v0.5.0

retro-skill v0.5.0

This release teaches retro to make routing decisions against the whole skill
catalogue, not just what happens to be installed.

✨ Highlights

Classify against all org skills — installed or not (#17)
A new scripts/find-org-skills.py reads every configured marketplace's
locally-synced catalogue (offline, and generic — never a hardcoded org) plus the
plugin cache, and emits every skill with {name, description, repo_url, marketplace, installed}. Skill-ownership discovery is now a mandatory first
step of classification: a learning routes to the skill that owns it even when
that skill isn't installed locally, new-skill is only proposed when nothing in
the catalogue covers it, and an available-but-not-installed skill is detectable
("a teammate is simply missing skill X"). This closes a real mis-route where a
skill-authoring lesson landed in user-memory instead of the skill that owns skill
authoring.

🐛 Fixes

• A14 (git_op_without_branch) no longer false-fires on release tag pushes
  (git push origin v1.2.3) or on commits after git checkout main && git checkout -b feat — a push now counts only when it targets the main branch,
  and the last branch reported in a block wins (#15)

📚 Docs

• /retro promote added to the README and AGENTS mode lists (#16)

Full changelog: v0.4.0...v0.5.0

v0.4.0

retro-skill v0.4.0

This release adds a new way to break memory out of per-project silos and makes
the repo's quality gates actually enforced in CI.

✨ Highlights

/retro promote — drain local memory upward (#11)
A new mode that inventories the memory Claude Code accumulates under
~/.claude/projects/<slug>/memory/ (across all slugs) and re-homes each note
to its correct, shareable destination — skill-update › project-rule ›
user-memory, never project-local memory — reusing the existing
classify → approve → materialize pipeline. The source note is drained only
after the upward write is verified (a reversible tombstone move, never rm,
with a sha256 race-check).

Trigger-coverage sweep (B15) + broadest-scope escalation (#9)
Systematically checks every installed skill's description against what the
session actually did, and biases every finding toward the broadest useful scope.

🐛 Fixes

• Route user-memory → global ~/.claude/CLAUDE.md and project-rule →
  <project>/AGENTS.md (no more cwd-scoped silos) (#8)
• Detector noise reduction: cut A1/A14 false positives, trust the is_error
  flag, fix git-flag/token boundaries; add self-review + detector-regression
  gates (#7)

🔧 CI / tooling

• Enforce skill-validation hooks in CI via the reusable validate.yml workflow
  (validate-skill, markdownlint, yamllint, ruff, ShellCheck, version parity,
  checkpoint schemas) — previously these ran only as uninstalled local hooks (#13)
• .markdownlint.jsonc encoding repo conventions + a broken-table fix (#12)

📚 Docs

• README rework for correctness and onboarding (#10), plus workflow /
  patch-workflow / spec updates

Full changelog: v0.3.0...v0.4.0

v0.3.0

What's Changed

• feat: add .pre-commit-config.yaml mirroring CI checks by @CybotTM in #5
• chore(release): v0.3.0 by @CybotTM in #6

Full Changelog: v0.2.0...v0.3.0

v0.2.0

Highlights

This is the first tagged release of retro-skill — an LLM-driven session retrospective that detects friction patterns, classifies them into one of six destinations (user memory, project rules, skill PRs, checkpoints, harness artefacts, or no-op), and materializes approved learnings as actionable artefacts.

Schicht A — deferred mechanical signals (#3). Implements the A4/A5/A11/A13/A18 signals used by the retro detector to identify friction at scale. These are the cheap, mechanical pre-pass signals that run before the LLM analysis layer and shave search-space cost on long sessions.

Schicht D + explicit audit mode (#2). Adds the deepest analysis layer (cross-session pattern audits) and a distinct audit-mode entrypoint — separate from the end-of-session retro flow, so periodic architecture audits can run on demand without polluting the session-end signal.

Maintenance

• session-end hook now pipes stdin directly to jq instead of going through a shell variable, fixing a class of stdin-truncation issues on long sessions.

Release pipeline

This is the first release cut against the standardized skill-repo-skill reusable release workflow. The new .github/workflows/release.yml caller wires up:

• signed-tag-only release trigger,
• cosign-keyless SHA256SUMS.txt signing,
• SLSA build-provenance attestations on every archive,
• GitHub native attestation API verification via gh attestation verify.

Verification

gh attestation verify retro-skill-v0.2.0.zip --repo netresearch/retro-skill

cosign verify-blob \
  --bundle SHA256SUMS.txt.bundle \
  --certificate-identity-regexp '^https://github\.com/netresearch/skill-repo-skill/\.github/workflows/release\.yml@' \
  --certificate-oidc-issuer 'https://token.actions.githubusercontent.com' \
  SHA256SUMS.txt
sha256sum --check SHA256SUMS.txt

Install

/plugin install retro@netresearch
composer require netresearch/retro-skill

Use the skill at the end of a Claude Code session, on demand for specific issues, or periodically for cross-session audits. Triggers: /retro command, "retrospective", "review the session", "fix this skill", "we keep hitting this", "audit the architecture".