How Naseem Works
The roles
Naseem splits a task between models, and always tells you which one did what. Choose them in Settings → Models. They fall into three groups: models that do the work, models that judge it, and models that take side jobs so the main model doesn't have to.
Doing the work
- Main model — does the work: reads, runs tools, edits files, and writes every answer you see. Runs: on every message. Every chat starts on it and can switch. Pick: your best model for real work.
- Sub-agents (Pro) — take delegated pieces of a task, several at once, each with only its own piece. Only their summaries come back to the main model. Runs: when the main model splits a job into independent parts. Pick: fast and affordable, at a lower effort — running in parallel, they can use many tokens. Automatic uses the helper model.
Judging the work
- Planner (Pro, optional) — before a task with several steps, writes a short plan — Goal · Steps · Files · Done when — that the main model then follows. It can look at the project but never changes anything. Runs: when the main model decides a task needs a plan, or when you press Plan first. Simple requests get no plan and cost nothing extra. Pick: a strong model. A strong planner with a cheaper main model is the classic saving.
- Reviewer (Pro, optional) — checks finished work against your request before it's called done, and sends it back with a reason when it falls short. It never writes the answer itself. Runs: once at the end of a task, at most twice per run. Pick: a careful model, ideally a different one from the main model, so it doesn't simply agree with itself.
- Brainstorm partner (Pro, optional) — another model the main model talks a question through with, for a few rounds, and brings back the ideas: a product choice, a plan before building, a fix that failed twice. Runs: when you ask — "brainstorm with opus 5.5", any model you've set up, "for 3 rounds" if you like — or in a chat with the Brainstorm mode on (the chip next to the model), at decision points only. The partner has no tools and sees only what the main model hands it; its ideas are shown as its own, with Naseem's view and the choice left to you. Pick: a strong model from a different vendor than your main one. 2 rounds by default (Settings → Models → Brainstorm).
Side jobs
- Helper — Naseem's own background jobs: project suggestions, project notes, MCP server descriptions. Runs: in the background, usually after a reply. Pick: leave it on Automatic — a fast, cheap sibling of your main model (a Flash, mini or Haiku class model). Avoid reasoning models here; they can take half a minute on a tiny job.
- Vision — looks at images when your main model can't see them. Runs: only when an image arrives and the chat model can't see it. Describe keeps your chat model in charge: the vision model writes one detailed description and the chat goes on. Switch hands that turn to the vision model, then switches back. Pick: any model that can see images.
- Reading — reads a long document once and writes notes the rest of the chat works from, instead of carrying the whole document along. Runs: when you attach a long document. Pick: a long context and a low price; cleverness matters less than room. Avoid reasoning models — they spend their budget thinking and hand back nothing. A local model keeps the document on this Mac.
- Image (Pro) — makes and edits pictures when you ask for one: a logo, a banner, an illustration. Runs: only when you ask, and it asks first unless the chat is Autonomous. Pick: Automatic uses your ChatGPT subscription if you're signed in (no key, no extra bill), then your OpenAI key, then Google. A chat can pick its own from the model menu → Image model, and every image is captioned with the model that made it.
Each role has its own effort. "Same as main model" follows your chat.
A good starting setup: your strongest model as Main (or as Planner, to save on Main), a different strong model as Reviewer, and Helper and Sub-agents left on Automatic.
Who did what
- Every reply names its model: Naseem · gpt-5.6-terra · Medium.
- A sub-agent card names its model: Sub-agent · glm-5.3.
- When a reviewer is on, every answer shows its verdict and who judged: ✓ Review · gpt-6-sol — passed, or sent back with the reason.
- Tools → Usage shows each model's tokens and the role it served.
What changed
- A task that edits files ends with a Changes card: every file it changed, lines added and removed, and each diff one click away. It's the thing to review — not committed, yours to decide.
- A sub-agent's card says what it actually did, from its own steps: 2 files changed · ran swift test, or read only.
- Naseem never commits unless you ask. A command that would commit, push, tag or hard-reset always asks first — even in a chat set to run without asking.
When a model fails
If a sub-agent's or the helper's model fails, that step runs again once on your main model, and the card says so: Sub-agent · glm-5.3 failed → gpt-5.6-terra. Stopping a task yourself never triggers a fallback. If the reviewer fails, the work stands, marked not reviewed.
One model for everything
In a chat, the model menu (chip icon) → Roles → This chat's model for all runs sub-agents and the reviewer on that chat's model and effort. It's useful for quick chats, local models, or when you want no surprises. The status line then says one model.
Who decides
1. The chat's "this chat's model for all", when on.
2. The chat's profile (Settings → Profiles → Roles for this profile).
3. Settings → Models.
4. Automatic — and the setting always shows what Automatic picked.
Planning
- The main model decides when a task needs a plan — several files or steps, or an unclear scope — and asks the planner. Simple requests get no plan, and cost nothing extra.
- Plan first — the clipboard button in the composer — asks for a plan for your next message.
- The planner can look at the project (read files, list folders, search) but never changes anything.
- The plan appears as a Plan card above the work, naming the model that wrote it. Its Done when list is what the reviewer checks.
- Show me the plan before it runs (Settings → Models → Planner): the plan waits for Run it / Edit / Cancel, and nothing changes until you choose. Rules and Telegram runs don't wait.
- A strong planner with a cheaper main model is the classic saving: think once on the expensive model, do the many steps on the cheap one.
Long conversations
- When a chat gets long, Naseem summarizes its older messages for the model and keeps the recent ones word-for-word, so long work stays fast and affordable and the model keeps track of the task.
- Your transcript never changes. A thin line marks where the summary ends, and View summary shows exactly what the model reads instead. Your own requests and exact file paths are kept as you wrote them.
- Every message re-sends the whole chat. Providers remember what they've read for a few minutes — the cache — so a chat you're actively using is fast and cheap; one you come back to later is read in full again.
- When Naseem compacts. With the recommended Auto strategy: at 250K tokens while you're working; at 100K when you come back to a chat after 10 minutes away (the cache has expired by then, so the summary is written while your message is answered, and the chat is short from then on); at 100K for models that don't cache (models running on this Mac, and Gemini); at three-quarters of small context windows.
- Choose a strategy in Settings → Agent → Long chats & cache: Auto, Keep more detail, Save cost, Custom (your own sizes and idle wait) or Off. Hover over the token counter in any chat to see the rule in force.
- Type /compact in any chat to do it now. Add words after it to say what to keep in detail, for example
/compact keep the API design. It's also in the model menu (chip icon) → Compact Conversation. If a model ever says a chat is too long, Naseem summarizes and tries once more on its own.