paqad-ai 1.34.0 → 1.36.0

package/CHANGELOG.md

@@ -1,5 +1,48 @@
 # paqad-ai
 
+## 1.36.0
+
+### Minor Changes
+
+- 69ab154: Add a native `aiassistant` adapter for JetBrains AI Assistant (#219).
+
+  - Onboarding and `paqad-ai refresh --providers` now generate and keep in sync `.aiassistant/rules/guidelines.md`, the lean bootstrap-pointer entry file AI Assistant auto-applies from `.aiassistant/rules/`.
+  - Soft, rules-only adapter (advisory hook coverage), like `junie`: AI Assistant exposes no hook/lifecycle system, so the sentinel gate cannot bind. MCP is intentionally omitted — AI Assistant configures MCP servers in the IDE, not a project file, so no dead artifact is written.
+  - The adapter is selectable everywhere the other ten are (onboarding prompt, `--providers`, factory), carries its Decision Pause Contract UI note, and is covered by tests parallel to `junie`.
+  - Also synced the `onboarding-manifest` schema `adapter` enum to the full adapter list (it had silently drifted to five, so a `cursor`/`aider`-onboarded manifest failed health-check validation).
+
+- 69ab154: Make the feature-development stage workflow **compel, not plead** (enforcement RCA, part 1 of the fix).
+
+  The stage-evidence recorder finally has a production caller, and a code edit is now blocked until the prior stages are recorded — closing the physics gaps that let every prior fix (#117–#261) regress.
+
+  - **Deterministic per-stage writer.** A new Claude PreToolUse hook (`stage-writer.mjs` → `src/stage-evidence/live-writer.ts`) script-mints a `live-mark` stage row on every Edit/Write/NotebookEdit, with real start/end times classified from the mutated file — no model involvement. The completion finalizer closes any stage left open at the turn boundary.
+  - **Block-forward deny.** A `stages` capability folded into the kernel evaluates at the pre-mutation seam and refuses an edit (exit 2, reason on stderr) until `planning` and `specification` each carry a recorded start+end pair. It reads the ledger, not a git delta, so a committed clean tree cannot bypass it. `se-mark` (a Bash script, ungated) is the escape hatch that clears the block.
+  - **Visible, non-vacuous verdict.** The Stop backstop now writes a failing verdict to **stderr** (the host surfaces only stderr to the model on exit 2), and the stage gate no longer skips on a zero-file working-tree diff — a committed-but-incomplete change still fails instead of passing silently.
+
+  Claude-only for the hard block (the only host with a pre-mutation seam); Codex/Gemini keep completion-time recording; other adapters are unchanged. Default `stages_mode` is strict.
+
+- 69ab154: Wire the last three enforcement consumers so the framework is felt, not just recorded (enforcement RCA, part 2 of the fix).
+
+  Building on the stage writer and block-forward deny, three contracts that shipped with a policy but no live consumer now bind through the kernel:
+
+  - **Delivery policy is enforced.** A `delivery` capability at the completion (Stop) seam reads HEAD branch/commit and — when `gh` can answer — the PR/CI state, then **warns** on a convention deviation (on the base branch, an off-convention branch name, red CI under `wait_for_green`) and appends a `delivery-evidence` row. Warn-floor by design: delivery is `mandatory:false`, so it surfaces a bad push one turn late but never blocks. Degrades cleanly when `gh` is absent.
+  - **The stage is printed on entry.** The stage-writer PreToolUse hook now prints a plain-English `▸ paqad · <stage>` line (via Claude's user-visible `systemMessage` channel) the first time a change enters each stage, so the developer sees the workflow running even when the model forgets to say it. Claude-only (the sole pre-mutation host); idempotent per change.
+  - **The Decision Pause Contract self-arms.** An opt-in pre-mutation capability reads the recent prompt from the turn transcript and, on a high-confidence create-vs-reuse fork with no decision pending or already made, mints ONE pending decision packet so the existing decision-pause gate blocks the _next_ edit — closing the "nothing mints the packet" gap. **Off by default** (enable with `PAQAD_DECISION_SELFARM` or a local `.config` `decision_selfarm` line); it only mints, never blocks, so the current edit is never interrupted.
+
+  The kernel gate now threads an optional tool/turn payload (edit target, transcript path, session id) to each capability; every existing capability ignores it. Coverage stays at the 95% branch gate.
+
+## 1.35.0
+
+### Minor Changes
+
+- fa171f3: Capability Kernel (F1–F8): unify the five enforcement contracts under one framework-owned, versioned kernel.
+
+  - Stage-evidence, rule-scripts, decision-pause, delivery, and narration are now rows in one `CapabilityRegistry`, each with per-capability policy/record schema versions.
+  - Evidence consumers (the dashboard and `paqad-ai audit export` / SIEM) read the always-on #249 session-ledger; the delivery, rule-compliance, and decision stores fold onto it via a shared project-ledger helper (dual-sink — the operational files the engine needs are preserved, so they stay in lockstep).
+  - `audit export` now unions the session-ledger doc types into the OCSF/ECS/CEF fold-view as a new `session` event kind, graded for a SOC.
+  - Compat spine (D2): the bless stamps a per-capability version vector into `capability-lock.json`; an install older than the schema a project was blessed under refuses to enforce cleanly (non-blocking) rather than misread a newer format.
+  - Removed the dead LaneRunner/SliceExecutor orchestration engine, its 19 pipeline phases, and the `paqad-ai plan resume` CLI command (they backed the trigger-independent engine that never ran at session time).
+
 ## 1.34.0
 
 ### Minor Changes

package/README.md

@@ -20,7 +20,7 @@
 [![Node](https://img.shields.io/node/v/paqad-ai.svg?style=flat-square)](https://nodejs.org)
 [![website](https://img.shields.io/badge/website-paqad.ai-0A7A5C?style=flat-square)](https://paqad.ai)
 
-Works with Claude Code, Cursor, GitHub Copilot, Gemini, Codex, Windsurf, Junie, Continue, Antigravity, and Aider.
+Works with Claude Code, Cursor, GitHub Copilot, Gemini, Codex, Windsurf, Junie, Continue, Antigravity, Aider, and JetBrains AI Assistant.
 
 </div>
 
@@ -140,18 +140,19 @@ Three steps turn AI coding from a prompt you hope works into a pipeline you can
 
 Run `paqad-ai onboard` one time. It writes a thin, native config file for each tool you use, and they all point at the same shared instruction bundle. Change a rule in one place, and every agent sees it. Switch tools next month without losing anything.
 
-| Tool                   | What it creates                                          |
-| ---------------------- | -------------------------------------------------------- |
-| **Claude Code**        | `CLAUDE.md` and MCP config                               |
-| **Codex CLI**          | `AGENTS.md` and MCP config                               |
-| **Google Antigravity** | `ANTIGRAVITY.md` and MCP config                          |
-| **Gemini CLI**         | `GEMINI.md` and MCP config                               |
-| **Junie**              | `.junie/AGENTS.md` and `.junie/mcp/mcp.json`             |
-| **Cursor**             | `.cursor/rules/paqad.mdc` plus MCP, skills, agents       |
-| **GitHub Copilot**     | `.github/copilot-instructions.md` and `.vscode/mcp.json` |
-| **Windsurf**           | `.windsurfrules` plus MCP, skills, agents                |
-| **Continue**           | `.continue/rules/paqad.md` plus MCP, prompts             |
-| **Aider**              | `CONVENTIONS.md`                                         |
+| Tool                       | What it creates                                          |
+| -------------------------- | -------------------------------------------------------- |
+| **Claude Code**            | `CLAUDE.md` and MCP config                               |
+| **Codex CLI**              | `AGENTS.md` and MCP config                               |
+| **Google Antigravity**     | `ANTIGRAVITY.md` and MCP config                          |
+| **Gemini CLI**             | `GEMINI.md` and MCP config                               |
+| **Junie**                  | `.junie/AGENTS.md` and `.junie/mcp/mcp.json`             |
+| **Cursor**                 | `.cursor/rules/paqad.mdc` plus MCP, skills, agents       |
+| **GitHub Copilot**         | `.github/copilot-instructions.md` and `.vscode/mcp.json` |
+| **Windsurf**               | `.windsurfrules` plus MCP, skills, agents                |
+| **Continue**               | `.continue/rules/paqad.md` plus MCP, prompts             |
+| **Aider**                  | `CONVENTIONS.md`                                         |
+| **JetBrains AI Assistant** | `.aiassistant/rules/guidelines.md`                       |
 
 Pick one or pick all. Onboarding is multi-select.