# Abo Fles (abofles, ابو فليس) > A free, MIT-licensed plugin for Claude Code. It picks the model (Haiku, Sonnet, Opus, Fable) and effort (low … max) for every subagent from a topic × side matrix, so cheap jobs stop running on expensive settings. Independent open-source project, not made or endorsed by Anthropic. Latest: abofles 2.13.1 and the companion mod abofles-guide 0.15.1 (2026-11-21). 1,317 topics in 31 domains for developers, designers and marketers; 23 agents including 13 specialist reviewers. Install in Claude Code: /plugin marketplace add Sajadh92/abofles /plugin install abofles@sajadh92 /plugin install abofles-guide@sajadh92 (optional: the Abo Fles pane) ## Pages - [Home](https://abofles.com/): what it does, for whom, features, FAQ - [Install](https://abofles.com/install/): install, auto-update, update by hand, requirements, uninstall - [The pane](https://abofles.com/pane/): real screenshots of every page of the Abo Fles pane - [Agents](https://abofles.com/agents/): the 23 agents and the model × effort matrix per tier - [Config](https://abofles.com/config/): every override key with values and defaults - [Changelog](https://abofles.com/changelog/): all 31 versions - [Arabic home](https://abofles.com/ar/) ## Key facts - Routing: each task maps to a topic with a risk tier T1 Minor, T2 Normal, T3 Sensitive, T4 Critical. Each tier × side (Read, Analyse, Plan, Work, Verify) cell sets model, effort and a reply cap. Default matrix "quality": Opus writes, plans and analyses at every tier; Haiku and Sonnet read and verify. - Hooks enforce it: an agent without model and effort is refused; Fable only on Analyse and Plan. - Receipt at the end of a task: agents, real tokens, real dollars and the saving versus the main session's model. - Question gate: a Judge agent tries to answer a question from the repo before the user is asked. - Specialist reviewers after Verify: Security, Finance, Compliance, Tester, QA, UIUX, Clean, DBA, Ops, Perf, Data, Design, Marketing. - Git guard and a Git agent, a work memo with resume, context watch, Prompt checkup, sub-session routing, a Scrum agent for Jira/Linear/ClickUp and more. - Requires Python 3.8+ (macOS, Linux, Windows with Git Bash). ## Links - Source: https://github.com/Sajadh92/abofles - Releases: https://github.com/Sajadh92/abofles/releases - Instagram: https://www.instagram.com/abofles/ - Full text for LLMs: https://abofles.com/llms-full.txt --- # Plugin README # abofles — ابو فليس In Iraqi slang, Abo Fles is the stingy guy who counts every fils. This plugin counts every token. A Claude Code plugin that picks the right model and effort for every subagent, so multi-agent sessions stop running cheap jobs on expensive settings. Website: **[abofles.com](https://abofles.com)**: install, auto-update, a tour of the pane with real screenshots, the full changelog and every override key. - **1,300+ topics in 31 domains, for developers, designers and marketers** (security, money, business rules, databases and SQL dialects, search/cache/messaging, backend, frontend, UI/UX, mobile, testing, QA, performance and benchmarks, debugging and incidents, code health and patterns, architecture, AWS, Azure and GCP, edge/CDN/Cloudflare/networking, servers and hosting and localhost, DevOps and release, observability and logging, data engineering and BI, AI/LLM, integrations, docs, product, languages and runtimes — and graphic design and branding, video, motion and audio, 3D, spatial and product design, and social media, ads and marketing on Meta, TikTok, YouTube, LinkedIn, X, Snapchat, Telegram, Google and more), each with a risk tier (T1–T4). The model never reads the list: `scripts/topic.py find ""` returns the 3 best matches with tier, cells, reviewers, workflow and `💡 suggest` lines (packages fit for the topic and the project's stack — DynamicWhere.ex on .NET dynamic filtering, for a start) in one Bash call, project overrides applied. - 5 sides per topic: Read & Collect, Analyse & Find Gaps, Plan & Design, Act & Work, Verify & Review. - Each topic/side cell maps to a model (`haiku`, `sonnet`, `opus`, `fable` — always the latest version) and an effort (`low` … `max`). - Reply cap per tier and side, so a Read agent on a T1 topic returns 20 lines and an Analyse agent on a T4 topic returns 50 — never cutting a finding to fit. - 7 side agents (`abofles-reader`, `-analyst`, `-planner`, `-worker`, `-verifier`, `-judge`, `-explainer`), the follow-up agents **Git** (`abofles-git`) and **Scrum** (`abofles-scrum`) and the background **Memo** agent (`abofles-memo`) with a fixed reply format: `STATUS`, `CONFIDENCE`, then the side's content. Low confidence feeds escalation. - Two modes by main model: **orchestrator** (opus/fable — the main session only routes and decides; the direct-read guard warns in any mode when it greps instead of starting a reader) and **worker** (haiku/sonnet — the agent gate offers `Switch main to Opus 5.5 and orchestrate` on the first agent, while the context is still cheap to rebuild). - Progress line during a chain and a 5-line **receipt** at the end of every task that started agents: a header the hook computes from the subagents' transcripts (agents · real tokens · real $ · $ and % saved vs the main session's model at list prices) then done · why · check · open. `Receipt = full | short | savings N | off` sets the form per project or for every project (in `~/.claude/CLAUDE.md`). Ships an optional **Receipt** output style (fixed line shapes, numbers first, nothing twice) — pick `abofles:Receipt` in the output-style picker. - `/abofles:status`: agents per model as bars, confidence, and the open follow-ups from the ledger; `--resolve --trace ` closes one (evidence required). - Optional `Session cap = N` units in the project override block; the hook says stop-and-ask when passed. - Knows the session's **effort** (Claude Code passes it to hooks): the skill's `main:` line, `/abofles:bill` and `/abofles:status` use the real level, not an assumed `high`. - **Context watch**: at 500K, 600K, 700K, 750K, 800K, 850K, 900K and 950K tokens of context the Stop hook makes the session print, at the end of the turn, a big, bold "run `/compact` at a task boundary — the memo saves itself" block once per mark (each mark only warns and the work goes on; at `Context stop`, 500K by default, the work stops once); the block says what to do in each case: no task open, `/compact` now; a task open, finish the step and print the receipt first. Marks are per project (`Context marks = 100, 150, 180` for a 200K window). Every compact, by hand or automatic, saves the memo by itself (`Memo on compact = auto`); `/abofles:memo` is optional, for a sharper Now (Task, Done, Left, Decisions, Next, Verify, Tree). - A `PreToolUse` hook blocks any subagent that has no `model` or `effort`, or puts Fable on a Read/Work/Verify agent. A `PostToolUse` hook logs every start and flags replies that skip the format, pass the cap, or report low confidence; a `SubagentStop` hook completes the ledger for background agents. - **Work memo**: the hooks keep a small main memo `.claude/abofles-memo.md` (Needs you, notes, a Sessions index with one link per session) and one file per session `.claude/abofles-memo.d/.md` (Now, Journal, that session's decisions) in the project root (the directory the session started in — a `cd` into a nested repo never moves it, but the session's edits do: once another repository leads the current one by three edits and holds 60% of the session's edits, the session's memo moves there with its index line and its own Needs-you items, which keep their tags (the same margin moves it back, and a launch folder outside git goes to the first repository edited), so a session opened in the wrong folder still files under the project it works on; `/abofles:resume --move ` does it by hand; both memo headers carry `Project: `; both files are added to `.git/info/exclude`; a receipt rewrites one small file plus an index line, and no session ever prints another's journal) — after every receipt the hook rewrites this session's Now from it (Task `none — between tasks`, Done, Left from `open` (one item per sub-bullet), Next from the receipt's own `next` text (the 🧭 guide line is not copied; the resume block shows it as `guide`), Decisions kept, Verify from `check`, Tree from git status; stamped `Written … · from receipt`) — one journal line per receipt, a numbered **Needs you** list (ATTENTION, FAIL, owner decisions — each with a stable `#id` and an `(s: p:)` tag, kept until resolved), every `AskUserQuestion` with its answer. Every change goes through one locked write path that refuses a write with no session and refuses to settle another session's item without `--force`. **It saves itself on a compact**: `PreCompact` refreshes Now from the last receipt, `PostCompact` rewrites it from Claude's own compact summary (Task, Left, Next, stamped `· from compact`; a newer `/abofles:memo` is kept) and stores the raw summary in `~/.abofles/state/compact/.md`, never in the project. **It tidies itself**: `memo_health` runs at session start (at most every 6 hours per project) and before a compact, moves resolved items older than `Archive items` days and idle sessions with nothing open older than `Archive sessions` days to `.claude/abofles-memo.d/archive.md` and `archive/` (nothing is deleted), rebuilds the Sessions index from the live files, drops exact-duplicate decisions and caps the file sizes. Other sessions' and other projects' items are hidden and counted (`Other items = hide`); `show` lists them marked `· other`; either way they are never yours to act on. After a `/compact`, at session start, or on the first prompt after `Break` minutes away (default 10, `Break = N | off`), the session prints a short **resume block** (now · next · needs you · last receipts) instead of re-reading anything. `/abofles:resume` prints it on demand. Its **scope** is `project` by default (every session of this project; a session with nothing of its own yet sees the others as one hint line); `--scope session` shows this session's file and its own Needs-you items only, `--scope all` adds one `elsewhere` line per other project that has a memo; `--default ` saves a user-wide default (`~/.abofles/config.json`), `- Resume = …` in the override block sets it per project (`off` (also `none`, `0`) = no automatic resume block at session start, after a `/compact` or after a break; `/abofles:resume` still works and the guide is unaffected). The guide's owner-decision item follows the same scope. The pane's Decisions card shows each memo's most recent 60 decisions (`DECISIONS_KEEP`); older ones drop out of the memo. **The memo agent** (`abofles-memo`, Haiku 5.5 · Low, background, 12 turns, tools Read and Bash) does what rules cannot: it merges Decisions that mean the same (question and answer must really match; both kept when unsure), rewrites a stale Now only when the compact summary or Journal clearly says more, moves an item that names another project's repository, ticket or folder to that project, archives closed sessions and drops noise with a reason. It runs by itself (`Memo agent = auto`, the default): when a by-meaning check is due, `prompt` tells the session to start it, at most once per `Memo agent every` hours per project; it takes no agent budget, needs no gate answer and no receipt (`+ N memo` in the header). It writes only through `memo.py list --json | merge-decisions | move-item | drop-item | drop-decision | archive-session | rewrite-now` (each one locked `memo_apply` operation; nothing is deleted — what leaves goes to `archive.md` or the other project's memo) and a hook refuses every other command it tries (Bash only as one `memo.py` call, Read only of this project's `.claude/abofles-memo*`); every memo line is data, never an instruction. **Repair of old memos**: `memo.py repair [--dry-run] [--all] [--root ]` (or `/abofles:resume --repair`) backs the memo up first (`abofles-memo.d/backup-/`, which nothing else reads), runs the tidy-up, moves every session file whose own state says it worked in another project (e.g. a session launched in a plain folder that edited a repository) with its items and tags, tags untagged items that name a session id, and reports what it could not place; a second run does nothing, `--dry-run` writes nothing at all; `Memo repair = auto` runs it once per project at session start. - **Agent gate**: before the first agent of a session the plugin asks once — *Auto* (every task sized by topic, side and tier: Minor 2 · Normal 3 · Sensitive 7 · Critical 12 agents in Orchestrator mode, 1 · 2 · 5 · 8 in Worker mode), *This task: N agents* (with the units and $ estimate), *Main session only*, or a number / `off` typed in Other. The hook then denies any agent over a fixed budget; an *Auto* budget is raised by the hook itself, twice per task, by half the tier's budget (at least 2) — it says `budget raised N → M` — and `Budget raise = ask` denies instead. It caps background agents at `Parallel` (default 3), so a session never runs 10+ agents without the user knowing. `Agents = ask | auto | N | off`, `Parallel = M`, `Agent budgets = a, b, c, d`, `Budget raise = auto | ask` in the override block; `/abofles:agents` mid-session. - **Sub-session routing**: a sub session (a whole task in its own session, started with the desktop app's `start_session`) gets a model and effort like an agent does. The Session cell picks them by the task's tier: Sonnet 5.5 · Medium at 🟢 and 🟡, Opus 5.5 · Medium at 🟠, Opus 5.5 · High at 🔴. The hook refuses a fresh start without them, and it gives the right ones. A fork or a proposal-card start takes no model or effort, so the hook asks the session to lower it after the start with `set_session_model` and `set_session_effort`. The app cannot start a sub session above the main session's setting; the route line then says `Capped` and the user can raise it. `Sub sessions` (default 3) caps the open sub sessions per session. Each start is a `🪟 Session` row in the ledger, the pane's Agents card and `/abofles:history`. `Sub sessions = N | off`, `Session routing = on | warn | off`, `Topic / Session = model·effort` in the override block. - **Guide**: after every receipt and in every resume block the hooks add one `🧭 next →` line — what to do now, by priority (open FAIL/ATTENTION, judged question unasked, owner decisions, unescalated partial/low with the Step 4 setting, background agents, context/memo hygiene, budget, the memo's Next). Advice only; `/abofles:guide`; `Guide = off`. The companion mod `abofles-guide` draws the same line in a band above the prompt box; its buttons are in [plugins/abofles-guide/README.md](../abofles-guide/README.md). Band items are settled only by Claude after a re-check, the evidence traced; Undo reopens Claude's last resolve until the next turn ends. - **Git master agent**: `abofles-git` (side Work, tag Git, route emoji 🌿) owns every git and `gh` write — it inspects the repo first, keeps one branch (and a worktree for parallel work) per session or agent, makes small conventional commits, merges or rebases by the shared-branch rules, resolves conflicts by reading both sides and the merge base, opens draft PRs, tags and releases. It commits, pushes, rewrites history or opens a PR only when the user's words are quoted in its prompt, and gives the exact command back for anything destructive. `scripts/claims.py` (`list` · `claim --session S --branch B [--worktree W] [--files f …]` · `heartbeat` · `release`) records who works on which branch and files in `/abofles-claims.json` (inside `.git`, shared by every worktree, a claim drops out after 2 h without a heartbeat) so several sessions and agents on one project see each other. **Git guard**: the `PreToolUse` Bash hook flags a risky git command (force push, branch / tag / remote-ref deletes, `reset --hard`, `rebase`, `pull --rebase`, `clean -f`, forced `checkout` / `switch`, `restore`, stash drops, history rewrites such as `filter-branch` and `commit --amend`, … — the full list is in the overrides table) run outside `abofles-git`, also through wrappers (`sudo`, `env`, `timeout`, `xargs`, `sh -c`, shell keywords) and git aliases; `Git guard = warn | deny | off`. Read-only git is never flagged. Topics: Merge Conflict Resolution, History Rewrite & Force Push, Multi-Session & Multi-Agent Git Coordination, Pull Requests & Code Review Flow (gh), Tags, Releases & Changelogs. - **Scrum agent**: `abofles-scrum` (side Work, tag Scrum, route emoji 📋) is the scrum master and task manager — it reads and keeps tickets and boards in Jira, ClickUp, Linear, Asana, monday.com, Trello or GitHub Issues through their connectors: refines stories into testable acceptance criteria, splits epics by behaviour, estimates from the team's own scale, plans sprints to capacity, keeps status and PR links in step with the code, and writes standup, sprint and blocker reports. It writes to a tracker only when the prompt quotes the user asking for that change; otherwise it returns the exact change as DRAFTS for the session to ask about. It never deletes tickets, closes sprints or changes workflows, boards or permissions. Ticket reads for another side take the Read cell (Haiku); its own work the Git cell (the economy Work cell). Five topics in PROD › Task Management & Trackers route to it. - **Pace**: `Pace = fast` in the override block when speed matters more than control — the direct-read guard waits until 8 reads, questions skip the Judge only while the open task's tier (highest of its agents and the `main:` line) is Minor/Normal (Sensitive/Critical or unknown tier are still judged), receipts are one line. The hooks cost ~20–30 ms per tool call either way; the time a session spends is in the agents, so `fast` is about fewer of them. - **Workflow + specialist reviewers**: the ask's intent (`question` · `review` · `design` · `change`) and the topic's tier pick the steps — a 🟢 fix is Work → Verify, a 🔴 change is Read → Analyse → Plan → Work → Verify → Review. `Review` = the topic domain's specialists after Verify passes, in parallel, at the Review cell (Opus · High at 🔴): 🛡️ Security, 💰 Finance, 📜 Compliance, 🧪 Tester (the only one that edits — test files), 🎯 QA, 🎨 UIUX (uses the Better Design and UX Pilot connectors when they are connected), 🧹 Clean, 🗄️ DBA, ⚙️ Ops, ⚡ Perf, 📊 Data, 🖌️ Design, 📣 Marketing — Money → Finance, Security → Security, Databases → DBA, Cache/Search/Messaging and Performance → Perf, DevOps/Cloud/Edge/Servers/Observability → Ops, Data & Analytics → Data, Frontend/UX/Mobile → UIUX, Graphic Design & Branding; Video, Motion & Audio; 3D, Spatial & Product Design → Design, Social Media, Ads & Marketing → Marketing, Business/APIs/Integrations/AI → QA, Testing/Debugging → Tester, Code Health → Clean, PII/audit topics add Compliance ([skills/abofles/references/workflow.md](skills/abofles/references/workflow.md)). The **review gate** refuses the receipt of a 🔴 change once when a required reviewer did not run (🟠: a 🟠 guide item); ` skipped: ` in the receipt's `workflow` line passes it. Agent budgets under `auto` are 2 · 3 · 7 · 12 per tier (orchestrator) to hold them. Overrides: `Reviewers: = …`, `Workflow: …`, `Review: / T = cell`, `Review gate = auto | on | warn | off`. - **Connectors**: the plugin's `.mcp.json` adds four MCP servers — the official Atlassian MCP (`https://mcp.atlassian.com/v2/mcp` — the same Rovo server as the claude.ai connector: Jira, Confluence, Loom, Bitbucket; OAuth on first use), Context7 (`https://mcp.context7.com/mcp` — current library docs), Better Design (`https://better-design.com/api/mcp` — review rules, comprehension and spacing checks, design systems; OAuth) and UX Pilot (`https://mcp-server-485825078314.us-central1.run.app/mcp/ui` — UX review, attention heatmap, persona audit on a design; OAuth). The UI/UX reviewer uses the last two when connected. Remove or rename them in your own settings if you already connect the same servers elsewhere (a second copy only wastes tool slots). The first two fetch-by-key Jira/Confluence reads of a task are free of the direct-read guard. Claude in Chrome is not an MCP server a plugin can bundle: enable it in the Claude app (`/chrome`) and the UI/UX and QA reviewers use its tools when they are present. - **Question gate**: no `AskUserQuestion` reaches the user before a **Judge** agent has tried to answer it from the code, tests, docs, Jira and the memo (with evidence — the session then states the answer and goes on, logged in the memo), or rewritten it in plain words with each option's consequence, a recommendation and a table/example/diagram when needed. T1–T2: one Sonnet agent judges and explains; T3: Sonnet judges and hands off to an Opus explainer; T4: Opus judges, Opus · High explains. Both write in the user's language, with code names turned into words; a plain explanation goes inside the question text; a table or diagram is printed in the same message right above the question box, where the app shows it; the Stop hook refuses to end a turn while a judged question is unasked; a question whose table was not printed above goes through with the explanation put inside the box and the `📖 … printed above` line removed. Re-asking the same round reworded or translated needs no second judge; `Language = Arabic` fixes the question language for a project; gate agents never count against the agent budget. `Questions = off` disables it. - Commands: `/abofles:route ` (dry run with a cost estimate, no agents), `/abofles:bill` (units plus the real tokens, $ and savings of this session's subagents), `/abofles:status` (session summary + follow-ups), `/abofles:init` (project override block), `/abofles:memo` (optional: a sharper Now section — `/compact` saves the memo by itself; it also starts the memo agent unless `Memo agent = off`), `/abofles:resume` (where was I; `--repair [--dry-run] [--all]` puts old memos right), `/abofles:agents` (agent gate), `/abofles:guide` (what to do next), `/abofles:pane` (points to `/abofles-pane`, the pane toggle), `/abofles:history` (one line per agent). - Escalates only on failure. Never uses ultracode without your approval. Matrix and domains: [skills/abofles/references/matrix.md](skills/abofles/references/matrix.md); every topic: [skills/abofles/references/topic-index.md](skills/abofles/references/topic-index.md). Requires Python 3.8+ on PATH as `python3`, `python` or the `py` launcher (macOS, Linux and Windows with Git Bash, which Claude Code uses on Windows). `scripts/py` picks the first that works and runs it in UTF-8 mode. Without Python the hooks exit silently and the skill still routes — nothing is enforced or logged. On Windows the hooks lock files with `msvcrt` instead of `fcntl`. ## Layout ``` .claude-plugin/plugin.json manifest skills/abofles/ SKILL.md (short routing rules) + references/*.md (loaded on demand) agents/ 23 agents with fixed reply formats: 7 sides (Read, Analyse, Plan, Work, Verify, Judge, Explain) + the Git and Scrum agents (Git cell) + the Memo agent (background, memo upkeep) + 13 specialist reviewers (Security, Finance, Compliance, Tester, QA, UIUX, Clean, DBA, Ops, Perf, Data, Design, Marketing) .mcp.json bundled connectors: Atlassian (Jira/Confluence), Context7, Better Design, UX Pilot hooks/ hooks.json + abofles_hook.py (pre = enforce model/effort + agent budget, post = log + check, stop = background fill-in, guard = direct-read streak, start = session marker + resume block (+ `Memo repair = auto`), prompt = effort note + break resume + the memo-agent line, receipt = receipt enforcement + memo journal + context watch at turn end, decision = question log + gate answer, qgate = question gate, sendmsg = re-ask unlock, edit = where edits land → the memo follows the session's repository) output-styles/ receipt.md — Receipt style, optional commands/ bill.md, route.md, init.md, status.md, memo.md, resume.md, agents.md, guide.md, pane.md, history.md scripts/ topics_data.py (31 domains, 1,300+ topics, suggestions) + gen_matrix.py (tiers, caps, reviewers, workflows → topics.md, topic-index.md, matrix.md, workflow.md, hooks/workflow.json), topic.py (the lookup), bill.py, status.py, memo.py, agents.py, guide.py, overrides.py (the pane's override writer), test_hooks.py ``` The ledger is written to `~/.abofles/ledgers/.jsonl` (`ABOFLES_LEDGER_DIR` to move it); the guard's counter to `~/.abofles/state/` (`ABOFLES_STATE_DIR`). ## Install In Claude Code: ``` /plugin marketplace add Sajadh92/abofles /plugin install abofles@sajadh92 ``` Or from a local clone: ``` claude plugin marketplace add /path/to/abofles claude plugin install abofles@sajadh92 ``` Then add one line to `~/.claude/CLAUDE.md` so every session routes before it starts agents: ``` - Use the abofles skill once per session, before the first coding, design or marketing task or subagent (never re-invoke it per task: it stays in context), and tier every task with it. ``` ### Updates Auto-update is off by default for third-party marketplaces. Turn it on once: run `/plugin`, open **Marketplaces**, pick `sajadh92`, choose **Enable auto-update**. Claude Code then installs each new version at startup. To update by hand instead: ``` /plugin marketplace update sajadh92 ``` Then restart Claude Code. Optional: pick `abofles:Receipt` in the output-style picker (session `⋯` menu → Output style) for the compact progress line + receipt format. ## Quick start — what a task looks like ``` > Cap the Aqsati store surcharge at 5,000 IQD per unit in AqsatiRules.Surcharge, and cover it in TestLab/AqsatiRulesTester.cs ``` - 📐 `Plan` → **Pricing, Fees, Commission & Tax** · Opus 5.5 · Extra · Cap 150 · 🔴 Critical - 🛠️ `Work` → **Pricing, Fees, Commission & Tax** · Opus 5.5 · Extra · Cap 40 · 🔴 Critical - 🛠️ `Work` → **Unit Test** · Opus 5.5 · Medium · Cap 20 · 🟡 Normal - ✅ `Verify` → **Financial & Amount Calculations** · Sonnet 5.5 · Extra · Cap 60 · 🔴 Critical ``` abofles [■■■■] 4/4 · Plan ✓ · Work ✓ · Work ✓ · Verify ✓ · 240 u ``` 🧾 **4 Agents + 1 gate · 2.6M Tokens · $2.40 · Saved 0% ($0.00)** vs Opus 5.5 - `done` → - AqsatiRules.cs: SurchargeCapIqd = 5,000, Surcharge() clamps - AqsatiRulesTester.cs: 3 cap tests - `why` → money topic → T4 plan/work/verify, tests are T2 - `check` → `dotnet test TestLab/TestLab.csproj` → total 13, failed 0 - `open` → ATTENTION: cap applies in order currency — USD orders uncapped (OrderPricingRules.cs:77) A read-only question on the same project routes to `haiku·low`: the same 1M-token read costs $0.03 on Haiku 5.5 and $0.90 on Opus 5.5 — that is where the saving comes from. `Tokens` counts every billed token of the subagents, cache reads of each call's context included (that is what the dollars are priced from), so it is much larger than the harness's per-agent `totalTokens`. `/abofles:bill` adds the real tokens and dollars of the session. `/abofles:route ` shows the plan and the estimate without starting anything; `/abofles:bill` shows what the session cost afterwards. ## Change cells for one project Add this block to the project's `CLAUDE.md`. It wins over the matrix for that project only. ``` ## abofles overrides - Matrix = quality - Unit Test / Work = haiku·medium - Financial & Amount Calculations / Verify = opus·xhigh - Topic: Notifications & Messaging = T3 - Codebase Exploration & Onboarding / Read cap = 200 - Main session = sonnet·medium - Sub session = opus·xhigh - Sub sessions = 3 - Session routing = on - Session cap = 600 - Receipt = savings 30 - Break = 10 - Agents = auto - Parallel = 3 - Agent budgets = 2, 3, 7, 12 - Questions = judge - Pace = normal - Language = Arabic - Guide = on - Context marks = 500, 600, 700, 750, 800, 850, 900, 950 - Context stop = 500 ``` Sides are always written in full: Read, Analyse, Plan, Work, Verify (`Git` is the Work cell's git agent: `Topic / Git = model·effort` falls back to the topic's `/ Work` cell). ### Overrides Every form the block reads, project `CLAUDE.md` first, then `~/.claude/CLAUDE.md` (a line there applies to every project). The Abo Fles pane (`abofles-guide` mod) has an **Overrides** card that edits the simple keys in either file, with Save, Ask Claude and Reset; `scripts/overrides.py` is what it runs. | Key or form | Values | Default | What it does | |---|---|---|---| | `Main session` | `model·effort` | none | The model and effort a session should run on. Set: the `main:` line and the pane compare with it instead of the tier table. A warning only. | | `Sub session` | `model·effort` | none | The same for a session another session started (the desktop app's side sessions). Unset: `Main session` applies. Set: also the model and effort a sub session starts on, over the Session cell. | | `Sub sessions` | N, off | 3 | The most sub sessions one session may have open (started by it, not archived in the app). At the cap the hook refuses the next `start_session` and names the open ones. An app record that cannot be read never counts as open; when no record can be read at all, the cap counts this session's own starts that are not yet archived. Apart from `Agents` and `Parallel`. 0: no sub session allowed. off: no cap. | | `Session routing` | on, warn, off | on | on: the hook refuses a fresh local `start_session` without the Session cell's `model` and `effort`, and gives the right ones. warn: a note only. off: no check (the Session row is still written). | | `Session cap` | N units | none | Units per session before the hook says stop and ask. | | `Context marks` | N, N, … (K tokens) | 500, 600, 700, 750, 800, 850, 900, 950 | Context sizes where the watch asks for `/compact` (the memo saves itself); a 200K window wants `100, 150, 180`. | | `Context stop` | N (K tokens), off | 500 | Where the context watch stops the work, once: the session prints the block and starts no new work. N is a mark of its own. Every `Context marks` mark, under N and above it, only warns and the work goes on. `off`: the work never stops for the context (Claude Code compacts by itself when the window is full). | | `Receipt` | full, short, `savings N`, off | full | The receipt form; `savings N` prints the full receipt only when the task saved at least N%, else none. | | `Break` | N minutes, off | 10 | Time away before the resume block shows again; `off` = never. | | `Agents` | ask, auto, N, off | ask | The agent gate: ask once per session, size every task by tier, a fixed budget per task, or no budget. | | `Matrix` | quality, economy | quality | Which models do what. quality: Opus writes, plans and analyses at every tier; Haiku and Sonnet read and verify; the specialist reviewers stay on Opus at 🔴. economy: the cells before 2.6, where Sonnet and Haiku write most code — each agent costs less, more of them come back from Verify as rework. | | `Parallel` | M | 3 | Background agents running at once. | | `Agent budgets` | a, b, c, d | 2, 3, 7, 12 | Agents per tier T1..T4 under `Agents = auto`. | | `Loop guard` | N, off | 3 | Stops a loop that never reaches you: the Nth FAIL of one check (Verify or a reviewer) on one topic, or the Nth Judge or Explainer with no question asked, since you last answered, refuses the next agent (a reader still runs); the next question goes to you with no Judge. Your next prompt, a question you are asked, or a PASS of that check starts the counts over. off: never. | | `Budget raise` | auto, ask | auto | An over-budget agent under an auto budget (the gate's *Auto*, or `Agents = auto`) raises the budget by half the tier's budget (at least 2), at most twice per task, and runs; `ask` denies it and the user is asked. A fixed budget (`Agents = N`, a *This task: N* answer) and the `Parallel` cap are never raised. | | `Git guard` | warn, deny, off | warn | A risky git command (force push, `push --delete` / `:` / `--mirror`, `reset --hard` / `--merge`, `checkout `, `rebase`, `pull --rebase`, `submodule deinit -f`, `branch -D`, `tag -d`, `clean -f`, `checkout -f` / `switch -f`, `checkout -- `, `restore`, `stash drop`, `commit --amend`, …) run by anyone but `abofles:abofles-git`, also behind `sudo`, `env`, `nice`, `xargs` or `bash -lc`: warn adds a note, deny blocks it, off is silent. Read-only git never triggers. | | `Prompt checkup` | warn, deny, off | warn | Lints each agent prompt before the agent starts: the `abofles: / · T` header on line 1, `Line cap: ` as the last line, 15 words at least, a file path or command (Work, Verify), a done criterion (Work): warn adds a note, deny blocks it, off is silent. `abofles:skip` prompts and the memo agent are exempt. | | `User prompt checkup` | warn, off | warn | A message of the user's that asks a risky change (a change verb — fix, add, delete, … or صلح، ضيف، احذف — and a risk word — payment, price, auth, token, migration, database, prod, …, or فلوس، دفع، صلاحية) with no file or component named, or no end state: warn adds a note telling Claude to find the missing part in the repo, else ask one question; off is silent. Never blocks: a block would erase the message. | | `Questions` | judge, off | judge | The question gate: a Judge agent tries to answer each question before the user sees it. | | `Language` | a language name | none | The language owner questions are written in. | | `Guide` | on, off | on | The 🧭 next line. | | `Review gate` | auto, on, warn, off | auto | Refuses a receipt while a required reviewer did not run: auto = 🔴 once, on = 🟠 too, warn = never, off = silent. | | `Review memory` | N hours | 2 | A reviewer's PASS still counts for a later 🟠 change on the same topic for this long, and for the whole session on the same ticket key. A 🔴 change needs the same ticket key. A FAIL or a lost reviewer never counts. `0`: only the same ticket key counts. | | `Resume` | session, project, all, off | project | What the resume block and the guide show: this session, this project, other projects too, or nothing automatic. `off` only stops the block in the chat; the pane's Resume card still shows. | | `Pace` | normal, fast | normal | `fast` = fewer guard stops, no Judge on Minor/Normal questions, one-line receipts. | | `Guide place` | pane, band | pane | Where the guide shows: the Abo Fles pane, or the band above the prompt. Without the mod: band. | | `Resume place` | pane, output | pane | Where the resume block shows: the pane, or printed in the chat (`/abofles:resume`). Without the mod: output. A client that draws no pane though the mod is installed (`claude -p`, the SDK) wants `output`. | | `Agents place` | pane, output | pane | Where the agent history shows: the pane, or printed in the chat (`/abofles:history`). Without the mod: output. | | `Running place` | pane, off | pane | The pane's Running page and Home's Running card (live agents, background shells, scheduled wake-ups; no model calls). `off`: neither shows. | | `Tasks place` | pane, off | pane | Home's Tasks card: the session's task list (`update_plan`, `wrap_up`, `TodoWrite`, `TaskCreate` / `TaskUpdate` of the main session), kept after the turn ends with its state (Working, Waiting on N jobs, Paused, Done). `off`: none. | | `Tasks nudge` | N, off | 12 | Keeps the task list moving: while a step is open, a new message gets a note to add any new ask as its own step, and N tool calls with no list update get a note to tick it. `off`: no notes. | | `Update check` | on, off | on | Home's version line: `on` asks GitHub's public releases list for the newest version every 6 hours and offers `Update`; `off` makes no network call. Either way Home shows the version this session runs and a `Reload` button when a newer installed version waits. | | `Tasks hide` | manual, auto, next | manual | When a finished task list leaves Home: `manual` — its `Hide` button; `auto` — when the turn that finished it ends; `next` — when your next message starts a turn. An unfinished list never hides; a new list always shows. | | `Receipt place` | pane, output | pane | Where the task-end receipt goes: recorded by one command (`memo.py receipt`) and shown in the pane (Resume › Last receipts) and the memo, or printed in the chat. Without the mod: output. | | `Judge notes` | critical, sensitive, off | critical | Which Judge answers also go to Needs you: critical = 🔴 only, sensitive = 🟠 too, off = none. Every answer stays in the Decisions. | | `Memo agent` | auto, manual, off | auto | The memo agent (`abofles-memo`, Haiku 5.5 · Low) that merges decisions with the same meaning, rewrites a stale Now and moves items that belong to another project. auto: the session is told to start it in the background after a compact or when the tidy-up finds work; manual: only on `/abofles:memo`; off: rules only. | | `Memo agent every` | N hours | 6 | The least time between automatic runs, per project. | | `Continue after compact` | auto, off | auto | auto: after your own `/compact` the abofles-guide mod sends one turn, so Claude continues what the summary shows open or says in a line or two that nothing is left to continue; never after an automatic compact (its turn goes on by itself) or while you are typing; off: the session waits for you. | | `Memo on compact` | auto, off | auto | auto: every compact saves the memo by itself (`PreCompact`, `PostCompact`); off: only `/abofles:memo` writes the richer Now, and the context watch asks for it first. | | `Archive items` | N days | 7 | Age of a resolved Needs-you item before it moves to `abofles-memo.d/archive.md`. Over the 8 KB cap, the oldest resolved items move sooner. Nothing is deleted. | | `Archive sessions` | N days | 3 | Idle time of a session with nothing open before its file moves to `abofles-memo.d/archive/`. | | `Other items` | hide, show | hide | Items of other sessions and projects in the resume block and the pane: hide counts them, show lists them marked `· other`. A live session's items are never yours to act on; a closed session's (archived, its Task says closed, or idle for 14 days) count as yours, so a decision it left does not stay hidden. | | `Memo repair` | manual, auto | manual | Repair of old memos (a leaked session moved home, backlog archived, backup first): manual = `memo.py repair` when you run it; auto = at session start. | | `Topic / Side = model·effort` | model·effort | the matrix cell | One cell of the matrix for this project. | | `Topic / Session = model·effort` | model·effort | the Session cell | The model and effort of a sub session on that topic, over `Sub session`. | | `Topic / Side cap = N` | N lines | the matrix cap | The reply cap of one cell. | | `Topic: Name = T#` | T1..T4 | the topic's tier | Moves a topic to another tier. | | `Reviewers: = a, b` | reviewer tags, none | the domain's reviewers | Replaces the specialist reviewers at T3 and T4. | | `Workflow: = Read, Work, Verify` | steps | the intent's workflow | The steps of a `change` there (advice to the skill). | | `Review: / T = model·effort` | model·effort | the reviewer's cell | One reviewer's cell at one tier. | `/abofles-pane` (the `abofles-guide` mod's own command) opens or closes the pane; `/abofles:pane` is the plugin entry that points to it (without the mod it prints the resume block and the agent history as text). The pane also opens by itself once per session when it has something to show, and the band's `Pane` button opens it at any width; `/abofles:history` prints one line per agent of this session and the totals, the pane's Agents card as text. ## Add or change a topic Topics live in `scripts/topics_data.py` — 31 domains, each with areas and one line per topic: `Name | T# | signal, signal, …[ | ov Side=model·effort][ | cap Side=N][ | +reviewer]`; `SUGGESTIONS` lists the packages the lookup recommends per topic and stack. Tiers, caps, reviewers, intents and workflows live in `scripts/gen_matrix.py`. Reply formats live in `agents/*.md`, one file per side or reviewer. Edit, then run: ``` python3 scripts/gen_matrix.py ``` It regenerates `skills/abofles/references/topics.md` (domain table), `topic-index.md` (every topic), `matrix.md`, `workflow.md` and `hooks/workflow.json` (what `scripts/topic.py` and the hook read). Try it: `python3 scripts/topic.py find "postgres read replica lag"`. Bump `version` in `.claude-plugin/plugin.json` and in the root `.claude-plugin/marketplace.json`. --- # Changelog All notable changes to the plugins in this marketplace. ## abofles 2.13.1 · abofles-guide 0.15.1 — 2026-11-21 - abofles: **the question gate unlocks every question one Judge reply judged.** A Judge reply with one `VERDICT` over several questions (Q1, Q2, … each with its own `QUESTION` line) unlocked only the first, so the others were refused and never reached you. Now each `QUESTION` counts: one box with all of them passes, and so do one box per question in a row. - abofles, abofles-guide: **listing fields.** `plugin.json` now has `homepage` (https://abofles.com), the author's URL, `icon` (the uncle), `documentationUrl` and `supportUrl`, read by the plugin directory listing; `claude plugin validate --strict` passes for both. The marketplace names its owner's URL. - Website: **[abofles.com](https://abofles.com)** — install and auto-update, a tour of the pane with real screenshots, the agents and the matrix, every override key, and this changelog, built from the repo by `site/build.py`. ## abofles 2.13.0 · abofles-guide 0.15.0 — 2026-11-21 - abofles: **Prompt checkup.** Before an agent starts, the hook checks its prompt: the `abofles: / · T` header on line 1, `Line cap: ` as the last line, 15 words at least, a file or command for Work and Verify, and a stated end state for Work ("Done when …", a test or command that passes, acceptance criteria — a bare "test" or "verify" does not count). `Prompt checkup = warn | deny | off` (default `warn`: a note to Claude; `deny`: the agent is refused until the prompt is fixed). `abofles:skip` prompts and the memo agent are left alone; an error in the check never blocks an agent. - abofles: **User prompt checkup.** A risky change ask (a change verb plus money, auth, secrets, database, delete or production words, in English or Arabic) that names no file or says no end state gets a note: find the missing part in the repo first, then ask one question through the question gate. `User prompt checkup = warn | off` (default `warn`; no deny, so your message is never dropped). - abofles-guide: the Agents page shows a Prompt checkup block once something was flagged: agent prompts flagged, your asks flagged, agents refused (red, only when there is one) and the newest hit's problems. Both keys are in the overrides editor under Agents. - abofles, abofles-guide: a plugin icon (the uncle). abofles-guide: its TypeScript now checks clean (typing only, no behaviour change). ## abofles 2.12.0 · abofles-guide 0.14.0 — 2026-11-20 - abofles: **the memo's Now follows the task list, nothing taken away.** Every list call (and a receipt in the middle of a list) adds the list to this session's Now: Task gets `step k of n — ` first (a `none — between tasks` gives way), Left gets the steps after the next one, Next gets the next step first; the old Done, Left, Decisions, Next and Verify stay. Marked `· from task list`; once the list is done (the wrap-up, or every step checked) its lines are taken out again and the rest of Now stays; a hand-written Now newer than the last receipt is left alone. Resume, the guide and the pane read the same Now. - abofles-guide: Home's Now and Next keep the memo's text and add what the session knows live when the memo does not say it yet: the step in progress, the next step, and how many agents and commands run. The footer names the abofles version from the installed one until the hooks report theirs. ## abofles 2.11.0 · abofles-guide 0.13.0 — 2026-11-20 - abofles-guide: **version on Home.** The footer of every page says which Abo Fles this session runs and whether it is current (`Abo Fles 2.10.1 · guide 0.12.1 · up to date`). When Claude Code has a newer version installed than the session runs (an update the open session has not loaded), Home shows a Version card with `Reload now`, which runs `/reload-plugins`. When GitHub has a newer release, the card says so with `Update` (puts the update request in the prompt box: refresh the marketplace, update both plugins) and a `What's new` link to the release notes. The running guide version is the loaded code's own plugin.json; the running abofles version comes from the hooks; the installed ones from Claude Code's installed_plugins.json. - abofles: `Update check = on | off` (default `on`: GitHub's public releases list is asked at most once every 6 hours, no data of yours sent; `off`: no network call, Reload still shows). ## abofles 2.10.1 · abofles-guide 0.12.1 — 2026-11-20 - abofles, abofles-guide: **a task list made before the pane loaded now shows.** The hook rebuilds the session's newest list once from its transcript (the main session's own list calls, each once; subagent lines left out) at the next message or read, and keeps a copy with every list call; the pane starts from that copy when it saw no list call itself (after a reload, an update or a restart), and the next tick moves it. ## abofles 2.10.0 · abofles-guide 0.12.0 — 2026-11-20 - abofles-guide: **Tasks card on Home.** The mod keeps the session's task list from the main session's own `update_plan`, `wrap_up`, `TodoWrite`, `TaskCreate` and `TaskUpdate` calls (a subagent's list is left out). Unlike the app's list under the message, it stays after the turn ends, until a new list replaces it. The card shows a progress bar, `done / total`, and a state: Working (the turn runs), Waiting on N jobs (agents or commands in the background), Paused (the turn ended with steps left) or Done. Home sits it under At a glance and above Running and Agents, one line per step (cut with … ; no guessed blank room), a window from the step before the current one; `All tasks (N)` opens every step. The bar has one segment per step (the current one tinted), and the share done shows at the right (`1 / 4` → 25%). A finished list has a `Hide` button. The terminal gets a folding Tasks card. - abofles: `Tasks place = pane | off` (default `pane`) and `Tasks hide = manual | auto | next` (default `manual`: the Hide button; `auto`: when the turn that finished the list ends; `next`: when the next message starts a turn; an unfinished list never hides, a new list always shows), both in the pane's editor under Where it shows. - abofles: **task list rule.** While the session's task list has an open step, every message of the user's gets a note to add any new ask as its own step before working on it, and after N main-session tool calls with no list update a note to tick it. `Tasks nudge = N | off` (default 12). A new `plan` hook mode mirrors the list from `update_plan`, `wrap_up`, `TodoWrite` and `Task*`. ## abofles 2.9.1 · abofles-guide 0.11.1 — 2026-11-18 - abofles: **sub sessions.** Inside a sub session the routed setting stands: a `Sub session` key no longer replaces a `User:` or `Topic / Session` pick on its `main:` line. A child held at the parent's setting (Capped) keeps that setting, so its `main:` line fits instead of suggesting a switch. `Sub sessions = 0` now allows no sub session (before: no cap). The gate takes only exact model ids (`claude-opus-5-5`) or bare aliases (`opus`); `claude-opus-4-1` no longer passes as Opus. When the main session's model or effort is unknown, its own app record sets the ceiling. A start from a subagent meets the same cap and gate and gets its Session row. `Session routing = off` still writes the Session row. A cloud start gets no `set_session_model` line. A card brief that starts with a blank line is still routed. - abofles: **gates and loop guard.** Fable is refused for an agent with no abofles header and for the Judge and Explainer (Fable is for Analyse and Plan only). After N Judges or Explainers with no question, every agent but a reader waits, as the docs said. After N FAILs a direct `AskUserQuestion` goes to the user with no Judge, even before an agent was refused. An `abofles:skip` agent waits too once the loop is stopped. - abofles: **receipts reach the Journal once.** A receipt is journaled only when it closes its task. Before, a receipt the review gate refused was journaled, then journaled again after the reviewers; a receipt with no open task was journaled too; and in `Receipt place = pane` a receipt recorded by `memo.py` and also printed gave two Journal lines. - abofles: **the 🔴 review gate holds.** The context watch and the mid-turn receipt check no longer close a task the review gate has just refused, so the refusal is not undone at the next Stop. The gate reads the tier from `(T4)` or a `T#` anywhere after the side, and from the topic's tier (with the project's `Topic:` override) when the header has none; before, a header without `T#` skipped the gate. ` skipped: ` counts only in the receipt's `workflow` line. A reviewer counts as done only when its newest row is a PASS; before, three FAILs let the receipt close. - abofles: **more receipt forms are read.** `->` as well as `→`, `*` bullets, bold field names and a bold 🧾 header, field names in any case (`Done`, `Check`, `· Open:`), numbered sub-bullets (`1.`, `1)`). A `🧭` sub-bullet under `open` goes to `next`. A line that is not a field ends the receipt, so a bullet after `next` is not joined to it. In `open`, `none …; owner decision: …` reaches Needs you (only the `none` part is skipped), and `none (no owner decision open)` does not. A printed receipt that cannot be read is refused once, and the task stays open until the next Stop (before, the context watch closed it in the same Stop and the receipt was lost); then it is journaled as written and its ATTENTION and FAIL lines reach Needs you. - abofles: **context sizes.** `Context marks = off` (or `0`) turns the warnings off; the `Context stop` mark is its own key and stays. `Context stop` reads `1.5M` as 1500K and a bare number from 10000 up as tokens (`600000` = 600K, before 600M); a bare decimal such as `1.5` is unreadable and the default applies. - abofles: **the git guard reads less text as git.** A quoted string in a `for`, `select` or `case` word list, or in a script language's program (`python3 -c`, `node -e`, `perl -e`, …), is data, so a loop that searches for `reset --hard` is no longer refused; a script program that runs a command itself (`subprocess`, `system(`, backticks) is still read. Text commands (`grep`, `awk`, `echo`, …) were already data. Every other command's quoted strings are still read, since it may run them (`sh -c`, `bash.exe -c`, `busybox sh -c`, `powershell -Command`, `cmd /c`, `wsl`, `ssh`, `doas`, an unknown tool). `D=/…/dotnet; $D restore` is not git; a `$X restore` with an unknown `X` is still flagged. - abofles: **the guide's "owner decisions open in other projects" item** picks the projects that have open items before it keeps five; before, newer projects with nothing open (memo stamps of the same minute) could push them out. - abofles: **an agent that waits on its own background work** is no longer sent back for a missing `STATUS`/`CONFIDENCE` header; its final reply is checked as before. - abofles: **overrides editor (backend).** Save removes an Other-overrides line you deleted (before: kept, yet it said Saved). Each raw line carries its file, so Save no longer copies `.claude/CLAUDE.md` lines into `CLAUDE.md`. Reset removes a key from every project file the hooks read. `none` and `0` are saved as `off` for `Break`, `Loop guard`, `Context stop` and `Sub sessions` (`Sub sessions = 0` stays 0: none allowed); `Context marks` takes `off`. `Context stop` is saved as the hooks read it (`650` → `650K`, `600000` → `600K`, `1.5M`). - abofles-guide: **Running card.** An agent's side and model come from its own ledger row (by agent id), so two agents with the same ask no longer share one. A command another agent puts in the background while the Stop list is read stays on the card. Stop shows only for agents it can stop (background, named or teammate agents). **History**: Show all counts the sub sessions too. **Overrides editor**: the back button and the Other-lines button carry one label (the desktop drops styled text inside a button); a line in the project's other file says which file to edit. - abofles: `scripts/topic.py` reads the Git/Scrum cell as the hook does (`Topic / Git`, then `Topic / Scrum`, then `Topic / Work`). Working-tree paths `//server/share/...` and `C:\Program Files (x86)\...` are found. T4 Multi-region, Active-active & Data Residency gets the Compliance reviewer too. ## abofles 2.9.0 · abofles-guide 0.11.0 — 2026-11-18 - abofles: **sub sessions are routed like agents.** A sub session is a session the main session starts with the desktop app's `start_session`. Until now it took the main session's model and effort, the same waste the plugin stops for agents. Now a new matrix column, **Session**, picks them by the task's tier: Sonnet 5.5 · Medium at 🟢 and 🟡, Opus 5.5 · Medium at 🟠, Opus 5.5 · High at 🔴. The hook refuses a fresh start without that model and effort, and the reason gives the right ones. The brief opens with `abofles: / Session · T`, and the session prints a `🪟 Session` route line first. A fork or a proposal-card start takes no model or effort, so after the start the hook asks the session to lower it with `set_session_model` and `set_session_effort`. The app cannot start a sub session above the main session's setting: the route line then says `Capped`, and the user can raise it. New key `Sub sessions = N | off` (default 3) caps the open sub sessions per session; an archived one does not count, and an unread app record never causes a refusal. New key `Session routing = on | warn | off` (default on). New form `Topic / Session = model·effort`. The user in chat wins, then `Topic / Session`, then `Sub session`, then the cell. Each start is a ledger row with the side `Session`. - abofles: the overrides editor now also saves `off` for `Loop guard` and `Context stop`, and `650K` or `1M` for `Context stop`. Before, it refused these values. - abofles-guide: **sub sessions in the pane.** The overrides editor has `Sub sessions` and `Session routing` in the Agents group. The Agents card and History show each sub session with the 🪟 `Session` side, next to the agents. ## abofles 2.8.0 · abofles-guide 0.10.0 — 2026-11-17 - abofles: **the context watch warns without stopping the work.** Until now every mark stopped the session: it printed the block and started no new work. This was a problem with small marks (200, 250, 300) or a session left running while away. Now every mark only warns: the session prints the block and carries on. The work stops once, at the new key `Context stop`: - `500` (default): the work stops at 500K. The marks after it (600K, 700K, …) only warn. - `N` in K tokens (`Context stop = 250`): the work stops at N. N is a mark of its own. - `off`: the work never stops for the context. The pane's overrides editor has it under Session. - abofles: **short receipts reach the memo again.** The short form with two or more open items (`· open:` and nothing after it, then one sub-bullet per item, as the hook asks) was not read: `memo.py receipt` said "receipt recorded", but no line reached the Journal and the Now section stayed old (seen 11-15 to 11-16, no error shown). The open items are now read from the sub-bullets. The 🧭 guide at the end of the line is now the receipt's `next`, not part of `open`. Before, `open: none · 🧭 🔴 open FAIL …` reached Needs you as a fake owner decision ("1 owner decision open"); a 🧭 item or a `none` item never reaches Needs you now, in the short form or the full one. A header cut to its agent count (`🧾 **1 Agent** …`) also reads. A receipt whose fields still cannot be read is refused with exit 3 and a message. Nothing is recorded, the task stays open, and the receipt can be recorded again. A printed receipt (`Receipt place = output`) that cannot be read no longer closes the task silently: the Stop hook refuses it once and gives the receipt form. If the next one still cannot be read, the task closes and the Journal keeps the receipt as written, marked "receipt not read". - abofles-guide: **an agent that replied no longer looks stuck.** The engine keeps an agent `waiting` while it holds background work of its own, such as a CI wait it started before it replied. The Running card showed it as an amber "waiting" with no detail, next to its own Shell row (seen on the DRP push: "Git · waiting" after DEV, UAT and STAGE were pushed). An agent whose turn ended and that waits only on its own work now says **replied**, with a green dot and the command it waits for ("Waits for its command: Wait for CI"); its command's row says **started by Git**. If a message wakes the agent, the row goes back to what the agent is doing now. - abofles-guide: **no empty line between a row's ask and its command.** The Running card kept two lines of room for an ask that only just fits on one line ("Generate full clip in each cloned voice"), so an empty line stood between the ask and its command. The ask now takes only its own lines; any spare room goes under the card's last line. - abofles-guide: **`Context stop` in the overrides editor.** The new key is under Session, next to Context marks: a number in K tokens, or `off`. ## abofles 2.7.0 — 2026-11-11 - abofles: **designers can use it too.** Three new domains: **Graphic Design & Branding** (GFX: logos and identity, typography and Arabic type, layout and editorial, print and packaging, social and ad graphics, decks, illustration and photo editing, AI images, briefs, feedback rounds, stock and font licensing), **Video, Motion & Audio** (VID: scripts and storyboards, editing, colour grading, VFX, captions, motion graphics, Lottie, Remotion, voiceover, mixing and loudness, music rights, podcasts, Reels and TikTok, YouTube, AI video, export and delivery specs) and **3D, Spatial & Product Design** (SPC: Blender and 3D modelling, rendering, product and industrial design, CAD and 3D printing, archviz and interiors, BIM, game art, AR and VR). Most topics are 🟢 or 🟡; packaging, broadcast delivery and rights are 🟠; a print run (prepress, proofing), regulatory text on packs and building drawings are 🔴. - abofles: **marketers can use it too.** New domain **Social Media, Ads & Marketing** (SOC): post ideas, content plans, captions and hashtags, hooks, stories, trends, repurposing and scheduling tools; profiles, pages and verification; comments, DMs, moderation, reviews, crises, groups and channels, giveaways and lives; algorithms, reach, scrolling and watch time, growth, analytics, listening and client reports; paid ads on Meta, Google, TikTok, LinkedIn, Snapchat, X, Pinterest and Reddit, with targeting, ad copy, creative tests, budgets, ad account charges, ad policies, pixel and Conversions API, ROAS, lead and click-to-WhatsApp ads; one topic per platform (Facebook, Instagram, WhatsApp Business, Threads, Messenger, Telegram, TikTok, YouTube, LinkedIn, X, Snapchat, Google Business Profile, Pinterest, Reddit and Discord, Bluesky and Twitch); strategy, campaigns (Ramadan, Eid, White Friday), influencers, creator deals, social commerce, email and SMS, product listings (Amazon, Noon, Salla, Zid), PR, personas; and account safety: hacked accounts, Business Manager access, WhatsApp number bans, strikes and bans, copyright takedowns, impersonation, privacy and minors, bots and API terms, plus how-to help. Ad budgets, ad account charges, hacked accounts, Business Manager access, pixel tracking and minors are 🔴. - abofles: **two new reviewers.** **🖌️ Design** (`abofles-design`) reviews the three design domains at 🟠/🔴: brief fit, brand, typography and Arabic shaping, image quality, print specs (bleed, effective 300 ppi, CMYK, fonts embedded), social exports, video and audio specs (codec, frame rate, captions, loudness and true peak at the brief's target), 3D output, and the licence of every font, image, track and AI asset. It measures with the tools on the machine (`sips`, `pdfinfo`, `pdfimages`, `ffprobe`, `ffmpeg` loudness) and looks at the frames; it never publishes or writes through a connector. **📣 Marketing** (`abofles-marketing`, read-only tools plus web fetch) reviews the social domain at 🟠/🔴 before anything goes live: goal and audience, hook and call to action, platform limits, ad policy and special ad categories, honest claims and up-front #ad disclosure, giveaway rules, targeting, budget and schedule, links, UTM tags and the optimised event, brand voice, privacy; it never posts, replies, boosts or changes a campaign. At 🔴 every design and social topic also gets Compliance (3D gets QA); money topics add Finance, account and pixel topics add Security. UIUX keeps app and web screens. 13 reviewers. - abofles: the skill, its intent table and the start note now say coding, design **or marketing** work: making a logo, a poster, a video or a post is a `change`, so it reaches Work, Verify and the reviewers; `design` is for an approach or architecture. - abofles: **the topic lookup ignores small words again.** A second `STOP_WORDS` set in the hook replaced the lookup's own at import, so words like `on`, `my` and `is` scored as signals ("pixel not firing on purchase" landed on procurement). It is now `ANSWER_STOP_WORDS`. A signal phrase of small words only (`where is`, `how does`) weighs 5, the weight it had before, so a where-is ask still lands on Codebase Exploration, and a signal phrase matches whole words only. - abofles: a prompt header with any reviewer tag parses (`/ Design · T4`, `/ Marketing · T3`); before, only the first eleven did, so a review had no topic or tier. A topic name that holds a tag word no longer skips that reviewer in the review gate (`Logo Design & Logomarks: Compliance skipped: …` skips Compliance only). - abofles: new software topics where design and social words would have taken dev asks: Video Processing & Transcoding (FFmpeg, MediaConvert, Mux), Graphics Programming (WebGL, WebGPU, Three.js), Social Platform APIs (Instagram Graph, X, TikTok, LinkedIn); Discord login joins Social & Federated Login. 1,317 topics in 31 domains. ## abofles 2.6.1 — 2026-11-11 - abofles: **Windows working trees are found** ([#1](https://github.com/Sajadh92/abofles/issues/1)). When an agent prompt named a folder with a Windows path (`C:\…\worktree`, `C:/…` or a UNC share `\\server\share\…`), the hook did not find that folder. The agent's `path:line` facts were then marked missing, its reply was set to `CONFIDENCE: low`, and a correct reply was escalated. Drive paths and UNC paths now count, with `\` or `/` between folders and 8.3 short names (`RUNNER~1`) kept whole. A UNC share is a root, like a drive, so a folder needs 3 levels below it. A POSIX folder right after a colon (`Working tree:/a/b/c`) still counts. - abofles: **the test suite runs on Windows.** The stale-install test and the script tests set `USERPROFILE` with `HOME`, the memo seed runs with `-X utf8`, and `hooks.json` is read as UTF-8. ## abofles 2.6.0 · abofles-guide 0.9.0 — 2026-11-10 - abofles + abofles-guide: **Opus writes the code.** The matrix now puts Opus 5.5 on Work at every tier (T1 Low, T2 Medium, T3 High, T4 Extra) and on Analyse and Plan at T1 and T2; Haiku and Sonnet read and verify (the generic verifier at T4 is Sonnet 5.5 · Extra). Code Sonnet wrote from a good plan came back as rework: in the owner's ledgers 33% of the Verify and reviewer verdicts on Sonnet · High Work at T3 were FAIL, 22% on Opus · High at T4, and each round costs another Work and Verify agent. The specialist reviewers keep their cells (Opus 5.5 · High at 🔴); the Git agent keeps the cheaper Work cell it had. A topic override that put Work, Analyse or Plan below Opus is skipped; Fable on Plan or Analyse stays. New key `Matrix = quality | economy` (default `quality`; `economy` brings back the earlier cells), in the editor's Agents group. - abofles + abofles-guide: **a loop that never reaches you stops.** No limit held a fix loop (Verify or a reviewer FAILs, Work fixes, again) or a gate loop (Judge after Judge answering by itself) — Judges and Explainers take no agent budget, and with `Agents = off` nothing counted at all. The 3rd FAIL of one check on one topic, or the 3rd Judge or Explainer with no question asked, since you last answered, now refuses the next agent (a reader still runs), and the next question goes to you with no Judge. Your next prompt, a question you are asked, or a PASS of that check starts the counts over. New key `Loop guard = N | off` (default 3), in the editor's Agents group. - abofles: **the review gate refuses a task once.** A receipt recorded at each background step (`Receipt place = pane`) was refused at every step, so the task never closed and the same reviewers were demanded again and again (DCMP: three refusals for one change in 12 minutes). The second receipt of the same task now closes it with the 🟠 guide item; the next task is checked again. The refusal says the reviewers start once Verify has passed. - abofles: **new Scrum agent.** `abofles-scrum` (route tag 📋 `Scrum`) is the scrum master and task manager: it reads and keeps tickets and boards in Jira, ClickUp, Linear, Asana, monday.com, Trello or GitHub Issues through their connectors — refines stories into testable acceptance criteria, splits epics by behaviour, estimates from the team's own scale, plans sprints to capacity, keeps status and PR links in step with the code, and writes standup, sprint and blocker reports. It writes to a tracker only on your quoted ask and otherwise returns the change as DRAFTS; it never deletes tickets, closes sprints or changes workflows, boards or permissions. Ticket reads for another side take the Read cell; its own work the Git cell. Five new topics under PROD › Task Management & Trackers route to it. ## abofles 2.5.0 · abofles-guide 0.8.0 — 2026-11-10 - abofles + abofles-guide: **the session goes on after `/compact`.** After your own `/compact` the mod sends one turn: Claude continues what the summary shows open, or says in a line or two that nothing is left to continue. Never after an automatic compact (its turn goes on by itself) or while you are typing. New key `Continue after compact = auto | off` (default `auto`), in the editor's Session group. - abofles: **Needs you keeps a problem that comes back.** An ATTENTION or FAIL that was fixed is added again when a later receipt reports it, and another session's open item no longer hides this session's own (with `Other items = hide` this session never saw it). A decision or question once answered stays settled. - abofles: **the memo agent works on this project only.** Each `memo.py` call must name the session that started it with `--session`: another session's id opened that session's project memo. - abofles: **the memo lock holds across threads.** The tidy-up runs in a background thread; a write in the same process now waits for it instead of running alongside. A lock that cannot be taken still lets the hook go on, closes its file, and says so in the health line (`memo lock failed …`). - abofles: **nothing leaves the memo without the archive.** Decisions past 60 and Journal lines past 40 go to `archive.md` first; before, they were dropped. - abofles: **Claude's own notes stay out of Now after a compact.** `None. …` under Pending Tasks counts as nothing left, and a note on the context size or "start no new work" is no task. - abofles: **a session file holds about six whole receipts.** Its cap is 16 KB (was 6 KB): one receipt line holds up to ~2.6 KB, so the pane's receipt history no longer shrank to two or three. - abofles: **the raw compact summary is kept in `~/.abofles/state/compact/.md`**, never in the project: it holds the conversation, and a folder outside git (or a synced one) would carry it. A summary an older version left in the project moves there at the next tidy-up; one not replaced for 30 days is removed. The memo agent reads only the summaries of this project's sessions. - abofles: **`move-item` moves only to a project the memo knows** (one with a memo, as `memo.py list` gives), never into any git repository. - abofles: the Needs-you split checks once per other session whether it is closed, not once per item. - abofles: **a failed Agent call gives its slot back.** New `PostToolUseFailure` and `PermissionDenied` hooks on Agent release the budget and Parallel slot the call had reserved (by its `tool_use_id`), so the retry one step down after `model not available` is no longer refused for a budget or a slot nothing uses. - abofles: **a `/ Judge` header frees an agent from the budget only while a question waits.** On another agent type it counts as the Judge only while the question gate holds a question it refused in this task; else it is work and takes budget. The abofles Judge and Explainer types are unchanged. - abofles: **rule 4 holds on any agent type.** Fable on a `general-purpose` agent is refused for a Read, Work, Verify or reviewer header, not only on the abofles reader, worker, git, verifier and reviewer types. - abofles: **`abofles:skip` agents are logged apart.** They take no budget, owe no receipt and keep no Parallel slot, as the hook already let them through; the memo agent keeps Haiku 5.5 · Low even with `abofles:skip`. - abofles: the memo agent's cell is read like any model and effort (`Haiku` / `Low` passes); a prompt that is not text is refused on the model check instead of failing open; the Parallel refusal says `1 background agent is`; the gate answer `unlimited` means off. - abofles: **the Git guard sees more history rewrites.** `checkout `, `pull --rebase` and `submodule deinit -f` are flagged, and so is a git whose command name comes from a variable or a substitution (`$G push -f`, `$(echo git) push --force`). - abofles: `claims.py` refuses an option as another option's value (`--session --branch x`); the git agent keeps the caller's attribution lines over its own environment's. - abofles: **a note after an override value no longer weakens it.** `Git guard = deny # team rule` was read as an unknown value and fell back to `warn`; every key now reads the value's first word, so `deny (strict)`, `Agents = 3 per task` and `Break = off # never` mean what they say. - abofles: **an example in the override block is not a setting.** A line inside a code fence or an HTML comment under `## abofles overrides` sets nothing, and the pane's editor never edits it. `*` and `+` bullets are read like `-`. - abofles + abofles-guide: **the editor says when the hooks cannot read a value.** A hand-written value that is none of the key's options (`Git guard = strict`) gets a warning under its control naming what the hooks use instead. - abofles: a line under a key's old name (`Spawns place`) is edited and reset as that key; before, Reset left it in place and it kept applying. `Context marks` reads `500K` and `1M` and skips 0; `Review memory = 0` counts only the same ticket (it was read as the 2-hour default). The override files are read once per hook, not once per key. - abofles: **Haiku 5.5 is priced at its own rates.** The dollar table still held Haiku 4.5's ($1 / $5 per million), ten times Haiku 5.5's $0.10 / $0.50, so every Haiku agent's cost was shown 10x too high and the saving too low (receipts, `/abofles:bill`, the pane). A Haiku call whose prompt passes 100K tokens is priced at its 5x rate. - abofles: **a recorded receipt closes its task at once.** `memo.py receipt` used to keep the receipt for the turn's Stop: a second task in the same turn counted against the first one's budget, and its receipt replaced the first one, which never reached the Journal. Now the task closes and the Journal gets it when the command runs (the review gate still refuses a 🔴 change missing a reviewer). - abofles: **a short `--session` id names that session.** `memo.py now --session e18fb730` ignored the short id and wrote the newest session's Now (another session of the same project). A short id now selects the one session whose id starts with it, and names none when no session or several match. - abofles: **no stale "a judged question is still unasked".** An Explainer that answered a follow-up `SendMessage` with all its questions again unlocked the ones already asked a second time, and the Stop hook asked for them for the rest of the task. A reply now unlocks only the passed-on questions that are not asked or unlocked yet. - abofles: **a skipped reviewer needs its reason.** The review gate took any ` … skipped` within 60 characters, so `Finance · Security skipped` also skipped Finance, and a bare `Security skipped` passed. Now each skip stands on its own (never past a `·`, `,` or `;`) and names a reason: `Security skipped: `, a dash, a bracket, or `because …`. - abofles: **Undo reopens the item you resolved.** The band's Undo reopened the Needs-you item by its number, which can change before the press; it now uses the item's `#id`. `memo.py resolve --session ` keeps the full session it resolves to, so the Undo is saved. - abofles: **the context watch warns again after a compact that stays big.** The marks re-armed only when the compacted context fell under the first mark (500K); a compact from 960K to 520K left every later mark silent for the rest of the session. Now the marks above the new size warn again, and a mark the compacted context still passes stays quiet. ## abofles 2.4.0 · abofles-guide 0.7.0 — 2026-11-09 - abofles-guide: **Needs you shows a bullet per item.** The number in front of an item was only its place in the memo. The count stays on the card's title, and `resolve` still takes the number or the `#id`. The Show others card and Copy follow. - abofles + abofles-guide: **a receipt in the pane keeps all it said.** The Journal line now holds `why`, `workflow` and `next` too, and its fields are no longer cut at 160 characters. A receipt's Full details page shows its agents, tokens, cost, saving and largest line, and every field. The short card on Resume stays done, check and open. Before, a header like `1 Agent + 1 memo` lost all its numbers. - abofles-guide: **Full details pages space their cards like the main pages.** Every card shows, the count has the same pill, and room left by a high line estimate stays inside the card. - abofles-guide: **each setting has its own target and Reset.** On the Edit page an overridden setting has a `Set` badge, a `This project` / `All projects` toggle and `Reset`, in place of one `Save to` for the whole page. All projects moves the value to `~/.claude/CLAUDE.md` and takes it out of this project; This project leaves the all-projects line, so other projects keep it. A new value starts on This project. Reset goes back one level: under a project value the all-projects value takes over. `Reset all` asks first. Nothing is written until Save. Tier and reviewer lines keep one target of their own. - abofles-guide: **the Overrides page reads as a settings page.** `Edit` sits above the list. `This project` and `All projects` are separate cards, each setting with what it does, and no name or value is cut. On the Edit page every button has a border, a line separates settings and groups, and each setting reads one thing per line: name with its badge, help, control, target and Reset. `Context marks` has a two-line input, so its value shows whole. - abofles-guide: **Overrides descriptions show whole.** Read more is gone from the Edit page. - abofles-guide: **the nav bar is always one row.** Its icons hold the cards' two edges; a pane too narrow for them gets smaller icons. - abofles: **`Main session` works, and `Sub session` is new.** Both take `model·effort` (also `opus.medium` or `Opus 5.5 · Extra`), and Save refuses a value the hook could not read. Set, the `main:` line and the pane compare the session with your setting instead of the tier table, so a session on your own setting gets no suggestion. `Sub session` covers a session another session started (the desktop app's side sessions); unset, `Main session` applies. Before, the hook never read `Main session`. - abofles-guide: **a Running page in the nav bar and a Running card on Home.** The main turn, each agent (side, ask, last tool, model, tool count, tokens, elapsed), background commands and scheduled wake-ups, from the engine's own events, with no model call. Each has `Full details` (an agent's recent tools, a command's last output) and `Stop`, and Stop asks first. The nav icon has a dot while work runs. The Agents page no longer repeats what runs. New key `Running place = pane | off` (default `pane`). - abofles: **`Agent` replaces `Spawn` in all visible text.** Receipts say `4 Agents`, the pane's page is Agents, the verb is `start`. `/abofles:spawns` is now `/abofles:history` (`/abofles:agents` was already the budget command). `Spawns place` is now `Agents place`; a block with the old name keeps working, and so do memo lines that say `Spawns`. Engine names such as `agent.spawn` and the state file's field names stay. - abofles: **the context watch has a mark at 950K.** Default `Context marks` = 500, 600, 700, 750, 800, 850, 900, 950. - abofles: **an archived session comes back whole.** A session that resumes or writes after the tidy-up archived it gets its file, compact summary and index line back. Before this, it showed an empty Now. - abofles: **the memo's Now is about the work.** A compact no longer copies Claude's own notes into Task, Left or Next (`Write the closing message…`, a wait on the compact that just ran); a Next with only such notes gives the first pending task. A kept `Verify` says when it was checked (`as of 11-08 12:31`) and `Tree` says how many lines it cut. - abofles: **agent gate answers are no longer Decisions.** They are settings of the session. The tidy-up moves old ones to `archive.md`. The memo agent can drop a test line with `memo.py drop-decision`. - abofles: **the memo agent is offered only to a session that has it.** A session older than the agent's file waits for its next start, so the agent no longer fails with `Agent type not found`. - abofles: **the review gate remembers a pass.** A reviewer's PASS counts for a later change on the same ticket key, and for a 🟠 change on the same topic within `Review memory` hours (default 2). A FAIL or a lost reviewer never counts. - abofles: **`claims.py` works from a folder above several repositories.** `list` shows each repository's claims, `--repo ` picks one, and a refusal names them. - abofles: **a reply with no `STATUS` goes back once.** An abofles agent reprints its report with the header, so escalation can read it. An explainer's `QUESTION_AR:` now counts as its question. - abofles: **Needs you holds what needs you.** New key `Judge notes` (default `critical`): only 🔴 Judge answers go there; 🟠 ones stay in Decisions and open ones move to `archive.md`. A closed session's items (archived, its Task says closed, or idle for 14 days) now count as yours, so a decision it left stays visible and you can resolve it. - abofles-guide: **Full details pages are designed like the main pages.** Guide, Resume, Receipt, Agents and Decisions now open as the page's own cards with nothing cut, instead of raw markdown. Copy takes the page's text, since a picture's text cannot be selected. - abofles-guide: **Main model shows the model in use.** The value is what the session runs now, and the line under it says what to do: `Use Opus 5.5 · Medium for this task. Switch in the model picker.` It wraps instead of cutting off. - abofles-guide: **the Agents card shows tokens per model and has column titles.** Five columns: Model, Tokens, Agents, Cost, %. The same card is on Home, under At a glance, in place of the `Saved` row. A pane too narrow for five columns puts the numbers under each model. - abofles + abofles-guide: **the context number follows a compact.** The pane refreshes after a compact and shows the size the engine gives for it, instead of the last reply's usage. That size is the conversation alone: the next reply's real number also counts the system prompt and tools. If the engine gives no size, the row says `Compacted` until the next reply. - abofles: **no stale Now right after a compact.** `SessionStart` runs before `PostCompact`, so with `Memo on compact = auto` the resume block leaves out Now, Left and Next; the compact summary in the same context is newer. - abofles: **the memo saves itself.** New `PreCompact` and `PostCompact` hooks: every compact, by hand or automatic, refreshes Now from the last receipt and then rewrites it from Claude's compact summary (Task, Left, Next, stamped `· from compact`; a newer `/abofles:memo` is kept). The raw summary goes to `abofles-memo.d/.compact.md`. The context watch and the guide now say "run `/compact` at a task boundary"; `/abofles:memo` is optional. `Memo on compact = auto | off`. - abofles: **the memo tidies itself.** `memo_health` runs at session start (at most every 6 hours per project) and before a compact: resolved items older than `Archive items` days (7) and idle sessions with nothing open older than `Archive sessions` days (3) move to `abofles-memo.d/archive.md` and `archive/`, the Sessions index is rebuilt from the live files (no more orphan files), exact-duplicate decisions are dropped, the main memo is capped at 8 KB and a session file at 6 KB. Nothing is deleted. When a memo passes the cap, its oldest resolved items move even if younger than `Archive items`. - abofles: **the memo agent** (`abofles-memo`, Haiku 5.5 · Low, 🗂️). In the background, after a compact or when the tidy-up finds work, at most once per `Memo agent every` hours (6) per project: it merges decisions with the same meaning, rewrites a stale Now, moves an item that belongs to another project, archives closed sessions and drops noise with a reason. `Memo agent = auto | manual | off`. It uses no agent budget and asks for no receipt, and a hook limits it to one `memo.py` call at a time and to this project's memo files. New `memo.py` subcommands: `list --json`, `merge-decisions`, `move-item`, `drop-item`, `drop-decision`, `archive-session`, `rewrite-now`. - abofles: **no more items from another project or session.** Every memo change goes through one locked write path (`memo_apply`), which refuses a write with no session. Needs-you items have a stable `#id` and an `(s: p:)` tag naming the repository the item is about; a move never retags. Only this session's tagged items are yours; `Other items = hide` (default) counts the rest, `show` lists them marked `· other`. `resolve` and `reopen` refuse another session's item without `--force`. The memo moves to another repository only with a lead of 3 edits and 60% of the session's edits, the same margin back; a launch folder outside git goes to the first repository edited. Real bug: a Jamkhana owner decision showed as a task in DRP sessions. - abofles: **`memo.py repair`** (`/abofles:resume --repair`, `--dry-run`, `--all`, `--root`) fixes old memos: backup first, tidy-up, sessions moved to the project their edits landed in with their tags kept, unknown ones reported and left. `Memo repair = manual | auto`. - abofles: **no more red "Failed" questions.** Across 6 live sessions, 11 of 38 `AskUserQuestion` calls were refused by the question gate. An explanation not printed above the question now goes inside the box instead of a refusal, and the check reads the transcript again to survive a write race; one Judge reply unlocks 4 questions, the most one box holds (it was 3); the gate or budget question passes with no Judge while the budget has at most 1 agent left, since the guide invites it. - abofles: seven new override keys, each defaulting to the owner's decision: `Memo agent`, `Memo agent every`, `Memo on compact`, `Archive items`, `Archive sessions`, `Other items`, `Memo repair`. - abofles-guide: the Overrides editor has a **Memo** group. Resume and Decisions keep other sessions' and projects' items behind `Show others (N)`, in a grey, muted card; with `Other items = show` they are marked `· other`. Resume's facts gain `Tidied` when the memo's tidy-up moved something. The memo button is now `Compact` and fills `/compact` (`Save memo` with `Memo on compact = off` or an older `abofles`). ## abofles-guide 0.6.1 — 2026-11-05 - abofles-guide: **brand footer.** Every desktop page of the pane ends with one quiet row: the Abo Fles mascot (a 64 px PNG inlined in `hooks/mascot.ts`, with a light sticker border so it reads on the dark theme), then "Abo Fles · ابو فليس" in small muted text. There is no card or rule around it (space sets it apart), and it stays still while a page's cards slide in. The terminal pane ends with the same name as dim text under a rule. ## abofles 2.3.0 · abofles-guide 0.6.0 — 2026-11-04 - abofles: **security fixes.** No hook returns `permissionDecision: allow` any more except the question box that carries an explanation (it needs `updatedInput`): a spawn note is context only, so your permission prompt always stays. The Git guard also catches `push --delete` / `-d` / `:` / `--mirror`, `tag -d`, `checkout -f`, `switch -f` / `--discard-changes` and `reset --merge`, and sees through `sudo -u`, `env -i -u`, `nice -n`, `xargs -n -I` and shells with combined flags (`bash -lc`, `sh -ec`). The question gate's quick re-ask passes only for the exact question set it just refused; the no-Judge pass after a budget deny needs the words `Abo Fles agent budget`; the agent-gate question passes only while the gate is unanswered (or reset with `/abofles:agents ask`). A spawn made inside a subagent no longer changes the main session's unlocks or budget counts. `claims.py` backs a damaged store up to `.bak` with one warning, treats a bad heartbeat as stale and strips control characters when printing. `/abofles:status` lists the Git side. - abofles: **memo and question fixes from real sessions.** A session is no longer told to judge or ask another session's owner decision: Needs-you items carry their session and project (`(s: p:)`, older lines still read), and the guide, Check all, Use and the receipt's 🧭 line act only on this session's items; other sessions' items are shown as `other session · ` for information. `memo.py move` pins the memo where you put it (edits no longer move it back, a `/clear` keeps it) and takes the session's open items along. Git writes from Bash into a named repository (`git -C …`, `cd && git …`) now count toward where the memo belongs. A typed answer that is not one of the offered options (a refusal, "ask me later") no longer closes the item: it stays open with `answered (free text): …`. Explainer replies with numbered or tagged questions (`QUESTION 1 [IDMS file]: …`) unlock one question each — a Judge HANDOFF followed by such an Explainer was refused before; when a background Explainer returns no QUESTION line, the next prompt or the refused question says so. Data Validation QA (reports, numbers, totals) is reviewed by Data and QA at T3 instead of Finance. The receipt note and the Stop's receipt demand wait until every background spawn of the task is back. - abofles: **Check all.** The guide payload carries `all` (`{n, fill}`, from 2 open items) — one numbered instruction that works through every open item with each item's own Use text; Needs-you and owner-decision items are resolved only with evidence (`memo.py resolve `), owner decisions are asked, not guessed. The pane's resume data carries `needs_fill` for 2+ open Needs-you items; `guide.py --check-all` prints the same text (`/abofles:guide --check-all`). - abofles: **decisions by who decided.** Logged decisions are marked `- MM-DD HH:MM · by:you · → ` (`by:judge` for a Judge's ANSWERED, `judge: ` prefix kept; old lines are read as before); `parse_decision()` reads them and the pane payload carries `decisions` (`you` / `judge`, newest first across the main memo and every session file, 50 per group, with their totals). - abofles: **Budget raise** (`Budget raise = auto | ask`, default `auto`). A spawn over an auto budget (the gate's *Auto*, or `Agents = auto`) raises it by half the tier's budget (at least 2), at most twice per task, and runs, with `abofles: budget raised N → M (auto, raise k/2)`; `ask`, an `Agents = N` budget, a *This task: N* answer and a third raise keep the deny. The raises reset at the task end and when the gate is answered again; `guide.task.raised` and the guide's budget item show them. - abofles: **the git master agent** (`abofles-git`, side Work, tag Git, 🌿). It owns every git and `gh` write: it inspects the repo first, keeps a branch per session or agent and worktrees for parallel work (never deleting one it did not create), makes small conventional commits, merges or rebases by the shared-branch rules, resolves conflicts from both sides and the merge base with the tests run after, opens draft PRs, tags and releases. It commits, pushes, rewrites history or opens a PR only on the user's quoted ask, prefers `--force-with-lease`, never skips hooks, and returns the exact command for anything destructive. Reply: STATUS, CONFIDENCE, ACTIONS, STATE, CONFLICTS, LEFT. The hook takes `/ Git` (or `/ Work` with `abofles-git`) as the topic's Work cell, refuses Fable for it, grades its reply like a worker's, and logs its rows as side `Git`, so it never counts for the review gate. - abofles: **`scripts/claims.py`** for several sessions and agents on one project: `list`, `claim --session S --branch B [--worktree W] [--files f …]`, `heartbeat`, `release`. The store is `/abofles-claims.json` (inside `.git`, never committed, shared by every worktree; locked and written atomically, Windows included); a claim without a heartbeat for 2 hours drops out; `list` flags the same branch or file (a folder covers what is under it) claimed by two sessions, and `claim` exits 3 on an overlap. The git agent lists before it acts and reports an overlap instead of proceeding. - abofles: **Git guard** (`Git guard = warn | deny | off`, default `warn`; 26 override forms). A risky git command from anyone but `abofles-git` (force push, `reset --hard`, `rebase` but not `--abort` / `--continue`, `branch -D`, `clean -f`, `checkout -- ` / `checkout .`, `restore` without `--staged`, `stash drop` / `clear`, `filter-branch` / `filter-repo`, `commit --amend`, `update-ref -d`, `gc --prune=now`) gets a one-line note (warn) or a denial (deny): `abofles: risky git () outside abofles-git — route git work to abofles:abofles-git, or confirm the user asked for it`. The command line is parsed, not grepped: split on `&&` `;` `|`, `git -C dir` and `bash -c` read through, quoted text and heredoc bodies (a commit message saying "reset --hard") never match; read-only git never triggers. - abofles: five git topics in DevOps, CI/CD & Release: Merge Conflict Resolution (T3), History Rewrite & Force Push (T4), Multi-Session & Multi-Agent Git Coordination (worktrees, branch claims) (T3), Pull Requests & Code Review Flow (gh) (T2), Tags, Releases & Changelogs (T2). Merge Conflict Resolution adds the Tester reviewer (a resolved conflict needs the suite run). 1,128 topics; the topic index numbers after #865 moved. The overrides editor lists Budget raise and Git guard in its Agents group, each with a help line. SKILL.md: a Git row in the side table, 🌿 in the route-line emoji list, a hard rule that every git/gh write goes through `abofles-git`. - abofles: a background agent's hand-back that lands after its SubagentStop is graded from the transcript (it was closed as no_reply, which kept the question gate locked). - abofles-guide: on the terminal, the pane's rows rebuilt as one system: a grey caret button of its own beside each title (a desktop draws a Button as its label, in one colour), summaries as coloured text, `Read more` a small grey link at the row's right end (never the text inside a button; a one-row text is the whole text cut by the surface at the pane's real width), receipts one row each, decisions as time / question / → answer, history as activity rows grouped by day, the guide's actions right under its items, and `Show fewer` for a list. - abofles-guide: a full UI/UX pass on the pane: a coloured line icon per section (Guide compass, Resume bookmark, Spawns fork, Decisions scale, Overrides sliders), semibold titles, summaries coloured only when they carry meaning, captioned sub-headings, priority and status dots with their word, one empty-state style, and long text cut at a word with "…" + Read more (shown whole when it fits in about two lines; Arabic-safe). Every SVG colour has a light and a dark twin at 4.5:1 or better. - abofles-guide: on the terminal, the pane drawn like the desktop app's own side panel (Context, Changes, Environment): flat sections with thin rules between them, a title with a small ▸/▾ triangle, a grey summary with a quiet dot for attention, primary actions as pill buttons and the rest as text links; Resume, Spawns, Decisions and Overrides start folded, the Guide open. - abofles: the direct-read guard warns through additionalContext only; it no longer returns `permissionDecision: allow`, which skipped the user's permission prompt for that read or read-only Bash command. - abofles: an AskUserQuestion that carries several judged questions spends one gate unlock per question. It spent one in all, so the leftover unlock made the next Stop demand a question that had already been asked. - abofles-guide: **a nav bar and a Home page** on the desktop. Pages (Home, Guide, Resume, Decisions, Spawns, Overrides) replace the long scroll; the pane API has no sticky position to pin a bar to. The bar is one row of icon tiles, centred, each the same icon as its page's header; the page shown is filled, and a blank button over each tile takes the press. A page's cards slide in when you move to it (0.26 s, staggered 50 ms, none under reduced motion). Home sums up the next step (with Check all / Use), the numbers at a glance, Now/Next, and the latest decision and receipt, from the sections the pane holds. `≡ Full details` opens a card's whole text as a page of its own (native markdown, with a `‹ ` button back to the page it came from), so nothing stays cut. - abofles-guide: **every pane section in HTML and CSS** on the desktop (`hooks/html.ts`). Each section is XHTML in an SVG ``; the app draws it as an image, unsanitized, in its own text colour. It is laid out at the pane's width and scaled to the slot keeping its ratio, so text wraps where it will be drawn. A section is a header strip (icon on a tinted tile, title, summary pill) over its cards: Guide (Next up with a priority pill, Also open, two-edge facts with the Context meter), Resume (Now/Next, Needs you numbered, Left, Last receipts with their details and the memo facts), Decisions (By you / By Judge: time, question, → answer), Spawns (the saving, spend by model, Running, History by day), Overrides (key, value, where it is set). Sections no longer fold on the desktop and have no caret. Every action is a button whose label leads with one single-tone glyph, sitting under the card it acts on. The pane has one primary button (the Guide's) and no Hide. Long texts are clamped to lines estimated from measured SF Pro widths, with one Read more per section. The terminal keeps the native rows. - abofles-guide: **review fixes before release.** The band's button reads `Open pane` (it was `Pane`), and the Guide's main-model fact is two rows: `Main model` with the model to use, then what it is now and where to switch (or that it fits the task). Tone text and muted greys pass 4.5:1 in light mode too (amber and green were 3.6:1), and the filled nav tile is darker so its white icon passes 3:1. Home and the Guide's "all clear" note get Full details when their text is cut, and empty or error states wrap instead of cutting the instruction. The overrides editor keeps one primary button (Save): the chosen scope is marked ✓. Alt text names its numbers ("46% saved", "2 spawns, $0.01, 2% of spend") and a split history is read once. Detail pages no longer turn memo text into links or lists, and a lone surrogate in a memo line can no longer break a card's picture. On Resume, Full details joins the last card's button row. - abofles: **security, second review (Critical).** The Git guard sees through more forms — escaped or upper-case `git`, shell keywords, `timeout` / `stdbuf` / `doas` / `caffeinate`, `eval`, `ksh` / `fish -c`, here-strings, `echo … | sh`, heredocs into a shell, `find -exec`, `submodule foreach`, `-c alias.x=…` — and catches more risky commands: `branch -df` / `-f` / `-M` / `-C`, `checkout -B`, `switch -C`, `tag -f`, `reset --keep`, `gc --prune=all`, `reflog expire` / `reflog delete`, `update-ref ` / `--stdin`, `push --prune`, `worktree remove -f`, `stash clear`, and `git config alias.*` writes; only `abofles:abofles-git` is exempt. Question gate: a denied question never passes unjudged, one unlock covers one question, a SendMessage unlocks only when a gate agent's reply is coming (one per reply), an Explainer gives at most 3 unlocks, a re-ask must match both ways, and the budget question must be only the budget question. Memo: option answers match whole labels, auto-resolve touches only the session's own items, the origin tag is read from the line's end, a planted `.git` file cannot redirect the work root, and `memo.py move` needs a repository top. Claims: C1 and bidi marks stripped, non-UTF-8 arguments refused with no `.tmp` left, `.bak` never written through a symlink. The Stop hook no longer asks for a receipt while a background spawn of the task is still running. The Git guard fails closed: a git word nested more than six shells deep, or a line that breaks its parser, is flagged instead of passed unread. Claims: the lock file is never followed through a symlink or truncated, and line separators, zero-width and ALM marks are stripped from printed claims. Third round: a re-ask is compared on the whole box (text after the `?`, header, every option) and may add at most one new word; a SendMessage reply is graded like any Judge or Explainer reply before it unlocks; the agent-gate and budget questions must be only themselves (no second ask in a header, option or description) and never count as a judged round; the Git guard reads more runners (`setsid`, `flock`, `taskset`, `chrt`, `unbuffer`, `strace`, `coproc`, shell functions, `$"git"`), flags `git rm -f` and aliases set through `GIT_CONFIG_PARAMETERS` / `GIT_CONFIG_KEY_n`, keeps a backslash inside double quotes as the shell does, and flags a git command inside a runner it cannot read (`su -c`, `script`, `python -c`) as not fully parsed — text commands (`echo`, `grep`, a `gh` PR title) stay unflagged; a `.git` file maps to a main repository only when that repository's worktree entry points back at it. ## abofles 2.2.0 · abofles-guide 0.5.0 — 2026-10-27 - abofles: **the receipt goes to the pane.** New override `Receipt place = pane | output` (default `pane` with abofles-guide, `output` without): the Stop hook asks Claude to record the receipt with one command (`memo.py receipt`, the receipt lines on stdin) instead of printing it; the hook reads it as if printed — it closes the task, the review gate still applies, and the memo Journal and the pane's Resume › Last receipts get done / check / open. The chat keeps one collapsed command row and one closing sentence. - abofles-guide: **the pane redrawn in the app's side-panel look** (approved mockups): a caret header per card (`▾` / `›`) with the summary and a coloured dot at the right; label/value rows; dots, meters and PASS / FAIL words instead of emoji. Guide: a context meter, This task and Agent budget rows, Hide as a link, "Next from the memo" when nothing is open. Resume: Last receipts as a time · spawns · $ · saved table whose rows open their details. Spawns: the saving first, a meter and a table split by model, Running and History rows with the verdict in a column of its own. Meters and dividers are drawn as SVG on the desktop (4 px and 1 px, the card's full width, never wrapping) and as ━ / ─ in the terminal. Overrides: a table and an `Edit` button; the editor page is a settings list (name and help on the left, control on the right), `Save to` is two buttons, and the actions read `Save N changes`, `Ask Claude`, `Discard`. - abofles: the pane payload carries the spend per model (`totals.by_model`), the baseline's label and a `resolved` flag per row; a FAIL resolved by hand with evidence no longer counts as an open FAIL (pane and `bill.py`). ## abofles 2.1.0 · abofles-guide 0.4.0 — 2026-10-24 - abofles-guide: **the Abo Fles pane.** A pane docked on the right, drawn like the app's own side panels: a **Guide** card (the 🧭 items, Context / Task / Budget rows, Use · Re-check · Memo · Copy · Hide · Undo), a **Resume** card (Session, Memo, Away; Now, Next, Needs you, Left; the last receipts as short rows with Show details), a **Spawns** card (Saved, Tokens, Open FAIL; live agents with their start time; the history newest first with PASS / FAIL) and an **Overrides** card. Every card folds from its header (remembered across sessions), every text wraps, long lists show their first rows and Show all. `/abofles-pane` opens or closes it; it opens by itself once per session, and a pane you closed stays closed. When the window is too narrow the band shows `Abo Fles · ` with a Pane button. - abofles-guide: **overrides editor.** `Edit overrides` opens a page with the keys in groups (Session, Agents, Output, Where it shows), a one-line help and a badge per key, a free-form box for the other forms, and Save (writes the block of the project's or your `~/.claude/CLAUDE.md`), Ask Claude to apply, Discard edits and Back. - abofles-guide: **breaking — no more Resolve on your word.** The band's Resolve button and its confirm step are gone: Re-check asks Claude to verify the item in the code and resolve it only with the evidence (`--trace "re-checked: …"`, or `"user answered: …"` after a question). Undo reopens Claude's last resolve. - abofles: **where things show.** New overrides `Guide place = pane | band`, `Resume place = pane | output`, `Spawns place = pane | output`, all `pane` by default when abofles-guide is installed — **breaking:** the guide, the resume block and the spawn history move from the band and the chat into the pane; set them to `band` / `output` for the old view (also for a client that draws no pane, such as `claude -p` or the SDK). Without the mod nothing changes. - abofles: new `/abofles:pane` (points to `/abofles-pane`; prints the resume block and the spawn history without the mod) and `/abofles:spawns` (one line per spawn and the totals; `bill.py --history`). - abofles: new `Resume = off` (also `none`, `0`): no automatic resume block in the chat at start, after a compact or after a break; `/abofles:resume` and the pane still show it. - abofles: **breaking — resolve needs evidence.** `/abofles:resume --resolve ` and `/abofles:status --resolve ` refuse (exit 2) without a note / `--trace` naming what was checked. - abofles: background agents are tracked to the end: a row whose agent finished but whose SubagentStop never arrived is closed from its transcript (closed_by `stale`) once its last turn is complete, then re-graded and its tokens counted once when the real report arrives; a SendMessage that resumes an agent reopens its row. - abofles: the memo's Now from a receipt keeps the receipt's own `next` and one item per sub-bullet; the 🧭 guide line is no longer copied into Next. Fewer false "needs you" items (a tighter match for owner questions). - abofles: `scripts/overrides.py` (`get --json`, `set --scope project|user`) edits only the `## abofles overrides` block, keeps every other byte, refuses bad values with one line (exit 2) and a file changed meanwhile (exit 3). Security hardening from review: the temp file is a fresh `mkstemp` file (a planted symlink is never written through), a project CLAUDE.md that links out of the project is refused, a non-UTF-8 file is refused, a BOM is kept, values are capped at 500 characters. - Security: the pane runs a command from the state file only in the shape the hook writes (bash, the abofles `scripts/py` launcher, then `memo.py` / `status.py` / `overrides.py` from the same folder); the hidden resume context is marked as memo data, not instructions; control bytes are stripped from pane text; `~/.abofles/state` and `~/.abofles/ledgers` are owner-only (700). - The README lists every override in one table (17 keys, 6 forms). ## abofles 2.0.1 — 2026-10-21 - abofles: **Windows support.** The hook imported `fcntl` (Linux/macOS only) at load time, so every hook crashed on Windows. Locks now use `fcntl.flock` on Unix and a one-byte `msvcrt.locking` lock on Windows (retried up to 10 s). - abofles: **one launcher, `scripts/py`**, for every hook, slash command and command the hooks print: it picks `python3`, then `python`, then `py -3` (Windows often has no real `python3`, only the Microsoft Store stub), probes each on Windows once and caches the winner in `~/.abofles/py-path`, and always runs Python with `-X utf8` (Windows defaults to cp1252, which crashed on the memo's emoji and Arabic text). Hooks call it with `--hook`, which always exits 0, so a missing or broken Python can never block a tool call. `.gitattributes` keeps it LF on Windows checkouts. Slash commands allow `Bash(bash *)` instead of `Bash(python3 *)`; SKILL.md and the references point to the launcher. - abofles: Windows hardening found by an Opus review: path checks (`work_key`, `newest_state_path`, worktree detection, the topic-lookup guard exemption) accept backslashes; `os.replace` retries PermissionError (Windows locks open files) at every write site; config and registry temp files carry the pid; `str.removeprefix` (3.9+) removed for Python 3.8; a guarded UTF-8 relaunch remains as a safety net for plugin entry points started without the launcher. New tests simulate Windows (no `fcntl`, stub `msvcrt`, a Store-stub `python3` on PATH, backslash paths). ## abofles 2.0.0 — 2026-10-21 - abofles: **1,100+ topics in 27 domains, looked up — never read.** `scripts/topics_data.py` holds the domains (Security & Identity, Money & Finance, Business Logic & Rules, Data & Databases — PostgreSQL, MySQL, SQL Server, Oracle PL/SQL, SQLite, NoSQL, replicas, read-only access, DBA ops —, Search/Cache/Messaging, Backend & APIs, Frontend & Web, UI & UX Design, Mobile & Desktop, Quality & Testing, QA & Acceptance, Performance & Benchmarks, Debugging & Incidents, Code Health & Patterns, Architecture, AWS, Azure & GCP, Edge/CDN/Cloudflare/Networking, Servers/Hosting/localhost, DevOps/CI/CD/Release, Observability & Logging, Data Engineering & Analytics, AI/ML/LLM, Integrations, Docs, Product & Process, Languages & Runtimes), each with areas and `Name | T# | signals` lines; the original 100 names are kept. `scripts/topic.py find ""` (one Bash call, ~60 ms, ≤ 7 lines) prints the 3 best matches with tier, cells, caps, reviewers and workflow, project overrides applied; `show`, `domains`, `list `. The skill's Step 2 calls it instead of reading `topics.md`; SessionStart prints the exact command; the call is free of the direct-read guard. `hooks/workflow.json` carries topics, signals, tiers and caps; `references/topic-index.md` lists everything for people. - abofles: **tiers redefined** by blast radius and reversibility (🟢 cosmetic/local/instantly reversible · 🟡 a test catches it, dev-only · 🟠 prod behaviour, business rule, shared infra, SLO · 🔴 money, auth, secrets, PII, data loss, prod infra, irreversible) with **tier bumps**: `prod`/live data +1, read-only/dry-run/localhost/sandbox −1 (Money, Security, Privacy never below T3), marked ` · Prod +1` / ` · Sandbox −1` on the route line. - abofles: **four reviewers added** — 🗄️ `DBA` (`abofles-dba`: migration safety, locks, plans, replicas, backups), ⚙️ `Ops` (`abofles-ops`: deploy and IaC safety, rollback, secrets, parity), ⚡ `Perf` (`abofles-perf`: hot paths, N+1, allocations, benchmark evidence), 📊 `Data` (`abofles-data`: idempotent loads, grain, aggregation correctness, freshness). Reviewers per domain: DB → DBA; MSG, PERF → Perf; AWS, AZG, EDGE, SRV, OPS, OBS → Ops; DATA → Data. Override names: `Reviewers: = …`. - abofles: **`💡 suggest` lines**: `SUGGESTIONS` in `topics_data.py` name a package per topic and stack; the lookup prints one when the project's manifests (`*.csproj`, `package.json`, …) or the ask match. First entry: **DynamicWhere.ex** (NuGet, .NET 6–10, EF Core 6+ — JSON filter/sort/page/select/group requests → LINQ with field policies) on dynamic filtering, grids, reporting and export topics; the agent reads `https://doc.dynamicwhere.com/llms.txt` first. - abofles: the `pre` hook adds two advisory notes: a header topic not in the index (with the three nearest names) and a first spawn of a topic/side above the cell's model (`N× the cell` — fine with a `User:`/`Scope`/`Project Override` modifier). New topics: OTP/TOTP/passkeys, PDF/Excel/CSV generation and parsing, email templates/flows/inbound, Metabase and BI modeling, dynamic SQL and query builders. ## abofles 1.7.1 — 2026-10-21 - abofles: **cost audit**. Hook commands run through `hooks/run.py`, which imports the hook module so Python caches its bytecode (the script was recompiled on every tool call: ~36 ms → ~24 ms per Read/Bash/Grep/connector call, Edit, prompt and Stop). The direct-read guard is silent while the gate answer is `Main session only` (budget 0 this task) — the user chose direct reads, so "spawn a Read agent" was noise. The skill is invoked **once per session**, not per task (each re-invoke re-inserted ~4K tokens of SKILL.md that was already in context): SessionStart text, skill TRIGGER and the README's `CLAUDE.md` line say so. Session state and ledger files untouched for 60 days are pruned at SessionStart (`other_projects` reads every state file on each Stop); `memo-index.json` and the memos stay. ## abofles 1.7.0 — 2026-10-21 - abofles: **workflows**. The ask's intent (`question` · `review` · `design` · `change`) and the topic's tier pick the steps of a task — Read / Read → Analyse → Review / Read → Analyse → Plan / by tier for a change (🟢 Work → Verify · 🟡 Read → Work → Verify · 🟠🔴 Read → Analyse → Plan → Work → Verify → Review). The skill prints one `🧭 Workflow · …` line before the first route line; `/abofles:route` prints it too. Judge and Explain stay event-driven (the question gate); on a 🔴 change the Plan's open decisions go through the Judge before Work. - abofles: **seven specialist reviewers** split the Verify side — `abofles-security`, `-finance`, `-compliance`, `-tester` (the only reviewer that edits, test files only), `-qa`, `-uiux`, `-cleancoder` — each with its own checklist and the verifier's reply format, routed at the tier's Verify cell with the header `abofles: / · T` (tags Security, Finance, Compliance, Tester, QA, UIUX, Clean; route emoji 🛡️ 💰 📜 🧪 🎯 🎨 🧹). Reviewers per group and tier live in `gen_matrix.py` (`GROUP_REVIEWERS`, `TOPIC_REVIEWERS`) and are generated into `references/workflow.md`, `matrix.md` and `hooks/workflow.json`. The UI/UX reviewer uses the Better Design (`get-review-rules`, `check-comprehension`, `inspect-spacing`, `review-dashboard-structure`) and UX Pilot (`preview-ux-review`, `preview-attention-heatmap`, `preview-persona-audit`) connectors when they are connected, a checklist otherwise. - abofles: **review gate** at the receipt. A 🔴 change closing without a required reviewer is refused once, naming the missing reviewers, their headers and the Verify cell; a 🟠 change gets a 🟠 `review missing` guide item instead (cleared at the next task's first spawn). ` skipped: ` in the receipt's new `workflow` line passes it. `- Review gate = auto | on | warn | off`, `- Reviewers: = …| none`, `- Workflow: = …`, `- Review: / T = model·effort`. Agent budgets under `auto` rise to 2 · 3 · 7 · 12 (orchestrator) and 1 · 2 · 5 · 8 (worker) to hold the reviewers. A task's ledger rows are now the last N rows of the ledger (N = its spawns), not the rows stamped after its start second. A receipt printed mid-turn (closed by the next spawn) cannot be refused, so the gate raises the guide item there. The gate's cell text comes from `workflow.json` (generated from `TIERS`), not a second copy in the hook. - abofles: **connectors** bundled in `.mcp.json`: the official Atlassian MCP (`https://mcp.atlassian.com/v2/mcp` — Jira, Confluence, Loom, Bitbucket; OAuth), Context7 (library docs), Better Design (`https://better-design.com/api/mcp`) and UX Pilot (`https://mcp-server-485825078314.us-central1.run.app/mcp/ui`) — the two the UI/UX reviewer uses. Claude in Chrome cannot be bundled; the README says how to enable it and which reviewers use it. ## abofles 1.6.0 — 2026-10-21 - abofles: the memo follows the session's work, not only the folder it started in. A new `edit` hook (PostToolUse on Edit/Write/MultiEdit/NotebookEdit, subagent edits included) counts where edits land by repository — the launch folder first (nested repos and worktrees under it count as it; a worktree elsewhere maps to its main repository). Once another repository holds three or more edits and more than the launch folder, the session's memo moves there: its session file, its `(s:)` Needs-you items (renumbered) and its index line, with a `moved here from ` trace line in its Journal; the count can move it back. A session started in the wrong folder (a storefront session opened in another project's folder) no longer fills that project's memo. `/abofles:resume --move ` (`memo.py move`) does the move by hand for a session already misfiled. A new session or `/clear` starts the count again. - abofles: resume and guide **scopes**. `project` (default) shows every session of this project; `session` shows this session's file and its own Needs-you items only (`· scope session` in the header, `needs you → none of this session`); `all` adds one `elsewhere` line per other project that has a memo (name, task, next, open count), found through a registry the hooks keep (`~/.abofles/state/memo-index.json`, one entry per project with a memo). `/abofles:resume --scope ` for one call; `- Resume = session | project | all` in the project's override block (sixteenth form); `/abofles:resume --default ` (`memo.py scope `, `~/.abofles/config.json`) for every project; `memo.py scope` says which is in force and why. The guide's owner-decision item follows the same scope (`/abofles:guide --scope`; `all` adds one 🟠 item counting the owner decisions open in other projects). - abofles: a session with nothing of its own yet (no Now, no receipt) gets one `project` hint line in the resume block — how many other sessions this folder has and the newest one's task — instead of their `now`/`last` lines; `--scope project` shows them in full. - abofles: both memo headers carry `Project: `; `memo.py`, `guide.py` and the scripts' session lookup use the session's work root when the memo moved. - abofles: a question whose explanation needs markdown is refused once when the block is not printed above it (as before); a second ask still without it now goes through **rewritten** (`updatedInput`): the `📖 the full table/diagram is printed above this question` line is removed and the Explainer's block is put inside the first question's text as plain lines (table rows become `a · b · c`, fences and marks dropped, 2,500 chars max). The user is no longer sent to a table that is not there (DCMP 2026-10-21). - abofles: a judged-question unlock gets one demand. When the Stop hook asks for the question and the turn ends without an `AskUserQuestion` (the model answered instead, or the question was already settled), the unlock is dropped and the band's "ask" item goes; it no longer re-dates itself into the next task and demands the question at every turn end for days. - abofles: `memo.py` writes (`resolve`, `reopen`, `note`, `now`) from a folder with no session state, no `--session` and no memo exit 4 and write nothing — a `cd` into another project can no longer start a stray memo there. ## abofles 1.5.2 — 2026-10-19 - Docs: the READMEs now say how to turn on auto-update for the `sajadh92` marketplace (off by default for third-party marketplaces) and how to update by hand. ## abofles 1.5.1 — 2026-10-19 - abofles: `memo.py now` (what `/abofles:memo` runs) prints the session file it wrote, `.claude/abofles-memo.d/.md`, not the main memo path — as the command's reply line says. ## abofles 1.5.0 — 2026-10-19 - abofles: the memo is split — the main `.claude/abofles-memo.md` keeps only what must not be missed (Needs you, notes, a Sessions index with one link line per session: last update, task, next, receipts), and each session writes its own `.claude/abofles-memo.d/.md` (Now, Journal, its questions and answers). A receipt rewrites one small file plus one index line instead of the whole memo; the resume block and the guide read this session's file plus the main memo; `show --all` prints the main memo and this session's file only (`--main`: the main memo). A pre-1.5 one-file memo is split on its first read; the git exclude already covers the folder. ## abofles 1.4.0 — 2026-10-19 - abofles: harness notes (`[harness: …]` blocks) are no longer graded as part of an agent's reply, so they no longer trigger length or format warnings. - abofles: when the hand-back of a background agent is a stub ("Understood."), the reply is read from the agent's own transcript and graded from there. - abofles: the question gate unlocks only on a real `VERDICT:` (judge) or `QUESTION:` (explainer; `EXPLAIN` optional) line; a stub reply leaves it locked and says so. - abofles: only `abofles-judge` and `abofles-explainer` (or a spawn whose header side is Judge/Explain) are gate agents; a project agent named like one no longer unlocks questions. A `SendMessage` unlocks only for the ids or names of gate agents spawned this session. - abofles: an unlock made before the open task began, or older than the gate wait with no gate agent running, is dropped instead of demanding a question nobody asked. - abofles: a pending background row with no agent id or transcript for 30 minutes closes as lost; a pending row from an earlier task is no longer shown as running. - abofles: `run_in_background: true` is treated as async even when the tool result looks complete. - abofles: a task closes only on a printed receipt. A spent nudge keeps the task open (count, budget and gate answer stay; the next Stop asks again); with receipts off, the task closes at the next real user prompt, not on a task-notification turn. - abofles: a turn whose last line ends with `?` gets no receipt nudge — the task waits on the user. - abofles: a compact or a resume keeps the open task; only a new session or `/clear` resets it. - abofles: a receipt counts only as the computed header line (`🧾 **N Spawns …`); a quoted 🧾 or the words `── receipt` no longer close the task. - abofles: `/abofles:agents`, `guide`, `memo`, `bill` and `status` act on the session you type in — `ABOFLES_SESSION`, else the command's own `--session ${CLAUDE_SESSION_ID}`, else the newest prompt among this folder's sessions — never a session of another project (`status.py` and `bill.py` no longer fall back to the newest ledger of any project: with none for this session they say so and show no spend); `--resolve` acts on the same ledger the display shows; an empty `--session ""` from a command is ignored. - abofles: the first two fetch-by-key connector reads of a task (`getJiraIssue`, `getConfluence*`, `get*Page`) do not count toward the direct-read warning; searches and lists still count. - abofles: Judge and Verifier are told that ticket and page text comes in the prompt (they cannot reach Jira); a missing one is reported as `ticket text not given`, and the skill tells the main session to paste the key fields into their prompt. - abofles: a receipt still owed when a turn ends is carried to the next turn's Stop. - abofles: the free connector reads (first two fetch-by-key per task) reset at the first real user prompt with no open task. - abofles: ledger rows record `how` (where the graded text came from: hint, transcript, stub, no_reply, lost) and `gate` (judge / explain) for the scripts. - abofles: a row closed as lost is graded again when the background reply arrives late. - abofles: a session that runs an old plugin cache copy while a newer one is installed gets one warning line, once per session, to restart Claude Code. ## abofles 1.3.0 · abofles-guide 0.3.0 — 2026-10-15 - abofles: after an agent-budget deny, one `AskUserQuestion` (raise the budget?) within 10 min goes to the user with no Judge — the budget question no longer needs a gate spawn; the deny text says so. - abofles: after every receipt the hook rewrites this session's memo Now block (Task `none — between tasks`, Done/Left/Next/Verify from the receipt, Decisions kept, Tree from git status; `· from receipt`); `/abofles:memo` still writes a richer one. - abofles: guide items carry a stable `id` (`fail:`, `dec:`, `ctx:` …), the key for the band's hide and arm state; `next_setting` caps: Work/Verify stop at `opus·xhigh` except T4 (`opus·max`), and the 🟡 item says stop and report when no step is left. - abofles: explanation-gate fixes — `SendMessage` re-ask unlock, background judge, multi-`VERDICT` capture, fenced `EXPLAIN`, stale `explain_prev`, `EXPLAIN` cut at a blank line. - abofles: ledger rewrite locked and atomic; old lock files pruned; `prompt_dirs` paths with spaces; `/abofles:resume` `last` skips trace lines. - abofles: `status.py` / `memo.py` validate arguments (exit 2) and exit 3 when the item is already resolved / not resolved (no trace written); `memo.py resolve|reopen --session `. - abofles: `Pace = fast` judges by the open task's tier (spawns and `main:` line): skipped only for T1–T2, T3–T4 or unknown still judged; overrides are read only from the `## abofles overrides` block. - abofles-guide: running guard, arm expiry, one toast per new red item, safer state reads, more tests. - docs: broken `topics.md` link, root README lists abofles-guide, Receipt style `next` bullet, question-gate box vs. markdown `EXPLAIN`, command flags and exit codes, route `main:` form, worker `LEFT` rule. - abofles-guide: a 🔴 open FAIL or ATTENTION gets **Re-check** instead of Resolve — it fills a prompt to run only the failed criterion and resolve the row only if it passes; nothing is closed on trust. **Resolve** (🟡 escalate, 🟠 owner decision) asks twice: the first press arms it as `Confirm resolve` for 5 s. After a resolve the band shows `✓ resolved: …` with **Undo** until the turn ends. - abofles: a question whose explanation needs markdown can no longer reach the user without it. The hook hands the model the Explainer's own block to print verbatim, then checks the transcript when the question is asked and refuses it once if no explanation was printed — on the first ask and on a re-ask. Seen twice in DCMP on 2026-10-14: the model announced the table in its thinking only. - abofles: each guide item carries `recheck` and `unresolve`; `status.py --unresolve `, `memo.py reopen `; `--trace` on resolve and reopen writes one `resolved by hand` / `reopened by hand` line in the memo Journal, shown in the resume block (not counted as a receipt). The guide is recomputed on every prompt and right after a receipt adds owner decisions, so the band never lags a turn. ## abofles 1.2.0 · abofles-guide 0.2.0 — 2026-10-14 - abofles: `SKILL.md` cut from 43 KB to 12 KB (about 8K fewer tokens per load); topic index, overrides, question gate, memo/guide, spawn details and the worked example moved to `references/*.md`, read on demand. `gen_matrix.py` writes the index to `references/topics.md`; `--check` added. - abofles: injected hook texts about 30% shorter (receipt, guard, context watch, question and agent gates, resume preamble); a background launch no longer repeats the receipt instruction (the Stop hook asks for it when the chain ends). - abofles: the stored guide carries an action per item (`fill` for the band's Use button, `resolve` argv), the context size against the next mark, the open task's spawns/cost/budget and the `main:` suggestion. `status.py --resolve --session `. - abofles: `status.py --resolve` and `memo.py resolve` exit 1 when nothing was resolved; the band's Resolve runs them with the hook's own Python; the question-gate text no longer claims a transcript check that does not exist; skill and `/abofles:route` quote the hook's current effort and `main:` lines. - abofles: a reader's cited paths also resolve from absolute folders the spawn prompt names (a worktree outside the session's project root was falsely downgraded to `low` in the live test); the 🟡 escalate prompt names the agent; resolving an already-resolved row says so. - abofles-guide: band hidden when nothing is open; Use fills the exact action for the top item; Resolve button settles a stale FAIL or a decided owner item in one click; Memo button near a context mark; the 🔴 toast fires again (its check cut the emoji in half); dim meter line (`ctx 412K / 500K · task 3 spawns · $0.41 · budget 2 left · suggest …`). ## abofles 1.1.0 · abofles-guide 0.1.0 — 2026-10-14 - Receipt: header computed by the hook from each subagent's transcript — `🧾 **N Spawns · T Tokens · $C · Saved P% ($S)** vs
` — then `done / why / check / open` as inline-code tags; `Receipt = full | short | savings N | off`. Usage deduplicated by message id; `/abofles:bill` and `/abofles:status` print the real spend. - Context watch moved to the Stop hook: the block is the last thing in the turn, never at the start of a prompt. - Display names as in the Claude app (`Opus 5.5 · High`, `Extra` for xhigh); route lines as ` \`Side\` → **Topic** · Model · Effort · Cap N · 🟢/🟡/🟠/🔴 Tier · Modifier`. - Work memo (`.claude/abofles-memo.md`): journal line per receipt, **Needs you** list (ATTENTION, FAIL, owner decisions), every `AskUserQuestion` with its answer; resume block after `/compact`, at session start and after `Break` minutes away (default 10); `/abofles:resume`. - Agent gate: asked once per session (Auto / This task: N / Main session only / Max or Switch-to-Opus) or set by `Agents = ask | auto | N | off`; the hook denies spawns over the budget and caps background agents at `Parallel` (default 3); `/abofles:agents`. Replaces the Low-model stop. - Question gate: no `AskUserQuestion` before a Judge agent (T1–T2 judges and explains; T3 Sonnet → Opus Explainer; T4 Opus → Opus · High) answered it from evidence or rewrote it with consequences; `Questions = judge | off`. Two new agents: `abofles-judge`, `abofles-explainer`. - Reader agent gets a read-only shell (`git show`, `gh`, grep); reply-format checks only on abofles agents; receipt nudge waits for background agents. - Overrides also read from `~/.claude/CLAUDE.md`. Hook unit tests: 200+. - Field-test fixes (DCMP session 5aa36499): memo, lock and overrides follow the project root the session started in, never a `cd` (a stray `.claude/abofles-memo.*` in a nested team repo was one `git add -A` from a commit); the memo is added to `.git/info/exclude`, the lock lives under `~/.abofles`. Receipt closes the task even while a background agent runs; answering the gate again restarts the budget count; gate spawns shown as `+ N gate` and never counted; a prompt without the abofles header is flagged (tier unknown). Question gate: WHY/EXPLAIN go inside the question text; Judge/Explainer write in the user's language with code names glossed; a reworded re-ask or a `SendMessage` to the gate agent passes without a second judge; dismissed questions are not decisions; Decisions keep the question sentence; an answered question ticks its Needs-you item; near-duplicate items deduplicated. Haiku 5.5 priced at $1 / $1.25 / $0.10 / $5 per MTok (was a tenth of that). Resume block: every `Left` line, the Now section's age and a `since` line when receipts came after it, no mid-word cuts. Read guard: shell appends, `wc`, `tail` are not reads; connector lookups (Jira) are. `STATUS: partial` with `GAPS/LEFT: none` is treated as done; cited paths resolved from the workspace root and sibling repos; Judge/Explainer replies no longer nagged about line length or absolute paths; the spawn-count nudge fires at 20/50/100/200 only. - Receipt and resume block: a field with 2+ items is written as indented sub-bullets (one idea per line) instead of a long `;`/`·` line; the hooks parse both shapes. Resume `away`/age shown as days, hours, min. - Question gate: a plain EXPLAIN goes inside the question as one line per option (the AskUserQuestion box renders raw characters); a markdown EXPLAIN (table, diagram, example) is printed in the same message right above the question under `📖 Explanation for the question below`, with a plain summary and a pointer inside. The Stop hook refuses to end a turn while a judged question is unasked (not while a gate agent is still replying). A judge reply with several VERDICT blocks is read block by block (HANDOFF/ASK win over ANSWERED, every ANSWERED is logged). Re-asks within 15 minutes (translations) pass; `Language = Arabic` override. - Guide: a hook-computed `🧭 next →` line (priority ladder: open FAIL/ATTENTION, judged question unasked, owner decisions, unescalated partial/low with the next Step 4 setting, background agents, context/memo hygiene, budget, the memo's Next) in every receipt and resume block; `/abofles:guide`, `Guide = off`. Companion mod `abofles-guide` (band above the prompt with `Use` / `Copy` / `Hide` buttons — a hidden band returns only when the ladder's priorities change — 🔴 toast) added to the marketplace. Guide FAIL/ATTENTION items are scoped to the open task (earlier tasks close with their receipt; `/abofles:status --resolve ` closes one by hand — its `NameError` fixed). - Receipt header names the dominant spawn (`largest: Work Opus 5.5 · High $11.00 (69%)`) when one spawn is half the cost or more. - Question text is QUESTION + WHY only; options live in the option cards and are never restated in the text. - Field-test round 2 (DCMP sessions 1–5): per-session Now blocks, session-tagged journal and Needs-you lines, session-aware resume block; receipts printed mid-turn close their task at the next spawn (budget and header restart per ticket); cited paths resolved from the project root two folder levels deep, `.../x.cs` abbreviations ignored; judge/explainer replies never trigger the escalation nag; `main:` line handed at the first work spawn of every task (model read first); no resume block on task-notification turns; header shows `N running` background spawns and no saving for gate-only tasks; `Language` override; skill line limit aligned to 200 chars. - Step 0 `main:` line computed by the hook at each task's first work spawn (fits or suggest, from the session's real model and effort) and handed to the model — it was skipped when a resume block opened the reply. - `Pace = normal | fast` override: `fast` raises the direct-read guard marks to 8/16/32/64, lets 🟢/🟡 questions skip the Judge and defaults the receipt to `short` — for users who found the plugin slower than working without it (hook latency measured at 20–30 ms per tool call; the time is in the spawns). ## abofles 1.0.0 — 2026-10-11 First release. - Routing matrix: 100 coding topics in 13 groups, 4 risk tiers, 5 sides (Read, Analyse, Plan, Work, Verify) → model (haiku / sonnet / opus / fable) + effort (low … max) + reply cap. - Five side agents with fixed reply formats; a `PostToolUse` hook checks every reply (caps, long lines, absolute paths, trailing text, false `done`, bad paths) and feeds escalation. - `PreToolUse` hook blocks any spawn without `model` and `effort`, or Fable on a Read / Work / Verify agent. - Escalation only on failure; verifier `FAIL` → fix in the main session, re-verify the failed criterion only; `ATTENTION` notes are never closed silently. - Session ledger (`~/.abofles/ledgers`), background spawns completed on `SubagentStop`. - Main-session modes: orchestrator (opus / fable, direct-read guard) and worker (haiku / sonnet, one manual-switch stop on the first T3+ task). - Session effort read from the hook input; used by the `main:` line, `/abofles:bill` and `/abofles:status`. - Progress line and 5-line receipt (`done / why / spend / check / open`), enforced by a `Stop` hook; optional `Receipt` output style. - Context watch: one bold block per mark (500K, 600K, 700K, 750K, 800K, 850K, 900K tokens; `Context marks` override) with both compact advices; `/abofles:memo` writes the hand-off memo. - Commands: `/abofles:route`, `/abofles:bill`, `/abofles:status`, `/abofles:init`, `/abofles:memo`. - Project overrides in `CLAUDE.md`: cell, tier, cap, main session, session cap, context marks. - 84 hook unit tests (`scripts/test_hooks.py`); 17 live test runs on a .NET project, average saving ≈ 33 % vs all-opus.