Chat

Bridge's chat is intentionally different from a generic chat-with-an-LLM tool. Every conversation has an explicit reference set — specific notes, a constellation, or a plan — and the AI works from those references.

Why scoping matters

When you chat with a general AI assistant, you have to either (a) paste your context in every turn or (b) trust the AI not to drift. Both are unreliable. Bridge fixes this by making the reference set a first-class part of the chat: you pick what's in scope and the AI keeps to it. When it finds other notes that look related to your question, it sees them labelled as not part of the chat, and says so if it uses them. You can swap the references mid-conversation without losing the thread.

What the AI sees

• The notes you picked.
• Each constellation you added, by name, with its current members. Members are read when you send a message, so the AI sees the constellation as your graph shows it now.
• Related notes: notes that look related to your message, marked as outside the chat.
• Anything you target with @ for that one message (see Targeting with @ below).

The AI can't change anything

The AI can't add, connect, move or edit notes, constellations, the graph or plans. Ask it for a change and it tells you which command or screen does it. On Refine, the one change it can prepare is a draft edit to a plan file, which you approve.

Three scope types

Note references (Discover and above)

Pick one or more individual notes. The AI sees those notes' raw text and expansion (if available). Useful for asking the AI to summarize a few related notes, or to surface contradictions across them.

Constellation references (Discover and above)

Pick a whole constellation. The AI sees every note in the cluster plus the constellation's AI-generated name and theme. Useful for "what's the through-line in this cluster" or "what's missing from this body of thinking" questions.

Plan references (Build and above)

Pick a generated Build Plan. The AI sees all eight plan files plus the AGENTS.md overlay. Useful for "explain this plan to me" or "what's the riskiest part of this plan" questions.

Refine-tier: chat that writes

On the Refine tier, chat can do more than read — it can propose edits to plan files when a plan is in the reference set. The flow:

1. You ask the AI to change something ("simplify the data model" / "add an authentication step to the user flow").
2. The AI proposes the change as an inline diff in the chat.
3. You approve or reject each diff.
4. Approved diffs write through to the plan files immediately.

You stay in control — nothing changes in your plan without your explicit approval per diff. Rejected diffs disappear; you can ask the AI to try again with different guidance.

Modifying references mid-conversation

You can swap which notes/constellation/plan the chat references at any point in the conversation without losing the thread. Open the chat's menu and tap Notes: the picker opens with the current references checked. Check or uncheck, then tap Update references.

The line above the message box sums up what the chat references, for example 1 constellation · 2 notes, and tapping it lists them. A referenced constellation's notes count toward the constellation, not the notes; the notes are the ones you picked one by one. If regrouping the graph dissolved a referenced constellation, it's listed as Removed constellation.

This is useful when a chat that started about one cluster of notes leads to a tangent in a different cluster. Instead of starting a new chat, add the new reference and keep going.

If you add a note the chat picked to a constellation the chat references (with /connect or a note's Connect sheet), the chat counts it under that constellation: "1 constellation · 3 notes" becomes "1 constellation · 2 notes".

If a constellation the chat references no longer exists (a regroup dissolved it, for example), Bridge removes it from the chat and says so. Notes you picked yourself, or that another constellation in the chat holds, stay. The chat's line reads: "c4 · Energy no longer exists in your graph, so it was removed from this chat."

Targeting with @

Type @ to target something for one message. The list shows the chat's constellations and notes first, then every other constellation and every plan. A target outside the chat is marked This message only: the answer uses it for that message, and the chat's references and header don't change. A mention you pick from the list always links.

You don't have to tap a suggestion. A typed @c4 (a constellation's label), @[Exact name], or a one-word @Name links as if you'd tapped it when exactly one constellation, plan or note in the chat matches it, ignoring case. A typed @ counts at the start of the message, after a space, or after an opening bracket or quote mark, and never straight after a letter or digit, so an email address is never a target.

Only a bracketed name such as @[Marathon], or a label such as @c4, can stop a message from sending. Anything right after a label is ignored, so @c4's, @c4는 and @c4, all mean @c4. When one of these matches nothing, or several things, Bridge doesn't send, keeps your text in the box, and says why:

• "No constellation, plan or note matches @c2."
• "@Marathon matches several items. Tap a suggestion to pick one."

Any other @word that doesn't match exactly one thing, such as a handle, is sent as plain text.

Slash commands

Slash commands are a Refine-tier capability. On Refine, drive chat with ten commands:

• /expand — queue a deeper research pass on the notes you @-pick
• /plan — generate a Build Plan for the referenced notes
• /findings — show cross-project connections and contradictions
• /summarize — summarize the referenced notes
• /note — draft a new note from the text you give (you approve before it persists)
• /help — list available commands
• /connect — connect a note to another note, or add it to a constellation
• /disconnect — remove a connection between two notes, or remove a note from a constellation
• /replan — regenerate a plan in place from its source notes
• /editplan — draft an edit to one of a plan's files (you approve before it's applied)

See Slash commands for the full reference.

When a reply fails

A reply that fails shows as an error with the reason. When your AI provider refused the call, it's the provider's reason, for example "Your OpenAI key is out of credit. Add credit with OpenAI or switch model in Settings.", with an AI & Models settings button that opens that screen. Retry on the latest reply asks again without repeating your message. No reply waits more than about three minutes.

Archived chats

An archived chat is read-only on every plan: sending, changing its references and resolving a drafted note all wait until you restore it. Sending to one says "This chat is archived. Restore it to keep chatting."

Below Refine you keep one active chat, so restoring an archived chat needs the active one archived first: "You can keep one active chat on your plan. Archive your current chat to restore this one, or upgrade to Refine for unlimited chats." On Refine you can restore any number of chats.

Tier gating

¹ Chat starts at Discover. Below Discover, opening a chat shows the upgrade notice, with or without your own API key.

Next

• See Slash commands for the full command reference.
• See Refining a plan for the Refine-tier diff-approval workflow.
• See Tier comparison.