Rytora Composer

Three install paths — pick whichever fits your environment:

Requirements

• VS Code 1.96+ (or Insiders / any 1.96+ compatible editor with Open VSX support)
• Network access to buildlabs.rytora.ai (AI features call our API)
• A Rytora BuildLabs account — free to sign up, Pro+ to use AI features

What's in the suite

The umbrella rytora.rytora-code bundles nine extensions:

The extension uses your existing Rytora BuildLabs account. AI features (Tab, Cmd+K, Composer, Chat) require Pro+.

1. Open the Rytora panel in the Activity Bar (or run Rytora: Sign in)
2. Your default browser opens to https://buildlabs.rytora.ai/code/auth
3. Complete sign-in (or sign up if needed)
4. The browser redirects back to your editor automatically. The extension passes its host's URI scheme (vscode, vscode-insiders, windsurf, etc.), so the deep link works regardless of which IDE you installed it in.
5. Token is stored in your OS keychain (Keychain on macOS, Credential Manager on Windows, libsecret on Linux)

Inline AI completions appear as ghost text while you type. Press Tab to accept, Esc to reject.

• Triggers automatically 200ms after you stop typing
• Multi-line completions with correct indentation
• Context: ~20 lines before the caret + filename + language
• Cache: same prefix asked twice doesn't hit the API
• Toggle: Rytora: Toggle Tab Completions

Select code → press ⌘K (macOS) or Ctrl+K (Win/Linux) → describe the change.

• Floating prompt input next to your selection
• History accessible with ↑ / ↓
• Streaming diff inline (added in green, removed in red)
• ⌘↵ accept · Esc reject · ⌘R regenerate

Open Composer from the Activity Bar (or ⌘L / Ctrl+L). Describe changes that span multiple files; the AI plans, edits, and shows diffs per file. Composer is the primary surface for the features below (conversation history, memory, rules, thinking).

Modes

Six modes — switch via the popover in the bottom bar (⇧Tab to cycle).

Effort + Extended Thinking

Two independent controls in the Modes popover:

Conversation history

The client owns the source of truth for conversation history. On every turn it sends the full prior-turns slice to the server, so the AI sees full context even when subsequent requests land on different pods in a multi-pod deployment. Works identically for all providers (Anthropic, Moonshot/Kimi, OpenAI).

Context usage + auto-compact

A small circular progress pill at the top of Composer shows estimated token usage vs the active model's context window. Tooltip on hover: "X% of context remaining until auto-compact". Click to run /compact — Haiku summarizes old turns. Auto-compact fires at ~85% usage. Colors: gray < 60%, amber 60-80%, red ≥ 80%.

Auto-attach editor context

Composer captures your active editor + selection automatically. A pill at the top toolbar shows the current state:

• Showing 👁 — file/selection will be attached to the next message
• Not showing 👁‍🗨 — click to toggle on/off (persisted)

@-mentions

Type @ in the input to attach context manually:

Model selection

Default is Auto (no credits — a Haiku classifier picks the model). Switch manually via the preset dropdown — Sonnet/Haiku/Opus/Kimi/GPT — but manual selection drains your credit pool. BYOK keys unlock the corresponding models without consuming Rytora credits.

Accept / reject

Each edit appears as a card with diff preview. Accept individually or batch all. Optional: enable auto-branch in settings to create agent-edits-{timestamp} git branch before applying.

Rules are markdown files that the AI reads as a constraint on every turn. They're injected as <user_rules> / <project_rules> XML blocks before your prompt.

File locations & precedence

Precedence: project primary > multi-file > local. All three layers concatenate. User rules append separately.

@include directive

Compose rules from other files with @<path> on its own line:

Memory is a markdown-based long-term store the AI maintains across sessions. Different from rules — memory holds learned facts, not enforced constraints.

Memory types

File locations

• User memory: ~/.rytora/memory/<name>.md — applies across all projects
• Project memory: .rytora/memory/<name>.md — workspace-scoped (consider gitignoring)
• Index: MEMORY.md in each directory — one-line pointers to each topic file

File format

---
name: tdd_preference
description: User prefers test-driven development
type: feedback
---

Always write the test first, watch it fail, then implement.

**Why:** caught a regression in payment flow because test was retrofitted
and ended up testing the bug, not the spec.

**How to apply:** when adding new behavior, write the failing test in the
same commit as the implementation.

Auto-injection

Both index files (user and project) are injected into every AI turn as a <memory_index> block — the AI sees the list of available topics with their one-line descriptions. When the AI needs the full content of a topic, it calls the memory_recall tool with the file name.

Working with memory

• From Composer: ask the AI to remember something — it'll call memory_save with the right type, description, and content automatically
• Slash commands: /memory list, /memory save <name>
• Auto-extraction: at the end of long sessions (or on session switch with 30+ messages), Haiku proposes new memories from the conversation. You approve/reject each.
• Manual edit: just open the .md file in your editor — changes take effect on the next turn

Tree-sitter parsers extract symbols (functions, classes, exports) from ~100 languages. Embeddings stored locally for semantic retrieval.

• Initial scan: ~30s for 1k files on SSD
• Incremental: <1s per file save (file watcher)
• Storage: .rytora/index.db in your workspace (~10MB per 1k files)

Run Rytora: Publish or click the rocket icon. Rytora Code detects your framework automatically:

Type / in the Composer input. Available commands:

Open ⌘, and search "rytora". Notable settings:

• rytora-code.tab.enabled — toggle Tab autocomplete
• rytora-code.tab.debounceMs — debounce time (default 200ms)
• rytora-code.cmdK.heuristicCutoffLines — Haiku/Sonnet routing threshold
• rytora-code.composer.defaultModel — Auto / Sonnet / Haiku / Opus / Kimi / GPT
• rytora-code.composer.autoBranch — git auto-branch before edits
• rytora-code.indexing.enabled — toggle codebase indexing
• rytora-code.indexing.embeddingProvider — Voyage / Nomic / OpenAI
• rytora-code.updates.channel — stable / beta / alpha
• rytora-code.environment — production / staging / local

BYOK (Bring Your Own Key)

Use your own provider keys instead of Rytora credits — open Rytora: Manage BYOK Keys:

• Anthropic (sk-ant-…): unlocks Claude Sonnet/Opus/Haiku
• Moonshot (sk-…): unlocks Kimi K2.6
• OpenAI (sk-…): unlocks GPT-5 series, o-series, Codex

"Cannot read the extension" on install

The Marketplace served an old corrupted version because your editor's version is older than the extension's engine requirement. Upgrade to VS Code 1.96+ and re-install:

code --uninstall-extension rytora.rytora-code
code --install-extension rytora.rytora-code

Tab not appearing

1. Check status bar — does it say "Tab Pro+"? You need Pro+ plan
2. Open settings, ensure rytora-code.tab.enabled is true
3. Type at least 3 characters on a line (default minimum)
4. Wait ~200ms after typing (debounce)

Sign-in browser doesn't redirect back

1. Some hosts don't register a URI protocol handler by default
2. The page falls back to a copy-token button: click it, then run Rytora: Paste Sign-in Token in your editor
3. Or use a manual sign-in via API token in Settings → Rytora Code → Auth

Composer agent loop hangs

1. Click Stop button in Composer panel (sends AbortSignal to API)
2. Check rytora-code.composer.maxIterations setting (default 25)
3. Open Output panel (View → Output) and select "Rytora Composer" channel for logs

AI "forgets" what we discussed earlier

1. Check the context-usage pill — if > 85%, it's auto-compacting. Click it to force-compact manually.
2. Verify rules / memory are loaded: run /doctor
3. If you opened a new Composer session, history doesn't carry over by design — use /clear intentionally only.

Bypass mode toggle has no effect on destructive tools

1. First-time activation requires confirming a warning modal — if you cancelled, the toggle never enabled
2. Look for the ⏵⏵ red indicator on the mode button: if absent, Bypass is OFF
3. Pre-tool hooks are skipped under Bypass; if a hook is still blocking you, reload Composer to pick up the 1.0.5 change

Publish fails detection

1. Manually choose strategy: command Rytora: Publish → "Choose strategy"
2. Verify your project has a recognized framework signature
3. Container fallback: add a Dockerfile at the project root

• Microsoft telemetry endpoints fully disabled in this fork
• Rytora telemetry is opt-in (default off, prompted in welcome flow)
• Codebase indexing is local — embeddings (if cloud provider chosen) only send chunks per request, never your full codebase
• AI requests carry only the context you explicitly include (selection / @-mentions / files referenced by Composer)
• BYOK keys travel via per-request headers, never persisted server-side
• Your authentication token is stored in your OS keychain (encrypted at rest)
• Code never leaves your machine unless you invoke an AI feature

See the full changelog page for all releases.