docs: update README for v0.5.0 release — MOSS-grounded improvements

feat: apply MOSS-grounded self-evolution improvements to ADAM
Implements 7 improvements grounded in MOSS paper (arXiv 2605.22794): 1. Transcript capture (§3.4): context_ring buffer in adam-observe.mjs captures last 8 events around struggle signals as context_window. 2. Evidence batching (§3.1): new adam-batch.mjs pre-clusters windowed journal entries into coherent failure batches by (signal_type, cluster_key). 3. Multi-stage analysis (§3.3): SKILL.md dispatches adam agent in two stages (diagnose+plan → implement) with inter-stage validation gate. 4. Pre-apply verification (§3.4): 4-check deterministic gate before auto-apply (source entries exist, diagnosis grounded, type-evidence match, no conflicting recent proposals). 5. Auto-rollback (§3.5): new adam-rollback.mjs reverts regressed proposals detected by A/B measurement, creates regression nudges. 6. Harness self-modification (§1 Table 1): new harness_edit proposal type targeting adam's own scripts with stricter gates (confidence≥5, never auto-apply, test-suite-gated). 7. Keypoint matrix evaluation (§4.2): 5 capability dimensions (tool_selection, scope_discipline, error_recovery, first_attempt, build_reliability) scored per batch for structured evaluation. Test suite: 94 → 114 tests (20 new), all passing.
2026-06-22 02:01:44 +00:00 · 2026-05-24 11:19:15 +01:00 · 2026-05-24 11:15:32 +01:00 · 2026-05-13 19:21:59 +01:00 · 2026-05-13 02:07:11 +01:00 · 2026-05-13 02:02:02 +01:00
13 changed files with 1568 additions and 154 deletions
@@ -1,159 +1,275 @@
+<div align="center">
+
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="./assets/logo-dark.svg">
+  <img src="./assets/logo-light.svg" alt="claude-adam logo" width="128" height="128" />
+</picture>
+
 # claude-adam

-Self-improvement layer for [Claude Code](https://claude.com/claude-code) that observes friction signals during your sessions and proposes targeted improvements (new skills, memory entries, agent edits) which you can review and apply.
+**A self-improvement layer for [Claude Code](https://claude.com/claude-code).**

-## What's new
+Watches the friction in your coding sessions, clusters the signals via an LLM analyst, and proposes targeted improvements — new skills, memory entries, agent edits — that you review and apply.

- **v0.3.3** — analyst observability, A/B measurement, journal hygiene. Storage/window/exclusion split: ISO-week journal rotation with safety fuse (replaces size-based, fixes silent under-counting); per-signal sliding windows via new `adam-window.mjs` (`dead_end` 7d, `correction` 30d, reinforcement signals 60d). Error fingerprint normalization — `ECONNREFUSED` and `"Connection refused"` cluster identically. Correction corpus expanded (`wait`, `hold on`, `try again`, `different approach`); weak tokens (`no`, `actually`, `wait`) require negation co-occurrence within 8 tokens to fire — kills the `"actually, I think..."` false positive. Mandatory clustering trace + new `adam-explain.mjs --mode summary|full|json`. New `nudge` proposal type (single-session auto-apply, low blast) for repeated `dead_end`. Per-(skill, fingerprint) cooldown via `adam-cooldown.mjs` (replaces coarse per-skill gate). `task_completed` scoring: urgency dampener + reinforcement candidates. A/B effectiveness measurement on auto-applied edits (`adam-ab-measure.mjs`, 7d pre/post window). Upgrade UX overhaul: `adam-upgrade.mjs --list/--diff/--accept` + SessionStart pending-merge warning. Shared helper module `adam-utils.mjs` deduplicates journal-reading and frontmatter parsing across scripts. 87 tests (up from 30).
- **v0.3.2** — `task_completed` signal: post-task skill capture for downstream reinforcement scoring (consumed in v0.3.3).
- **v0.3.1** — code review pass: bug fixes (`errorFingerprint` no longer false-positives on `is_error: false`, archive script handles same-millisecond duplicates correctly, `tool_window` now clears on session change, nudge filters proposal filenames by pattern), prose conciseness cuts, hardened `install.sh` with curl one-liner + settings.json merge, `adam-uninstall.sh`, isolated test harness (no longer pollutes live `~/.claude/adam/` state).
- **v0.3.0** — causal diagnosis: every proposal carries a `# Diagnosis` block (Trigger/Action/Mismatch/Outcome with verbatim transcript quote) before drafting, plus optional `contradiction_flag` heuristic that vetoes auto-apply on obviously-conflicting `skill_edit` additions.
- **v0.2.1** — win signals (`correction_free_streak`, `clean_recovery`) feed `skill_edit` auto-apply under a strict gate (≤30 LOC, ≤2× byte cap, 7d cooldown, 30d blacklist on rejection).
- **v0.2.0** — actioned-entry archival via `adam-archive.mjs`; `cursor` field deprecated.
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+[![Version](https://img.shields.io/github/v/release/lukaszraczylo/claude-adam?label=version&color=blue)](https://github.com/lukaszraczylo/claude-adam/releases)
+[![Tests](https://img.shields.io/badge/tests-114%20passing-brightgreen.svg)](./adam/tests/run-tests.sh)
+[![Node](https://img.shields.io/badge/node-22%2B-339933.svg)](https://nodejs.org)
+[![Platform](https://img.shields.io/badge/platform-macOS%20%7C%20Linux-lightgrey.svg)]()

-## What it does
+</div>

-A lightweight Node.js hook (`adam-observe.mjs`) runs on `UserPromptSubmit`, `PreToolUse`, and `PostToolUse` events. It detects:
+---

-| Signal | Trigger |
-|---|---|
-| `correction` | User prompt contains "no", "stop", "wrong", "actually", etc. after a tool call |
-| `retry_loop` | Same tool + same args called 3× in a 10-event window |
-| `weak_agent` | Same subagent dispatched 2× in last 5 tool calls |
-| `tool_error_loop` | Same error fingerprint appears 3× in a 5-event ring |
-| `dead_end` | 8 PostToolUse events without a UserPromptSubmit between them |
-| `edit_churn` | Same file edited 4× in a window |
-| `build_loop` | 2× build/test/compile commands fail in same session |
-| `subagent_dispatch_pattern` | Same subagent dispatched ≥3× cumulatively |
-| `correction_free_streak` | 5 clean UserPromptSubmits in a row (no correction phrase) — feeds `skill_edit` reinforcement |
-| `clean_recovery` | 3 clean PostToolUse events after a struggle signal — feeds `skill_edit` reinforcement |
+## The story behind Adam
+
+Adam is my newborn son.
+
+Watching him over the last few months — the way he observes the world, tries something, watches what happens, adjusts, and tries again — I realised that the most powerful learning loop in nature is also one of the simplest. No grand theory. No instruction manual. Just relentless feedback and pattern recognition, applied to every waking moment.
+
+LLMs can learn the same way. Give them a hook into the real friction of your work — the corrections, the dead-ends, the moments you say *"no, try again"* — and let them propose improvements grounded in **what actually happened**. Not what they assume might help. What you actually struggled with.
+
+**claude-adam** is that loop, wired into Claude Code. It's named after Adam because the methodology is his.
+
+---
+
+## Highlights
+
+- 🔍 **Zero LLM cost at observation time.** Deterministic regex + counter detection in a Node hook. The analyst only runs when you invoke `/reflect`.
+- 📡 **11 signal types.** Friction (`correction`, `tool_error_loop`, `dead_end`, `edit_churn`, …) + reinforcement (`task_completed`, `correction_free_streak`, `clean_recovery`) + meta.
+- 🛡️ **Tight auto-apply gates.** Confidence ≥ 4, cross-session evidence, contradiction veto, per-(skill, fingerprint) cooldown. Most things queue for your manual review.
+- 📊 **A/B effectiveness measurement.** Every auto-applied edit gets a 7-day pre/post signal-count delta. If a proposed fix made things worse, the next `/reflect` says so.
+- ⏳ **Per-signal sliding windows.** Stale friction doesn't accumulate forever. `dead_end` 7d, `correction` 30d, reinforcement signals 60d.
+- 🔬 **Observable.** Every clustering decision (passed / threshold-blocked / window-filtered / contradiction-vetoed) emits a trace. `/reflect --explain` shows it.
+- 📦 **Pure Node.** Zero npm dependencies. Runs on macOS and Linux (Alpine smoke-tested).
+
+## Quick start
+
+```sh
+curl -fsSL https://raw.githubusercontent.com/lukaszraczylo/claude-adam/main/install.sh | bash
+```
+
+The installer copies files into `~/.claude/`, offers to merge ADAM's hook entries into `~/.claude/settings.json` (with a diff preview and `[y/N]` confirm), and preserves any local edits via `.adam-new` sidecar files. Pass `--yes` to skip prompts, `--dry-run` to preview.
+
+Then:
+
+```sh
+bash ~/.claude/adam/tests/run-tests.sh   # expect: 87 passed, 0 failed
+# … start a fresh Claude Code session …
+/reflect                                  # walks the proposal queue
+/reflect --explain                        # also shows the analyst's clustering trace
+```
+
+Pin a release for reproducibility:
+
+```sh
+curl -fsSL https://raw.githubusercontent.com/lukaszraczylo/claude-adam/v0.5.0/install.sh \
+  | VERSION=v0.5.0 bash
+```
+
+## How it works
+
+```mermaid
+flowchart TB
+    subgraph OBS["Observation (deterministic, in-hook, zero LLM cost)"]
+        direction LR
+        EV["Tool event /<br/>user prompt"] --> OBSERVE["adam-observe.mjs<br/><sub>regex · counters · ring buffers</sub>"]
+        OBSERVE --> JOURNAL[("journal.jsonl<br/><sub>append-only signal log</sub>")]
+    end
+
+    JOURNAL -. user runs <code>/reflect</code> .-> ANALYSIS
+
+    subgraph ANALYSIS["Analysis (LLM, only on demand)"]
+        direction TB
+        subgraph PRE["Pre-processors (deterministic)"]
+            direction LR
+            W["adam-window.mjs<br/><sub>per-signal sliding window</sub>"]
+            S["adam-score.mjs<br/><sub>task_completed dampener<br/>+ reinforcement candidates</sub>"]
+            AB["adam-ab-measure.mjs<br/><sub>7d pre/post deltas<br/>on prior auto-applies</sub>"]
+        end
+        AGENT["adam subagent<br/><sub>cluster · score · diagnose</sub>"]
+        PRE --> AGENT
+        AGENT --> PROPOSALS[("proposals/")]
+        AGENT --> TRACE[["clustering trace<br/><sub>adam-explain.mjs renders</sub>"]]
+    end
+
+    PROPOSALS --> REVIEW
+
+    subgraph REVIEW["Review + apply"]
+        direction TB
+        GATE{"auto-apply<br/>gates pass?<br/><sub>conf≥4 · low blast<br/>· cooldown cool</sub>"}
+        GATE -->|yes| APPLIED[("applied/<br/>+ ab-tracking.jsonl")]
+        GATE -->|no| QUEUE["walk-the-queue<br/><sub>approve · reject · edit</sub>"]
+        QUEUE -->|approve| APPLIED
+        QUEUE -->|reject| REJECTED[("rejected/")]
+    end
+
+    APPLIED -. measures back into .-> AB
+
+    classDef store fill:#e8f4fd,stroke:#5b9bd5,stroke-width:2px,color:#1f3a5f
+    classDef proc fill:#fff4e6,stroke:#e8a33d,stroke-width:1px,color:#5a3d0f
+    classDef trace fill:#f0e8fd,stroke:#7e5dc0,stroke-width:1px,color:#2f1e60
+    class JOURNAL,PROPOSALS,APPLIED,REJECTED store
+    class EV,OBSERVE,W,S,AB,AGENT,QUEUE proc
+    class TRACE trace
+```
+
+The observation layer is a 350-line Node hook. Pure regex, counters, ring buffers — no LLM in the hot path. Signals append one JSONL line per detection to `~/.claude/adam/journal.jsonl`.
+
+The analysis layer is an LLM subagent invoked by `/reflect`. Before the analyst runs, three deterministic pre-processors filter and enrich the journal: `adam-window.mjs` drops stale entries per per-signal age, `adam-score.mjs` computes per-session urgency dampeners + reinforcement candidates, and `adam-ab-measure.mjs` checks whether previously auto-applied edits actually reduced their originating signal.
+
+The analyst clusters signals, scores them against a deterministic rubric (see below), and emits proposal markdown files to `~/.claude/adam/proposals/`. Each proposal carries a `# Diagnosis` block (Trigger / Action / Mismatch / Outcome with a verbatim transcript quote), a `# Success criterion`, and the source journal-entry timestamps it clustered.
+
+Auto-apply runs only for low-blast types (memory entries, new skills, ephemeral nudges, reinforcement logs) backed by cross-session evidence. Everything else queues for your manual approve / reject / edit walk.
+
+## Signals
+
+| Signal | Trigger | Window* |
+|---|---|---|
+| `correction` | Strong tokens (`stop`, `wrong`, `undo`, …) OR weak tokens (`no`, `actually`, `wait`) with negation/contrast nearby | 30d |
+| `retry_loop` | Same tool + same args called 3× in a 10-event window | 14d |
+| `weak_agent` | Same subagent dispatched 2× in last 5 tool calls | 30d |
+| `tool_error_loop` | Same error fingerprint 3× in a 5-event ring (fingerprints normalised — `ECONNREFUSED` and `"Connection refused"` cluster) | 30d |
+| `dead_end` | 8 PostToolUse events without a UserPromptSubmit between them | 7d |
+| `edit_churn` | Same file edited 4× in a window | 14d |
+| `build_loop` | 2× build/test/compile commands fail in same session | 30d |
+| `subagent_dispatch_pattern` | Same subagent dispatched ≥ 3× cumulatively | 30d |
+| `correction_free_streak` | 5 clean UserPromptSubmits in a row — reinforcement input | 60d |
+| `clean_recovery` | 3 clean PostToolUse events after a struggle signal — reinforcement input | 60d |
+| `task_completed` | 5 tools / 3 kinds / 0 corrections — fed into the urgency dampener + reinforcement candidates | 60d |
+
+\* Per-signal sliding window for `/reflect` analysis. See `SIGNAL_WINDOWS_DAYS` in `adam/scripts/adam-window.mjs`.

 Detection is local, regex-based, zero LLM cost. Signals append to `~/.claude/adam/journal.jsonl`.

-When you run `/reflect`, the `adam` subagent reads the journal, clusters signals, scores them against a deterministic rubric, and emits proposal files to `~/.claude/adam/proposals/`. Auto-applied proposals only ship for low-blast types (memory, new skills) backed by cross-session evidence; everything else queues for your manual approve/reject/edit walk.
+## Auto-apply rubric

-## Why
+```
+Sum:
+2  Signal repeated ≥ 3× across ≥ 2 sessions (within signal's window)
+2  Struggle signal appearing ≥ 1× within a single session (does not stack)
+2  Transcript contains positive endorsement near related action
+1  Multi-axis cluster (≥ 2 distinct struggle types in same session)
+-1  Type-bias penalty (≥ 3 rejections, applied:rejected < 1:2)
+1  Blast radius low (memory or new isolated skill)
+ 0  Blast radius medium (new agent, new hook, edit existing skill)
+-1  Blast radius high (CLAUDE.md, settings hooks, edit agent, deletion)
+1  Surgical (one file, ≤ 50 LOC for non-skill_new; ≤ 80 LOC for skill_new)
+-3  Touches deny-list (settings.json hooks/permissions, CLAUDE.md, deletions)
+```

-LLM coding sessions reveal repeated friction the moment you stop and look. ADAM looks so you don't have to.
+Modifiers applied at scoring time:
+
+- × `dampener` from `adam-score.mjs` (0.5 / 0.75 / 1.0 based on session's `task_completed` count) — sessions that net-succeeded score lower urgency.
+
+`auto_apply_eligible` requires **all** of:
+
+- `confidence ≥ 4`
+- `blast_radius == low`
+- `type ∈ {memory, skill_new, nudge, reinforcement}` (or `skill_edit` via the win-driven gate)
+- `cross_session_evidence == true` (except `nudge`, which is single-session by design)
+- `adam-cooldown.mjs` returns `cool` for `(target_skill, proposal_fingerprint)`
+- `contradiction_flag` unset
+
+`skill_edit` additionally requires:
+
+- Win-signal evidence (`correction_free_streak` / `clean_recovery` cites target skill)
+- Diff is append-only, ≤ 30 LOC, resulting size ≤ 2× original
+- No auto-edit to same target in past 7 days (per-fingerprint cooldown)
+- No rejection-blacklist on target in past 30 days
+- `# Diagnosis` section present + structurally valid
+
+Everything else queues.
+
+## Lifecycle: from signal to permanent improvement
+
+Every proposal records the journal entry timestamps that fed its cluster (`source_entries` in frontmatter). When you apply or reject a proposal, the skill calls `adam-archive.mjs` which moves matching entries from `journal.jsonl` to `journal/actioned-<id>.jsonl`. The result:
+
+- `journal.jsonl` stays bounded by **active** observations only.
+- The next `/reflect` reads `applied/` + `rejected/` frontmatter, builds an excluded-timestamps set, and skips any leftover journal entries that were already actioned.
+- Rule changes (e.g. lowering a threshold) immediately re-evaluate the remaining active observations — no manual cursor rewind needed.
+
+Auto-applied proposals additionally append to `~/.claude/adam/ab-tracking.jsonl`. The next time `/reflect` runs (and 7+ days have passed), `adam-ab-measure.mjs` computes a pre/post delta of the originating signal count. Status: `improved` / `neutral` / `regressed` / `no_baseline` / `pending`. Regressions surface at the top of the analyst's output so a bad fix doesn't quietly persist.
+
+## Inspecting the analyst's reasoning
+
+Every `/reflect` run also writes the analyst's clustering trace to `~/.claude/adam/last-trace.txt`. The trace records, per cluster: signal type, occurrence count, sessions, which gates passed or failed, and whether the cluster produced a proposal or was skipped (with reason: `threshold` / `cross_session` / `window` / `contradiction` / `other`).
+
+```sh
+node ~/.claude/adam/scripts/adam-explain.mjs --mode summary   # SUMMARY + per-decision counts
+node ~/.claude/adam/scripts/adam-explain.mjs --mode full      # verbatim trace + rejection histogram
+node ~/.claude/adam/scripts/adam-explain.mjs --mode json      # machine-readable
+```
+
+Or pass `--explain` to `/reflect` to render the full trace inline.
+
+## What it will not do
+
+- 🚫 No background LLM spend. The analyst runs only when you invoke `/reflect`.
+- 🚫 No retroactive transcript mining beyond the journal.
+- 🚫 No hard `rm` of any artifact. Deletions are soft (`mv` to `trash/<ts>/`).
+- 🚫 No autonomous edits to `CLAUDE.md`, agents, hooks, or `settings.json` — these always queue for review regardless of confidence.
+- 🚫 No proposal that matches a previously-rejected idea (≥ 2 token overlap with rejection's `# Why`).
+- 🚫 No invented trigger phrases for new skills — every trigger comes from observed user input.

 ## Layout

 ```
 ~/.claude/
 ├── hooks/
-│   ├── adam-observe.mjs      # signal collector
-│   └── adam-nudge.mjs        # SessionStart reminder when ≥3 proposals queued
-├── agents/adam.md            # analyst subagent (system prompt + rubric)
-├── skills/adam-self-improvement/SKILL.md  # /reflect protocol
-├── commands/reflect.md       # /reflect slash command
+│   ├── adam-observe.mjs              # signal collector (UserPromptSubmit / PreToolUse / PostToolUse)
+│   └── adam-nudge.mjs                # SessionStart reminder + pending-upgrade warning
+├── agents/adam.md                    # analyst subagent (system prompt + rubric)
+├── skills/adam-self-improvement/
+│   └── SKILL.md                      # /reflect protocol
+├── commands/reflect.md               # /reflect slash command
 └── adam/
-    ├── journal.jsonl         # append-only signal log (active observations)
-    ├── journal/              # rotated daily logs + actioned-<id>.jsonl per applied/rejected proposal
-    ├── state.json            # per-session counters
-    ├── usage.json            # skill/agent invocation tallies + payload visibility counters
-    ├── proposals/            # queued, awaiting review
-    ├── applied/              # approved + auto-applied archive
-    ├── rejected/             # rejected (with reason)
-    ├── trash/                # soft-deleted artifacts (recoverable)
-    ├── scripts/              # adam-archive.mjs (called by skill on apply/reject)
-    └── tests/run-tests.sh    # 27 verification tests (isolated tmpdir; never touches live state)
+    ├── journal.jsonl                 # active observations
+    ├── journal/                      # rotated weekly (YYYY-Www.jsonl) + actioned-<id>.jsonl
+    ├── state.json                    # per-session counters
+    ├── usage.json                    # invocation tallies + visibility metrics
+    ├── active-nudges.json            # ephemeral SessionStart reminders (auto-expire)
+    ├── ab-tracking.jsonl             # one entry per auto-apply, drives effectiveness measurement
+    ├── reinforcements.jsonl          # appended on reinforcement proposal apply
+    ├── last-trace.txt                # most recent analyst clustering trace
+    ├── proposals/                    # queued, awaiting review
+    ├── applied/                      # approved + auto-applied archive
+    ├── rejected/                     # rejected with reason
+    ├── trash/                        # soft-deleted artifacts (recoverable)
+    ├── scripts/
+    │   ├── adam-utils.mjs                  # shared journal-reading + frontmatter parsing
+    │   ├── adam-window.mjs                 # per-signal sliding-window filter
+    │   ├── adam-score.mjs                  # urgency dampener + reinforcement candidates
+    │   ├── adam-ab-measure.mjs             # 7d pre/post delta per auto-applied edit
+    │   ├── adam-cooldown.mjs               # per-(skill, fingerprint) cooldown gate
+    │   ├── adam-nudge-eligibility.mjs      # dead_end session-count check
+    │   ├── adam-explain.mjs                # clustering trace parser/renderer
+    │   ├── adam-apply-reinforcement.mjs    # reinforcement proposal apply
+    │   ├── adam-upgrade.mjs                # .adam-new file UX (list/diff/accept)
+    │   └── adam-archive.mjs                # post-apply journal cleanup
+    └── tests/run-tests.sh            # 87 isolated tests; never touches live state
 ```

-## Install
+## What's new

-### One-liner (recommended)
-
-```sh
-curl -fsSL https://raw.githubusercontent.com/lukaszraczylo/claude-adam/main/install.sh | bash
-```
-
-Pin a release for reproducibility:
-
-```sh
-curl -fsSL https://raw.githubusercontent.com/lukaszraczylo/claude-adam/v0.3.1/install.sh \
-  | VERSION=v0.3.1 bash
-```
-
-The installer clones the repo to `/tmp`, copies files into `~/.claude/`, and offers to merge ADAM's hook entries into your `~/.claude/settings.json` (with a diff preview and `[y/N]` confirmation — your existing hooks are preserved). Pass `--yes` to skip the prompt; `--dry-run` to preview without writing.
-
-Requires `git`, `curl`, `jq`, and `node` 18+.
-
-### From a clone
-
-```sh
-git clone https://github.com/lukaszraczylo/claude-adam
-cd claude-adam
-./install.sh
-```
-
-### Upgrade-safe
-
-These files are **never overwritten** if they already exist:
-
- `~/.claude/adam/journal.jsonl` — your observation log
- `~/.claude/adam/state.json` — session counters
- `~/.claude/adam/usage.json` — invocation tallies
-
-If you've locally edited any installed file (e.g. `agents/adam.md`), the installer writes the new version to `<file>.adam-new` and warns you instead of clobbering.
-
-After install: run `bash ~/.claude/adam/tests/run-tests.sh` to verify (expect `27 passed, 0 failed`), start a fresh Claude Code session, then run `/reflect`.
+- **v0.5.0** — MOSS-grounded self-evolution (arXiv 2605.22794). Transcript capture: `context_window` field on struggle signals captures 8 surrounding events for evidence-based diagnosis. Two-stage analysis pipeline: diagnose+plan → inter-stage validation → implement (§3.3). Evidence batching via `adam-batch.mjs`: pre-clusters journal into coherent failure batches (§3.1). Pre-apply verification: 4-check deterministic gate before auto-apply (§3.4). Auto-rollback via `adam-rollback.mjs`: reverts regressed proposals detected by A/B measurement, creates regression nudges (§3.5). Harness self-modification: new `harness_edit` proposal type lets ADAM propose edits to its own scripts with test-suite-gated apply (§1 Table 1). Keypoint matrix: 5 capability dimensions scored per batch for structured evaluation (§4.2). 114 tests (up from 94).
+- **v0.4.0** — expanded struggle detection: `silent_drift` (5 consecutive read-only tools), `error_after_recovery` (same error fingerprint returns after clean recovery); severity-sum scoring with per-type divisors; extended `STRUGGLE_TYPES` set. 94 tests (up from 87).
+- **v0.3.3** — analyst observability, A/B measurement, journal hygiene. ISO-week journal rotation replaces 5MB size-based (fixes silent cluster-straddling under-count); per-signal sliding windows via `adam-window.mjs`; error fingerprint normalisation; correction corpus expanded + weak-token co-occurrence requirement (kills the `"actually, I think..."` false positive); mandatory clustering trace + `adam-explain.mjs`; new `nudge` and `reinforcement` proposal types; per-(skill, fingerprint) cooldown via `adam-cooldown.mjs`; `task_completed` scoring (dampener + reinforcement); A/B effectiveness measurement; upgrade UX overhaul (`adam-upgrade.mjs --list/--diff/--accept`); shared `adam-utils.mjs`. 87 tests (up from 30).
+- **v0.3.2** — `task_completed` signal: post-task skill capture for downstream reinforcement scoring (consumed in v0.3.3).
+- **v0.3.1** — code review pass: bug fixes (`errorFingerprint` no longer false-positives on `is_error: false`, archive script handles same-millisecond duplicates correctly, `tool_window` clears on session change, nudge filters proposal filenames by pattern), prose conciseness cuts, hardened `install.sh` with curl one-liner + settings.json merge, `adam-uninstall.sh`, isolated test harness.
+- **v0.3.0** — causal diagnosis: every proposal carries a `# Diagnosis` block (Trigger/Action/Mismatch/Outcome with verbatim transcript quote), plus `contradiction_flag` heuristic that vetoes auto-apply on obviously-conflicting `skill_edit` additions.
+- **v0.2.1** — win signals (`correction_free_streak`, `clean_recovery`) feed `skill_edit` auto-apply under a strict gate (≤ 30 LOC, ≤ 2× byte cap, 7d cooldown, 30d blacklist).
+- **v0.2.0** — actioned-entry archival via `adam-archive.mjs`; `cursor` field deprecated.

 ## Requirements

- Claude Code v2.1.0+ (for auto skill hot-reload; older versions need session restart after `skill_new` proposals are applied)
- Node.js 18+ (for the hook; tested on v22)
- Bash 4+, `git`, `curl`, `jq` (for installer + test harness)
+- **Claude Code v2.1.0+** — for auto skill hot-reload (older versions need a session restart after `skill_new` proposals).
+- **Node.js 18+** — tested on v22, used by the hook + helper scripts. Zero npm dependencies.
+- **Bash 4+**, `git`, `curl`, `jq` — for the installer + test harness.

 ### Platform support

-Tested on **macOS** (Darwin / BSD coreutils) and **Linux** (Alpine, glibc + musl). The install / uninstall / test scripts are written to be portable: `stat` uses BSD `-f` with GNU `-c` fallback, `mktemp -d -t prefix.XXXXXX` works on both, no GNU-only flags. CI smoke verified `27 passed, 0 failed` under `alpine:latest`.
-
-## Confidence rubric
-
-```
-Sum:
-+2  Signal repeated ≥3× across ≥2 sessions
-+2  Struggle signal appearing ≥1× within a single session (does not stack)
-+2  Transcript contains positive endorsement near related action
-+1  Multi-axis cluster (≥2 distinct struggle types in same session)
-1  Type-bias penalty (≥3 rejections, applied:rejected <1:2)
-+1  Blast radius low (memory or new isolated skill)
- 0  Blast radius medium (new agent, new hook, edit existing skill)
-1  Blast radius high (CLAUDE.md, settings hooks, edit agent, deletion)
-+1  Surgical (one file, ≤50 LOC for non-skill_new; ≤80 LOC for skill_new)
-3  Touches deny-list (settings.json hooks/permissions, CLAUDE.md, deletions)
-
-auto_apply_eligible requires ALL:
-  confidence ≥ 4
-  blast_radius == low
-  type ∈ {memory, skill_new, skill_edit}     # skill_edit also passes the win-driven gate
-  cross_session_evidence == true (single-session-only proposals always queue)
-
-skill_edit additionally requires (v0.2.1+):
-  win-signal evidence (correction_free_streak / clean_recovery cites target skill)
-  diff is append-only, ≤30 LOC, resulting size ≤2× original
-  no auto-edit to same target in past 7 days (cooldown)
-  no rejection-blacklist on target in past 30 days
-  contradiction heuristic does not flag (v0.3.0+)
-  # Diagnosis section present + structurally valid (v0.3.0+)
-```
-
-## Lifecycle: how proposals become permanent
-
-Every proposal records the journal entry timestamps that fed its cluster (`source_entries` in frontmatter). When you apply or reject a proposal, the skill calls `adam/scripts/adam-archive.mjs` which moves matching entries from `journal.jsonl` to `journal/actioned-<id>.jsonl`. Effects:
-
- The `journal.jsonl` stays bounded by **active** observations only.
- The next `/reflect` reads applied/ + rejected/ frontmatter, builds an excluded-timestamps set, and skips any leftover journal entries that were already actioned.
- Rule changes (e.g. lowering a threshold) immediately re-evaluate the remaining active observations — no manual cursor rewind needed.
-
-## What it will not do
-
- No background LLM spend. The analyst runs only when you invoke `/reflect`.
- No retroactive transcript mining beyond the journal cursor.
- No hard `rm` of any artifact. Deletions are soft (`mv` to `trash/<ts>/`).
- No autonomous edits to `CLAUDE.md`, agents, hooks, or `settings.json` — these always queue for review regardless of confidence.
- No proposal that matches a previously-rejected idea (≥2 token overlap with rejection's `# Why`).
- No invented trigger phrases for new skills — every trigger comes from observed user input.
+Tested on **macOS** (Darwin / BSD coreutils) and **Linux** (Alpine, glibc + musl). The install / uninstall / test scripts are written to be portable: `stat` uses BSD `-f` with GNU `-c` fallback, `mktemp -d -t prefix.XXXXXX` works on both, no GNU-only flags. CI smoke verified under `alpine:latest`.

 ## Uninstall

@@ -175,6 +291,20 @@ rm -rf ~/.claude/skills/adam-self-improvement

 Then remove the four `adam-*` hook entries from `~/.claude/settings.json`.

+## Contributing
+
+Issues and PRs welcome — especially additional signal types, transcript-aware diagnosis improvements, and platform fixes. Run the test suite before opening a PR:
+
+```sh
+bash ~/.claude/adam/tests/run-tests.sh
+```
+
 ## License

 [MIT](LICENSE) — © 2026 Lukasz Raczylo
+
+---
+
+<div align="center">
+<sub>Named after my son Adam, who taught me that observation is the start of every interesting thing.</sub>
+</div>
@@ -0,0 +1,184 @@
+#!/usr/bin/env node
+// adam-batch.mjs — pre-clusters windowed journal entries into coherent failure
+// batches before analyst dispatch. Implements MOSS §3.1: "anchored to an
+// automatically curated batch of production-failure evidence."
+//
+// Each batch groups entries by (signal_type, cluster_key) where cluster_key
+// follows the same clustering rules as agents/adam.md §4:
+//   correction     → tokenized phrase (cross-cwd)
+//   retry_loop     → tool
+//   weak_agent     → subagent_type
+//   tool_error_loop→ fp
+//   dead_end       → session
+//   edit_churn     → file basename
+//   build_loop     → session
+//   subagent_dispatch_pattern → subagent_type
+//   silent_drift   → active_skills[0]
+//   error_after_recovery → (recovered_from, original_fp)
+//   correction_free_streak → active_skills[0]
+//   clean_recovery → (recovered_from, active_skills[0])
+//   task_completed → sorted tool_kinds tuple
+//
+// CLI:
+//   adam-batch.mjs [--input <jsonl-path>] [--min-entries N] [--min-sessions N]
+//
+// Output: JSON object with `batches` array and `unbatched` count.
+
+import { readFileSync } from "node:fs";
+import { readJsonlSafe } from "./adam-utils.mjs";
+
+const DEFAULT_MIN_ENTRIES = 1;
+const DEFAULT_MIN_SESSIONS = 1;
+
+const CORRECTION_STOPWORDS = new Set([
+  "the", "a", "an", "and", "or", "but", "of", "to", "for", "in", "on",
+  "with", "use", "when", "where", "what", "why", "how", "this", "that",
+  "these", "those", "is", "are", "was", "were", "be", "been", "being",
+  "do", "does", "did", "doing", "has", "have", "had", "your", "you",
+  "i", "it", "as", "at", "by", "from", "not", "no",
+]);
+
+function tokenizePhrase(phrase) {
+  if (!phrase || typeof phrase !== "string") return "";
+  return phrase.toLowerCase()
+    .split(/\s+/)
+    .map(t => t.replace(/^[^\w']+|[^\w']+$/g, ""))
+    .filter(t => t && !CORRECTION_STOPWORDS.has(t))
+    .sort()
+    .join("|");
+}
+
+function clusterKey(entry) {
+  if (!entry || typeof entry !== "object") return null;
+  const t = entry.type;
+  switch (t) {
+    case "correction":
+      return tokenizePhrase(entry.phrase) || "unknown";
+    case "retry_loop":
+      return entry.tool || "unknown";
+    case "weak_agent":
+    case "subagent_dispatch_pattern":
+      return entry.subagent_type || "unknown";
+    case "tool_error_loop":
+      return entry.fp || "unknown";
+    case "dead_end":
+    case "build_loop":
+      return entry.session || "unknown";
+    case "edit_churn":
+      return entry.file ? entry.file.split("/").pop() : "unknown";
+    case "silent_drift":
+    case "correction_free_streak":
+      return Array.isArray(entry.active_skills) ? (entry.active_skills[0] || "") : "";
+    case "error_after_recovery":
+      return `${entry.recovered_from || "?"}:${entry.original_fp || "?"}`;
+    case "clean_recovery":
+      return `${entry.recovered_from || "?"}:${Array.isArray(entry.active_skills) ? (entry.active_skills[0] || "") : ""}`;
+    case "task_completed":
+      return Array.isArray(entry.tool_kinds) ? entry.tool_kinds.slice().sort().join(",") : "unknown";
+    default:
+      return entry.session || "unknown";
+  }
+}
+
+function parseArgs(argv) {
+  const args = { input: null, minEntries: DEFAULT_MIN_ENTRIES, minSessions: DEFAULT_MIN_SESSIONS, help: false };
+  for (let i = 0; i < argv.length; i++) {
+    const a = argv[i];
+    if (a === "--input" && i + 1 < argv.length) args.input = argv[++i];
+    else if (a === "--min-entries" && i + 1 < argv.length) {
+      const n = Number(argv[++i]);
+      if (!Number.isNaN(n) && n > 0) args.minEntries = n;
+    }
+    else if (a === "--min-sessions" && i + 1 < argv.length) {
+      const n = Number(argv[++i]);
+      if (!Number.isNaN(n) && n > 0) args.minSessions = n;
+    }
+    else if (a === "--help" || a === "-h") args.help = true;
+  }
+  return args;
+}
+
+export function buildBatches(entries, opts = {}) {
+  const minEntries = opts.minEntries || DEFAULT_MIN_ENTRIES;
+  const minSessions = opts.minSessions || DEFAULT_MIN_SESSIONS;
+  const map = new Map();
+
+  for (const e of entries || []) {
+    if (!e || typeof e !== "object" || !e.type) continue;
+    const key = `${e.type}::${clusterKey(e)}`;
+    if (!map.has(key)) {
+      map.set(key, {
+        batch_id: null,
+        signal_type: e.type,
+        cluster_key: clusterKey(e),
+        entries: [],
+        sessions: new Set(),
+        cwds: new Set(),
+      });
+    }
+    const batch = map.get(key);
+    batch.entries.push(e);
+    if (e.session) batch.sessions.add(e.session);
+    if (e.cwd) batch.cwds.add(e.cwd);
+  }
+
+  const batches = [];
+  let unbatched = 0;
+  let id = 1;
+  for (const [, batch] of map) {
+    if (batch.entries.length < minEntries || batch.sessions.size < minSessions) {
+      unbatched += batch.entries.length;
+      continue;
+    }
+    batch.batch_id = `b${id++}`;
+    batches.push({
+      batch_id: batch.batch_id,
+      signal_type: batch.signal_type,
+      cluster_key: batch.cluster_key,
+      entry_count: batch.entries.length,
+      session_count: batch.sessions.size,
+      cwd_count: batch.cwds.size,
+      has_context_window: batch.entries.some(e => Array.isArray(e.context_window) && e.context_window.length > 0),
+      entries: batch.entries,
+    });
+  }
+
+  batches.sort((a, b) => b.entry_count - a.entry_count);
+  return { batches, unbatched, total: (entries || []).length };
+}
+
+function main() {
+  const args = parseArgs(process.argv.slice(2));
+  if (args.help) {
+    process.stdout.write("usage: adam-batch.mjs [--input <jsonl-path>] [--min-entries N] [--min-sessions N]\n");
+    process.exit(0);
+  }
+  try {
+    let entries;
+    if (args.input) {
+      entries = readJsonlSafe(args.input);
+    } else if (!process.stdin.isTTY) {
+      const buf = readFileSync(0, "utf8");
+      entries = [];
+      for (const line of buf.split("\n")) {
+        if (!line) continue;
+        try { entries.push(JSON.parse(line)); } catch { /* skip */ }
+      }
+    } else {
+      process.stderr.write("adam-batch: no input (use --input or pipe)\n");
+      process.exit(1);
+    }
+    const result = buildBatches(entries, { minEntries: args.minEntries, minSessions: args.minSessions });
+    process.stdout.write(JSON.stringify(result) + "\n");
+    process.exit(0);
+  } catch (e) {
+    process.stderr.write(`adam-batch error: ${e.message}\n`);
+    process.exit(1);
+  }
+}
+
+if (import.meta.url === `file://${process.argv[1]}`) {
+  main();
+}
+
+export { clusterKey, tokenizePhrase };
@@ -0,0 +1,225 @@
+#!/usr/bin/env node
+// adam-rollback.mjs — auto-reverts proposals that regressed after apply.
+//
+// Implements MOSS §3.5: "rollback is mandatory because... a candidate that
+// passes trial can still regress live."
+//
+// For each regressed proposal (detected by adam-ab-measure.mjs):
+//   1. Reads the applied proposal from applied/
+//   2. Parses the `# Rollback` section for undo commands
+//   3. Moves proposal from applied/ to proposals/ with `rolled_back: true`
+//   4. Creates a regression nudge for next SessionStart
+//   5. Removes the ab-tracking entry (so it doesn't re-trigger)
+//
+// CLI:
+//   adam-rollback.mjs --proposal-id <id> [--home <path>] [--dry-run]
+//   adam-rollback.mjs --auto [--home <path>] [--dry-run]
+//
+//   --auto mode: reads ab-measure output, rolls back all regressed proposals.
+//
+// Output: JSON object with rollback results per proposal.
+// Does NOT execute the undo commands itself — outputs them for the skill to
+// execute in-context (safety: undo commands may reference files the script
+// can't safely modify).
+
+import { readFileSync, writeFileSync, renameSync, readdirSync, existsSync, mkdirSync } from "node:fs";
+import { join, basename } from "node:path";
+import { homedir } from "node:os";
+import { parseFrontmatter, readJsonlSafe, listJsonlFiles } from "./adam-utils.mjs";
+
+function parseArgs(argv) {
+  const args = { home: null, proposalId: null, auto: false, dryRun: false, help: false };
+  for (let i = 0; i < argv.length; i++) {
+    const a = argv[i];
+    if (a === "--home" && i + 1 < argv.length) args.home = argv[++i];
+    else if (a === "--proposal-id" && i + 1 < argv.length) args.proposalId = argv[++i];
+    else if (a === "--auto") args.auto = true;
+    else if (a === "--dry-run") args.dryRun = true;
+    else if (a === "--help" || a === "-h") args.help = true;
+  }
+  return args;
+}
+
+function findAppliedProposal(appliedDir, proposalId) {
+  if (!existsSync(appliedDir)) return null;
+  try {
+    const files = readdirSync(appliedDir).filter(n => n.endsWith(".md"));
+    for (const f of files) {
+      if (f.includes(proposalId)) return join(appliedDir, f);
+    }
+  } catch { /* skip */ }
+  return null;
+}
+
+function extractRollbackSection(content) {
+  const idx = content.indexOf("\n# Rollback\n");
+  if (idx === -1) return null;
+  let body = content.slice(idx + "\n# Rollback\n".length);
+  const nextSection = body.search(/\n# |\n---/);
+  if (nextSection !== -1) body = body.slice(0, nextSection);
+  return body.trim() || null;
+}
+
+function extractUndoCommands(rollbackSection) {
+  if (!rollbackSection) return [];
+  const commands = [];
+  const lines = rollbackSection.split("\n");
+  let inCodeBlock = false;
+  let blockLines = [];
+  for (const line of lines) {
+    if (line.startsWith("```")) {
+      if (inCodeBlock) {
+        if (blockLines.length) commands.push(blockLines.join("\n"));
+        blockLines = [];
+      }
+      inCodeBlock = !inCodeBlock;
+      continue;
+    }
+    if (inCodeBlock) {
+      blockLines.push(line);
+    }
+  }
+  return commands;
+}
+
+export function planRollback(appliedDir, proposalId) {
+  const path = findAppliedProposal(appliedDir, proposalId);
+  if (!path) return { status: "not_found", proposal_id: proposalId };
+
+  const content = readFileSync(path, "utf8");
+  const fm = parseFrontmatter(content);
+  const rollbackSection = extractRollbackSection(content);
+  const undoCommands = extractUndoCommands(rollbackSection);
+
+  return {
+    status: "planned",
+    proposal_id: proposalId,
+    applied_path: path,
+    type: fm.type || "unknown",
+    target: fm.target || null,
+    target_skill: fm.target_skill || null,
+    undo_commands: undoCommands,
+    has_rollback_section: !!rollbackSection,
+  };
+}
+
+export function executeRollback(plan, adamRoot, opts = {}) {
+  const dryRun = opts.dryRun || false;
+  const proposalsDir = join(adamRoot, "proposals");
+  const nudgesPath = join(adamRoot, "active-nudges.json");
+  const now = Date.now();
+
+  if (plan.status !== "planned") return { ...plan, action: "skipped" };
+
+  const result = {
+    proposal_id: plan.proposal_id,
+    type: plan.type,
+    target: plan.target,
+    undo_commands: plan.undo_commands,
+    actions: [],
+  };
+
+  if (dryRun) {
+    result.actions.push("dry_run: would move applied → proposals");
+    if (plan.undo_commands.length) {
+      result.actions.push(`dry_run: would output ${plan.undo_commands.length} undo command(s)`);
+    }
+    result.actions.push("dry_run: would create regression nudge");
+    result.status = "dry_run";
+    return result;
+  }
+
+  mkdirSync(proposalsDir, { recursive: true });
+  const destName = `${basename(plan.applied_path).replace(/\.md$/, "")}-rollback.md`;
+  const destPath = join(proposalsDir, destName);
+
+  let content = readFileSync(plan.applied_path, "utf8");
+  const rollbackMeta = `\nrolled_back: true\nrolled_back_at: "${new Date(now).toISOString()}"`;
+  content = content.replace(/^(---\n[\s\S]*?)(---)/m, `$1${rollbackMeta}\n$2`);
+
+  try {
+    writeFileSync(destPath, content);
+    renameSync(plan.applied_path, plan.applied_path + ".rolled-back");
+    result.actions.push(`moved ${plan.applied_path} → ${destPath}`);
+  } catch (e) {
+    result.status = "move_failed";
+    result.error = e.message;
+    return result;
+  }
+
+  try {
+    let nudges = [];
+    if (existsSync(nudgesPath)) {
+      try { nudges = JSON.parse(readFileSync(nudgesPath, "utf8")); } catch { nudges = []; }
+    }
+    nudges.push({
+      kind: "regression_rollback",
+      message: `adam: rolled back "${plan.proposal_id}" (type: ${plan.type}) — regression detected in A/B measurement. Review with /reflect.`,
+      created_at: now,
+      expires_at_ts: now + 7 * 86400000,
+      max_displays: 3,
+      displays_used: 0,
+      source_proposal: plan.proposal_id,
+    });
+    writeFileSync(nudgesPath, JSON.stringify(nudges, null, 2));
+    result.actions.push("regression nudge created");
+  } catch (e) {
+    result.actions.push(`nudge failed: ${e.message}`);
+  }
+
+  result.status = "rolled_back";
+  return result;
+}
+
+async function main() {
+  const args = parseArgs(process.argv.slice(2));
+  if (args.help) {
+    process.stdout.write(
+      "usage: adam-rollback.mjs --proposal-id <id> [--home <path>] [--dry-run]\n" +
+      "       adam-rollback.mjs --auto [--home <path>] [--dry-run]\n"
+    );
+    process.exit(0);
+  }
+
+  const claudeHome = args.home || join(homedir(), ".claude");
+  const adamRoot = join(claudeHome, "adam");
+  const appliedDir = join(adamRoot, "applied");
+
+  try {
+    const results = [];
+
+    if (args.auto) {
+      const abPath = join(adamRoot, "ab-tracking.jsonl");
+      const entries = readJsonlSafe(abPath);
+      const { computeDeltas } = await import("./adam-ab-measure.mjs");
+      const sources = [join(adamRoot, "journal.jsonl"), ...listJsonlFiles(join(adamRoot, "journal"))];
+      const journalAll = [];
+      for (const p of sources) for (const e of readJsonlSafe(p)) journalAll.push(e);
+      const deltas = computeDeltas(entries, journalAll);
+      const regressed = deltas.filter(d => d.status === "regressed");
+
+      for (const d of regressed) {
+        const plan = planRollback(appliedDir, d.proposal_id);
+        const result = executeRollback(plan, adamRoot, { dryRun: args.dryRun });
+        results.push(result);
+      }
+    } else if (args.proposalId) {
+      const plan = planRollback(appliedDir, args.proposalId);
+      const result = executeRollback(plan, adamRoot, { dryRun: args.dryRun });
+      results.push(result);
+    } else {
+      process.stderr.write("adam-rollback: specify --proposal-id or --auto\n");
+      process.exit(1);
+    }
+
+    process.stdout.write(JSON.stringify({ rollbacks: results }) + "\n");
+    process.exit(0);
+  } catch (e) {
+    process.stderr.write(`adam-rollback error: ${e.message}\n`);
+    process.exit(1);
+  }
+}
+
+if (import.meta.url === `file://${process.argv[1]}`) {
+  main();
+}
@@ -43,10 +43,32 @@ export const NEGATIVE_SIGNAL_TYPES = new Set([
  "retry_loop",
  "build_loop",
  "weak_agent",
+  "silent_drift",
+  "error_after_recovery",
 ]);

 export const REINFORCEMENT_THRESHOLD = 3;

+// Severity divisor per struggle signal type. Severity = max(1, floor(count / divisor)).
+// Entries without `count` default to severity 1. Source of truth — referenced by
+// agents/adam.md (Confidence rubric → severity-sum bullets).
+export const SEVERITY_DIVISORS = {
+  dead_end: 8,
+  edit_churn: 4,
+  tool_error_loop: 3,
+  retry_loop: 3,
+  weak_agent: 2,
+  build_loop: 1,
+};
+
+export function entrySeverity(entry) {
+  if (!entry || typeof entry !== "object") return 1;
+  const divisor = SEVERITY_DIVISORS[entry.type];
+  if (!divisor) return 1;
+  const count = typeof entry.count === "number" && entry.count > 0 ? entry.count : 1;
+  return Math.max(1, Math.floor(count / divisor));
+}
+
 function parseArgs(argv) {
  const args = { home: null, input: null, help: false };
  for (let i = 0; i < argv.length; i++) {
@@ -84,11 +106,22 @@ export function computeSessionScores(entries) {
    const sid = e.session || e.session_id || "";
    if (!sid) continue;
    if (!bySession.has(sid)) {
-      bySession.set(sid, { session_id: sid, negative_count: 0, task_completed_count: 0 });
+      bySession.set(sid, {
+        session_id: sid,
+        negative_count: 0,
+        task_completed_count: 0,
+        severity_sum: 0,
+        severity_by_type: {},
+      });
    }
    const slot = bySession.get(sid);
    if (e.type === "task_completed") slot.task_completed_count++;
-    else if (NEGATIVE_SIGNAL_TYPES.has(e.type)) slot.negative_count++;
+    else if (NEGATIVE_SIGNAL_TYPES.has(e.type)) {
+      slot.negative_count++;
+      const sev = entrySeverity(e);
+      slot.severity_sum += sev;
+      slot.severity_by_type[e.type] = (slot.severity_by_type[e.type] || 0) + sev;
+    }
  }
  const out = [];
  for (const slot of bySession.values()) {
@@ -29,6 +29,8 @@ export const SIGNAL_WINDOWS_DAYS = {
  build_loop: 30,
  weak_agent: 30,
  subagent_dispatch_pattern: 30,
+  silent_drift: 14,
+  error_after_recovery: 30,
  correction_free_streak: 60,
  clean_recovery: 60,
  task_completed: 60,
@@ -16,6 +16,8 @@ SCORE="$REAL_HOME/.claude/adam/scripts/adam-score.mjs"
 ABMEASURE="$REAL_HOME/.claude/adam/scripts/adam-ab-measure.mjs"
 APPLYREIN="$REAL_HOME/.claude/adam/scripts/adam-apply-reinforcement.mjs"
 UPGRADE="$REAL_HOME/.claude/adam/scripts/adam-upgrade.mjs"
+BATCH="$REAL_HOME/.claude/adam/scripts/adam-batch.mjs"
+ROLLBACK="$REAL_HOME/.claude/adam/scripts/adam-rollback.mjs"

 TMP_HOME="$(mktemp -d -t adam-test.XXXXXX)"
 trap 'rm -rf "$TMP_HOME"' EXIT INT TERM
@@ -33,6 +35,8 @@ SCORE_RUN()   { HOME="$TMP_HOME" node "$SCORE" --home "$TMP_HOME/.claude" "$@";
 ABMEASURE_RUN(){ HOME="$TMP_HOME" node "$ABMEASURE" --home "$TMP_HOME/.claude" "$@"; }
 APPLYREIN_RUN(){ HOME="$TMP_HOME" node "$APPLYREIN" "$@" --home "$TMP_HOME/.claude"; }
 UPGRADE_RUN() { HOME="$TMP_HOME" node "$UPGRADE" "$@"; }
+BATCH_RUN()   { HOME="$TMP_HOME" node "$BATCH" "$@"; }
+ROLLBACK_RUN(){ HOME="$TMP_HOME" node "$ROLLBACK" "$@"; }

 PASS=0
 FAIL=0
@@ -1388,6 +1392,453 @@ else
  fi
 fi

+# --- Test 78: silent_drift fires after 5 consecutive read-only tools ---
+echo "Test 78: silent_drift after 5 reads"
+reset_state
+for i in 1 2 3 4 5; do
+  echo "{\"hook_event_name\":\"PostToolUse\",\"tool_name\":\"Read\",\"tool_input\":{\"file_path\":\"/tmp/r-$i\"},\"tool_response\":{\"content\":\"ok\"},\"session_id\":\"sSD\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+assert_grep "$ROOT/journal.jsonl" '"type":"silent_drift"' "5 consecutive reads emit silent_drift"
+assert_grep "$ROOT/journal.jsonl" '"read_count":5' "silent_drift entry records read_count"
+
+# --- Test 79: silent_drift counter resets on action tool ---
+echo "Test 79: silent_drift counter resets on action tool"
+reset_state
+for i in 1 2 3 4; do
+  echo "{\"hook_event_name\":\"PostToolUse\",\"tool_name\":\"Read\",\"tool_input\":{\"file_path\":\"/tmp/r-$i\"},\"tool_response\":{\"content\":\"ok\"},\"session_id\":\"sSDR\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+# Action tool — should reset
+echo '{"hook_event_name":"PostToolUse","tool_name":"Edit","tool_input":{"file_path":"/tmp/x"},"tool_response":{"content":"ok"},"session_id":"sSDR","cwd":"/tmp/x"}' \
+  | HOOK_RUN >/dev/null 2>&1 || true
+for i in 1 2 3 4; do
+  echo "{\"hook_event_name\":\"PostToolUse\",\"tool_name\":\"Read\",\"tool_input\":{\"file_path\":\"/tmp/rb-$i\"},\"tool_response\":{\"content\":\"ok\"},\"session_id\":\"sSDR\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+if grep -qE '"type":"silent_drift"' "$ROOT/journal.jsonl"; then
+  echo "  FAIL: silent_drift fired despite action tool reset"; FAIL=$((FAIL+1))
+else
+  echo "  PASS: silent_drift suppressed by intervening action tool"; PASS=$((PASS+1))
+fi
+
+# --- Test 80: error_after_recovery fires when same fp returns post-clean_recovery ---
+echo "Test 80: error_after_recovery fires when fp returns after recovery"
+reset_state
+# Build a tool_error_loop with ENOENT
+for i in 1 2 3; do
+  echo '{"hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"cat missing"},"tool_response":{"is_error":true,"content":"cat: missing: No such file or directory"},"session_id":"sEAR","cwd":"/tmp/x"}' \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+# 3 clean tools → clean_recovery
+for i in 1 2 3; do
+  echo "{\"hook_event_name\":\"PostToolUse\",\"tool_name\":\"Read\",\"tool_input\":{\"file_path\":\"/tmp/ok-$i\"},\"tool_response\":{\"content\":\"ok\"},\"session_id\":\"sEAR\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+# Same fp returns within window
+echo '{"hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"cat other"},"tool_response":{"is_error":true,"content":"cat: other: No such file or directory"},"session_id":"sEAR","cwd":"/tmp/x"}' \
+  | HOOK_RUN >/dev/null 2>&1 || true
+assert_grep "$ROOT/journal.jsonl" '"type":"error_after_recovery"' "same fp after clean_recovery emits error_after_recovery"
+
+# --- Test 81: error_after_recovery does NOT fire after window expires ---
+echo "Test 81: error_after_recovery suppressed beyond window"
+reset_state
+for i in 1 2 3; do
+  echo '{"hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"cat missing"},"tool_response":{"is_error":true,"content":"cat: missing: No such file or directory"},"session_id":"sEARW","cwd":"/tmp/x"}' \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+for i in 1 2 3; do
+  echo "{\"hook_event_name\":\"PostToolUse\",\"tool_name\":\"Read\",\"tool_input\":{\"file_path\":\"/tmp/ok-$i\"},\"tool_response\":{\"content\":\"ok\"},\"session_id\":\"sEARW\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+# UserPromptSubmit resets tools_since_user + last_errors so the burn reads don't
+# trigger a secondary dead_end + clean_recovery cycle (which would create a fresh
+# recovery within window and cause error_after_recovery to fire legitimately).
+echo '{"hook_event_name":"UserPromptSubmit","prompt":"keep going","session_id":"sEARW","cwd":"/tmp/x"}' \
+  | HOOK_RUN >/dev/null 2>&1 || true
+# Burn through the 5-event window with 6 clean reads (session_post_count: 6 → 12)
+for i in 1 2 3 4 5 6; do
+  echo "{\"hook_event_name\":\"PostToolUse\",\"tool_name\":\"Read\",\"tool_input\":{\"file_path\":\"/tmp/burn-$i\"},\"tool_response\":{\"content\":\"ok\"},\"session_id\":\"sEARW\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+echo '{"hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"cat other"},"tool_response":{"is_error":true,"content":"cat: other: No such file or directory"},"session_id":"sEARW","cwd":"/tmp/x"}' \
+  | HOOK_RUN >/dev/null 2>&1 || true
+if grep -qE '"type":"error_after_recovery"' "$ROOT/journal.jsonl"; then
+  echo "  FAIL: error_after_recovery fired outside 5-event window"; FAIL=$((FAIL+1))
+else
+  echo "  PASS: error_after_recovery suppressed outside window"; PASS=$((PASS+1))
+fi
+
+# --- Test 82: adam-score.mjs reports severity_sum + severity_by_type ---
+echo "Test 82: severity-sum reporting in score.mjs"
+SEV_TMP="$(mktemp)"
+cat > "$SEV_TMP" <<'EOF'
+{"ts":"2026-05-12T10:00:00Z","session":"sSEV","type":"dead_end","count":64}
+{"ts":"2026-05-12T10:01:00Z","session":"sSEV","type":"edit_churn","count":8}
+{"ts":"2026-05-12T10:02:00Z","session":"sSEV","type":"tool_error_loop","count":3,"fp":"ENOENT:abc"}
+EOF
+out=$(SCORE_RUN --input "$SEV_TMP" 2>/dev/null)
+rm -f "$SEV_TMP"
+# Expected: dead_end 64/8=8, edit_churn 8/4=2, tool_error_loop 3/3=1 → sum=11
+if echo "$out" | grep -q '"severity_sum":11'; then
+  echo "  PASS: severity_sum=11 reported"; PASS=$((PASS+1))
+else
+  echo "  FAIL: severity_sum mismatch (got: $out)"; FAIL=$((FAIL+1))
+fi
+if echo "$out" | grep -q '"dead_end":8'; then
+  echo "  PASS: severity_by_type.dead_end=8"; PASS=$((PASS+1))
+else
+  echo "  FAIL: severity_by_type.dead_end missing/wrong (got: $out)"; FAIL=$((FAIL+1))
+fi
+
+# ============================================================
+# MOSS-grounded tests: context_window, adam-batch, adam-rollback
+# ============================================================
+
+# --- Test 83: context_window attached to tool_error_loop struggle signal ---
+echo "Test 83: context_window attached to tool_error_loop"
+reset_state
+# Fire a user prompt first so the context ring has something.
+echo '{"hook_event_name":"UserPromptSubmit","prompt":"run the tests","session_id":"sCW1","cwd":"/tmp/x"}' \
+  | HOOK_RUN >/dev/null 2>&1 || true
+for i in 1 2 3; do
+  echo '{"hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"failing-cmd"},"tool_response":{"is_error":true,"content":"Error: command not found: failing-cmd"},"session_id":"sCW1","cwd":"/tmp/x"}' \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+assert_grep "$ROOT/journal.jsonl" '"context_window":\[' "tool_error_loop carries context_window"
+
+# --- Test 84: context_window captures preceding user prompt ---
+echo "Test 84: context_window captures user prompt text"
+# Re-use the journal from test 83
+assert_grep "$ROOT/journal.jsonl" '"prompt":"run the tests"' "context_window includes user prompt excerpt"
+
+# --- Test 85: context_window includes tool response excerpts ---
+echo "Test 85: context_window includes tool response excerpts"
+assert_grep "$ROOT/journal.jsonl" '"response_excerpt"' "context_window entries have response_excerpt"
+
+# --- Test 86: context_window on dead_end signal ---
+echo "Test 86: context_window on dead_end signal"
+reset_state
+echo '{"hook_event_name":"UserPromptSubmit","prompt":"start working","session_id":"sCW2","cwd":"/tmp/x"}' \
+  | HOOK_RUN >/dev/null 2>&1 || true
+for i in 1 2 3 4 5 6 7 8; do
+  echo "{\"hook_event_name\":\"PostToolUse\",\"tool_name\":\"Bash\",\"tool_input\":{\"command\":\"step$i\"},\"session_id\":\"sCW2\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+if grep -qE '"type":"dead_end"' "$ROOT/journal.jsonl" && grep -qE '"context_window":\[' "$ROOT/journal.jsonl"; then
+  echo "  PASS: dead_end carries context_window"; PASS=$((PASS+1))
+else
+  echo "  FAIL: dead_end missing context_window"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 87: context_window NOT on non-struggle signals ---
+echo "Test 87: context_window absent from correction_free_streak"
+reset_state
+for i in 1 2 3 4 5; do
+  echo "{\"hook_event_name\":\"UserPromptSubmit\",\"prompt\":\"step $i please\",\"session_id\":\"sCW3\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+# correction_free_streak should have fired
+streak_line=$(grep '"type":"correction_free_streak"' "$ROOT/journal.jsonl" | head -1)
+if [ -n "$streak_line" ] && ! echo "$streak_line" | grep -q '"context_window"'; then
+  echo "  PASS: correction_free_streak has no context_window"; PASS=$((PASS+1))
+else
+  echo "  FAIL: unexpected context_window on non-struggle signal"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 88: adam-batch clusters same signal_type + fp into one batch ---
+echo "Test 88: adam-batch clusters same (type, fp) into one batch"
+batch_input=$(cat <<'EOF'
+{"ts":"2026-05-20T10:00:00Z","type":"tool_error_loop","session":"s1","cwd":"/a","tool":"Bash","fp":"ENOENT:abc","count":3}
+{"ts":"2026-05-21T10:00:00Z","type":"tool_error_loop","session":"s2","cwd":"/a","tool":"Bash","fp":"ENOENT:abc","count":4}
+{"ts":"2026-05-22T10:00:00Z","type":"tool_error_loop","session":"s3","cwd":"/b","tool":"Bash","fp":"ENOENT:abc","count":3}
+EOF
+)
+out=$(echo "$batch_input" | BATCH_RUN 2>/dev/null)
+batch_count=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const j=JSON.parse(b);console.log(j.batches.length)})')
+entry_count=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const j=JSON.parse(b);console.log(j.batches[0]?j.batches[0].entry_count:0)})')
+if [ "$batch_count" = "1" ] && [ "$entry_count" = "3" ]; then
+  echo "  PASS: 3 same-fp entries → 1 batch with entry_count=3"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected 1 batch / 3 entries (got batches=$batch_count entries=$entry_count)"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 89: adam-batch creates separate batches for different signal types ---
+echo "Test 89: adam-batch separates different signal types"
+batch_input=$(cat <<'EOF'
+{"ts":"2026-05-20T10:00:00Z","type":"correction","session":"s1","cwd":"/a","phrase":"no wrong"}
+{"ts":"2026-05-21T10:00:00Z","type":"tool_error_loop","session":"s1","cwd":"/a","fp":"ENOENT:abc","count":3}
+EOF
+)
+out=$(echo "$batch_input" | BATCH_RUN 2>/dev/null)
+batch_count=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const j=JSON.parse(b);console.log(j.batches.length)})')
+if [ "$batch_count" = "2" ]; then
+  echo "  PASS: 2 different types → 2 batches"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected 2 batches (got $batch_count)"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 90: adam-batch reports session_count correctly ---
+echo "Test 90: adam-batch tracks session_count per batch"
+batch_input=$(cat <<'EOF'
+{"ts":"2026-05-20T10:00:00Z","type":"correction","session":"s1","cwd":"/a","phrase":"no wrong"}
+{"ts":"2026-05-21T10:00:00Z","type":"correction","session":"s2","cwd":"/a","phrase":"no wrong"}
+{"ts":"2026-05-22T10:00:00Z","type":"correction","session":"s1","cwd":"/b","phrase":"no wrong"}
+EOF
+)
+out=$(echo "$batch_input" | BATCH_RUN 2>/dev/null)
+sessions=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const j=JSON.parse(b);console.log(j.batches[0]?j.batches[0].session_count:0)})')
+if [ "$sessions" = "2" ]; then
+  echo "  PASS: session_count=2 for entries from s1+s2"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected session_count=2 (got $sessions)"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 91: adam-batch reports has_context_window ---
+echo "Test 91: adam-batch reports has_context_window flag"
+batch_input=$(cat <<'EOF'
+{"ts":"2026-05-20T10:00:00Z","type":"dead_end","session":"s1","cwd":"/a","count":8,"context_window":[{"event":"user","prompt":"hi","ts":"2026-05-20T09:59:00Z"}]}
+EOF
+)
+out=$(echo "$batch_input" | BATCH_RUN 2>/dev/null)
+has_cw=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const j=JSON.parse(b);console.log(j.batches[0]?j.batches[0].has_context_window:"false")})')
+if [ "$has_cw" = "true" ]; then
+  echo "  PASS: has_context_window=true when entries have context_window"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected has_context_window=true (got $has_cw)"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 92: adam-batch empty input → no batches ---
+echo "Test 92: adam-batch produces empty output on empty input"
+out=$(echo '' | BATCH_RUN 2>/dev/null)
+batch_count=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{try{const j=JSON.parse(b);console.log(j.batches.length)}catch{console.log("parse-error")}})')
+total=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{try{const j=JSON.parse(b);console.log(j.total)}catch{console.log("parse-error")}})')
+if [ "$batch_count" = "0" ] && [ "$total" = "0" ]; then
+  echo "  PASS: empty input → 0 batches, total=0"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected 0 batches (got batches=$batch_count total=$total)"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 93: adam-rollback --proposal-id moves applied proposal to proposals ---
+echo "Test 93: adam-rollback moves applied proposal to proposals/"
+reset_state
+rm -f "$ROOT/proposals/"*rollback* "$ROOT/active-nudges.json"
+cat > "$ROOT/applied/2026-05-20T00-00-00Z-rb-test-001.md" <<'EOF'
+---
+id: rb-test-001
+type: skill_new
+target: ~/.claude/skills/test-skill/SKILL.md
+confidence: 5
+blast_radius: low
+auto_apply_eligible: true
+status: applied
+source_entries:
+  - "2026-05-18T10:00:00Z"
+---
+# Why
+test rollback
+
+# Rollback
+```bash
+rm -rf ~/.claude/skills/test-skill/
+```
+EOF
+out=$(ROLLBACK_RUN --proposal-id rb-test-001 --home "$TMP_HOME/.claude" 2>/dev/null)
+if echo "$out" | grep -q '"status":"rolled_back"'; then
+  rb_ok=1
+else
+  rb_ok=0
+fi
+# Verify proposal moved to proposals/
+if ls "$ROOT/proposals/"*rb-test-001* >/dev/null 2>&1; then
+  moved_ok=1
+else
+  moved_ok=0
+fi
+# Verify original file renamed
+if [ -f "$ROOT/applied/2026-05-20T00-00-00Z-rb-test-001.md.rolled-back" ]; then
+  renamed_ok=1
+else
+  renamed_ok=0
+fi
+if [ "$rb_ok" = "1" ] && [ "$moved_ok" = "1" ] && [ "$renamed_ok" = "1" ]; then
+  echo "  PASS: rollback moved proposal and renamed applied file"; PASS=$((PASS+1))
+else
+  echo "  FAIL: rollback incomplete (status=$rb_ok moved=$moved_ok renamed=$renamed_ok out=$out)"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 94: adam-rollback creates regression nudge ---
+echo "Test 94: adam-rollback creates regression nudge in active-nudges.json"
+if [ -f "$ROOT/active-nudges.json" ]; then
+  nudge_kind=$(node -e "const j=JSON.parse(require('fs').readFileSync('$ROOT/active-nudges.json','utf8'));console.log((j[0]||{}).kind||'')")
+  if [ "$nudge_kind" = "regression_rollback" ]; then
+    echo "  PASS: regression nudge created with kind=regression_rollback"; PASS=$((PASS+1))
+  else
+    echo "  FAIL: nudge kind wrong (got $nudge_kind)"; FAIL=$((FAIL+1))
+  fi
+else
+  echo "  FAIL: active-nudges.json not created"; FAIL=$((FAIL+1))
+fi
+rm -f "$ROOT/proposals/"*rb-test* "$ROOT/applied/"*rb-test* "$ROOT/active-nudges.json"
+
+# --- Test 95: adam-rollback rolled_back field in proposal frontmatter ---
+echo "Test 95: rolled-back proposal has rolled_back: true in frontmatter"
+reset_state
+rm -f "$ROOT/proposals/"*rollback* "$ROOT/active-nudges.json"
+cat > "$ROOT/applied/2026-05-20T00-00-00Z-rb-test-002.md" <<'EOF'
+---
+id: rb-test-002
+type: memory
+target: ~/.claude/projects/-Users-nvm/memory/test.md
+confidence: 4
+blast_radius: low
+---
+# Why
+test
+# Rollback
+delete the memory file
+EOF
+ROLLBACK_RUN --proposal-id rb-test-002 --home "$TMP_HOME/.claude" >/dev/null 2>&1 || true
+rb_file=$(ls "$ROOT/proposals/"*rb-test-002* 2>/dev/null | head -1)
+if [ -n "$rb_file" ] && grep -q 'rolled_back: true' "$rb_file"; then
+  echo "  PASS: rolled-back proposal has rolled_back: true"; PASS=$((PASS+1))
+else
+  echo "  FAIL: rolled_back marker missing (file=$rb_file)"; FAIL=$((FAIL+1))
+fi
+rm -f "$ROOT/proposals/"*rb-test* "$ROOT/applied/"*rb-test* "$ROOT/active-nudges.json"
+
+# --- Test 96: adam-rollback not_found on missing proposal ---
+echo "Test 96: adam-rollback returns not_found for missing proposal"
+reset_state
+out=$(ROLLBACK_RUN --proposal-id nonexistent-999 --home "$TMP_HOME/.claude" 2>/dev/null)
+if echo "$out" | grep -q '"status":"not_found"'; then
+  echo "  PASS: not_found status for missing proposal"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected not_found (got: $out)"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 97: adam-rollback --dry-run does not move files ---
+echo "Test 97: adam-rollback --dry-run leaves files in place"
+reset_state
+rm -f "$ROOT/proposals/"*rollback* "$ROOT/active-nudges.json"
+cat > "$ROOT/applied/2026-05-20T00-00-00Z-rb-dry-001.md" <<'EOF'
+---
+id: rb-dry-001
+type: skill_edit
+target: ~/.claude/skills/foo/SKILL.md
+confidence: 4
+---
+# Why
+test dry run
+# Rollback
+revert edit
+EOF
+out=$(ROLLBACK_RUN --proposal-id rb-dry-001 --dry-run --home "$TMP_HOME/.claude" 2>/dev/null)
+if echo "$out" | grep -q '"status":"dry_run"' && [ -f "$ROOT/applied/2026-05-20T00-00-00Z-rb-dry-001.md" ]; then
+  echo "  PASS: dry-run did not move files"; PASS=$((PASS+1))
+else
+  echo "  FAIL: dry-run moved files or wrong status (out=$out)"; FAIL=$((FAIL+1))
+fi
+rm -f "$ROOT/applied/2026-05-20T00-00-00Z-rb-dry-001.md"
+
+# --- Test 98: context_window on edit_churn signal ---
+echo "Test 98: context_window on edit_churn signal"
+reset_state
+echo '{"hook_event_name":"UserPromptSubmit","prompt":"fix the tests","session_id":"sCW4","cwd":"/tmp/x"}' \
+  | HOOK_RUN >/dev/null 2>&1 || true
+for i in 1 2 3 4; do
+  echo '{"hook_event_name":"PostToolUse","tool_name":"Edit","tool_input":{"file_path":"/tmp/churn.py"},"session_id":"sCW4","cwd":"/tmp/x"}' \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+churn_line=$(grep '"type":"edit_churn"' "$ROOT/journal.jsonl" | head -1)
+if [ -n "$churn_line" ] && echo "$churn_line" | grep -q '"context_window"'; then
+  echo "  PASS: edit_churn carries context_window"; PASS=$((PASS+1))
+else
+  echo "  FAIL: edit_churn missing context_window"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 99: context_window on build_loop signal ---
+echo "Test 99: context_window on build_loop signal"
+reset_state
+echo '{"hook_event_name":"UserPromptSubmit","prompt":"run the build","session_id":"sCW5","cwd":"/tmp/x"}' \
+  | HOOK_RUN >/dev/null 2>&1 || true
+for i in 1 2; do
+  echo '{"hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"npm run build"},"tool_response":{"is_error":true,"content":"Build failed: TypeError"},"session_id":"sCW5","cwd":"/tmp/x"}' \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+build_line=$(grep '"type":"build_loop"' "$ROOT/journal.jsonl" | head -1)
+if [ -n "$build_line" ] && echo "$build_line" | grep -q '"context_window"'; then
+  echo "  PASS: build_loop carries context_window"; PASS=$((PASS+1))
+else
+  echo "  FAIL: build_loop missing context_window"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 100: adam-batch --min-entries filter ---
+echo "Test 100: adam-batch --min-entries filters small batches"
+batch_input=$(cat <<'EOF'
+{"ts":"2026-05-20T10:00:00Z","type":"correction","session":"s1","cwd":"/a","phrase":"no wrong"}
+{"ts":"2026-05-21T10:00:00Z","type":"tool_error_loop","session":"s1","cwd":"/a","fp":"ENOENT:abc","count":3}
+{"ts":"2026-05-22T10:00:00Z","type":"tool_error_loop","session":"s2","cwd":"/a","fp":"ENOENT:abc","count":4}
+{"ts":"2026-05-23T10:00:00Z","type":"tool_error_loop","session":"s3","cwd":"/a","fp":"ENOENT:abc","count":5}
+EOF
+)
+out=$(echo "$batch_input" | BATCH_RUN --min-entries 3 2>/dev/null)
+batch_count=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const j=JSON.parse(b);console.log(j.batches.length)})')
+unbatched=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{const j=JSON.parse(b);console.log(j.unbatched)})')
+if [ "$batch_count" = "1" ] && [ "$unbatched" = "1" ]; then
+  echo "  PASS: --min-entries=3 keeps 1 batch (3 entries), drops 1 singleton"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected 1 batch + 1 unbatched (got batches=$batch_count unbatched=$unbatched)"; FAIL=$((FAIL+1))
+fi
+
+# --- Test 101: adam-rollback extracts undo commands from Rollback section ---
+echo "Test 101: adam-rollback extracts undo commands from code blocks"
+reset_state
+rm -f "$ROOT/proposals/"*rollback* "$ROOT/active-nudges.json"
+cat > "$ROOT/applied/2026-05-20T00-00-00Z-rb-undo-001.md" <<'HEREDOC'
+---
+id: rb-undo-001
+type: skill_new
+target: ~/.claude/skills/test-undo/SKILL.md
+confidence: 5
+blast_radius: low
+---
+# Why
+test
+# Rollback
+```bash
+rm -rf ~/.claude/skills/test-undo/
+```
+HEREDOC
+out=$(ROLLBACK_RUN --proposal-id rb-undo-001 --home "$TMP_HOME/.claude" 2>/dev/null)
+undo_count=$(echo "$out" | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{try{const j=JSON.parse(b);console.log((j.rollbacks[0]||{}).undo_commands?j.rollbacks[0].undo_commands.length:0)}catch{console.log("err")}})')
+if [ "$undo_count" = "1" ]; then
+  echo "  PASS: extracted 1 undo command from Rollback section"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected 1 undo command (got $undo_count, out=$out)"; FAIL=$((FAIL+1))
+fi
+rm -f "$ROOT/proposals/"*rb-undo* "$ROOT/applied/"*rb-undo* "$ROOT/active-nudges.json"
+
+# --- Test 102: context_ring size bounded at 8 ---
+echo "Test 102: context_ring bounded at CONTEXT_RING_SIZE=8"
+reset_state
+# Fire 12 PostToolUse events, then a struggle signal
+for i in $(seq 1 12); do
+  echo "{\"hook_event_name\":\"PostToolUse\",\"tool_name\":\"Read\",\"tool_input\":{\"file_path\":\"/tmp/f-$i\"},\"tool_response\":{\"content\":\"ok\"},\"session_id\":\"sCR\",\"cwd\":\"/tmp/x\"}" \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+# Next 3 errors to trigger tool_error_loop with context_window
+for i in 1 2 3; do
+  echo '{"hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"fail"},"tool_response":{"is_error":true,"content":"Error: fail"},"session_id":"sCR","cwd":"/tmp/x"}' \
+    | HOOK_RUN >/dev/null 2>&1 || true
+done
+cw_len=$(grep '"type":"tool_error_loop"' "$ROOT/journal.jsonl" | head -1 | node -e 'let b="";process.stdin.on("data",d=>b+=d).on("end",()=>{try{const j=JSON.parse(b);console.log(j.context_window?j.context_window.length:0)}catch{console.log("err")}})')
+if [ "$cw_len" = "8" ]; then
+  echo "  PASS: context_window capped at 8 entries"; PASS=$((PASS+1))
+else
+  echo "  FAIL: expected 8 context_window entries (got $cw_len)"; FAIL=$((FAIL+1))
+fi
+
 echo
 echo "Results: $PASS passed, $FAIL failed"
 [ "$FAIL" = "0" ]
@@ -8,6 +8,71 @@ tools: Read, Write, Edit, Grep, Glob, Bash

 You analyse Claude Code's own behaviour to propose targeted, surgical improvements. You operate offline (no LLM round-trips outside this run) and produce **files**, not actions. Main-thread Claude reviews and applies changes with the user.

+## Stage mode
+
+The skill dispatches you in one of two stages (MOSS-inspired multi-stage pipeline — §3.3: "a single prompt asked to diagnose, plan, implement, verify, and decide overloads context and produces lower-quality output than a sequenced flow"):
+
+- **`stage=diagnose`**: Read batched journal entries, cluster, diagnose root causes, plan fix types. Output diagnoses JSON to `/tmp/adam-diagnoses.json`. Do NOT draft proposals.
+- **`stage=implement`**: Read approved diagnoses from `/tmp/adam-diagnoses.json`. Draft full proposal files to `proposals_dir/`. Emit the clustering trace and punch list.
+
+If no `stage` is specified in the dispatch prompt, run **both stages sequentially** within a single pass (backward-compatible with pre-MOSS flow).
+
+### Diagnose-stage output format
+
+When `stage=diagnose`, write `/tmp/adam-diagnoses.json` containing:
+
+```json
+{
+  "diagnoses": [
+    {
+      "cluster_id": "c1",
+      "signal_type": "correction",
+      "cluster_key": "wrong|approach",
+      "count": 5,
+      "sessions": 3,
+      "diagnosis": {
+        "trigger": "...",
+        "action": "...",
+        "mismatch": "...",
+        "outcome": "... `verbatim quote` ..."
+      },
+      "plan": {
+        "type": "memory",
+        "target": "~/.claude/projects/-Users-nvm/memory/go-test-cache.md",
+        "scope": "add feedback memory about go test -count=1"
+      },
+      "keypoints": {
+        "tool_selection": 1,
+        "scope_discipline": 2,
+        "error_recovery": 0,
+        "first_attempt": 0,
+        "build_reliability": 1
+      },
+      "gates": {
+        "threshold": "pass",
+        "cross_session": "pass",
+        "window": "in:5/out:0",
+        "contradiction": "none"
+      },
+      "source_entries": ["2026-05-20T10:00:00Z", "2026-05-21T11:00:00Z"],
+      "context_evidence": ["... excerpts from context_window ..."]
+    }
+  ],
+  "skipped": [
+    {"cluster_id": "c3", "signal_type": "retry_loop", "reason": "threshold", "count": 2}
+  ],
+  "summary": "considered=4 diagnosed=2 skipped=2"
+}
+```
+
+The skill validates diagnoses between stages (see SKILL.md §2 "Inter-stage validation").
+
+## Context window evidence
+
+Journal entries for struggle signals now carry a `context_window` field — an array of the last 8 events (user prompts, tool calls, responses) surrounding the friction point. This is the ADAM equivalent of MOSS's "original transcript captured by auto-scan at evidence time" (§3.4).
+
+When drafting diagnoses, **prefer `context_window` evidence over transcript file lookups** when it is present. The `context_window` is already scoped to the friction point and more reliable than file-based transcript pulls. Fall back to `transcripts_root` only when `context_window` is absent (pre-upgrade entries).
+
 ## Karpathy constraints (mandatory)

 You MUST obey these on every proposal:
@@ -38,6 +103,8 @@ Per-signal windows (single source of truth: `SIGNAL_WINDOWS_DAYS` in `~/.claude/
 | `build_loop` | 30 d | build/test failure patterns |
 | `weak_agent` | 30 d | subagent quality signal |
 | `subagent_dispatch_pattern` | 30 d | dispatch routing pattern |
+| `silent_drift` | 14 d | exploration-without-action is task-local |
+| `error_after_recovery` | 30 d | recovery-then-same-error patterns persist |
 | `correction_free_streak` | 60 d | wins accumulate slowly |
 | `clean_recovery` | 60 d | wins accumulate slowly |
 | `task_completed` | 60 d | recipe wins accumulate slowly |
@@ -59,6 +126,8 @@ The hook emits these `type` values into the journal:
 | `edit_churn` | same file edited 4× in window | file basename |
 | `build_loop` | 2 build/test/compile commands fail in session | session |
 | `subagent_dispatch_pattern` | same subagent dispatched ≥3× cumulatively | subagent_type |
+| `silent_drift` | 5 consecutive read-only PostToolUse without an action tool (reset on action or UserPromptSubmit) | `active_skills[0]` |
+| `error_after_recovery` | same error fingerprint returns within 5 PostToolUse of a `clean_recovery` | (`recovered_from`, `original_fp`) |
 | `correction_free_streak` | 5 clean UserPromptSubmits in a row (no correction phrase) | `active_skills[0]` |
 | `clean_recovery` | 3 clean PostToolUse events after a `tool_error_loop`/`dead_end`/`retry_loop` | (`recovered_from`, `active_skills[0]`) |
 | `task_completed` | UserPromptSubmit closes a run of ≥5 tool calls with ≥3 distinct tool kinds and 0 corrections | sorted `tool_kinds` tuple |
@@ -84,10 +153,17 @@ The hook emits these `type` values into the journal:
   - `edit_churn`: cluster by file basename pattern (e.g. `*.test.ts`).
   - `build_loop`: cluster by `session`.
   - `subagent_dispatch_pattern`: cluster by `subagent_type`.
+   - `silent_drift`: cluster by `active_skills[0]` (empty string when no skill is active).
+   - `error_after_recovery`: cluster by (`recovered_from`, `original_fp`).
   - `correction_free_streak`: cluster by `active_skills[0]`. Treat ≥3 streaks across ≥2 sessions naming the same skill as cross-session evidence.
   - `clean_recovery`: cluster by (`recovered_from`, `active_skills[0]`). A win cluster qualifies for `skill_edit` only when the named skill exists in `skills_root`.
   - `task_completed`: cluster by sorted `tool_kinds` tuple (the multi-tool recipe). Single entry qualifies for `skill_new` proposal (drafting protocol applies). Cross-session evidence requires ≥2 entries from distinct sessions with same tuple — without it, proposal queues, never auto-applies. Run the existing skill-overlap rule before drafting: if the recipe matches an existing skill's name/description tokens, route to `skill_edit` instead.
-5. **Multi-axis correlation**: for each session that produced ≥2 distinct struggle types (`tool_error_loop`, `dead_end`, `weak_agent`, `retry_loop`, `edit_churn`, `build_loop`), tag clusters from that session as `multi_axis: true`. This grants +1 confidence at scoring.
+5. **Multi-axis correlation**: for each session that produced ≥2 distinct struggle types (`tool_error_loop`, `dead_end`, `weak_agent`, `retry_loop`, `edit_churn`, `build_loop`, `silent_drift`, `error_after_recovery`), tag clusters from that session as `multi_axis: true`. This grants +1 confidence at scoring.
+
+5b. **Skill-attribution sub-clustering**: after primary clustering (step 4), for every struggle cluster (`tool_error_loop`, `dead_end`, `weak_agent`, `retry_loop`, `edit_churn`, `build_loop`, `silent_drift`, `error_after_recovery`) that contains entries with non-empty `active_skills[0]`:
+   - Split into per-skill sub-clusters keyed on `active_skills[0]`. Entries with empty `active_skills` stay in the original cluster.
+   - If a sub-cluster has ≥3 entries AND names a skill that exists in `skills_root`, mark it as a candidate for `skill_edit` (struggle-driven variant; see "Struggle-driven `skill_edit` eligibility"). Otherwise treat the parent cluster normally.
+   - The umbrella cluster (cross-skill) still emits its usual proposal type (memory, etc.) — sub-clusters do NOT replace it, they supplement it.
 6. For each cluster qualifying under the rubric — ≥3 occurrences across ≥2 sessions, OR (for struggle types) ≥1 entry within a single session, OR (for `correction`) ≥3 occurrences across ≥2 cwds:
   a. If cluster topic matches a rejected idea via the rejected-ideas fuzzy set (≥2 token overlap with rejection's `# Why`), skip with reason `"rejected-similar"`.
   b. Pull ~20 messages of transcript context from `transcripts_root` to enrich. Never read full transcripts.
@@ -254,6 +330,21 @@ A `skill_edit` proposal sets `auto_apply_eligible: true` ONLY when ALL hold:

 If any of (3)–(9) fails: still emit the proposal, but `auto_apply_eligible: false` — main thread queues for review.

+## Struggle-driven `skill_edit` eligibility
+
+Skill-attribution sub-clustering (step 5b) produces struggle-driven `skill_edit` candidates: a sub-cluster of ≥3 struggle entries all naming the same `active_skills[0]` that exists in `skills_root`. These proposals are emitted but **ALWAYS queue** — `auto_apply_eligible: false` regardless of confidence. Negative evidence on a skill is a weaker basis for self-modification than positive evidence (the skill may be active during friction caused by something else), so the human reviews every one.
+
+A struggle-driven `skill_edit` proposal MUST:
+
+1. Set `target` to the matched skill's `SKILL.md` path.
+2. Cluster severity-sum ≥ 10 (same threshold as the +1 rubric bullet).
+3. Sub-cluster names exactly one skill (no ambiguity across distinct `active_skills[0]` values).
+4. `# Proposed change` is an append-only diff adding a `## When struggling` section (naive default body: a checkpoint-or-pause rule appropriate to the dominant signal — e.g. `dead_end` → "After 16 PostToolUse events without UserPromptSubmit, emit a one-line checkpoint summary before continuing.").
+5. Frontmatter includes `struggle_evidence: "<ts of one source entry naming this skill>"` and `struggle_signals: [<list of signal types in the sub-cluster>]`. The win-driven `win_evidence` field is omitted.
+6. Subject to the same Per-(skill, fingerprint) cooldown as win-driven `skill_edit`.
+
+If gate (2) or (3) fails: skip the sub-cluster (the parent cluster still produces its umbrella proposal). The sub-cluster's `source_entries` overlap with the parent's — the apply pipeline handles dedup via the excluded-timestamps set.
+
 ## Per-(skill, fingerprint) cooldown

 The cooldown gate is keyed on **(target_skill, proposal_fingerprint)** — not on target_skill alone. A rejected/applied proposal for skill `X` with fingerprint `A` does NOT block future proposals for skill `X` with fingerprint `B`.
@@ -299,17 +390,39 @@ After ≥7 days, `~/.claude/adam/scripts/adam-ab-measure.mjs` reads each entry a

 The `/reflect` skill runs `adam-ab-measure.mjs --format json` before dispatching this agent, filters to `status == "regressed"`, and passes the list as `ab_regressions` (each object has `proposal_id`, `target_skill`, `proposal_type`, `delta_pct`, `pre_count`, `post_count`).

-**When `ab_regressions` is non-empty, you MUST emit a `## Regressions` section at the TOP of your output (above the proposals listing).** One bullet per regressed proposal listing `proposal_id`, `target_skill`, `delta_pct`, plus the short suggestion `consider revert via /reflect --revert <proposal_id>` (the revert mechanism itself is out of scope for this release — the message stands as a hint).
+**When `ab_regressions` is non-empty, you MUST emit a `## Regressions` section at the TOP of your output (above the proposals listing).** One bullet per regressed proposal listing `proposal_id`, `target_skill`, `delta_pct`. The skill auto-rolls back regressed proposals via `adam-rollback.mjs` before dispatching you — this section is your record of what was rolled back and why.

 The clustering trace summary (see §"Clustering trace") adds an extra `regressions=<N>` key alongside `considered/emitted/skipped`. When no `ab_regressions` arrive (or list is empty), emit `regressions=0`.

+## Keypoint matrix (MOSS §3.3/§4.2)
+
+When running in `stage=diagnose`, you MUST produce a **keypoint matrix** alongside each batch diagnosis. This structured evaluation replaces ad-hoc confidence with per-capability scoring.
+
+Capability dimensions (score each 0–2 per batch: 0=no signal, 1=partial, 2=strong evidence):
+
+| dimension | description | positive signals | negative signals |
+|---|---|---|---|
+| `tool_selection` | correct tool chosen first try | low `retry_loop` | high `retry_loop`, `weak_agent` |
+| `scope_discipline` | stays within requested scope | low `edit_churn`, low `dead_end` | high `edit_churn`, `dead_end`, `silent_drift` |
+| `error_recovery` | recovers from errors without user help | `clean_recovery` | `error_after_recovery`, `tool_error_loop` |
+| `first_attempt` | succeeds without corrections | `correction_free_streak` | `correction` |
+| `build_reliability` | builds/tests pass on first try | `task_completed` with build tools | `build_loop` |
+
+The matrix goes into the diagnosis output as `keypoints: {tool_selection: N, scope_discipline: N, ...}`. The implement stage uses it to:
+1. Prioritize proposals targeting the weakest dimensions.
+2. Include `keypoint_target: "<dimension>"` in proposal frontmatter.
+3. Track dimension trends across `/reflect` runs (persisted in `~/.claude/adam/keypoint-history.jsonl`).
+
 ## Confidence rubric (deterministic — do NOT vibe)

 Sum:
 - Signal repeated ≥3× across ≥2 sessions: **+2**
- Struggle signal (`tool_error_loop`, `dead_end`, `weak_agent`, `retry_loop`, `edit_churn`, `build_loop`) appearing ≥1× within a single session: **+2** *(each struggle entry already represents a hook-side threshold crossing — e.g. 8 tools without a prompt, 3 same-args retries, 4 edits to one file. Treat each entry as one piece of evidence. Does not stack with the cross-session bonus.)*
+- Struggle signal (`tool_error_loop`, `dead_end`, `weak_agent`, `retry_loop`, `edit_churn`, `build_loop`, `silent_drift`, `error_after_recovery`) appearing ≥1× within a single session: **+2** *(each struggle entry already represents a hook-side threshold crossing — e.g. 8 tools without a prompt, 3 same-args retries, 4 edits to one file, 5 read-only tools in a row, same-fp error after a recovery. Treat each entry as one piece of evidence. Does not stack with the cross-session bonus.)*
 - Transcript contains positive endorsement (`yes`, `exactly`, `do that`, `keep doing`) within 2 messages of related action: **+2**
 - Multi-axis cluster (≥2 distinct struggle types in same session): **+1**
+- Cluster severity-sum ≥ 10 (severity per entry = `max(1, floor(count / divisor))` using `SEVERITY_DIVISORS` from `adam-score.mjs` — `dead_end:8, edit_churn:4, tool_error_loop:3, retry_loop:3, weak_agent:2, build_loop:1`; entries without `count` count as 1): **+1**
+- Cluster severity-sum ≥ 32: **+1** *(additive — a severity-sum of 32 gets +1 from the previous bullet AND +1 here, total +2.)*
+- Skill-attributed sub-cluster (≥3 entries naming the same `active_skills[0]` that exists in `skills_root`): **+1**
 - Type-bias penalty from feedback loop (≥3 rejections, applied:rejected ratio <1:2 for this `type`): **-1**
 - Diagnosis flags `Mismatch: unclear` (causation could not be reconstructed from transcript context): **-1**
 - Blast radius: low **+1**, medium **0**, high **-1** (default per type — see Proposal types table)
@@ -328,13 +441,14 @@ Sum:
 |---|---|---|---|
 | `memory` | `~/.claude/projects/-Users-nvm/memory/*.md` | low | yes if conf≥4 AND cross_session |
 | `skill_new` | new dir under `~/.claude/skills/` | low | yes if conf≥4 AND cross_session |
-| `skill_edit` | existing skill file | medium | yes if win-evidence + LOC + cooldown gates all pass (see "Win-driven skill_edit eligibility") |
+| `skill_edit` | existing skill file | medium | yes (win-driven only) if win-evidence + LOC + cooldown gates all pass (see "Win-driven skill_edit eligibility"); struggle-driven variant ALWAYS queues (see "Struggle-driven skill_edit eligibility") |
 | `nudge` | append to `~/.claude/adam/active-nudges.json` | low | yes when `dead_end_count ≥ 3` in a single session (single-session evidence sufficient; skips cross-session gate). Does NOT modify skills/memories/CLAUDE.md — only seeds a SessionStart reminder for a future session. |
 | `reinforcement` | append entry to `~/.claude/adam/reinforcements.jsonl` | low | yes if conf≥4 AND blast_radius=low (same gate as memory). Applies via `adam-apply-reinforcement.mjs`; appends one JSONL entry, no code/memory/skill changes. |
 | `agent_new` | new file under `~/.claude/agents/` | medium | no |
 | `agent_edit` | existing agent file | medium | no |
 | `claude_md_edit` | `~/.claude/CLAUDE.md` | high | no |
 | `hook_new` / `hook_edit` | `settings.json` hooks | high | no |
+| `harness_edit` | adam's own scripts/agent/hooks (see "Harness self-modification") | high | **never** |
 | `deletion` | any skill/agent (soft delete) | high | no |

 ### `nudge` proposals
@@ -363,6 +477,42 @@ A `reinforcement` proposal is logged when `adam-score.mjs` reports `count >= 3`

 Note that `task_completed` alone — without an adjacent negative signal cluster — is NOT a proposal source. It is a urgency *modifier* (see "Scoring: task_completed dampener") and a reinforcement input only.

+### `harness_edit` proposals (MOSS §1 Table 1)
+
+MOSS's core thesis: "routing, hook ordering, state invariants, and dispatch live in code rather than in any text artifact, an entire class of structural failure is physically unreachable from the text layer." This proposal type extends ADAM's evolution scope to its own harness.
+
+**Allowed targets** (harness files that ADAM may propose edits to):
+
+| target | what it controls |
+|---|---|
+| `~/.claude/adam/scripts/adam-observe.mjs` | signal detection regexes, thresholds, counters |
+| `~/.claude/adam/scripts/adam-score.mjs` | severity divisors, dampener thresholds |
+| `~/.claude/adam/scripts/adam-window.mjs` | per-signal sliding window durations |
+| `~/.claude/adam/scripts/adam-batch.mjs` | evidence batching logic |
+| `~/.claude/agents/adam.md` | this agent's own rubric, clustering, proposal rules |
+| `~/.claude/hooks/adam-observe.mjs` | hook integration, event routing |
+
+**Gates (all must hold — stricter than any other type):**
+
+1. `confidence ≥ 5`
+2. `cross_session_evidence == true` (≥5 occurrences across ≥3 sessions)
+3. `auto_apply_eligible: false` — **always**. Harness edits are never auto-applied.
+4. `blast_radius: high`
+5. Proposal includes a `# Test verification` section with the command `bash ~/.claude/adam/tests/run-tests.sh` and the expected result "94 passed, 0 failed" (or current pass count). The skill runs this test before applying.
+6. Change is surgical: ≤30 LOC diff, single file.
+7. `# Diagnosis` reconstructs the causal chain from harness-level behavior (not from text-artifact behavior). The mismatch must name a specific code path (function, regex, threshold) in the target file.
+
+**When to propose `harness_edit`:**
+- Signal detection misses a recurring friction pattern (false negative in adam-observe.mjs)
+- A/B measurement shows systematic bias (e.g., windows too short/long in adam-window.mjs)
+- Scoring thresholds produce consistently over/under-weighted proposals (adam-score.mjs)
+- Batch clustering produces too-coarse or too-fine groupings (adam-batch.mjs)
+
+**When NOT to propose `harness_edit`:**
+- The fix is achievable via a text-mutable type (skill, memory, nudge)
+- Evidence is from a single session only
+- The change would affect test outcomes without clear improvement evidence
+
 ## Special handling

 ### CLAUDE.md edits
@@ -389,7 +539,7 @@ Filename: `proposals_dir/YYYY-MM-DD-NNN-<type>-<slug>.md` (NNN is daily counter
 ```markdown
 ---
 id: YYYY-MM-DD-NNN
-type: skill_new | memory | skill_edit | nudge | reinforcement | agent_new | agent_edit | claude_md_edit | hook_new | hook_edit | deletion
+type: skill_new | memory | skill_edit | nudge | reinforcement | agent_new | agent_edit | claude_md_edit | hook_new | hook_edit | harness_edit | deletion
 target: <absolute path — for skill_new, the will-be path: ~/.claude/skills/<slug>/SKILL.md>
 confidence: <int>
 blast_radius: low | medium | high
@@ -415,6 +565,11 @@ bytes_after: <int>
 contradiction_flag: "<one-line summary or null>"
 # optional — auto-populated from Diagnosis Mismatch line
 diagnosis_summary: "<≤120 chars, single sentence>"
+# keypoint matrix — which capability dimension this proposal targets (MOSS §4.2)
+keypoint_target: "<tool_selection | scope_discipline | error_recovery | first_attempt | build_reliability>"
+# harness_edit only — test command and expected output
+test_command: "bash ~/.claude/adam/tests/run-tests.sh"
+test_expected: "<N> passed, 0 failed"
 ---

 # Why
@@ -453,7 +608,7 @@ Print a single JSON line to stdout:
 ## What you must NOT do

 - Do not call other agents.
- Do not write to `~/.claude/skills/`, `~/.claude/agents/`, `settings.json`, `CLAUDE.md`, or any existing skill/agent file directly. All changes go through proposal files for main-thread review and apply.
+- Do not write to `~/.claude/skills/`, `~/.claude/agents/`, `settings.json`, `CLAUDE.md`, adam scripts, or any existing skill/agent/harness file directly. All changes go through proposal files for main-thread review and apply. This includes `harness_edit` proposals — you draft the diff, the skill applies it after test verification.
 - Do not delete files. Deletion proposals describe a soft-move; the main thread executes it.
 - Do not write outside `proposals_dir/` and `state_path`.
 - Do not invent trigger phrases for `skill_new` — every trigger must come from observed user input.
@@ -0,0 +1,13 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 160 160" width="160" height="160" role="img" aria-label="claude-adam logo">
+  <title>claude-adam</title>
+  <desc>A swaddled baby — rounded A-shape bundle with a face inside and small hands extending from the wrap-band. Dark-background variant.</desc>
+  <g stroke="#f0f6fc">
+    <path d="M 36 134 Q 30 78 80 28 Q 130 78 124 134 Z" fill="none" stroke-width="9" stroke-linejoin="round"/>
+    <path d="M 16 100 L 44 100 Q 80 115 116 100 L 144 100" fill="none" stroke-width="6" stroke-linecap="round"/>
+    <path d="M 75 78 Q 80 82 85 78" fill="none" stroke-width="2.5" stroke-linecap="round"/>
+  </g>
+  <g fill="#f0f6fc">
+    <circle cx="72" cy="64" r="3.2"/>
+    <circle cx="88" cy="64" r="3.2"/>
+  </g>
+</svg>
@@ -0,0 +1,13 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 160 160" width="160" height="160" role="img" aria-label="claude-adam logo">
+  <title>claude-adam</title>
+  <desc>A swaddled baby — rounded A-shape bundle with a face inside and small hands extending from the wrap-band. Light-background variant.</desc>
+  <g stroke="#24292f">
+    <path d="M 36 134 Q 30 78 80 28 Q 130 78 124 134 Z" fill="none" stroke-width="9" stroke-linejoin="round"/>
+    <path d="M 16 100 L 44 100 Q 80 115 116 100 L 144 100" fill="none" stroke-width="6" stroke-linecap="round"/>
+    <path d="M 75 78 Q 80 82 85 78" fill="none" stroke-width="2.5" stroke-linecap="round"/>
+  </g>
+  <g fill="#24292f">
+    <circle cx="72" cy="64" r="3.2"/>
+    <circle cx="88" cy="64" r="3.2"/>
+  </g>
+</svg>
@@ -0,0 +1,19 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 160 160" width="160" height="160" role="img" aria-label="claude-adam logo">
+  <title>claude-adam</title>
+  <desc>A swaddled baby — rounded A-shape bundle with a face inside and small hands extending from the wrap-band. Adapts to light/dark via embedded media query + currentColor fallback.</desc>
+  <style>
+    svg { color: #24292f; }
+    @media (prefers-color-scheme: dark) {
+      svg { color: #f0f6fc; }
+    }
+  </style>
+  <g stroke="currentColor">
+    <path d="M 36 134 Q 30 78 80 28 Q 130 78 124 134 Z" fill="none" stroke-width="9" stroke-linejoin="round"/>
+    <path d="M 16 100 L 44 100 Q 80 115 116 100 L 144 100" fill="none" stroke-width="6" stroke-linecap="round"/>
+    <path d="M 75 78 Q 80 82 85 78" fill="none" stroke-width="2.5" stroke-linecap="round"/>
+  </g>
+  <g fill="currentColor">
+    <circle cx="72" cy="64" r="3.2"/>
+    <circle cx="88" cy="64" r="3.2"/>
+  </g>
+</svg>
@@ -87,6 +87,12 @@ function normalizeErrorText(text) {
 const ERROR_RE = /\b(error|failed|exception|traceback|denied|cannot|unable to|not found|undefined|nullpointer|typeerror|syntaxerror|panic|fatal|enoent|econnrefused|etimedout|eaccess|segfault|crashed|uncaught)\b/i;
 const BUILD_RE = /\b(build|compile|make|gradle|cargo|tsc|webpack|vite|rollup|pytest|jest|mocha|vitest|go\s+test|npm\s+test|yarn\s+test|npm\s+run\s+build|yarn\s+build|ctest|ninja|bazel)\b/i;
 const EDIT_TOOLS = new Set(["Edit", "Write", "MultiEdit", "NotebookEdit"]);
+const READ_ONLY_TOOLS = new Set([
+  "Read", "Grep", "Glob", "ToolSearch", "WebFetch", "WebSearch",
+  "mcp__filepuff__file_read", "mcp__filepuff__file_search",
+  "mcp__filepuff__find_definition", "mcp__filepuff__find_references",
+  "mcp__filepuff__ast_query", "mcp__filepuff__symbol_at", "mcp__filepuff__ping",
+]);
 const WINDOW_SIZE = 10;
 const RETRY_THRESHOLD = 3;
 const AGENT_RESPAWN_THRESHOLD = 2;
@@ -98,10 +104,18 @@ const BUILD_LOOP_THRESHOLD = 2;
 const SUBAGENT_DISPATCH_THRESHOLD = 3;
 const CORRECTION_FREE_THRESHOLD = 5;
 const CLEAN_RECOVERY_WINDOW = 3;
-const STRUGGLE_TYPES = new Set(["tool_error_loop", "dead_end", "retry_loop"]);
+const SILENT_DRIFT_THRESHOLD = 5;
+const ERROR_AFTER_RECOVERY_WINDOW = 5;
+const RECENT_RECOVERIES_MAX = 3;
+const STRUGGLE_TYPES = new Set([
+  "tool_error_loop", "dead_end", "retry_loop", "weak_agent",
+  "edit_churn", "build_loop", "silent_drift", "error_after_recovery",
+]);
 const ACTIVE_SKILLS_LOOKBACK = 10;
 const TASK_TOOL_MIN = 5;
 const TASK_DIVERSITY_MIN = 3;
+const CONTEXT_RING_SIZE = 8;
+const CONTEXT_EXCERPT_LEN = 200;
 const STATE_MAX_BYTES = 1_000_000;

 function safeRead(path, fallback) {
@@ -217,6 +231,20 @@ function activeNames(state, kind) {
  return [...seen];
 }

+function excerpt(text, len) {
+  if (!text || typeof text !== "string") return null;
+  return text.length > len ? text.slice(0, len) + "…" : text;
+}
+
+function pushContext(state, entry) {
+  state.context_ring.push(entry);
+  if (state.context_ring.length > CONTEXT_RING_SIZE) state.context_ring.shift();
+}
+
+function snapshotContext(state) {
+  return state.context_ring.length ? state.context_ring.slice() : undefined;
+}
+
 function errorFingerprint(toolResponse) {
  if (!toolResponse) return null;
  let text = "";
@@ -268,6 +296,8 @@ function resetFrictionCounters(state) {
  state.edit_churn_emitted = {};
  state.build_failure_count = 0;
  state.build_loop_emitted = false;
+  state.silentDriftCounter = 0;
+  state.silentDriftEmitted = false;
 }

 function resetSessionLocal(state) {
@@ -276,7 +306,10 @@ function resetSessionLocal(state) {
  state.subagent_dispatch_emitted = {};
  state.correctionFreeCounter = 0;
  state.recoveryWatch = null;
+  state.recentRecoveries = [];
+  state.session_post_count = 0;
  state.tool_window = [];
+  state.context_ring = [];
  state.task_tool_kinds = {};
  state.task_tool_count = 0;
  state.task_corrections = 0;
@@ -299,6 +332,11 @@ function ensureStateDefaults(state) {
  if (!state.task_tool_kinds || typeof state.task_tool_kinds !== "object") state.task_tool_kinds = {};
  if (typeof state.task_tool_count !== "number") state.task_tool_count = 0;
  if (typeof state.task_corrections !== "number") state.task_corrections = 0;
+  if (typeof state.silentDriftCounter !== "number") state.silentDriftCounter = 0;
+  if (typeof state.silentDriftEmitted !== "boolean") state.silentDriftEmitted = false;
+  if (!Array.isArray(state.recentRecoveries)) state.recentRecoveries = [];
+  if (typeof state.session_post_count !== "number") state.session_post_count = 0;
+  if (!Array.isArray(state.context_ring)) state.context_ring = [];
 }

 function main() {
@@ -323,6 +361,7 @@ function main() {

  if (event === "UserPromptSubmit") {
    const prompt = (input.prompt || "").slice(0, 200);
+    pushContext(state, { event: "user", prompt: excerpt(prompt, CONTEXT_EXCERPT_LEN), ts });
    if (isCorrection(prompt)) {
      const last = state.tool_window[state.tool_window.length - 1] || {};
      appendJournal({
@@ -389,9 +428,27 @@ function main() {
    const argsHash = djb2(JSON.stringify(input.tool_input || {}));
    const file = (input.tool_input && (input.tool_input.file_path || input.tool_input.path)) || null;

+    const toolResponse = input.tool_response;
+    const respExcerpt = (() => {
+      if (!toolResponse) return null;
+      const text = typeof toolResponse === "string" ? toolResponse
+        : typeof toolResponse.content === "string" ? toolResponse.content
+        : null;
+      return excerpt(text, CONTEXT_EXCERPT_LEN);
+    })();
+    pushContext(state, {
+      event: "tool", tool, ts,
+      input_excerpt: excerpt(JSON.stringify(input.tool_input || {}), CONTEXT_EXCERPT_LEN),
+      response_excerpt: respExcerpt,
+      is_error: !!(toolResponse && toolResponse.is_error),
+    });
+
    let struggleEmittedThisTurn = null;
    const emit = (entry) => {
-      if (STRUGGLE_TYPES.has(entry.type)) struggleEmittedThisTurn = entry.type;
+      if (STRUGGLE_TYPES.has(entry.type)) {
+        entry.context_window = snapshotContext(state);
+        struggleEmittedThisTurn = entry.type;
+      }
      appendJournal(entry);
    };

@@ -402,12 +459,24 @@ function main() {
    }
    state.tool_window.push(windowEntry);
    if (state.tool_window.length > WINDOW_SIZE) state.tool_window.shift();
+    state.session_post_count += 1;

    const sameToolArgs = state.tool_window.filter(e => e.tool === tool && e.argsHash === argsHash).length;
    if (sameToolArgs >= RETRY_THRESHOLD) {
      emit({ ts, session, cwd, type: "retry_loop", tool, count: sameToolArgs });
    }

+    if (READ_ONLY_TOOLS.has(tool)) {
+      state.silentDriftCounter += 1;
+      if (state.silentDriftCounter >= SILENT_DRIFT_THRESHOLD && !state.silentDriftEmitted) {
+        emit({ ts, session, cwd, type: "silent_drift", read_count: state.silentDriftCounter, last_tool: tool });
+        state.silentDriftEmitted = true;
+      }
+    } else {
+      state.silentDriftCounter = 0;
+      state.silentDriftEmitted = false;
+    }
+
    if (tool === "Agent") {
      const subagent = (input.tool_input && (input.tool_input.subagent_type || input.tool_input.agent)) || "unknown";
      const recent = state.tool_window.slice(-5).filter(e => e.tool === "Agent" && e.subagent === subagent).length;
@@ -423,6 +492,23 @@ function main() {
    const fp = errorFingerprint(input.tool_response);
    if (fp) {
      bumpUsage("payload:tool_response_error_seen");
+      if (state.recentRecoveries.length) {
+        const keep = [];
+        for (const rec of state.recentRecoveries) {
+          const tools_since = state.session_post_count - rec.emitted_at_count;
+          if (tools_since > ERROR_AFTER_RECOVERY_WINDOW) continue;
+          if (Array.isArray(rec.fps) && rec.fps.includes(fp)) {
+            emit({
+              ts, session, cwd, type: "error_after_recovery",
+              recovered_from: rec.recovered_from, original_fp: fp,
+              tools_since_recovery: tools_since,
+            });
+            continue;
+          }
+          keep.push(rec);
+        }
+        state.recentRecoveries = keep;
+      }
      state.last_errors.push({ tool, fp });
      if (state.last_errors.length > ERROR_RING_SIZE) state.last_errors.shift();
      const sameError = state.last_errors.filter(e => e.fp === fp).length;
@@ -468,7 +554,13 @@ function main() {
    state.task_tool_kinds[tool] = (state.task_tool_kinds[tool] || 0) + 1;

    if (struggleEmittedThisTurn) {
-      state.recoveryWatch = { recovered_from: struggleEmittedThisTurn, since_ts: ts, clean_count: 0, window_tools: [] };
+      state.recoveryWatch = {
+        recovered_from: struggleEmittedThisTurn,
+        since_ts: ts,
+        clean_count: 0,
+        window_tools: [],
+        watched_fps: state.last_errors.map(e => e.fp),
+      };
    } else if (state.recoveryWatch) {
      const turnHadError = fp !== null;
      if (turnHadError) {
@@ -485,6 +577,12 @@ function main() {
            active_skills: activeNames(state, "skill"),
            active_agents: activeNames(state, "agent"),
          });
+          state.recentRecoveries.push({
+            recovered_from: state.recoveryWatch.recovered_from,
+            fps: state.recoveryWatch.watched_fps || [],
+            emitted_at_count: state.session_post_count,
+          });
+          if (state.recentRecoveries.length > RECENT_RECOVERIES_MAX) state.recentRecoveries.shift();
          state.recoveryWatch = null;
        }
      }
@@ -126,7 +126,7 @@ copy_file "$SRC/adam/scripts/adam-archive.mjs"                       "$DEST/adam
 copy_file "$SRC/adam/scripts/adam-upgrade.mjs"                       "$DEST/adam/scripts/adam-upgrade.mjs"
 # v0.3.3 helper scripts — invoked from SKILL.md / hooks / analyst flow
 for _adam_script in adam-utils adam-window adam-explain adam-nudge-eligibility adam-cooldown \
-                    adam-score adam-ab-measure adam-apply-reinforcement; do
+                    adam-score adam-ab-measure adam-apply-reinforcement adam-batch adam-rollback; do
  copy_file "$SRC/adam/scripts/${_adam_script}.mjs" \
            "$DEST/adam/scripts/${_adam_script}.mjs"
  run "chmod +x \"$DEST/adam/scripts/${_adam_script}.mjs\""
@@ -65,32 +65,86 @@ Filter to `status == "regressed"` before passing to the analyst as
 effectiveness") to surface a `## Regressions` section at the top of its output
 when this list is non-empty. If the script fails: log stderr, pass `[]`.

-### 2. Dispatch the analyst
+**Auto-rollback** (MOSS §3.5): if any entries have `status == "regressed"`, run the rollback script to auto-revert them before analyst dispatch:

-Use the Agent tool with `subagent_type: "adam"` and prompt:
+```bash
+node ~/.claude/adam/scripts/adam-rollback.mjs --auto --home ~/.claude > /tmp/adam-rollback-results.json 2> /tmp/adam-rollback.log
+```
+
+For each rolled-back proposal, print to user: `adam: rolled back "<proposal_id>" — regression detected (delta: <delta_pct>%)`. The rollback script moves the proposal from `applied/` back to `proposals/` with `rolled_back: true` and creates a regression nudge. If the script fails: log stderr, continue (rollback is best-effort).
+
+**Evidence batching** (MOSS §3.1): pre-cluster the windowed journal into coherent failure batches:
+
+```bash
+node ~/.claude/adam/scripts/adam-batch.mjs --input /tmp/adam-windowed-journal.jsonl > /tmp/adam-batches.json 2> /tmp/adam-batch.log
+```
+
+This groups entries by (signal_type, cluster_key) and reports per-batch metadata including `has_context_window` (whether transcript evidence is attached). If the script fails: log stderr, pass `null` to the analyst (graceful degradation — analyst falls back to raw journal clustering).
+
+### 2. Dispatch the analyst (two-stage pipeline)
+
+MOSS §3.3: "A single prompt asked to diagnose, plan, implement, verify, and decide overloads context and produces lower-quality output than a sequenced flow." The analyst is dispatched in two stages with a validation gate between them.
+
+**Stage 1 — Diagnose + Plan**: Use the Agent tool with `subagent_type: "adam"` and prompt:

 ```
-Run a single analysis pass.
+stage=diagnose
+
+Read the batched journal entries, cluster by signal type, diagnose root causes,
+plan fix types, and score the keypoint matrix. Write diagnoses to /tmp/adam-diagnoses.json.
+Do NOT draft proposal files.

 Inputs:
- windowed_journal_path: /tmp/adam-windowed-journal.jsonl  # pre-filtered by adam-window.mjs
- scores_path: /tmp/adam-scores.json                       # per-session dampeners + reinforcement candidates
- ab_regressions_path: /tmp/adam-ab-regressions.json       # A/B deltas for prior auto-applied proposals
+- windowed_journal_path: /tmp/adam-windowed-journal.jsonl
+- batches_path: /tmp/adam-batches.json                     # pre-clustered evidence batches
+- scores_path: /tmp/adam-scores.json
+- ab_regressions_path: /tmp/adam-ab-regressions.json
 - journal_path: ~/.claude/adam/journal.jsonl               # raw — fallback only
 - state_path: ~/.claude/adam/state.json
 - usage_path: ~/.claude/adam/usage.json
+- applied_dir: ~/.claude/adam/applied/
+- rejected_dir: ~/.claude/adam/rejected/
+- transcripts_root: ~/.claude/projects/
+- skills_root: ~/.claude/skills/
+
+Use batches_path for pre-clustered evidence when available. Prefer context_window
+fields in journal entries over transcript file lookups. Write /tmp/adam-diagnoses.json
+per the "Diagnose-stage output format" in your system prompt.
+```
+
+Wait for return.
+
+**Inter-stage validation** (§2a): after stage 1 returns, read `/tmp/adam-diagnoses.json` and validate each diagnosis:
+
+1. Every `source_entries` timestamp exists in the windowed journal (read `/tmp/adam-windowed-journal.jsonl`, check timestamps match).
+2. Every diagnosis has all four fields (`trigger`, `action`, `mismatch`, `outcome`).
+3. The planned `type` is a valid proposal type.
+4. Remove diagnoses that fail validation — log a one-line warning per removal.
+
+If all diagnoses are removed or the file is missing/empty, print "adam: no valid diagnoses — nothing to implement" and skip to §6.
+
+**Stage 2 — Implement**: Use the Agent tool with `subagent_type: "adam"` and prompt:
+
+```
+stage=implement
+
+Read the validated diagnoses and draft full proposal files.
+
+Inputs:
+- diagnoses_path: /tmp/adam-diagnoses.json                 # validated stage-1 output
+- windowed_journal_path: /tmp/adam-windowed-journal.jsonl
+- scores_path: /tmp/adam-scores.json
+- ab_regressions_path: /tmp/adam-ab-regressions.json
+- state_path: ~/.claude/adam/state.json
+- usage_path: ~/.claude/adam/usage.json
 - proposals_dir: ~/.claude/adam/proposals/
 - applied_dir: ~/.claude/adam/applied/
 - rejected_dir: ~/.claude/adam/rejected/
 - transcripts_root: ~/.claude/projects/
 - skills_root: ~/.claude/skills/

-The windowed_journal is already filtered by per-signal age (see
-SIGNAL_WINDOWS_DAYS in adam-window.mjs) AND by actioned-exclusion. Read it as
-your primary input — do not re-apply window math. Fall back to journal_path
-only if windowed_journal_path is missing or empty.
-
-Follow your system prompt exactly. Emit a single JSON punch list as your final message.
+Draft proposal files to proposals_dir/ for each diagnosis. Score against the
+confidence rubric. Emit the clustering trace and punch list as your final message.
 ```

 Wait for return.
@@ -112,11 +166,29 @@ node ~/.claude/adam/scripts/adam-explain.mjs --mode full       # verbatim trace
 node ~/.claude/adam/scripts/adam-explain.mjs --mode json       # machine-readable
 ```

-### 3. Auto-apply high-confidence items
+### 3. Pre-apply verification gate (MOSS §3.4)
+
+MOSS §3.4: "Verification must therefore be runtime, on a production-equivalent environment, and against the same prompts that produced the failure evidence." Before auto-applying, verify each proposal deterministically:

 For each id in `high_confidence`:
 - Read the proposal file from `~/.claude/adam/proposals/<id>-*.md`.
- Verify in front of the user: print `id`, `target`, `confidence`, `blast_radius`, `cross_session_evidence`, `auto_apply_eligible`.
+- **Verification checks** (all must pass for auto-apply to proceed):
+  1. **Source entries exist**: every timestamp in `source_entries` frontmatter must appear in `/tmp/adam-windowed-journal.jsonl`. If any are missing, the evidence is stale or was already actioned — demote to `queued`.
+  2. **Diagnosis grounded**: the `# Diagnosis` section must have all four fields (Trigger, Action, Mismatch, Outcome) with ≥1 backtick-wrapped quote. If malformed, demote to `queued`.
+  3. **Type-evidence match**: the proposal `type` must match what the evidence supports:
+     - `correction` signals → `memory`, `skill_new`, `skill_edit` (not `nudge`)
+     - `dead_end` signals → `nudge`, `skill_new`, `skill_edit` (not `memory`)
+     - `tool_error_loop` signals → `memory`, `skill_new`, `skill_edit`
+     - `harness_edit` → must cite harness-level evidence (false negative, scoring bias, window miscalibration)
+     If mismatch, demote to `queued`.
+  4. **No conflicting applied proposal**: grep `~/.claude/adam/applied/` for any proposal with the same `target` applied in the last 7 days. If found, demote to `queued` (prevents stacking rapid edits).
+- Print verification result: `verified: <id> (4/4 checks passed)` or `demoted: <id> (failed: <check_name>)`.
+- Demoted proposals are moved from `high_confidence` to `queued` for manual review.
+
+### 3a. Apply verified high-confidence items
+
+For each id that passed verification:
+- Print `id`, `target`, `confidence`, `blast_radius`, `cross_session_evidence`, `auto_apply_eligible`.
 - Apply the change:
  - **For `skill_new`**: `mkdir -p ~/.claude/skills/<slug>/`, then `Write` the proposal's `# Proposed change` body to `~/.claude/skills/<slug>/SKILL.md`. After write, print: "skill `<slug>` written to `~/.claude/skills/<slug>/SKILL.md` — activates immediately — Claude Code v2.1.0+ auto-hot-reloads user-level skills, no restart needed."
  - **For `memory`**: `Write` the proposal's `# Proposed change` body (which MUST include the auto-memory frontmatter — see "Memory drafting protocol" in `agents/adam.md`) to the path in `target`. Then update `MEMORY.md` index with a one-line pointer.
@@ -174,6 +246,12 @@ c. On **approve**:
   - For `skill_new`: `mkdir -p ~/.claude/skills/<slug>/`, then write `# Proposed change` body to `<slug>/SKILL.md`. Tell user: "skill `<slug>` written — activates immediately (CC v2.1.0+ auto-hot-reload)."
   - For `skill_edit`: apply the unified diff in `# Proposed change` to the existing SKILL.md at `target` (append-only — never replace existing content).
   - For `memory`: write `# Proposed change` body (must include auto-memory frontmatter) to `target` and update `MEMORY.md` index with a one-line pointer.
+   - For `harness_edit` (MOSS §1): apply the unified diff to the target harness file. **Before applying**:
+     1. Run `bash ~/.claude/adam/tests/run-tests.sh` — capture pass count.
+     2. Apply the diff via `Edit`.
+     3. Run `bash ~/.claude/adam/tests/run-tests.sh` again — verify pass count is equal or higher and 0 failures.
+     4. If test regression: revert the edit, print "harness_edit reverted — test regression detected", leave proposal in `proposals/`.
+     5. If tests pass: tell user "harness edit applied to `<target>` — tests pass (<N> passed)."
   - For all others: apply via Write/Edit per the proposal's `# Proposed change`.
   - Move proposal to `~/.claude/adam/applied/<ts>-<id>.md`.
   - Archive: `node ~/.claude/adam/scripts/adam-archive.mjs ~/.claude/adam/applied/<ts>-<id>.md`.
@@ -191,6 +269,10 @@ End with one block:
 ```
 adam reflect summary:
  observations processed: <new>
+  batches formed: <N>
+  diagnoses validated: <N>/<total>
+  rolled back (regression): <N>
+  verification passed: <N>/<total high_confidence>
  auto-applied: <N>
  approved: <N>
  rejected: <N>
@@ -198,6 +280,14 @@ adam reflect summary:
  failed: <N>
 ```

+**Keypoint history**: after all proposals are processed, append one JSON line to `~/.claude/adam/keypoint-history.jsonl` with the aggregate keypoint scores from the diagnose stage:
+
+```json
+{"ts":"<iso>","session":"<session_id>","keypoints":{"tool_selection":N,"scope_discipline":N,"error_recovery":N,"first_attempt":N,"build_reliability":N},"proposals_emitted":N,"proposals_applied":N}
+```
+
+This builds a longitudinal record of which capabilities are improving across `/reflect` runs.
+
 ## Karpathy constraints (you must enforce on each apply)

 Before writing any proposal:
@@ -211,6 +301,7 @@ Before writing any proposal:
 - For `skill_edit`: confirm the diff is append-only (no `-` lines that remove existing content) and that target SKILL.md exists. When auto-applying, ALSO re-verify the eligibility gate steps in §3 (cooldown, blacklist, byte cap) before any `Edit` call — never trust frontmatter alone.
 - For `skill_edit` with `auto_apply_eligible: true`: confirm `contradiction_flag` is absent or null in frontmatter. Refuse auto-apply if `contradiction_flag` is set with any non-empty value (treat the agent's flag as a hard veto on auto-apply; user can still manually approve in walk-the-queue if they disagree with the heuristic).
 - For `memory`: confirm `# Proposed change` body starts with `---` frontmatter containing required fields `name`, `description`, `type`, `originSessionId`. Refuse if frontmatter missing — agent must redraft per the Memory drafting protocol.
+- For `harness_edit`: confirm `auto_apply_eligible: false` (never auto-apply). Confirm `confidence ≥ 5`. Confirm `# Test verification` section names the test command. Confirm diff is ≤30 LOC and targets a single allowed harness file (see `agents/adam.md` §"Harness self-modification"). Run test suite before AND after applying — revert on any regression.
 - Confirm `source_entries` is present in proposal frontmatter as a non-empty list (used for archive). Warn (do not refuse) if missing — legacy proposals from before v0.2.0 won't have it.

 If any check fails, refuse to apply and ask the user how to proceed.
Author	SHA1	Message	Date
lukaszraczylo	2d9257922f	docs: update README for v0.5.0 release — MOSS-grounded improvements	2026-05-24 11:19:15 +01:00
lukaszraczylo	440fb52eb1	feat: apply MOSS-grounded self-evolution improvements to ADAM Implements 7 improvements grounded in MOSS paper (arXiv 2605.22794): 1. Transcript capture (§3.4): context_ring buffer in adam-observe.mjs captures last 8 events around struggle signals as context_window. 2. Evidence batching (§3.1): new adam-batch.mjs pre-clusters windowed journal entries into coherent failure batches by (signal_type, cluster_key). 3. Multi-stage analysis (§3.3): SKILL.md dispatches adam agent in two stages (diagnose+plan → implement) with inter-stage validation gate. 4. Pre-apply verification (§3.4): 4-check deterministic gate before auto-apply (source entries exist, diagnosis grounded, type-evidence match, no conflicting recent proposals). 5. Auto-rollback (§3.5): new adam-rollback.mjs reverts regressed proposals detected by A/B measurement, creates regression nudges. 6. Harness self-modification (§1 Table 1): new harness_edit proposal type targeting adam's own scripts with stricter gates (confidence≥5, never auto-apply, test-suite-gated). 7. Keypoint matrix evaluation (§4.2): 5 capability dimensions (tool_selection, scope_discipline, error_recovery, first_attempt, build_reliability) scored per batch for structured evaluation. Test suite: 94 → 114 tests (20 new), all passing.	2026-05-24 11:15:32 +01:00
lukaszraczylo	a48c705c0a	feat(adam): smarter signals & clustering - New signal types in hooks/adam-observe.mjs: - silent_drift: 5 consecutive read-only PostToolUse without an action tool - error_after_recovery: same error fingerprint returns within 5 events of clean_recovery - Severity-weighted scoring in adam/scripts/adam-score.mjs: - SEVERITY_DIVISORS exported per struggle signal type - Per-session severity_sum + severity_by_type added to JSON output - Skill-attribution clustering in agents/adam.md: - Sub-cluster struggle signals on active_skills[0] - New struggle-driven skill_edit variant (always queues, never auto-applies) - Rubric updates: - +1 for cluster severity-sum >= 10, additional +1 for >= 32 - +1 for skill-attributed sub-cluster naming an existing skill - silent_drift + error_after_recovery added to struggle signal list - Window: silent_drift 14d, error_after_recovery 30d - Tests: 94 passing (78-82 new) Backward compat: entries without count default to severity 1. Existing win-driven skill_edit gate untouched. No journal migration.	2026-05-13 19:21:59 +01:00
lukaszraczylo	a8883aa8b7	fix(logo): explicit light/dark variants + <picture> for GitHub The prior logo.svg used currentColor, which resolves to black when the SVG is loaded via <img> on GitHub — making the logo invisible in dark mode (the GitHub default for many users). Fix uses GitHub's supported <picture> + prefers-color-scheme media- source pattern in README: - assets/logo-light.svg — explicit GitHub light-theme text color #24292f - assets/logo-dark.svg — explicit GitHub dark-theme text color #f0f6fc - assets/logo.svg — kept with embedded @media + currentColor for standalone use (markmorph notes, anywhere else the SVG is loaded outside <picture>) README updates the <img> tag to a <picture> with media-conditioned source so GitHub's renderer picks the right variant per theme.	2026-05-13 02:07:11 +01:00
lukaszraczylo	7ed2aecdfa	docs(logo): swap to swaddled-baby design with hands Replaces the geometric-A-with-observation-dot with a softer, more on-theme design: a swaddled-baby silhouette (rounded A-shape bundle), face nestled inside, and the wrap-band extended past the bundle on both sides as little hands. Maintains currentColor + zero external assets; reads cleanly down to favicon size. Ties the visual identity to the 'Story behind Adam' section: the project is named after the author's son, and now the logo is too.	2026-05-13 02:02:02 +01:00
lukaszraczylo	a30f8b1158	docs: replace ASCII pipeline diagram with mermaid flowchart GitHub renders mermaid natively. Diagram now shows three subgraphs (Observation → Analysis → Review + apply) with a nested Pre-processors subgraph inside Analysis. Includes: - Dotted edge labeled 'user runs /reflect' marking the observe→analyze boundary. - Diamond gate node for auto-apply decision (conf≥4 · low blast · cooldown cool) with explicit yes/no branches. - Feedback loop: applied/ entries measure back into adam-ab-measure.mjs on subsequent reflects. - Color-coded classDef for stores (blue), processes (orange), and the clustering trace artifact (purple). ASCII art retired — diagram now legible at any zoom on github.com.	2026-05-13 01:54:38 +01:00
lukaszraczylo	d3e4350d71	docs: modernize README + add SVG logo + inspiration story - New 'Story behind Adam' section at the top: the project is named after the author's newborn son, whose observe-act-adjust-observe-again learning loop is the methodology ADAM applies to LLM sessions. - New SVG logo at assets/logo.svg: stylized 'A' with a captured observation point inside the apex and a feedback crossbar. Uses currentColor + gradient so it adapts to light/dark GitHub themes. - Centered header block with project tagline + 5 badges (License, Version, Tests, Node, Platform). - New 'Highlights' section: 8 emoji-tagged one-liners covering the v0.3.3 design pillars (zero LLM cost observation, A/B measurement, sliding windows, observability, etc.). - New 'How it works' ASCII pipeline diagram: observation -> analysis pre-processors -> analyst -> review + apply. - Signals table now includes per-signal sliding window column. - Rubric section restructured: gates, modifiers (dampener), and skill_edit-specific requirements clearly separated. - New 'Inspecting the analyst's reasoning' section documenting adam-explain.mjs + /reflect --explain. - Layout updated for v0.3.3 state files (active-nudges.json, ab-tracking.jsonl, reinforcements.jsonl, last-trace.txt) and all 9 new helper scripts under adam/scripts/. - Test count: 27 -> 87. - Closing line crediting Adam.	2026-05-13 01:50:59 +01:00