release: v2.3.0 — first fully stable release

Bump version to 2.3.0 across package.json, openclaw.plugin.json.
Add comprehensive CHANGELOG documenting all bugs found and fixed
during the v2.x development cycle.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

CHANGELOG.md

@@ -1,75 +1,73 @@
 # Changelog
 
+## v2.3.0 — 2026-03-16
+
+### Summary
+First **fully stable** release. All previous versions had critical bugs causing silent failure on gateway deployments.
+
+### Fixed
+- **Workspace resolution** (`_ctx.workspaceDir`) — Hooks now read workspace path from the hook context parameter instead of `api.runtime.workspaceDir` (which was always `undefined` in gateway mode). This was the root cause of the plugin silently doing nothing on all gateway/Telegram/Discord deployments.
+- **Channel metadata stripping** — User messages from Telegram/Discord are cleaned of `Conversation info (untrusted metadata)` prefixes before saving, preventing garbage in the state file.
+- **Recovery death spiral** — Short conversations (< 2 real user messages) no longer overwrite existing meaningful state. Previously: tell secret → `/new` → model ignores injected context → "I don't remember" → `agent_end` overwrites secret with failure → permanent data loss.
+- **State staleness** — Removed "skip if existing state is meaningful" guard that prevented state from ever updating after the first write.
+
+### Debugging timeline
+1. Hooks not firing → `api.runtime.workspaceDir` was `undefined` → fix: use `_ctx.workspaceDir`
+2. State file full of Telegram metadata → fix: strip `Conversation info` / `Sender` prefixes
+3. State never updating → "Existing state is meaningful, skipping" → fix: always update
+4. Good state overwritten by bad → fix: require ≥ 2 real user messages to overwrite
+
+### Tested on
+- macOS: GPT-5.4 (openai-codex) ✅
+- Oracle Cloud Linux: MiniMax M2.5 ✅
+- Multiple OpenClaw agents via Telegram ✅
+
+### Known limitations
+- State extraction uses last user/assistant message only (not full summary)
+- Models that ignore `prependSystemContext` won't surface recovered state (state file is preserved, not overwritten). Adding recovery instructions to BOOTSTRAP.md helps.
+
+---
+
+## v2.2.0 — 2026-03-16
+
+### Added
+- `package.json` with npm-style metadata and `openclaw.type: "plugin"` provenance
+- `plugins.allow` auto-configuration in `post-install.sh`
+- GitHub release with comparison table vs vector-DB plugins
+
+### Fixed
+- "plugins.allow is empty" gateway warning
+- "loaded without install/load-path provenance" gateway warning
+
+---
+
 ## v2.1.0 — 2026-03-16
 
-### Summary
-Clean install experience: eliminates gateway provenance warnings for fresh installations.
-
 ### Added
-- `package.json` — proper npm-style metadata with `openclaw.type: "plugin"` provenance field
-- `plugins.allow` auto-configuration in `post-install.sh` — registers plugin in OpenClaw trust list
+- `openclaw.plugin.json` manifest with `configSchema`
+- `scripts/post-install.sh` one-command installer
 
 ### Changed
-- `post-install.sh` now copies `package.json` to extensions directory
-- `openclaw.plugin.json` now includes `source` field pointing to GitHub repo
-- README updated with `plugins.allow` documentation and install step
-
-### Fixed
-- "plugins.allow is empty; discovered non-bundled plugins may auto-load" warning
-- "loaded without install/load-path provenance; treat as untracked local code" warning
+- README updated with "Why this plugin?" section
 
 ---
 
 ## v2.0.0 — 2026-03-15
 
 ### Summary
-Major architecture upgrade: from prompt-based skill to **lifecycle plugin**.
-State injection and extraction now happen automatically via OpenClaw hooks,
-making recovery model-agnostic and reliable across all providers.
+Architecture upgrade: skill (prompt-based) → **plugin** (lifecycle hooks).
 
 ### Added
-- `index.js` — plugin entry point with 5 lifecycle hooks
-- `openclaw.plugin.json` — plugin manifest with configSchema
-- Automated installer (`scripts/post-install.sh`) that handles copy + config + restart
-
-### Changed
-- Architecture: skill (prompt-based) → plugin (hook-based)
-- `before_agent_start` hook injects `CURRENT_STATE.md` into system context automatically
-- `agent_end` hook auto-extracts working state from conversation
-- `before_compaction` hook preserves state through context compression
-- `before_reset` hook archives state before `/new`
-- `session_end` hook ensures state file exists
-- README fully rewritten for plugin architecture
-- post-install.sh rewritten for plugin installation flow
+- `index.js` with 5 lifecycle hooks: `before_agent_start`, `before_compaction`, `before_reset`, `agent_end`, `session_end`
+- Automated state injection via `prependSystemContext`
+- Auto-extraction of working state from conversation
 
 ### Removed
 - Dependency on model cooperation for state read/write
-- Reliance on `skillsSnapshot` cache clearing
-
-### Validated
-- Tested with MiniMax M2.5 and GPT-5.4 on OpenClaw 2026.3.12
-- `/new` recovery: agent correctly surfaces recovered state
-- Multi-turn state: conversation context (secrets, scheduled events) persists across resets
 
 ---
 
-## v0.3.0-probe — 2026-03-13
+## v1.0.0 — 2026-03-14
 
 ### Summary
-Transition from skill-only to dual-form package (skill + lifecycle plugin probe).
-
-### Added
-- `plugin/lifecycle-prototype.ts`
-- `references/phase2-hook-validation.md`
-
-### Changed
-- Repository direction clarified: primary path is lifecycle plugin
-- ContextEngine documented as future option, not v1 default
-
-### Validated
-- Startup continuity injection on resident subagents
-- Hook wiring confirmed without `read` tool dependency
-
-### Known limitation
-- Skill-based approach unreliable: models can ignore recovery instructions
-- This limitation is resolved in v2.0.0 by moving to hooks
+Initial release as a skill (SKILL.md only). Required model to voluntarily read/write `CURRENT_STATE.md`. Did not work reliably with weaker models.