pi-skill-shiori 0.3.0 → 0.3.2

package/CHANGELOG.md

@@ -0,0 +1,67 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+This project follows semantic versioning.
+
+## [0.3.2] - 2026-06-24
+
+### Added
+
+- `ROADMAP.md` with maintenance-first phased goals (Month 1–3), template compliance checklist, and maintenance schedule.
+- README links to `ROADMAP.md`.
+- npm package `files` now includes `ROADMAP.md`.
+
+## [0.3.1] - 2026-06-19
+
+### Added
+
+- `SECURITY.md` with vulnerability reporting guidance and Pi package security notes.
+- `CHANGELOG.md` for version history.
+
+### Changed
+
+- README links to `SECURITY.md` and `CHANGELOG.md`.
+- npm package files now include `SECURITY.md` and `CHANGELOG.md`.
+
+## [0.3.0] - 2026-06-08
+
+### Changed
+
+- Shiori slash commands use UI flows for doctor, bootstrap, reload, recommend, and stats.
+
+## [0.2.1] - 2026-06-02
+
+### Changed
+
+- Patch release after the recommendation flow updates in `0.2.0`.
+
+## [0.2.0] - 2026-06-02
+
+### Added
+
+- Recommendation-first `/shiori:recommend` flow with query expansion and agent task queuing.
+- Pre-loading of multiple recommended skills in one agent turn.
+- Planning-intent handling that returns a dev-plan intake template without dumping skill bodies.
+
+### Fixed
+
+- Skill retrieval for block descriptions and Japanese queries.
+
+### Changed
+
+- README install snippets aligned with the published package version.
+
+## [0.1.1] - 2026-05-26
+
+### Added
+
+- npm publish workflow and auto-release on merge to `main`.
+- npm publish commands and unpinned install examples in README.
+
+## [0.1.0] - 2026-05-24
+
+### Added
+
+- Initial Pi Skill Shiori release.
+- Zero-Catalog mode, policy-based retrieval, SQLite FTS indexing, and on-demand `shiori_load_skill` tool.

package/README.md

@@ -4,6 +4,8 @@ Pi Skill Shiori is a [Pi](https://pi.dev/) package that keeps large Agent Skill
 
 It is built for vaults or projects with many Agent Skills where the default catalog becomes noisy and expensive.
 
+For maintenance priorities and the phased plan, see [`ROADMAP.md`](ROADMAP.md).
+
 ## What it does
 
 - **Zero-Catalog Mode**: hides the normal skill catalog when Pi exposes one in the system prompt.
@@ -15,7 +17,7 @@ It is built for vaults or projects with many Agent Skills where the default cata
 
 ## Status
 
-`0.3.0` is an early working release. It is useful for local Pi workflows, but the prompt-boundary suppression logic is intentionally conservative: if Shiori cannot safely recognize a catalog boundary, it leaves the prompt untouched and warns instead of deleting too much.
+`0.3.1` is an early working release. It is useful for local Pi workflows, but the prompt-boundary suppression logic is intentionally conservative: if Shiori cannot safely recognize a catalog boundary, it leaves the prompt untouched and warns instead of deleting too much.
 
 ## Requirements
 
@@ -37,7 +39,7 @@ Pi core packages are peer dependencies and should be supplied by the Pi runtime:
 After the package is published to npm:
 
 ```bash
-pi install npm:pi-skill-shiori@0.3.0
+pi install npm:pi-skill-shiori@0.3.1
 ```
 
 Without a version pin:
@@ -51,7 +53,7 @@ pi install npm:pi-skill-shiori
 Use `-l` to write the package to the current project’s `.pi/settings.json`:
 
 ```bash
-pi install -l npm:pi-skill-shiori@0.3.0
+pi install -l npm:pi-skill-shiori@0.3.1
 ```
 
 Without a version pin:
@@ -65,7 +67,7 @@ Equivalent pinned manual `.pi/settings.json` entry:
 ```json
 {
   "packages": [
-    "npm:pi-skill-shiori@0.3.0"
+    "npm:pi-skill-shiori@0.3.1"
   ]
 }
 ```
@@ -83,7 +85,7 @@ Equivalent unpinned manual `.pi/settings.json` entry:
 ### Global install from GitHub
 
 ```bash
-pi install git:github.com/eiei114/pi-skill-shiori@v0.3.0
+pi install git:github.com/eiei114/pi-skill-shiori@v0.3.1
 ```
 
 Without a tag pin:
@@ -95,7 +97,7 @@ pi install git:github.com/eiei114/pi-skill-shiori
 ### Project-local install from GitHub
 
 ```bash
-pi install -l git:github.com/eiei114/pi-skill-shiori@v0.3.0
+pi install -l git:github.com/eiei114/pi-skill-shiori@v0.3.1
 ```
 
 Without a tag pin:
@@ -267,6 +269,8 @@ npm whoami
 npm run release:npm
 ```
 
+For version history, see [`CHANGELOG.md`](CHANGELOG.md).
+
 For a first-time publish, login first:
 
 ```bash
@@ -276,7 +280,7 @@ npm login
 After publishing, verify the Pi install path:
 
 ```bash
-pi install npm:pi-skill-shiori@0.3.0
+pi install npm:pi-skill-shiori@0.3.1
 pi install npm:pi-skill-shiori
 pi list
 /shiori:doctor
@@ -290,12 +294,14 @@ git push --follow-tags
 npm run release:npm
 ```
 
-## Security notes
+## Security
 
 Pi packages execute local code with the same permissions as Pi. Review third-party packages before installing them.
 
 Shiori itself does not sandbox skills. It only changes how skill candidates are discovered and loaded. A loaded skill can still instruct the model to run tools, edit files, or execute commands according to your Pi/tool permissions.
 
+For vulnerability reporting, see [`SECURITY.md`](SECURITY.md).
+
 ## License
 
 MIT

package/ROADMAP.md

@@ -0,0 +1,132 @@
+# Roadmap
+
+Pi Skill Shiori keeps large Agent Skill catalogs out of the model prompt and lets
+the agent load only the skills that match the current task. This roadmap states
+**what we prioritize next** and how the package is maintained.
+
+## Direction
+
+`0.3.x` is a working release. The near-term priority is **stabilization and
+public quality over new feature surface**. Before widening functionality, we
+harden what already ships:
+
+- **Stabilize existing behavior** — make catalog suppression and retrieval reliable.
+- **Speed and token efficiency** — startup cost, index rebuild, and candidate injection footprint.
+- **Explicit design boundaries** — name the contracts (policy file, hook activation, suppression).
+- **Template compliance** — close gaps against the shared Pi extension OSS policy.
+- **Publish quality** — README, CHANGELOG, SECURITY, CI, `npm pack`, and release handoff.
+
+Constraints:
+
+- Existing commands, tools, and the main UX stay unless a compatibility fix is required.
+- `shiori:*` commands stay argument-less; detail is collected after launch via Pi UI.
+- npm publish stays Trusted Publishing / OIDC. Secrets and production actions remain human-owned.
+
+## Skill features vs. template requirements
+
+Two kinds of work are tracked separately so feature work is not confused with
+packaging/hygiene work:
+
+- **Skill features** — Zero-Catalog suppression, policy-based retrieval, candidate
+  injection, on-demand loading, hybrid retrieval. These change what the package does.
+- **Template requirements** — README badges, CI gates, `npm pack` cleanliness, action
+  pinning, version-bump checks. These keep the package publishable and aligned with
+  `pi-extension-template`.
+
+Template gaps are addressed first because they protect every later release.
+
+## Maintenance schedule
+
+- **Patch release** — as needed for fixes, doc/template compliance, and small behavior changes.
+- **Minor release** — retrieval quality or candidate-flow improvements.
+- **Release cadence** — ship per ready slice; no fixed calendar release. Every mergeable
+  change bumps `package.json` and `CHANGELOG.md` so npm can publish.
+- **Dependency review** — monthly; bump `devDependencies`/Pi peers without behavior change.
+- **Release flow** — `npm version <patch|minor>` → push → `auto-release.yml` tags and
+  releases → `publish.yml` publishes via Trusted Publishing (provenance verified).
+
+`/shiori:stats` reports operational counters (token saved, candidate hits, loads,
+zero-candidate turns, explicit invocations) that feed back into this schedule.
+
+## Phased goals (Month 1–3)
+
+Each item is a small, independently verifiable slice (roughly 30–90 minutes) with a
+semver classification. Patches default to `patch`; `none` means no package publish.
+
+### Month 1 — Public quality and template compliance
+
+| # | Slice | Bump | Status |
+|---|---|---|---|
+| 1 | Add README status badges (CI, Publish, npm version, npm downloads, License, Pi package, Trusted Publishing) | patch | planned |
+| 2 | Add `version:check` CI gate (block publishable-path changes without a semver increase) | none | planned |
+| 3 | Pin GitHub Actions to commit SHAs in `ci.yml`, `auto-release.yml`, `publish.yml` | none | planned |
+
+### Month 2 — Stabilization and design boundaries
+
+| # | Slice | Bump | Status |
+|---|---|---|---|
+| 4 | Harden catalog boundary detection and add boundary tests (reduce `failed-pattern-not-found`) | patch | planned |
+| 5 | Integration test for Dual Hook Activation (input-hook explicit pass-through + `before_agent_start` candidate injection) | none | planned |
+
+### Month 3 — Speed and token efficiency
+
+| # | Slice | Bump | Status |
+|---|---|---|---|
+| 6 | Add startup/operation timing instrumentation to `/shiori:stats` (Full Index Rebuild + retrieval baseline) | patch | planned |
+| 7 | Startup index strategy (incremental / lazy / cached) — needs a design decision before implementation | none | blocked (design) |
+
+## Integration testing for skill hooks
+
+Dual Hook Activation is the core discovery path and must be covered end to end:
+
+- **Input hook** — explicit invocation pass-through: `/skill:name` and explicit names load
+  regardless of activation policy, including under Zero-Catalog Mode.
+- **`before_agent_start` hook** — automatic candidate injection for triggerable skills only,
+  bounded by `maxCandidates` and `minScore`, with explicit skills excluded.
+
+The Month 2 integration test asserts both paths together (explicit load still works while
+triggerable candidates are injected) so a regression in one hook surfaces immediately.
+
+## Template compliance checklist
+
+Current status against the shared Pi extension OSS policy. Gaps become Month 1 work.
+
+- [x] `README.md`, `LICENSE`, `SECURITY.md`, `CHANGELOG.md` at repo root
+- [x] CI runs `npm run typecheck`, `npm test`, and `npm pack --dry-run`
+- [x] npm Trusted Publishing (`publish.yml` uses `id-token: write`, no `NPM_TOKEN`)
+- [x] `auto-release.yml` → `publish.yml` release handoff
+- [x] `package.json` `files` is explicit (no `node_modules`, tests, or local state)
+- [x] No template `example-skill` / `example` / `example-theme` resources shipped
+- [ ] **README status badges** (CI, Publish, npm version, npm downloads, License, Pi package, Trusted Publishing) — *Month 1 #1*
+- [ ] **`version:check` CI gate** so publishable-path changes cannot merge without a bump — *Month 1 #2*
+- [ ] **Pin GitHub Actions to commit SHAs** (currently floating `@v4`) — *Month 1 #3*
+
+## Current backlog integration
+
+Feature-track items already shipped (kept here so the roadmap reflects reality, not a
+fresh to-do list):
+
+- Recommendation-first `/shiori:recommend` flow with query expansion and planning intent — shipped (`0.2.0`)
+- CI + auto-release + publish handoff — shipped (`0.1.1`+)
+- `shiori:*` commands use Pi UI flows (argument-less) — shipped (`0.3.0`)
+- `SECURITY.md` and `CHANGELOG.md` — shipped (`0.3.1`)
+
+Remaining feature-track item, deferred behind the maintenance-first direction:
+
+- README/docs alignment with the minimal public-docs policy — open; the Month 1 badge
+  and CI slices advance this track without blocking on the broader docs-policy decision.
+
+## Later (feature track)
+
+- Retrieval quality improvements (embedding candidate added beside SQLite FTS).
+- Cross-vault skill index.
+- ResourceLoader-side catalog replacement (stable-stage Two-stage Catalog Suppression).
+
+These are intentionally **after** the Month 1–3 maintenance work.
+
+## Non-goals
+
+- Rewriting existing commands, tools, or the candidate/recommend UX.
+- Auto-approving npm publish, permission changes, or production actions.
+- Removing the conservative catalog-suppression safety boundary (`not-needed` /
+  `failed-pattern-not-found` stay safe-by-default).

package/SECURITY.md

@@ -0,0 +1,22 @@
+# Security Policy
+
+## Supported versions
+
+Only the latest published version receives security fixes.
+
+## Reporting a vulnerability
+
+Open a private security advisory on GitHub, or contact the maintainer by the preferred channel listed in the repository profile.
+
+Please include:
+
+- Affected version
+- Impact
+- Reproduction steps
+- Suggested fix, if known
+
+## Pi package security note
+
+Pi packages can execute code with local user permissions. Review installed packages and avoid running untrusted extensions.
+
+Pi Skill Shiori does not sandbox skills. It changes how skill candidates are discovered and loaded. A loaded skill can still instruct the model to run tools, edit files, or execute commands according to your Pi and tool permissions.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-skill-shiori",
-  "version": "0.3.0",
+  "version": "0.3.2",
   "description": "Pi extension that reduces Agent Skill catalog tokens with policy-based skill retrieval and on-demand skill loading.",
   "type": "module",
   "license": "MIT",
@@ -23,7 +23,10 @@
   },
   "files": [
     "README.md",
+    "ROADMAP.md",
     "LICENSE",
+    "SECURITY.md",
+    "CHANGELOG.md",
     "src"
   ],
   "scripts": {