Every knob the uncle has.
Write a few lines in CLAUDE.md — or click them in the pane's Overrides editor. A project's block wins over your global ~/.claude/CLAUDE.md. 50 keys and forms, all optional: the defaults are already stingy.
A block to start from
## 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
Or let Claude write one: /abofles:init reads your code and suggests tiers for what it does.
All keys
| 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 / :<ref> / --mirror, reset --hard / --merge, checkout <commit> <paths>, rebase, pull --rebase, submodule deinit -f, branch -D, tag -d, clean -f, checkout -f / switch -f, checkout -- <paths>, 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: <Topic> / <Side> · T<n> header on line 1, Line cap: <N> 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: <Domain or Topic> = a, b | reviewer tags, none | the domain's reviewers | Replaces the specialist reviewers at T3 and T4. |
Workflow: <Domain or Topic> = Read, Work, Verify | steps | the intent's workflow | The steps of a change there (advice to the skill). |
Review: <Tag> / T<n> = model·effort | model·effort | the reviewer's cell | One reviewer's cell at one tier. |
Generated from the plugin's own README at build time, so it never drifts from the code.