continuous-improvement 3.11.0 → 3.12.3

package/.claude-plugin/marketplace.json

@@ -7,8 +7,8 @@
   "plugins": [
     {
       "name": "continuous-improvement",
-      "description": "Stops Claude Code from skipping research, claiming 'done' without verifying, and repeating yesterday's mistakes. The 7 Laws of AI Agent Discipline — 25 bundled skills, gating hooks, the Mulahazah auto-leveling instinct engine, and a GitHub Action transcript linter.",
-      "version": "3.11.0",
+      "description": "Makes Claude Code reason harder, recall past corrections, and learn from every session so its competence compounds run over run. The Mulahazah engine turns each fix into a reusable instinct, so a lesson learned once is applied automatically next time — no re-teaching. Built on the 7 Laws of AI Agent Discipline (research, plan, verify, reflect, learn) and shipped as 25 bundled skills, instinct-aware hooks, an MCP toolset for recall and reflection, and a GitHub Action transcript linter that feeds real work history back into sharper instincts.",
+      "version": "3.12.3",
       "source": "./plugins/continuous-improvement",
       "author": {
         "name": "naimkatiman"
@@ -53,7 +53,7 @@
     },
     {
       "name": "oh-my-claudecode",
-      "description": "Multi-agent orchestration system for Claude Code — 39 skills + 19 agents covering audit, plan, build, verify, ship, release, retrospective, and ops. Includes ralph (autonomous PRD loop), release, ultrawork, ultraqa, team, trace, visual-verdict, debug, deep-dive, deep-interview, autopilot, autoresearch. Vendored snapshot at third-party/oh-my-claudecode/ pinned to upstream SHA aacde3e (v4.13.6). Heavy overlap with continuous-improvement /ralph and /superpowers — pick per task.",
+      "description": "Multi-agent orchestration system for Claude Code — 38 skills + 19 agents covering audit, plan, build, verify, ship, release, retrospective, and ops. Includes ralph (autonomous PRD loop), release, ultrawork, ultraqa, team, trace, visual-verdict, debug, deep-dive, deep-interview, autopilot, autoresearch. Vendored snapshot at third-party/oh-my-claudecode/ pinned to upstream SHA aacde3e (v4.13.6). Heavy overlap with continuous-improvement /ralph and /superpowers — pick per task.",
       "version": "4.13.6",
       "source": "./third-party/oh-my-claudecode",
       "author": {

package/CHANGELOG.md

@@ -6,6 +6,40 @@ All notable changes to this skill are documented here.
 
 ## [Unreleased]
 
+## [3.12.3] — 2026-06-08
+
+### Fixed
+
+- **Completed the OIDC trusted-publishing pipeline by restoring setup-node `registry-url`** — #211 had removed it to kill the dummy-token `.npmrc`, but that left npm with no registry anchor for the OIDC token exchange (`ENEEDAUTH` on 3.12.1/3.12.2). With `registry-url` back, Node 22 (#213), npm ≥ 11.5.1, and a configured trusted publisher, npm mints and uses the OIDC token. **3.12.3 is the first version actually published to npm** carrying the 3.12.0 changes (#198 positioning, #199 recall hook, #202 OIDC, #203 doc-drift, #204 tool-count); 3.12.0–3.12.2 were never published. (#215)
+
+## [3.12.2] — 2026-06-08
+
+### Fixed
+
+- **Release workflow now runs on Node 22 so OIDC trusted publishing actually authenticates** — npm trusted publishing requires npm ≥ 11.5.1 **and Node ≥ 22.14.0** ([docs.npmjs.com/trusted-publishers](https://docs.npmjs.com/trusted-publishers)). `release.yml` ran on Node 20, below the floor, so the OIDC token exchange never engaged: the `v3.12.0` publish fell back to setup-node's dummy `.npmrc` token (`E404`) and `v3.12.1` had no auth at all (`ENEEDAUTH`). Bumping `setup-node` to Node 22 (plus #211 dropping the token-bearing `.npmrc`) completes the OIDC pipeline. **3.12.2 is the first npm artifact to carry the 3.12.0 changes** (#198 positioning, #199 recall hook, #202 OIDC pipeline, #203 doc-drift, #204 tool-count invariant); `3.12.0` and `3.12.1` were never published to npm. (#211, #213)
+
+## [3.12.1] — 2026-06-08
+
+### Fixed
+
+- **npm release pipeline now authenticates via OIDC instead of a dummy token** — the `v3.12.0` tag publish failed with `E404` on the registry PUT: `actions/setup-node` with `registry-url` wrote an `.npmrc` carrying `_authToken=${NODE_AUTH_TOKEN}` set to its placeholder value, so npm used that dummy token for the publish instead of OIDC trusted publishing (provenance still signed, since the sigstore OIDC token is separate). Dropping `registry-url` from `release.yml` lets `npm publish --provenance` fall back to OIDC as intended. **3.12.1 is the first npm artifact to carry the 3.12.0 changes** (intelligence-amplifier positioning #198, recall-briefing hook #199, OIDC pipeline #202, doc-drift fixes #203, `verify:tool-count` invariant #204); the `3.12.0` version was never published to npm. (#211)
+
+## [3.12.0] — 2026-06-07
+
+### Added
+
+- **Proactive recall-briefing hook (opt-in episodic memory)** — `hooks/recall-briefing.mjs` runs on `UserPromptSubmit`, queries the BM25 observation index in-process, and surfaces the most relevant prior corrections as a briefing before the agent acts. First capability increment of the intelligence-amplifier reframe: a lesson learned once is recalled automatically on the next related prompt instead of being re-taught. Opt-in — disabled unless wired into the hook config.
+- **`verify:tool-count` content invariant (12th in `verify:all`)** — `bin/check-tool-count.mjs` pins MCP tool-count claims in docs and source to the generated `plugins/{expert,beginner}.json` `tools[].length`, so a count can't drift when a tool is added — the gap that let "12 tools" pass `verify:all` green in the 2026-06-07 audit. Pins four claims across `docs/skills.md`, `README.md`, `QUICKSTART.md`, and the `mcp-server` banner. Follow-up to #203 (PR #204).
+
+### Changed
+
+- **Positioning reframed from "seatbelt" to intelligence amplifier across every user-facing surface** — the 7 Laws of AI Agent Discipline keep their names but each is reframed from a restriction into a capability the agent gains. README h1 "A seatbelt for Claude Code" → "Claude Code that gets sharper every session"; landing page title/hero/CTA, `SHARED_PLUGIN_DESCRIPTION` (the source of truth that propagates to `package.json` + all generated manifests + `llms.txt`), SKILL.md, and CONTRIBUTING.md all reframed from "Stops Claude Code from…" loss-framing to capability-led copy. Honesty held: recall stays lexical/BM25, instincts decay, GateGuard still blocks (framed as forced grounding) — no over-claiming. Plan: `docs/plans/2026-06-07-intelligence-amplifier-reframe.md`.
+- **npm release now publishes via OIDC trusted publishing instead of `NPM_TOKEN`** — the account enforces 2FA-on-writes, which makes long-lived granular tokens hit EOTP in CI. `release.yml` upgrades npm to ≥ 11.5.1 and publishes with `npm publish --access public --provenance` (signed provenance from the OIDC claims); the `NPM_TOKEN` secret dependency is gone. See `docs/RELEASING.md` for the one-time trusted-publisher setup.
+
+### Fixed
+
+- **Doc/count drift surfaced by a post-merge audit** — corrected prose/count claims that drifted from the v3.11.0 implementation and that no `verify:all` invariant covers: MCP expert surface is 18 tools (not 12); the CONTRIBUTING release checklist now defers to `docs/RELEASING.md` instead of describing the retired manual publish flow; `agents/README.md` repointed off a non-existent reference doc; the oh-my-claudecode vendored snapshot count corrected to 38 skills; `harvest` observe-path corrected to `instincts/bin/observe.mjs`; and reports/update-card test counts synced to 793.
+
 ## [3.11.0] — 2026-06-07
 
 ### Added

package/LICENSE

@@ -1,21 +1,21 @@
-MIT License
-
-Copyright (c) 2026 Naim Katiman
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
+MIT License
+
+Copyright (c) 2026 Naim Katiman
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/QUICKSTART.md

@@ -130,7 +130,7 @@ Corrections drop instinct confidence. Unused instincts decay. The system self-co
 
 ## Expert (npx) — only if you want MCP, hooks, or instinct packs
 
-The Beginner path above is enough for most users. Pick this only if you want the MCP tools (12 of them, including `ci_plan_init` / `ci_plan_status` for `task_plan.md`-style planning), the session hooks that feed Mulahazah, or the starter instinct packs.
+The Beginner path above is enough for most users. Pick this only if you want the MCP tools (18 of them, including `ci_plan_init` / `ci_plan_status` for `task_plan.md`-style planning), the session hooks that feed Mulahazah, or the starter instinct packs.
 
 Do not run both paths against the same `~/.claude/` — that produces duplicated state. Pick one and stick with it.
 

package/README.md

@@ -4,14 +4,14 @@
   <img src="assets/combined.gif" alt="Before vs After — The 7 Laws of AI Agent Discipline" width="700" />
 </p>
 
-<h1 align="center">A seatbelt for Claude Code</h1>
+<h1 align="center">Claude Code that gets sharper every session</h1>
 
 <p align="center">
-  <b>Research first. Edit safely. Verify before done. Remember what worked.</b>
+  <b>Reasons deeper. Recalls what it already solved. Verifies before "done". Keeps every lesson.</b>
 </p>
 
 <p align="center">
-  <i>The 7 Laws of AI Agent Discipline — runtime hooks, enforcement skills, and project memory.</i>
+  <i>The 7 Laws of AI Agent Discipline — runtime hooks, instinct memory, and skills that compound what it learns.</i>
 </p>
 
 <p align="center">
@@ -26,13 +26,13 @@
   <b>New here?</b> → <a href="QUICKSTART.md">QUICKSTART.md</a> (2 minutes) · <a href="https://continuous-improvement.dev">continuous-improvement.dev</a>
 </p>
 
-> **What this is *not*:** a prompt template, a `CLAUDE.md`, or a vibes-based reminder. It is a runtime hook (`hooks/gateguard.mjs`) plus a bundled skill set that physically blocks `Edit` / `Write` / destructive `Bash` until the agent has done the work.
+> **What this is *not*:** a prompt template, a `CLAUDE.md`, or a vibes-based reminder. It is a runtime hook (`hooks/gateguard.mjs`) plus a bundled skill set that makes the agent ground every change in real facts — it physically blocks `Edit` / `Write` / destructive `Bash` until the investigation is done, so edits land on understanding instead of guesses.
 
 ---
 
 ## What this does
 
-Claude Code is powerful but skips the boring discipline: it edits before reading, guesses instead of checking, stacks five concerns into one commit, and says "done" without running tests. Continuous Improvement adds three layers that stop that:
+Claude Code is powerful but leaves intelligence on the table: it edits before reading, guesses instead of checking, stacks five concerns into one commit, and says "done" without running tests. Continuous Improvement adds three layers that make it sharper:
 
 1. **Before an edit** — [`gateguard`](skills/gateguard.md) ships as a `PreToolUse` hook (`hooks/gateguard.mjs`) that physically blocks `Edit` / `Write` / `MultiEdit` and destructive `Bash` until the agent presents a fact-list investigation.
 2. **During work** — bundled skills enforce planning, one-thing-at-a-time execution, TDD ([`tdd-workflow`](skills/tdd-workflow.md)), and a six-phase verification ladder ([`verification-loop`](skills/verification-loop.md)) before "done".
@@ -52,7 +52,7 @@ With Continuous Improvement, the same prompt is forced through the gate:
 
 > `gateguard` blocks the first `Edit` until Claude presents a fact list. Claude reads `useAuth.ts`, finds the existing `redirectAfterLogin` helper, traces *why* the redirect loops (a stale `from` query param), and edits one line in one file. `verification-loop` runs the tests. The reply names the file, the line, the cause.
 
-Same agent. Same model. Different discipline.
+Same agent. Same model. Different intelligence.
 
 ---
 
@@ -138,7 +138,7 @@ V1 honest limitations: the runtime gate is honor-system once the agent flips `_g
 
 ### Expert — adds MCP server, observation hooks, and instinct packs
 
-Pick this if you want the MCP tools (12 of them, including `ci_plan_init` / `ci_plan_status` for `task_plan.md`-style planning), the session hooks that feed Mulahazah, and starter packs.
+Pick this if you want the MCP tools (18 of them, including `ci_plan_init` / `ci_plan_status` for `task_plan.md`-style planning), the session hooks that feed Mulahazah, and starter packs.
 
 Preconditions: Node 18 / 20 / 22, plus bash on Windows (Git Bash or WSL — `hooks/observe.sh` is a bash script and silently no-ops without it). **`jq` is no longer required**: as of v3.6.0, `observe.sh` prefers the Node observer (`bin/observe.mjs`) which writes the rich event schema natively without external dependencies. The bash thin-schema path is kept as a two-phase shim, so legacy installs that have not re-run `npx continuous-improvement install` since v3.5.x will still degrade silently without `jq` (`winget install jqlang.jq` on Windows, `brew install jq` on macOS, `apt install jq` on Debian/Ubuntu) — re-running the installer is the cleaner fix and removes the dependency entirely. See [CHANGELOG.md](CHANGELOG.md) `[3.6.0]` for the migration details.
 
@@ -171,20 +171,21 @@ The framework has documented operator-level modes that change hook behavior with
 |---|---|---|
 | `CLAUDE_THREE_SECTION_CLOSE_DISABLED=1` | `three-section-close.mjs` short-circuits before any enforcement or telemetry. Use when end-of-turn reflection should run as internal thinking rather than visible "What has been done / What is next / Recommendation" sections. Public default unchanged — the rule still fires for everyone else. | bash/zsh: `export CLAUDE_THREE_SECTION_CLOSE_DISABLED=1` in `~/.bashrc` / `~/.zshrc`. PowerShell: `$env:CLAUDE_THREE_SECTION_CLOSE_DISABLED=1` (session) or `[Environment]::SetEnvironmentVariable('CLAUDE_THREE_SECTION_CLOSE_DISABLED','1','User')` (persistent). |
 | `CLAUDE_GOAL_DRIFT_GATE` | `goal-drift-stop.mjs` (a `Stop` hook) scores each turn's activity against the `## Goal` in `task_plan.md` and acts on drift. `warn` (default) prints a one-line stderr notice and never blocks; `block` re-prompts a substantive wrap-up that has drifted off-goal so the goal gates the close; `off` disables it. Reads the same observation feed as Mulahazah; fails open on any error. | bash/zsh: `export CLAUDE_GOAL_DRIFT_GATE=block` in `~/.bashrc` / `~/.zshrc`. PowerShell: `$env:CLAUDE_GOAL_DRIFT_GATE='block'` (session) or `[Environment]::SetEnvironmentVariable('CLAUDE_GOAL_DRIFT_GATE','block','User')` (persistent). |
+| `CLAUDE_RECALL_BRIEFING=1` | `hooks/recall-briefing.mjs` (a UserPromptSubmit hook) makes episodic memory proactive: on the first substantive prompt of a session it searches this project's past observations (BM25) and injects a one-time `<system-reminder>` with the most relevant prior activity, so the agent reuses a past fix instead of re-deriving it. Opt-in and default off; it is an amplifier, never a gate — it cannot block a prompt and fails open. The `ci_recall` MCP tool stays available for explicit, deeper searches. | bash/zsh: `export CLAUDE_RECALL_BRIEFING=1` in `~/.bashrc` / `~/.zshrc`. PowerShell: `$env:CLAUDE_RECALL_BRIEFING=1` (session) or `[Environment]::SetEnvironmentVariable('CLAUDE_RECALL_BRIEFING','1','User')` (persistent). |
 
 ---
 
 ## The 7 Laws
 
-| # | Law | Without it, agents... |
+| # | Law | What it gives the agent |
 |---|-----|----------------------|
-| 1 | **Research Before Executing** | reinvent what already exists |
-| 2 | **Plan Is Sacred** | scope-creep and overbuild |
-| 3 | **One Thing at a Time** | stack untested changes |
-| 4 | **Verify Before Reporting** | lie about being "done" |
-| 5 | **Reflect After Sessions** | repeat the same failures |
-| 6 | **Iterate One Change** | debug 5 changes at once |
-| 7 | **Learn From Every Session** | lose knowledge when context ends |
+| 1 | **Research Before Executing** | builds on what already exists instead of reinventing it |
+| 2 | **Plan Is Sacred** | lands effort on the goal with success defined up front |
+| 3 | **One Thing at a Time** | ships each change on a known-good base |
+| 4 | **Verify Before Reporting** | backs every "done" with evidence you can trust |
+| 5 | **Reflect After Sessions** | turns each session into a captured lesson |
+| 6 | **Iterate One Change** | keeps debugging isolated and the signal clean |
+| 7 | **Learn From Every Session** | compounds knowledge so next week's agent is sharper |
 
 ```
 Research -> Plan -> Execute (one thing) -> Verify -> Reflect -> Learn -> Iterate

package/SKILL.md

@@ -6,7 +6,7 @@ description: "Install structured self-improvement loops with instinct-based lear
 
 # continuous-improvement
 
-You follow the continuous-improvement framework. These 7 laws govern all your work.
+You follow the continuous-improvement framework. These 7 laws make every task sharper — each is a capability that compounds: research deeper, plan tighter, verify with evidence, reflect, and learn so the same lesson is never re-taught.
 
 ## Law 1: Research Before Executing
 

package/bin/analyze.sh

@@ -1,161 +1,161 @@
-#!/usr/bin/env bash
-# analyze.sh — Read observations.jsonl, detect patterns, create instinct YAML files
-# Runs via /continuous-improvement command. Uses claude CLI with Haiku for cost-efficient analysis.
-
-set -euo pipefail
-
-INSTINCTS_DIR="${HOME}/.claude/instincts"
-
-# ---------------------------------------------------------------------------
-# Detect project
-# ---------------------------------------------------------------------------
-PROJECT_ROOT=""
-if [[ -n "${CLAUDE_PROJECT_DIR:-}" && -d "${CLAUDE_PROJECT_DIR}" ]]; then
-  PROJECT_ROOT="${CLAUDE_PROJECT_DIR}"
-fi
-if [[ -z "$PROJECT_ROOT" ]]; then
-  PROJECT_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || true)"
-fi
-if [[ -z "$PROJECT_ROOT" ]]; then
-  PROJECT_ROOT="global"
-fi
-
-# Use shasum on macOS, sha256sum on Linux
-if command -v sha256sum &>/dev/null; then
-  PROJECT_HASH="$(printf '%s' "$PROJECT_ROOT" | sha256sum | cut -c1-12)"
-elif command -v shasum &>/dev/null; then
-  PROJECT_HASH="$(printf '%s' "$PROJECT_ROOT" | shasum -a 256 | cut -c1-12)"
-else
-  # Fallback: use project root basename if no hashing available
-  PROJECT_HASH="$(basename "$PROJECT_ROOT" | tr -cd 'a-zA-Z0-9' | cut -c1-12)"
-fi
-PROJECT_NAME="$(basename "${PROJECT_ROOT%.git}")"
-PROJECT_DIR="${INSTINCTS_DIR}/${PROJECT_HASH}"
-OBS_FILE="${PROJECT_DIR}/observations.jsonl"
-
-# ---------------------------------------------------------------------------
-# Check observations exist
-# ---------------------------------------------------------------------------
-if [[ ! -f "$OBS_FILE" ]] || [[ ! -s "$OBS_FILE" ]]; then
-  echo "No observations found at ${OBS_FILE}"
-  echo "Use Claude Code with hooks installed to generate observations."
-  exit 0
-fi
-
-OBS_COUNT=$(wc -l < "$OBS_FILE")
-echo "Found ${OBS_COUNT} observations in ${OBS_FILE}"
-
-if (( OBS_COUNT < 20 )); then
-  echo "Need at least 20 observations for meaningful analysis. Keep using Claude Code (${OBS_COUNT}/20)."
-  exit 0
-fi
-
-# ---------------------------------------------------------------------------
-# Read existing instincts to avoid duplicates
-# ---------------------------------------------------------------------------
-EXISTING_INSTINCTS=""
-for f in "${PROJECT_DIR}"/*.yaml "${INSTINCTS_DIR}/global"/*.yaml; do
-  [[ -f "$f" ]] && EXISTING_INSTINCTS="${EXISTING_INSTINCTS}$(cat "$f")"$'\n'
-done
-
-# ---------------------------------------------------------------------------
-# Take last 500 observations
-# ---------------------------------------------------------------------------
-RECENT_OBS=$(tail -500 "$OBS_FILE")
-
-# ---------------------------------------------------------------------------
-# Build analysis prompt
-# ---------------------------------------------------------------------------
-ANALYSIS_PROMPT="Analyze these Claude Code session observations and extract behavioral patterns as instinct YAML files.
-
-OBSERVATIONS (JSONL — each line is a tool call):
-${RECENT_OBS}
-
-EXISTING INSTINCTS (already learned — do NOT duplicate these):
-${EXISTING_INSTINCTS}
-
-YOUR TASK:
-1. Look for REPEATED PATTERNS — same tool sequence used 3+ times
-2. Look for ERROR-THEN-FIX sequences — tool fails, next tools fix it
-3. Look for TOOL PREFERENCES — one tool consistently chosen over alternatives
-4. Look for WORKFLOW PATTERNS — consistent ordering of operations
-5. Look for USER CORRECTIONS — user says no/stop/don't after an action
-
-OUTPUT FORMAT — output ONLY new instincts as YAML blocks, separated by ---:
-
-id: descriptive-kebab-case-id
-trigger: \"when [specific situation]\"
-confidence: 0.5
-domain: workflow|tooling|testing|patterns
-source: observation
-scope: project
-project_id: ${PROJECT_HASH}
-created: \"$(date -u +%Y-%m-%d)\"
-last_seen: \"$(date -u +%Y-%m-%d)\"
-observation_count: [number of times pattern was seen]
----
-[One sentence describing the specific actionable behavior]
-
-Rules:
-- Only create instincts for patterns seen 3+ times
-- Start confidence at 0.5 (suggest level)
-- Be specific and actionable (not vague advice)
-- Different from existing instincts listed above
-
-If no new patterns are found, output exactly: NO_NEW_PATTERNS
-
-Output ONLY the YAML blocks or NO_NEW_PATTERNS. No explanation, no preamble."
-
-# ---------------------------------------------------------------------------
-# Run analysis with Haiku
-# ---------------------------------------------------------------------------
-echo "Analyzing patterns with Haiku..."
-RESULT=$(echo "$ANALYSIS_PROMPT" | claude --model haiku --print -p - 2>/dev/null) || {
-  echo "Analysis failed — claude CLI error. Try running manually."
-  exit 1
-}
-
-if [[ "$RESULT" == "NO_NEW_PATTERNS" ]] || [[ -z "$RESULT" ]]; then
-  echo "No new patterns detected yet. Keep using Claude Code — patterns emerge over time."
-  exit 0
-fi
-
-# ---------------------------------------------------------------------------
-# Write instinct YAML files
-# ---------------------------------------------------------------------------
-mkdir -p "$PROJECT_DIR"
-
-NEW_COUNT=0
-INSTINCT_ID=""
-BLOCK=""
-
-while IFS= read -r line; do
-  if [[ "$line" == "---" ]] && [[ -n "$BLOCK" ]]; then
-    if [[ -n "$INSTINCT_ID" ]]; then
-      DEST="${PROJECT_DIR}/${INSTINCT_ID}.yaml"
-      printf '%s\n' "$BLOCK" > "$DEST"
-      echo "  + ${INSTINCT_ID} -> ${DEST}"
-      NEW_COUNT=$((NEW_COUNT + 1))
-    fi
-    INSTINCT_ID=""
-    BLOCK=""
-  else
-    BLOCK="${BLOCK}${line}"$'\n'
-    if [[ "$line" =~ ^id:\ (.+) ]]; then
-      INSTINCT_ID="${BASH_REMATCH[1]}"
-      INSTINCT_ID="${INSTINCT_ID//\"/}"
-      INSTINCT_ID="${INSTINCT_ID//\'/}"
-    fi
-  fi
-done <<< "$RESULT"
-
-# Handle last block
-if [[ -n "$INSTINCT_ID" ]] && [[ -n "$BLOCK" ]]; then
-  DEST="${PROJECT_DIR}/${INSTINCT_ID}.yaml"
-  printf '%s\n' "$BLOCK" > "$DEST"
-  echo "  + ${INSTINCT_ID} -> ${DEST}"
-  NEW_COUNT=$((NEW_COUNT + 1))
-fi
-
-echo ""
-echo "Created ${NEW_COUNT} new instinct(s) in ${PROJECT_DIR}/"
+#!/usr/bin/env bash
+# analyze.sh — Read observations.jsonl, detect patterns, create instinct YAML files
+# Runs via /continuous-improvement command. Uses claude CLI with Haiku for cost-efficient analysis.
+
+set -euo pipefail
+
+INSTINCTS_DIR="${HOME}/.claude/instincts"
+
+# ---------------------------------------------------------------------------
+# Detect project
+# ---------------------------------------------------------------------------
+PROJECT_ROOT=""
+if [[ -n "${CLAUDE_PROJECT_DIR:-}" && -d "${CLAUDE_PROJECT_DIR}" ]]; then
+  PROJECT_ROOT="${CLAUDE_PROJECT_DIR}"
+fi
+if [[ -z "$PROJECT_ROOT" ]]; then
+  PROJECT_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || true)"
+fi
+if [[ -z "$PROJECT_ROOT" ]]; then
+  PROJECT_ROOT="global"
+fi
+
+# Use shasum on macOS, sha256sum on Linux
+if command -v sha256sum &>/dev/null; then
+  PROJECT_HASH="$(printf '%s' "$PROJECT_ROOT" | sha256sum | cut -c1-12)"
+elif command -v shasum &>/dev/null; then
+  PROJECT_HASH="$(printf '%s' "$PROJECT_ROOT" | shasum -a 256 | cut -c1-12)"
+else
+  # Fallback: use project root basename if no hashing available
+  PROJECT_HASH="$(basename "$PROJECT_ROOT" | tr -cd 'a-zA-Z0-9' | cut -c1-12)"
+fi
+PROJECT_NAME="$(basename "${PROJECT_ROOT%.git}")"
+PROJECT_DIR="${INSTINCTS_DIR}/${PROJECT_HASH}"
+OBS_FILE="${PROJECT_DIR}/observations.jsonl"
+
+# ---------------------------------------------------------------------------
+# Check observations exist
+# ---------------------------------------------------------------------------
+if [[ ! -f "$OBS_FILE" ]] || [[ ! -s "$OBS_FILE" ]]; then
+  echo "No observations found at ${OBS_FILE}"
+  echo "Use Claude Code with hooks installed to generate observations."
+  exit 0
+fi
+
+OBS_COUNT=$(wc -l < "$OBS_FILE")
+echo "Found ${OBS_COUNT} observations in ${OBS_FILE}"
+
+if (( OBS_COUNT < 20 )); then
+  echo "Need at least 20 observations for meaningful analysis. Keep using Claude Code (${OBS_COUNT}/20)."
+  exit 0
+fi
+
+# ---------------------------------------------------------------------------
+# Read existing instincts to avoid duplicates
+# ---------------------------------------------------------------------------
+EXISTING_INSTINCTS=""
+for f in "${PROJECT_DIR}"/*.yaml "${INSTINCTS_DIR}/global"/*.yaml; do
+  [[ -f "$f" ]] && EXISTING_INSTINCTS="${EXISTING_INSTINCTS}$(cat "$f")"$'\n'
+done
+
+# ---------------------------------------------------------------------------
+# Take last 500 observations
+# ---------------------------------------------------------------------------
+RECENT_OBS=$(tail -500 "$OBS_FILE")
+
+# ---------------------------------------------------------------------------
+# Build analysis prompt
+# ---------------------------------------------------------------------------
+ANALYSIS_PROMPT="Analyze these Claude Code session observations and extract behavioral patterns as instinct YAML files.
+
+OBSERVATIONS (JSONL — each line is a tool call):
+${RECENT_OBS}
+
+EXISTING INSTINCTS (already learned — do NOT duplicate these):
+${EXISTING_INSTINCTS}
+
+YOUR TASK:
+1. Look for REPEATED PATTERNS — same tool sequence used 3+ times
+2. Look for ERROR-THEN-FIX sequences — tool fails, next tools fix it
+3. Look for TOOL PREFERENCES — one tool consistently chosen over alternatives
+4. Look for WORKFLOW PATTERNS — consistent ordering of operations
+5. Look for USER CORRECTIONS — user says no/stop/don't after an action
+
+OUTPUT FORMAT — output ONLY new instincts as YAML blocks, separated by ---:
+
+id: descriptive-kebab-case-id
+trigger: \"when [specific situation]\"
+confidence: 0.5
+domain: workflow|tooling|testing|patterns
+source: observation
+scope: project
+project_id: ${PROJECT_HASH}
+created: \"$(date -u +%Y-%m-%d)\"
+last_seen: \"$(date -u +%Y-%m-%d)\"
+observation_count: [number of times pattern was seen]
+---
+[One sentence describing the specific actionable behavior]
+
+Rules:
+- Only create instincts for patterns seen 3+ times
+- Start confidence at 0.5 (suggest level)
+- Be specific and actionable (not vague advice)
+- Different from existing instincts listed above
+
+If no new patterns are found, output exactly: NO_NEW_PATTERNS
+
+Output ONLY the YAML blocks or NO_NEW_PATTERNS. No explanation, no preamble."
+
+# ---------------------------------------------------------------------------
+# Run analysis with Haiku
+# ---------------------------------------------------------------------------
+echo "Analyzing patterns with Haiku..."
+RESULT=$(echo "$ANALYSIS_PROMPT" | claude --model haiku --print -p - 2>/dev/null) || {
+  echo "Analysis failed — claude CLI error. Try running manually."
+  exit 1
+}
+
+if [[ "$RESULT" == "NO_NEW_PATTERNS" ]] || [[ -z "$RESULT" ]]; then
+  echo "No new patterns detected yet. Keep using Claude Code — patterns emerge over time."
+  exit 0
+fi
+
+# ---------------------------------------------------------------------------
+# Write instinct YAML files
+# ---------------------------------------------------------------------------
+mkdir -p "$PROJECT_DIR"
+
+NEW_COUNT=0
+INSTINCT_ID=""
+BLOCK=""
+
+while IFS= read -r line; do
+  if [[ "$line" == "---" ]] && [[ -n "$BLOCK" ]]; then
+    if [[ -n "$INSTINCT_ID" ]]; then
+      DEST="${PROJECT_DIR}/${INSTINCT_ID}.yaml"
+      printf '%s\n' "$BLOCK" > "$DEST"
+      echo "  + ${INSTINCT_ID} -> ${DEST}"
+      NEW_COUNT=$((NEW_COUNT + 1))
+    fi
+    INSTINCT_ID=""
+    BLOCK=""
+  else
+    BLOCK="${BLOCK}${line}"$'\n'
+    if [[ "$line" =~ ^id:\ (.+) ]]; then
+      INSTINCT_ID="${BASH_REMATCH[1]}"
+      INSTINCT_ID="${INSTINCT_ID//\"/}"
+      INSTINCT_ID="${INSTINCT_ID//\'/}"
+    fi
+  fi
+done <<< "$RESULT"
+
+# Handle last block
+if [[ -n "$INSTINCT_ID" ]] && [[ -n "$BLOCK" ]]; then
+  DEST="${PROJECT_DIR}/${INSTINCT_ID}.yaml"
+  printf '%s\n' "$BLOCK" > "$DEST"
+  echo "  + ${INSTINCT_ID} -> ${DEST}"
+  NEW_COUNT=$((NEW_COUNT + 1))
+fi
+
+echo ""
+echo "Created ${NEW_COUNT} new instinct(s) in ${PROJECT_DIR}/"

package/bin/backfill.mjs

@@ -27,25 +27,25 @@ const args = process.argv.slice(2);
 const DRY_RUN = args.includes("--dry-run");
 const HELP = args.includes("--help") || args.includes("-h");
 if (HELP) {
-    console.log(`
-backfill — Tag legacy thin-schema observation rows.
-
-Usage:
-  npx continuous-improvement backfill            Tag rows in place
-  npx continuous-improvement backfill --dry-run  Report counts, write nothing
-
-What it does:
-  Walks ~/.claude/instincts/<hash>/observations.jsonl files and adds a
-  schema field ("thin" or "rich") to every row that lacks one. Thin rows
-  came from the legacy bash fallback (jq missing) and have no input_summary;
-  rich rows came from the Node observer and have full tool_input.command /
-  Edit.file_path / etc.
-
-Output:
-  - observations.jsonl rewritten in place (original kept as .bak)
-  - backfill-summary.json with per-project counts
-
-Idempotent. Always exits 0 — failures go to stderr only.
+    console.log(`
+backfill — Tag legacy thin-schema observation rows.
+
+Usage:
+  npx continuous-improvement backfill            Tag rows in place
+  npx continuous-improvement backfill --dry-run  Report counts, write nothing
+
+What it does:
+  Walks ~/.claude/instincts/<hash>/observations.jsonl files and adds a
+  schema field ("thin" or "rich") to every row that lacks one. Thin rows
+  came from the legacy bash fallback (jq missing) and have no input_summary;
+  rich rows came from the Node observer and have full tool_input.command /
+  Edit.file_path / etc.
+
+Output:
+  - observations.jsonl rewritten in place (original kept as .bak)
+  - backfill-summary.json with per-project counts
+
+Idempotent. Always exits 0 — failures go to stderr only.
 `);
     process.exit(0);
 }