Argoforgebeta
Argoforgebeta
Nexus adds a task card to every working message. Your Claude first says in a few live words whether you are right and what it would do differently, then sizes the task after reading the project map. A simple task is done at once. A medium one gets the card — the goal and the items, each with a sign that it is done — a team and a short plan. A large one first goes through options, and you approve its plan. The answer is your Claude's own — no report template at the end.
Step by step
The server gathers, selects and checks; your Claude thinks.
The app cleans it and sends it with the project's counts and hashes. The project map stays on your Mac.
By fixed rules, with no model: the rules and memory that fit, a pointer to the project map and the task card. The full card comes once per session (and again after Claude Code compacts the conversation), then a short reminder. A question or a short reply gets no card.
Claude opens the answer with its view — whether you are right and what would be better — and sizes the task by the map and START.md: simple, medium or large. A simple one is done right away. For a medium or large one Claude restates the goal and lists the items, each with a sign that it is done. If it seriously disagrees, it asks and waits.
For a medium or large task Claude runs a script from the app. It sends the cleaned goal and the size — for a large task also the chosen option, cleaned. Which MCP servers are connected, it checks right on your Mac and does not send.
Up to 10 agents, 8 skills and 5 MCP servers from your project's pick, best for the goal first. Each says what it does and whether it is installed; an agent also says how to call it. Plus the plan template. No pick yet — Claude works without a team.
Claude picks the team. A medium task gets a short plan of 3–5 lines — steps, who does them, the check — and the work starts. A large task first goes through Metis's options and your choice; then Claude fills in the full plan — the steps, the files, the agent and skills for each step, what runs in parallel, how to check, the risks — and waits for your “yes”.
Claude answers in its own words. The app sends only what the count needs: the cleaned opening and closing of the answer (the whole cleaned answer only for a plan checked item by item), the agents and skills used as hashes, and whether files changed.
The size, the team used and whether the answer opened with the view — for the figures on the Nexus page. Nothing comes back to Claude. The text is not stored.
Templates
The server sends these templates and Claude Code fills them in, beside your files and in your language. They are the same files the turn uses.
The first lines of the answer to a working message. It sizes the task and says what comes next.
▍Nexus · Task card — the start of the answer, before any edit
- View: <1–3 live sentences to the human, not a template: they are right, almost right or not — and why; what you would do differently and why that is better. A serious disagreement — ask and wait for the answer, do not start the work>
- Size: simple | medium | large — <why, in one line>
The card — when your size's steps below call for it:
- Goal: <what should come out — 1–3 lines in your own words, not a retelling of the message>
- Items:
1. <what to do> — done when <a checkable sign>
2. …
- Where: <places from the map and START.md in .argoforge/brain/ the task touches>
- Unclear: <what to ask the human before the work; nothing — "—">
Then, by size
- simple — no card and no plan: View and Size → the work, leaning on the map, START.md and the project's rules.
- medium — the card → the team, always before the plan: "{{ARGOFORGE_TEAM}}" --session {{SESSION}} --size medium --task "<goal>" →
the plan from the template in its answer: the Team section (agents, skills, MCP servers and for which steps) is required, one line per step →
the work.
- large — the card → scout the code (subagents only read, in parallel) →
the method and options template: "{{ARGOFORGE_METIS}}" --method --session {{SESSION}} --kind code|ui →
the human chooses → the team for the chosen option: "{{ARGOFORGE_TEAM}}" --session {{SESSION}} --task "<goal>" --size large --option "<option>" →
the full plan from the template, with its Team section, in plan mode → "yes" → the work: independent steps by subagents in parallel.
The human decides the plan. Do not change an agreed plan silently. If it proves wrong — stop and say: "changing step N: it was X, it turned out Y, I propose Z". A plan in plan mode — wait for "yes"; otherwise go on. A small detail inside a step is not a deviation.
Rework. The human's message is rework of the previous task (wrong, broken, redo)? First mark the reason:
`"{{ARGOFORGE_REWORK}}" --session {{SESSION}} --reason <goal (misread the goal) | missed (an item was missed) | broke (broke something nearby) | tests (tests missing or failing) | place (the wrong place in the code) | style (against the project's conventions) | changed (the human changed their mind)>`A large task gets it after you choose among Metis's options and the team is chosen; Claude shows it in plan mode and waits for your “yes”. A medium task gets a short plan of 3–5 lines instead.
▍Nexus · Plan · <project> · <medium | large> - Goal: <from the task card> - Option: <the one chosen in Metis, with what was taken from the others; for medium — "—"> - Not touched: <what is nearby but outside the task> Team — picked before the first edit from the candidates above - Agents: <role> → <subagent_type> — steps <N, M>; no agent — "—, doing it myself: <why>" - Skills: <name> — what for — step <N>; call with the Skill tool before the work; none needed — "—" - MCP servers: <name> — what for — step <N>; not connected — "needs <name>, not connected"; none needed — "—" - Not taken: <candidate — why it is not needed; all taken — "—"> Steps 1. <what is done> — files: <path — place in the file — what changes> — who: <an agent from the team or "myself"> Depends on: <steps or "—"> · In parallel with: <steps or "—"> · Done when: <a checkable sign> 2. … Checks and risks - Items → steps: 1 → step 1; 2 → steps 2–3 - Check: <commands, tests, review and what counts as "passed"; independent ones in parallel> - Risks: <risk> → <how we reduce it> (from the chosen Metis option and from scouting) - Recommendations: <what to do along the way so we do not come back to it tomorrow>
Your data
Nowhere on the server: it gathers, picks and checks by fixed rules. The thinking — the task card, the size, the options, the plan — is done by your Claude Code, beside your files.
Rules
Nexus has 120 ms per message, counted from the moment your request reaches the server. If it runs late, it does not send the task straight to options; the rules and the card still arrive.
Your Claude sizes the task after reading the map. The wording only decides whether a message is a question or a short reply and, in Auto mode, whether it asks for ideas.
Agents, skills and MCP come only for a medium or large task, when Claude asks for them.
Options and the plan of a large task wait for your choice and your “yes”.
Turn on the A/B and half of the sessions work without the task card, then the two halves are compared. A verdict needs 20 sessions on each side.
How to use
“Unclear” lists what Claude needs from you; a question before the work is cheaper than redoing it.
Claude adds an important place it finds missing; you can add one too.
Auto sends “how best…” requests to options first. Plan does not — only a large task gets options, through the card. Options starts every message with them.
Then you see whether the card and the plans actually help your work.
The Nexus page shows tasks a day over 30 days and the 30- and 90-day totals: tasks, done without rework, came back, messages. These counts stay when the decision log is trimmed.
The card works only as well as your Claude follows it. The report template at the end is paused for now, so the answer stays your Claude's own; nothing checks it on the way out. The figures on the app's Nexus page show how often the card is followed.
Nexus works only in folders added in the app. In a Claude Code session started elsewhere, messages get only your rules and memory: no card, and neither the Nexus page nor the message figures on the Kairos page change.
Nexus can be switched off on its page in the app: messages then go without the card, the plan and Metis's options, while your rules, memory and the project's start notes still arrive.
A shorter reminder after the full card — one line, without the routes — is being measured as a server experiment on part of the sessions; it becomes the default only if it does no harm.