Edit prediction

Active contributors: rtfeldman, benbrandt

Purpose

Edit prediction is Zed's inline AI completion. As you edit, a model proposes a multi-line change; you accept (Tab) or reject (Esc). It is distinct from LSP completions — it can rewrite code in arbitrary ways, not just append at the cursor.

Crates

Crate	Role
crates/edit_prediction	Core: provider trait, request lifecycle
crates/edit_prediction_context	Builds the context window sent to the model
crates/edit_prediction_types	Shared types between provider implementations
crates/edit_prediction_ui	Inline ghost-text and accept-toolbar
crates/edit_prediction_metrics	Quality + latency metrics
crates/edit_prediction_cli	CLI for testing edit-prediction outside the editor
crates/zeta_prompt	First-party Zeta model prompt
crates/copilot	Copilot adapter (legacy completions)

Key abstractions

Type	File	Description
EditPredictionProvider	crates/edit_prediction/src/...	Trait every backend implements
Suggestion	crates/edit_prediction_types/src/...	One proposed edit
Context	crates/edit_prediction_context/src/...	Surrounding buffer + cursor + recent edits
Inline view	crates/edit_prediction_ui/src/...	The ghost-text element rendered in the editor

How it works

sequenceDiagram
    participant E as Editor
    participant CTX as edit_prediction_context
    participant P as EditPredictionProvider
    participant UI as edit_prediction_ui

    E->>CTX: cursor moved / text changed
    CTX->>CTX: build context (buffer slice, recent edits, language, …)
    CTX->>P: request_suggestion(context)
    P-->>CTX: stream tokens
    CTX->>UI: render ghost text
    E->>UI: Tab pressed
    UI-->>E: apply suggestion

Providers

The default provider is Zeta, Zed's first-party model, with prompts in crates/zeta_prompt. Other providers wired through:

• crates/copilot — Microsoft Copilot completions.
• (Other adapters can be added by implementing EditPredictionProvider.)

The choice is a setting; the registry is wired in crates/zed/src/zed/edit_prediction_registry.rs.

Context construction

crates/edit_prediction_context decides what to feed the model:

• The current buffer slice around the cursor.
• Recent edits (so the model sees the trajectory).
• Language and file path.
• Surrounding open files (configurable).
• Project-level .rules content.

UI

When a suggestion arrives, crates/edit_prediction_ui renders it as ghost text inline in the editor with a small toolbar (Accept / Reject / Next / Previous). It composes with the editor's DisplayMap so wrap, fold, and scroll states stay correct.

Metrics

crates/edit_prediction_metrics records acceptance rate, latency, and per-language stats. These flow through telemetry only when the user opts in.

Integration points

• Reads: open buffers, recent edit history, project .rules.
• Writes: ghost text in the editor; on accept, applies a regular edit.
• Connects to: language-model providers (Zeta runs through the cloud gateway; Copilot uses the GitHub auth flow).

Entry points for modification

• New provider — implement EditPredictionProvider, register in crates/zed/src/zed/edit_prediction_registry.rs.
• Tweak context — crates/edit_prediction_context.
• Tweak prompt — crates/zeta_prompt.
• Inline UI changes — crates/edit_prediction_ui.
• CLI evaluation — crates/edit_prediction_cli.

Related pages

• AI agent — separate AI surface
• Editor