick reference. Python-specific guidance for Django views.py / models.py / admin.py / serializers.py splits (including Meta.app_label requirement when models move into subpackages), JavaScript/TypeScript barrel files and component-folder layouts, hook extraction for React, and partial / {% include %} / component patterns for Django / Jinja / React / Vue / Svelte HTML. Each section ends with a concrete "avoid" — e.g., don't multiply 1 file into 20 trivial files; the tooling overhead eats the win.
  • Common mistakes section. Splitting on length alone, premature multiplication, breaking API silently, mixing categories per commit, ignoring circular imports, refactoring "while you're there", skipping the test run between commits.

Added — repository

  • New plugin entry in marketplace.json and a new "Code quality & refactoring" section in README.md, plus install + usage entries for the new skill.
  • Lockstep version bump: every plugin and the marketplace move to 2026.05.8.

2026.05.7 — 2026-05-12

readme-guardian quality pass. Frontmatter description, badge handling, and the Django × Python matrix are all reworked to remove drift hazards and bring the skill closer in line with superpowers:writing-skills guidance.

Changed — readme-guardian

  • Description rewritten to triggering conditions only. Previous description summarized the workflow ("converts RST to Markdown, ensures proper badges..., short description, rationale, features list..."), which writing-skills warns becomes a shortcut Claude follows instead of reading the body. New description names only the situations that should trigger the skill (missing badges/install/version-support/rationale/features, RST README, post-python-upgrade-package, pre-publish).
  • Django × Python matrix extracted to canonical reference file. New django-python-matrix.md sibling file is the single owner of (Django, Python) compatibility data — snapshot date, upstream URLs (Django FAQ + downloads page), the matrix itself, and a step-by-step regeneration procedure. SKILL.md no longer inlines the table; it points to the reference and enforces a 90-day freshness gate before emission. Other skills that referenced the inline matrix (e.g., python-upgrade-package Step 2b) now have a single file to point at instead of a section in SKILL.md.
  • Badge identifier derivation rules added (new Step 0f). Previously the target template carried literal OWNER, REPO, WORKFLOW, and PACKAGE placeholders with no instruction for how to resolve them. New section walks each one to its source (pyproject.toml [project].name, [project.urls].Repository / git remote, .github/workflows/ test workflow filter), with fallbacks. Includes a curl https://pypi.org/pypi/${PACKAGE}/json presence check that gates which badge set to emit.
  • Two badge variants for PyPI vs non-PyPI packages. Was previously a single set of shields routed through img.shields.io/pypi/..., which renders as "invalid" badges for packages not published to PyPI. Variant A keeps the existing PyPI shields; Variant B swaps in github/license/OWNER/REPO, a static Python-versions badge, and drops pypi/v entirely. Agent picks based on the Step 0f presence check.
  • Django-mode skip note moved out of the README target's code fence. A blockquote ("Skip this subsection entirely for Django packages...") had been added inside the fenced markdown block that represents the emitted README — risk of it being copied into a user's output. Now an HTML comment outside the fence + comment-style <!-- only if … --> markers on the matrix headings, so the agent's instructions don't leak into rendered output. Also dropped the awkward empty-corner cell in the Python-only matrix table.
  • Step 5 commit is now opt-in. Title changed from "Apply and Commit" to "Apply (and commit only on request)". Explicit "do not commit on your own initiative" instruction, with the commit-message template kept but conditional on the user asking. RST cleanup also clarified: git rm if tracked.
  • Version-support analysis row updated. Step 2 checklist now distinguishes Django-mode (Django × Python only) from non-Django (Python only) and adds a DUPLICATED status so the analysis flags READMEs that emit both tables.
  • Common Mistakes row added for the duplicate-matrix pitfall in Django projects, and a Content generation rule (#5) forbidding a standalone Python matrix when Django is detected.

Iron law clarified

The compatibility data lives in exactly one place (django-python-matrix.md); skills that emit a per-project matrix derive from that file, never inline a copy. Freshness check is mandatory before emission — a stale matrix in a user's README is a failure mode, not a minor inaccuracy.

2026.05.6 — 2026-05-11

New plugin release: django-extract-app adds a guided workflow for extracting a Django app embedded in a monolithic project into a standalone reusable package. Designed to chain into the existing readme-guardian + oss-github-publisher pipeline so the extracted package ships with the same level of polish as a greenfield OSS project.

Added — django-extract-app plugin (two skills)

  • django-extract-app — the main extraction skill. 10-step procedure, one commit per step in the new package's repo:

    1. Reconnaissance — locates the app inside the monolith, confirms it's a real Django app (apps.py / INSTALLED_APPS membership / settings module detection), inventories source files (models / views / admin / signals / commands / templatetags / templates / static / migrations / existing tests).
    2. Extractability audit — eight independent checks, each producing OK / WARN / BLOCK: cross-app FKs + dynamic apps.get_model refs, cross-app imports, settings.X usages (becomes the package's settings contract), migration dependencies=[...] cross-app deps, reverse() + {% url %} cross-app coupling, template {% extends %} / {% include %} cross-app, @receiver(..., sender=other_app.X) signals, admin.site.register of cross-app models. ≥1 BLOCK pauses and asks the user to abort or knowingly continue.
    3. Decisions (via AskUserQuestion, one at a time) — OSS vs internal package, PyPI distribution name (suggests django-<foo>), importable name, target package path, git history strategy (clean init or git filter-repo --subdirectory-filter with preflight check), DB engine in tests (auto-detects Postgres-specific features like ArrayField, JSONField PG ops, django.contrib.postgres, pg_trgm and recommends pytest-testcontainers-django; defaults to SQLite otherwise), example/demo project (default Yes for apps with views/urls/admin/templates), license choice.
    4. Repo bootstrap — mkdir + git init (or git filter-repo on a clone of the monolith), copies app source into src/<importable>/, first commit "Initial extraction from <monolith> @ <sha>".
    5. pyproject.toml + uv — detects third-party imports vs stdlib/Django/same-package, generates a complete [project] block (Django × Python matrix from readme-guardian's canonical table — 5.2 LTS + 6.0 as of 2026-05-08), [tool.setuptools.packages.find] for src/ layout, [tool.pytest.ini_options] with DJANGO_SETTINGS_MODULE = "tests.settings", runs uv sync --all-extras.
    6. Test scaffolding with TDD stubs — generates tests/settings.py (SQLite or Postgres-via-testcontainers fixture), conftest.py, optional tests/urls.py, and per-construct test stubs (test_models.py, test_views.py, test_admin.py, test_signals.py, test_commands.py, test_templatetags.py, test_apps.py) — every stub pytest.skip("TODO: ...") so failure-to-fill-in is visible. Existing tests from the monolith are migrated with import paths rewritten (from <monolith>.<app>.X → from <importable>.X); cross-app imports get flagged as TODO comments rather than auto-resolved.
    7. GitHub Actions — .github/workflows/tests.yml with include:-style Django × Python matrix (5.2 LTS / 6.0 × Python 3.10–3.14, intersected with readme-guardian's canonical compatibility table), separate lint job (ruff check + format, non-blocking), optional example-check job for the demo project (manage.py check + migrate --run-syncdb). Postgres mode adds a docker info smoke step (testcontainers manages its own container).
    8. Tooling — .pre-commit-config.yaml (ruff lint E/F/W + ruff-format + pyupgrade --py310-plus derived from requires-python + django-upgrade --target-version 5.2 derived from Django floor), LICENSE (MIT default), .gitignore (Python + Django + uv), README.md starter (install / INSTALLED_APPS / URLs / requirements / settings contract / dev setup), CHANGELOG.md in Keep-a-Changelog format.
    9. Optional example/demo project — example/ with manage.py + minimal settings.py (SQLite + admin) + urls.py wiring the package's URLs. Doubles as a CI smoke test for the package.
    10. Local verification + chaining — runs uv sync && uv run pytest && uv run pre-commit run --all-files && uv build && uv run --with twine twine check dist/*, surfaces any failure to the user (no silent auto-fix). Then asks (default Yes) whether to chain into /readme-guardian:readme-guardian and /oss-github-publisher:oss-github-publisher, both invoked via the Skill tool in the new package's directory. Final report enumerates commits, audit findings, verification results, and next-steps (push to GitHub, publish to PyPI, run the cleanup sub-skill).

  • django-extract-app-cleanup — phase-2 sub-skill, invoked separately after the package is published or installed editable. Wires the new package back into the monolith and removes the original app directory, in two commits with three verifications:

    1. Wiring — adds the package as a dependency via one of four strategies (AskUserQuestion): PyPI released (uv sync-immediately), PyPI placeholder (added with TODO, no install until publish), editable local install (uv add --editable <path>), or git URL with explicit @ref. Verifies the package imports and manage.py check passes BEFORE removing the original app — so removal can't strand the monolith if wiring is broken.
    2. Removal — before deletion, runs a migration safety check: compares the AppConfig label between monolith and installed package, and diffs the migration filename sets in <monolith>/<app>/migrations/ vs the installed package's <importable>/migrations/. If either mismatches, stops without deleting. Otherwise git rm -r <app>/, then re-verifies all three: manage.py check, makemigrations --check --dry-run (the critical drift detector — if this finds anything, the source-of-truth has diverged and the user must resolve manually, not paper over with a fresh migration the package won't have), and the monolith's full test suite.

Added — repository

  • New plugin entry in marketplace.json and README.md (plugin table, Python project lifecycle skill graph with extraction sub-graph, install + usage sections, cross-references).
  • Cross-reference: django-extract-app reads the canonical Django × Python compatibility matrix from readme-guardian's SKILL.md — same single source of truth as python-upgrade-package, no duplicated copies of the table.

Iron laws encoded in the extraction skill

  1. Never modify the monolith in the main skill. Read-only access only. Wiring + removal happens in the separate cleanup sub-skill, with a clean working-tree precondition.
  2. One commit per step in the new repo. Every step is independently revertible.
  3. Every decision goes through AskUserQuestion. No silent defaults for naming, location, history strategy, DB engine, license, or example project.
  4. Audit BLOCKers stop the workflow. A blocking issue triggers an explicit abort/continue prompt — never silently shipping a known-broken package.

2026.05.5 — 2026-05-08

Security hardening release. Three iterative codex self-reviews of the repo surfaced 9 distinct sandbox/secret-handling issues in code-review-opencode (and adjacent skills). Every finding was reproduced empirically before fixing — git diff --no-index /etc/hosts /dev/null, git log --output=PATH, malicious-filename command injection via xargs -I {}, and nested-.env pathspec leakage all confirmed live. Round 3 returned with no new findings against the previously-fixed surfaces, so this release closes the audit. Also folds in stale Django version data from a parallel review.

Fixed (security — code-review-opencode)

  • Bash cat/find/grep/rg/head/tail/wc wildcards removed. The previous policy paired read.*.env: deny with bash."cat *": "allow", so the model could cat .env straight through the bash channel and bypass the read-tool path policy. Replaced with read/glob/grep opencode tools (which respect path policy) for any file access.
  • git * wildcards removed. Blanket "git *": "allow" accepted destructive verbs like git checkout HEAD~5 -- /etc/passwd, git clean -fdx, git reset --hard, git config user.email evil@x. Whitelist now lists only explicit read-only verbs.
  • git diff and git diff * removed entirely. git diff --no-index FILE1 FILE2 reads any file the user can read, fully bypassing external_directory: deny. Verified empirically: git diff --no-index /etc/hosts /dev/null dumps /etc/hosts content. Glob can't express "any flag except --no-index", so the verb is dropped — diff for uncommitted/commit modes is now pre-computed in the wrapper instead.
  • git log * and git show * removed. Both verbs accept --output=PATH, which writes a file (verified: git log --output=/tmp/x -1 creates /tmp/x). That bypasses the edit tool policy via the bash channel. git log and git show are still allowed without arguments (HEAD-only, no flag injection surface).
  • git branch * and git tag * removed. Wildcards accepted mutating forms — git branch -D foo, git tag -d v1, git branch new, git tag new — which modify .git/. No-args forms (list-only) are still allowed.
  • Command injection in untracked-file pre-compute. The previous git ls-files --others ... | xargs -I {} sh -c '... cat -- "{}"' template literal-substituted filenames into the sh -c template, so a malicious filename like $(touch /tmp/INJECTED) executed the substitution before reaching sh. Verified live: old pattern created /tmp/INJECTED, new pattern (NUL-delimited bash while read -d '' loop) does not. The new loop also doesn't cat content at all — opencode now reads each file via its own read tool under the path-deny policy.
  • .env pathspec was anchored to repo root only. :(exclude).env and :(exclude).env.* excluded ./env only; sub/.env and sub/.env.local slipped through into git diff HEAD. Replaced with :(exclude,glob).env, :(exclude,glob)**/.env, :(exclude,glob).env.*, :(exclude,glob)**/.env.*. Verified: with the old form sub/.env content leaked into the diff; new form catches all four nesting variants.
  • Tracked-diff exclusion extended to all secret families. Was filtering .env* only; now mirrors the full read denylist (*.pem, *.key, id_rsa*, id_dsa*, id_ed25519*, id_ecdsa*, *.p12, *.pfx, *.keystore, *.jks, credentials.json, *credentials*.json, service-account*.json, .npmrc, .pypirc, .netrc) — root + every nested directory. Smoke-tested with 16 secret files across 3 directory levels: all filtered, only the benign tracked change remains.
  • Read-tool denylist was root-only. Patterns like id_rsa, .npmrc, service-account*.json matched only the repo root because opencode pattern matching does not cross / with a single *. secrets/id_rsa and packages/app/.npmrc were going through. Every secret pattern is now duplicated — once without **/ (root) and once with **/ (any depth).

Fixed (other skills)

  • github-build-fixer — short-poll loop now actually waits. Step 6's for _ in 1..6; do gh run list ...; done had no delay, so it fired 6 queries in well under a second and returned "No run found" before GitHub had time to register the workflow run. Added sleep 5 between iterations and clarified the Step 1 sleep prohibition: it covers the long CI watch (use gh run watch instead), not the short post-push materialization wait.
  • readme-guardian — Django × Python compatibility matrix updated to 2026-05-08. Snapshot was missing Django 6.0 and treated 4.2 LTS as still in extended support (which expired Apr 2026). Verified against djangoproject.com/download/ and the dev/faq/install/ page: only 5.2 LTS (Python 3.10–3.14) and 6.0 (Python 3.12–3.14) are upstream-supported as of 2026-05-08; 4.2/5.0/5.1 are EOL. Table updated and dated; explicit "currently supported" callout added so consumers know which rows to filter to.
  • python-upgrade-package — Django matrix examples and badges propagated. Sample CI matrix (Step 4.3), classifiers list (Step 1.2), Django badge example (Step 4.5), and the "drop EOL versions" guidance (Step 4.2b) all updated to drop 4.2/5.0/5.1 and add 6.0. The canonical table still lives in readme-guardian — this skill only mirrors the examples.
  • python-upgrade-package — Django conftest.py example fixed. The Step 5.5 sample assigned django.conf.settings.DJANGO_SETTINGS_MODULE = "..." on the LazySettings proxy, which silently does nothing useful and would surface as ImproperlyConfigured on the next django.setup(). Replaced with os.environ.setdefault("DJANGO_SETTINGS_MODULE", ...) and demoted the snippet — the pyproject.toml [tool.pytest.ini_options] form is now presented as the preferred path.

2026.05.4 — 2026-05-08

Patch release: closes two MEDIUM findings from a code-review-opencode self-review of the repo. Fixes a real .env exfiltration risk in the opencode review wrapper and propagates the tmux observability/stuck-detection pattern from code-review-external into the premortem-multiple family.

Fixed

  • code-review-opencode — .env no longer leaks into prompts in uncommitted mode. The pre-computed DIFF (git diff HEAD + cat of every untracked file) ran in the main agent before the opencode restrictive config existed, so any local .env / .env.local was being read into the prompt body sent to the opencode API regardless of the in-session permission deny. Replaced with git diff HEAD -- ':(exclude).env' ':(exclude).env.*' plus a grep -Ev '(^|/)\.env($|\.)' filter on the untracked file list.

Changed

  • premortem-codex — migrated to tmux pattern. Codex now runs inside a detached pm-codex-{TS} tmux session (analogous to cr-codex-{TS} in code-review-external). Pane output captured via tmux pipe-pane to /tmp/premortem-codex-{TS}.log; codex still writes the final markdown report to /tmp/premortem-codex-{TS}.md via its write tool (artifact-file pattern preserved). Switched invocation from codex exec ... > $RUN_LOG 2>&1 (raw pipe) to codex exec --skip-git-repo-check --sandbox workspace-write ... </dev/null inside the runner.
  • premortem-opencode — migrated to tmux pattern. Opencode now runs inside a detached pm-opencode-{TS} tmux session. Restrictive project-local .opencode/opencode.json setup + cleanup trap kept intact, with one fix: the trap now also calls tmux kill-session so a Ctrl-C in the parent agent cleans up both the session and the config. Added --print-logs so the bootstrap stage is visible (and stuck-detectable) instead of buffered into silence. Combined config-setup + tmux-launch + 90 s stuck-detector + hard 600 s deadline live in one bash command (trap fires after polling, never before).
  • premortem-multiple wrapper — surfaces tmux attach commands. After dispatching the three background tasks, the wrapper prints tmux attach -t pm-codex-{TS} and tmux attach -t pm-opencode-{TS} so users can watch either reviewer live. The Claude subagent stays on the Agent tool (no tmux) — it has its own task lifecycle and doesn't need pane observability.

Added

  • New shared file plugins/premortem-multiple/shared/tmux-runner.md — premortem-flavored variant of the code-review-external tmux runner pattern (different naming: pm-{tool}-{TS} sessions, /tmp/premortem-{tool}-{TS}.{md,log} files). Referenced by both leaf skills.

Requires

  • tmux ≥ 3.0 in $PATH for premortem-codex and premortem-opencode (preflight which tmux; abort with install hint if missing). Same requirement as code-review-external.

2026.05.3 — 2026-05-08

Patch release: code-review-external family rewritten to dispatch every reviewer (codex / opencode / claude) inside its own attachable tmux session, with a stuck-detector that aborts after 90 s of no log growth instead of hanging for 15+ min on a wedged provider.

Changed

  • code-review-external family — all four skills (code-review-codex, code-review-opencode, code-review-claude, code-review-external) now run their underlying CLI inside a detached tmux session named cr-{tool}-{TS}. Users can tmux attach -t cr-{tool}-{TS} at any point to watch a reviewer live. Pane output is captured to /tmp/code-review-{tool}-{TS}.log via tmux pipe-pane while the reviewer writes its final markdown to /tmp/code-review-{tool}-{TS}.md (artifact-file pattern preserved).
  • code-review-claude — switched from Agent-tool subagent dispatch to a headless claude -p process inside tmux. Uses --permission-mode auto, --add-dir /tmp, --add-dir <project>, --allowedTools "Read Grep Glob Bash Write Edit". Three reviewers now share identical lifecycle (tmux session) instead of mixing tmux + Agent.
  • code-review-external wrapper — replaced Bash run_in_background + TaskOutput x3 orchestration with a single bash command that creates three tmux sessions, prints three attach commands up-front, and runs one combined polling loop (tmux has-session × 3, per-session stuck detector, hard 600 s deadline). Simpler model, no pkill, deterministic kill via tmux kill-session.
  • code-review-codex — switched from codex review to codex exec --skip-git-repo-check --sandbox workspace-write. The old codex review lacks --skip-git-repo-check and crashed in non-git directories with Not inside a trusted directory; codex exec works in both git and non-git contexts.

Added

  • New shared file plugins/code-review-external/shared/tmux-runner.md — universal tmux launch + polling pattern referenced by all four leaf/wrapper skills (naming convention, printf %q-based runner script, tmux pipe-pane capture, stuck-detector loop, combined-poll variant for the wrapper).
  • --print-logs flag added to opencode invocations so the bootstrap stage is visible in the pane (and in RUN_LOG) instead of buffered into silence — makes auth/network hangs diagnosable in real time.

Fixed

  • Opencode reviews on machines where the model API is slow or wedged no longer wait 15+ min in silence — stuck detector kills the tmux session after 90 s of zero log growth and reports last log lines to the user.
  • Codex reviews of design docs in fresh / non-git directories (e.g. SPEC.md in a brand-new project folder) no longer crash with Not inside a trusted directory and --skip-git-repo-check was not specified.

Requires

  • tmux ≥ 3.0 in $PATH for all code-review-external skills (preflight which tmux; abort with install hint if missing). Install: brew install tmux / apt install tmux.

2026.05.2 — 2026-05-08

Major release: one new plugin, big rewrites across python-upgrade-package and oss-github-publisher, DRY refactor across the multi-CLI families, HTML reports for premortem-multiple, and a setuptools.build_meta build-backend bug fix that quietly broke every previous run of python-upgrade-package.

Plugins included in this release:

  • code-review-external — parallel external code review (codex + opencode + Claude subagent)
  • github-build-fixer — diagnoses and fixes failing GitHub Actions CI builds
  • oss-github-publisher — pre-flight audit before publishing a repo as open source
  • premortem — Klein-style premortem on plans, launches, hires, pricing, or strategy
  • premortem-multiple — three parallel premortems with meta-synthesis
  • python-upgrade-package — modernizes legacy Python packages
  • python2-cleanup ← NEW — removes Py2 compatibility cruft from a Py3 codebase
  • readme-guardian — analyzes and improves Python project READMEs

Added

  • New plugin python2-cleanup — 12 categories (__future__, six, unicode()/basestring/xrange/long/unichr, dict iter/view methods, dict.has_key(), python_2_unicode_compatible, u'' prefix, custom s2u/u2s/compat.py helpers, __metaclass__/with_metaclass/add_metaclass, stdlib renames, dunder renames, optional super() simplification). Ripgrep-first detection, one commit per category, tests after every change. Iron Law: touch only Py2 cruft, never reformat.
  • oss-github-publisher rewritten with 10 new audit areas:
    • D.2 extended secret regex set: Stripe (sk_live_/rk_live_/sk_test_), Twilio (SK/AC SIDs), HuggingFace, DigitalOcean, Heroku HRKU, Mailchimp, Mailgun, Shopify (shpat/shpss/shpca), Square, Facebook EAACE, Google ya29 + refresh tokens, JWT, database connection URLs (postgres/mysql/mongodb/redis/amqp/mssql/cockroachdb).
    • D.7 hardcoded internal URLs / hostnames (.internal, .intra, .corp, staging-/qa-/dev- patterns, internal SaaS instances, AWS ELB DNS, private IP ranges).
    • D.8 TODO/FIXME comments with internal ticket references (JIRA-style, GitHub-style, Asana/Linear/Trello).
    • D.9 IDE / editor configs tracked (.idea/.vscode/.history/.fleet/.nova/sublime-workspace) with per-path decision table; cross-content scan for embedded /Users/<name>/ paths.
    • D.10 GPG/SSH key material and config (private keys → FAIL, host keys → FAIL, public keys/known_hosts → WARN).
    • D.11 OS / IDE turds (.DS_Store, Thumbs.db, desktop.ini, etc.).
    • Section E (new) — GitHub Actions security: SHA-pinning third-party actions (post tj-actions Mar 2025 incident), pull_request_target misuse, deprecated actions (upload-artifact@v3 etc.), permissions blocks, secrets in logs, run-name injection.
    • Section F (new) — PyPI name availability check (curl https://pypi.org/pypi/<name>/json, normalisation rules).
    • Section G (new) — PyPI metadata audit ([project.urls], classifiers, Development Status, three-way license consistency, no deprecated [tool.setuptools] entries).
    • Section H (new) — README content audit (length, Installation/Usage sections, no lorem ipsum, code blocks present, badges resolve).
  • python-upgrade-package Step 1: build smoke-test — uv build + twine check before commit, catches broken metadata before it hits PyPI.
  • python-upgrade-package Step 1: [project.urls], keywords, and classifiers blocks added to the pyproject.toml template, with extraction rules and a "no fabricated placeholders" rule.
  • python-upgrade-package Step 1: full enumeration of requirements*.txt variants (requirements/, *-requirements.txt, requirements-{test,docs,dev,prod}.txt) with find discovery and per-variant routing into [project.dependencies] vs [project.optional-dependencies] extras.
  • python-upgrade-package Step 3: pyupgrade + django-upgrade hooks in .pre-commit-config.yaml, with --py-plus derived from requires-python floor and --target-version derived from the project's lowest Django.
  • python-upgrade-package Step 4: dynamic test matrix derivation — Python versions from requires-python, Django × Python pairs from canonical compatibility matrix (now sourced from readme-guardian). include:-style YAML for Django projects with explicit (python, django) pairs.
  • python-upgrade-package Step 6: Makefile substitution table — 21 concrete substitutions (e.g., python setup.py install → uv sync, flake8 → uv run ruff check, python setup.py sdist bdist_wheel → uv build) plus a list of what NOT to auto-substitute.
  • premortem-multiple HTML visual report (parity with single premortem): dark theme, synthesis above the fold, per-agent accent colors (codex cyan / opencode green / claude purple), consensus / divergence / contradiction sections with pill-badges showing which agents converged. Saved alongside markdown synthesis and transcript.
  • code-review-external family: free-form mode (5th detection mode) — when argument doesn't match file/dir/commit, treat as a free-form hint passed verbatim to all three reviewers (e.g., "całe repo", "security audit src/auth/").
  • README: skill graph with three Mermaid diagrams (Python project lifecycle / decision-making / code review), recommended sequence ("python-upgrade-package → python2-cleanup → readme-guardian → oss-github-publisher → publish"), cross-references between skills.

Changed

  • code-review-external family: DRY refactor. Extracted 3 shared files (shared/standard-review-prompt.md, shared/target-detection.md, shared/write-directive.md). The 4 leaf SKILL.md files reference these instead of duplicating ~80 lines × 3. Editing review criteria is now a single-file change. Net: −161 lines.
  • premortem-multiple family: same DRY refactor. Extracted 2 shared files (shared/standard-premortem-prompt.md, shared/write-directive.md). Net: −77 lines.
  • All external review and premortem skills: switched from tee-pipeline to artifact-file pattern. Each tool (codex / opencode / claude subagent) writes its final review or premortem directly to a designated $OUT file via its own write tool; verbose stdout / stderr go to a separate .log file. The wrapper reads only the clean $OUT files (1–3 KB of markdown each, instead of 50+ KB of banners and reasoning).
  • opencode skills (review + premortem): temporary project-local .opencode/opencode.json config with explicit allow / deny rules (read repo, bash limited to git/ls/find/cat/etc., write only to /tmp/code-review-* or /tmp/premortem-*), restored via trap on EXIT/INT/TERM. Avoids --dangerously-skip-permissions and prevents 60+ minute hangs on permission asks in non-TTY mode.
  • code-review-claude and premortem-claude: switched from "subagent returns text + main agent Writes" to "subagent writes directly to $OUT via Write tool" — now consistent with codex / opencode and with wrapper expectations.
  • Django × Python compatibility matrix: now lives canonically in readme-guardian (single source of truth across the plugin family). python-upgrade-package reads from there instead of maintaining a parallel copy. The readme-guardian snapshot was also updated to match the canonical version (added Django 5.2 LTS, fixed two factual errors that mismarked Django 5.0 support).
  • README: plugins regrouped under three purpose categories (Python project lifecycle / Decision-making / Code review).

Fixed

  • python-upgrade-package BUG: pyproject.toml template specified build-backend = "setuptools.backends._legacy:_Backend", which is not a real backend identifier. Every project modernised by previous versions of this skill ended up with a pyproject.toml that fails to build. Replaced with setuptools.build_meta (the canonical default) in both spots where it appeared (Step 1 template + Step 2 setuptools-scm template). The Step 1 build smoke-test added in this release would have caught this — and now does.
  • python-upgrade-package ruff target-version was hardcoded "py310" despite the skill's own "DO NOT hardcode" rule for the CI matrix. Now derived from requires-python floor, with an explicit derivation table mirroring Step 4.2a.
  • python-upgrade-package lint-job example used uv python install 3.13 literal — replaced with HIGHEST_PY placeholder so the workflow run fails loudly if not replaced (rather than silently using a wrong matrix-misaligned version).
  • github-build-fixer Step 6 used sleep 5 after git push to "wait for GitHub to register the push" — both a Bash antipattern blocked by the harness and unnecessary (gh run watch polls). Replaced with a SHA-based poll loop (up to 30 s, exits early when the run for HEAD's SHA appears) and an explicit Bash timeout: 1800000 directive (30 min) on gh run watch — previously the default 2-min Bash timeout would kill the watch on any CI run longer than 2 minutes.
  • code-review-codex dropped --uncommitted / --commit <SHA> flags — codex (≥ 0.129) rejects those when [PROMPT] is given (the argument '--uncommitted' cannot be used with '[PROMPT]'). Codex now runs git diff / git show itself via its bash tool, instructed by the prompt body, which also lets the artifact-file write directive co-exist with the target-specifying invocation.

2026.05.0 — 2026-05-07

Initial CalVer cut. Versioning scheme moved from SemVer (1.0.0) to CalVer (YYYY.0M.MICRO).

Plugins included in this release:

  • code-review-external — parallel external code review (codex + opencode + Claude subagent)
  • github-build-fixer — diagnoses and fixes failing GitHub Actions CI builds
  • oss-github-publisher — pre-flight audit before publishing a repo as open source
  • premortem — Klein-style premortem on plans, launches, hires, pricing, or strategy
  • python-upgrade-package — modernizes legacy Python packages
  • readme-guardian — analyzes and improves Python project READMEs

Added

  • New premortem plugin (migrated from a personal ~/.claude/skills/premortem.md to a full plugin layout).
  • Top-level version field in .claude-plugin/marketplace.json.
  • This CHANGELOG.md.
  • scripts/bump-version.sh — helper to advance every version field across the repo to the next CalVer.

Changed

  • All plugin.json and marketplace entries: 1.0.0 → 2026.05.0.
  • README.md tagline expanded to cover code review and decision-making workflows.