GhostPrompt
Ghost-text completions for your Copilot prompts — write faster, think clearer.
Version 0.5.3 (host/ desmantelado en api/ + vscode/ + core/pipeline/; última publicada de referencia: 0.4.0)
Why GhostPrompt
Writing prompts in Copilot Chat is often repetitive and context-switch heavy.
GhostPrompt gives you an inline mini-composer inside VS Code with ghost-text suggestions, so you can draft faster and send to Copilot Chat without breaking flow.
Benefits at a glance
- Faster prompt drafting with inline continuation suggestions
- Safer usage controls (included-model policy, request governor)
- Better writing flow with keyboard-first interactions (
Tab, Enter)
- Compact controls directly inside the composer (policy/style/context/debug)
Preview
Screenshots are collapsed by default so the README stays scannable; click a title to expand. Images are shown at native resolution (no width scaling in markup) so text stays sharp; swap the PNG files if you want smaller repo size.
Video demo
Watch GhostPrompt in action on YouTube: GhostPrompt demo.
1 · Icon and inline composer
2 · Inline continuation in action
3 · Next to Copilot Chat
4 · Quick controls (model, style, context)
Key features
- Inline ghost-text completions — suggestions appear as continuation of your current text. Press
Tab to accept.
- One session, two views — the Activity Bar sidebar and bottom Panel mirror the same draft, settings, and suggestion state. Open both at once without drift; edits and chips stay in sync in real time.
- Always visible — choose sidebar, panel tab, or both; layout is preference, not duplicated state.
- One-key send —
Enter sends your prompt to Copilot Chat. Shift+Enter adds a newline.
- Request safety controls — dedupe, cache, cooldown, rate-limit, and session budget to prevent over-calling.
- Model policy controls — force included (
0x) models by default, with optional override.
- Non-intrusive — completions run via
vscode.lm; no draft editor tabs, no focus stealing.
- Private logs — accepted suggestions and sent prompts are saved in the extension's private storage (not inside your project).
- Project context (v0.4) — with
contextMode: project and project memory enabled, GhostPrompt can attach short excerpts from your workspace (bootstrap files + optional editor-ingest) to the LM instruction. Data stays under the extension’s global storage (see Privacy and on-disk data).
Privacy and on-disk data (v0.4)
GhostPrompt never writes project-memory JSON inside your repository by default. Indexed snippets and registry metadata live only under the extension host’s storage:
| Location |
Contents |
ExtensionContext.globalStorageUri/ghostPrompt/projectMemory/v1/ |
registry.json (workspace keys + lastSeenAt), stores/<workspaceKeySha>/entries.json (bootstrap + editor-ingest items), manifest.json. |
- What may be stored: truncated text derived from files you open or from standard bootstrap paths (
README*, package.json), plus metadata (relativePath, mtime, content hash, LRU timestamps). Nothing is uploaded by GhostPrompt itself beyond what you already send to your chosen completion backend (Copilot LM / OpenCode) when you request a suggestion.
- Opt-out / disable: set
ghostPrompt.projectMemoryEnabled to false to stop persisting and merging this store into suggestions (volatile bootstrap read may still occur depending on version; global toggle is the supported off switch for disk-backed memory).
- Editor ingest & watchers:
ghostPrompt.projectMemoryEditorIngestEnabled controls indexing focused files; ghostPrompt.projectMemoryFileWatcherEnabled controls proactive invalidation watchers (disable to reduce background file subscriptions).
- Delete locally: command
GhostPrompt: Clear Project Memory (This Workspace) removes the store for the workspace folder tied to the active editor (or the first folder in a multi-root setup when resolving that command).
- Stale stores: GhostPrompt may delete entire per-repo folders under global storage if unused longer than
ghostPrompt.projectMemoryUnusedStoreTtlDays (default 30).
For the full settings matrix, search ghostPrompt.projectMemory in VS Code Settings.
Requirements
- VS Code 1.90 or later
- GitHub Copilot extension installed and signed in (default completion backend)
Optional: OpenCode backend
If you set ghostPrompt.completionProvider to opencode (or include "opencode" in ghostPrompt.enabledCompletionSources), GhostPrompt connects to an already-running OpenCode instance via HTTP (default http://127.0.0.1:4096). You need the OpenCode CLI installed and running (opencode serve). Configure models and provider credentials in OpenCode as described in the upstream docs (OpenCode SDK). Authentication for third-party APIs is handled by OpenCode, not by GhostPrompt.
Connection settings: ghostPrompt.opencodePort (default 4096) and ghostPrompt.opencodeAuthToken (Bearer token, optional). GhostPrompt performs a lightweight health check (config.get()) before each suggestion request and maintains a session pool (TTL 5 min, max 4) to avoid create+delete overhead per request.
Model catalog: The webview lists models from OpenCode's config.providers() on demand. If you change providers or models in OpenCode externally, reload the window so the list and model routing match the new configuration.
Inline session reuse: GhostPrompt reuses pooled sessions for repeated inline suggestions. Stale sessions (unused > 5 min) are evicted automatically. If you notice odd context bleed, reload the window to force a new session cycle.
Completion provider (Copilot LM vs OpenCode)
| Setting |
Behavior |
ghostPrompt.completionProvider = copilot (default) |
Inline suggestions use vscode.lm (GitHub Copilot chat models). |
ghostPrompt.completionProvider = opencode |
Suggestions use your OpenCode installation via GhostPrompt’s embedded server (port 17433 by default). The webview lists models from OpenCode’s catalog; model ids use providerID/modelID. |
ghostPrompt.enabledCompletionSources (optional) |
Array such as [ "copilot", "opencode" ] to show both catalogs in one dropdown and route by selected model (providerID/modelID → OpenCode; Copilot chat model id → Copilot LM). If you never save this key, GhostPrompt keeps using completionProvider only (backward compatible). |
Switch providers in Settings (search ghostPrompt.completionProvider or enabledCompletionSources). With both sources enabled, the composer shows Motor: Copilot + OpenCode and Auto (Copilot primero) uses Copilot when available.
Related settings: ghostPrompt.opencodeExcludedModelIds (hide specific providerID/modelID rows), ghostPrompt.selectedModelId (auto or an explicit id), ghostPrompt.agentDestination (copilotChat | vsOpenCodeX — see Agent destination above).
Agent destination (who owns Send + chat thread): Setting ghostPrompt.agentDestination separates the suggestion motor (Copilot LM / OpenCode, chips in GhostPrompt) from where the final prompt is sent. Effective default: if VSOpenCodeX is installed and you have never saved agentDestination in User/Workspace settings, GhostPrompt behaves as vsOpenCodeX until you set it explicitly. The GhostPrompt view shows a Destino dropdown next to Motor when VSOpenCodeX is installed (Copilot Chat | VSOpenCodeX). With vsOpenCodeX, GhostPrompt hides the inline composer and Send to Copilot in the webview but keeps the configuration strip. VSOpenCodeX drives suggestions via ghostPrompt.runSuggestPipeline and receives UI via vsopencodex.ghostPromptInlineUi (see roadmap). If you choose vsOpenCodeX but the extension is not installed, GhostPrompt shows a one-time-per-session hint. Matriz motor/destino: GhostPrompt-motor-destino-matrix.md; roadmap v0.5.0.
OpenCode model tiers (dropdown): Tiers come only from catalog metadata (pricing multipliers such as 0x / 1x, free, or the provider id for the built-in opencode / typical local backends like Ollama). GhostPrompt does not infer premium vs free from model names. Rows without usable signals stay Sin clasificar / UNKNOWN. With nonPremiumOnly, models classified as premium are hidden from the list and excluded from Auto selection.
Quick start (30 seconds)
- Open GhostPrompt from the Activity Bar (chat-bubble icon) and/or the bottom Panel tab — both show the same in-progress prompt.
- Start typing your prompt — after a short pause, a ghost-text continuation appears inline (and matches if both views are open).
- Press
Tab to accept the suggestion and append it to your prompt.
- Press
Enter to send the final prompt to Copilot Chat.
Screenshots: expand 2 · Inline continuation in Preview above.
Integrated with Copilot
You can keep GhostPrompt near Copilot Chat and move quickly between drafting and sending prompts.
See 3 · Next to Copilot Chat in Preview.
Keyboard shortcuts
| Key |
Action |
Tab |
Accept the current ghost-text suggestion |
Enter |
Send prompt to Copilot Chat |
Shift+Enter |
Insert a newline in the prompt |
Settings
Completion provider & models
ghostPrompt.completionProvider: copilot (default) or opencode — legacy single-backend switch when enabledCompletionSources has not been saved; see Completion provider.ghostPrompt.enabledCompletionSources: e.g. [ "copilot", "opencode" ] for a unified model list (optional).ghostPrompt.selectedModelId: auto or a concrete model id; with OpenCode, ids are providerID/modelID.ghostPrompt.opencodeExcludedModelIds: array of OpenCode model ids to hide from the dropdown (only applies when the provider is OpenCode).
Model policy (Copilot LM)
ghostPrompt.suggestionModelPolicy = nonPremiumOnly (default): only included (0x) models are allowed when pricing metadata is available.ghostPrompt.suggestionModelPolicy = anyModel: uses the first available model (may consume premium quota).
Suggestion quality
ghostPrompt.maxSuggestionChars (default 180)ghostPrompt.suggestionStyle (concise | balanced | detailed, default balanced)ghostPrompt.contextMode (off | basic | project, default basic)
Request governor (cost/frequency protection)
ghostPrompt.minCharsForSuggestion (default 6)ghostPrompt.suggestionDebounceMs (default 800) — wait after you stop typing before requesting a suggestion (reduces LM calls while drafting; higher default helps avoid races with VSOpenCodeX)ghostPrompt.requestCooldownMs (default 500) — host-side cooldown when the same normalized text is requested againghostPrompt.cacheTtlMs (default 45000)ghostPrompt.rateLimitMaxRequests (default 90)ghostPrompt.rateLimitWindowMs (default 600000)ghostPrompt.sessionRequestBudget (default 300)
Debug mode
- Run command:
GhostPrompt: Toggle Debug
- Output channel:
GhostPrompt Log
- Settings:
ghostPrompt.debugSuggestions, ghostPrompt.logLevel
- With
ghostPrompt.completionProvider set to opencode (or OpenCode-routed models when using multi-source mode), suggestion requests also emit timing lines prefixed [opencode-perf] in GhostPrompt Log, tied to each webview captureId: providers snapshot (providers, providers-network-fetch-ms), pooled session reuse vs session-create, prompt round-trip, optional stream-first-delta / sse-consumer-settled, and opencode-lm-total. Only when ghostPrompt.debugSuggestions is on (Toggle Debug / chips). Same channel still shows [capture:…] stages for Copilot/OpenCode orchestration (request-start, request-success, etc.).
Inside the webview mini-input, you can also change policy, style, context, and debug from the control strip (chips at the top). See 4 · Quick controls in Preview.
Project memory (v0.4)
Requires an open workspace folder for paths relative to that folder. Summary — tune in Settings (ghostPrompt.projectMemory*):
| Setting |
Role |
projectMemoryEnabled |
Master switch for persisted reconcile + disk store. |
projectMemoryEditorIngestEnabled |
Focused-file excerpts (extension/size filters). |
projectMemoryMaxTotalBytes / projectMemoryMaxEditorSources |
LRU caps for editor-ingest pool. |
projectMemoryUnusedStoreTtlDays |
GC for whole store folders left unused. |
projectMemoryFileWatcherEnabled / projectMemoryFileWatcherThrottleMs |
Invalidate indexed paths on external change/delete (optional). |
Command: GhostPrompt: Clear Project Memory (This Workspace). Details: Roadmap-v0.4-project-context-store.md.
Manual QA checklist (before publishing v0.4.x)
- Multi-root: open a workspace with two folders; confirm each folder gets an isolated store (different hashes under
globalStorageUri/.../stores/); suggestions in project mode do not mix excerpts across roots.
- Clear command: run
GhostPrompt: Clear Project Memory (This Workspace); confirm the active folder’s store is removed and suggestions no longer include old excerpts until files are re-indexed.
- Context modes: with
ghostPrompt.contextMode: off — no project/bootstrap lines; basic — session-only context; project — bootstrap ± editor lines per settings above.
Developing and tests
From the repo root:
| Command |
What it does |
npm run validate |
ESLint on src + no circular deps (deps:circular) + extension tsc + webview typecheck (webview/tsconfig.json) + React webview build + verify:webview-bundle (src/system/build → out/build) |
npm run build:webview |
Build webview only: React app in src/ui/webview/react/ → src/ui/webview/dist/react/index.html + assets. The extension host loads this bundle via CSP-safe URI substitution in src/ui/provider/webviewHtml.ts. |
npm run typecheck:webview |
tsc --noEmit for the webview tree only |
npm run test |
Vitest unit tests (OpenCode is mocked; safe for CI, no network) |
npm run check |
validate + test — run before shipping or opening a PR |
Edit webview behavior in TypeScript under src/ui/webview/react/ (not hand-edit src/ui/webview/dist/react/index.html; it is regenerated). After changing webview sources, npm run validate or npm run build:webview refreshes the bundle.
Webview ↔ host message contracts
Checklist when you change message shapes: edit zschemWebviewMessages.ts first; update parseWebviewInbound.ts / webviewProtocols.ts if needed; run npm run check (includes zschemWebviewMessages.test.ts, webviewProtocols.test.ts, and the webview bundle build).
GhostPrompt registers two WebviewViewProvider instances (ghostPrompt.input and ghostPrompt.inputPanel) backed by the same HTML/JS/CSS bundle. Changes to the toolbar, chips, or protocol must stay symmetric unless a roadmap explicitly documents a divergence.
| Rule |
Detail |
| Single bundle |
HTML is built only in src/ui/provider/webviewHtml.ts → src/ui/webview/dist/react/index.html + React assets. Do not maintain separate templates per view. |
| Capabilities |
MiniInputViewProvider passes viewContributionId into window.__ghostPromptCapabilities; use it for layout flags only—keep suggestion/settings behavior identical across views. |
| Shared copy |
User-visible empty/error strings for the status line live in webview/src/lib/userErrorMessage.ts. Optional host toasts (src/host/suggestionHostNotification.ts) should stay aligned for the same actionable cases. |
| Regression tests |
Before merging webview or toolbar edits, run npm run check (includes src/ui/webview/webviewToolbarParity.test.ts, src/api/boundary/webviewProtocols.test.ts, src/ui/webview/webviewThemeTokens.test.ts). |
Architecture and dependencies (v0.3.2)
High-level roadmap: Docs/Plans/Roadmaps/Roadmap-v0.3.2-host-refactor-webview-tooling.md.
| Command |
What it does |
npm run deps:graph |
Lists the dependency tree from src/extension/extension.ts (uses madge; optional ad‑hoc inspection — deps:circular is what runs in validate). |
npm run deps:circular |
Fails with exit code 1 if circular imports are found (same entrypoint). Also runs automatically as part of npm run validate / npm run check. |
npm run verify:webview-bundle |
Runs the compiled smoke script out/build/verifyWebviewBundle.js (source: src/system/build/verifyWebviewBundle.ts). Checks that src/ui/webview/dist/react/index.html exists after the React build. Not shipped in the VSIX — .vscodeignore excludes out/build/**; this is dev/CI tooling only, not extension runtime. |
Soft size guideline: prefer keeping new host modules under ~400 lines per file unless the content is mostly data; split extractors before crossing ~800 lines without a strong reason (same spirit as roadmap Phase A).
OpenCode integration tests (optional)
These exercises use your real OpenCode CLI, GhostPrompt’s API client connecting to http://127.0.0.1:4096, and a configured model. They are off by default (describe.skipIf) so npm run test stays fast without CLI or API keys.
Prerequisites: opencode --version succeeds; providers and models are configured in OpenCode (same as using the extension with ghostPrompt.completionProvider: opencode).
PowerShell
$env:GHOST_PROMPT_OPENCODE_INTEGRATION = "1"
npm run test:integration
bash
export GHOST_PROMPT_OPENCODE_INTEGRATION=1
npm run test:integration
Optional — force a specific catalog id if auto picks the wrong model (providerID/modelID):
$env:GHOST_PROMPT_OPENCODE_MODEL = "your-provider/your-model-id"
npm run test:integration
The suite lives in tests/opencodeSuggestions.integration.test.ts.
Release (OpenCode): checklist mantenedor y notas de CI — Docs/Plans/Releasing-opencode-integration.md. Resumen: npm run check siempre; si tocaste el motor OpenCode, además GHOST_PROMPT_OPENCODE_INTEGRATION=1 npm run test:integration en una máquina con CLI. En GitHub: el workflow OpenCode integration es solo manual y requiere opencode en el PATH del runner (p. ej. self-hosted); el CI estándar (.github/workflows/ci.yml) ejecuta npm run check en PR/push a main sin OpenCode.
Roadmap
v0.2 is shipped; roadmap documents:
Full release history is maintained in CHANGELOG.md.
Release notes
0.5.3
- Desmantelado
host/ → api/ + vscode/ + core/pipeline/: api/ gestiona protocolos Zod, handlers inbound, settings y workspace getters; vscode/ contiene WebviewViewProvider, HTML/CSP y notifications; core/pipeline/ orquesta el flujo de suggestion (governor → LM → broadcast). 226 tests passing.
0.5.2
- Reorganización
src/ → dominios canónicos: core/ (lógica pura de suggestions: types, instruction, normalize, streaming, language, loading, sources, catalog, governor, session, context) y system/ (infra transversal: debug, log, contracts, build). Carpetas eliminadas: completion/, governor/, session/, debug/, log/, shared/, build/. 226 tests passing, npm run check verde.
0.5.0 (pendiente de publicación)
- VSOpenCodeX: coexistencia OpenCode (reintentos, sin servidor embebido en el puerto compartido si VSX está instalada y prefer activo), destino agente (
ghostPrompt.agentDestination), comandos ghostPrompt.runSuggestPipeline / vsopencodex.ghostPromptInlineUi, selector Destino en webview, debounce suggestions por defecto 800 ms.
- OpenCode observability: con debug activo, líneas
[opencode-perf] en GhostPrompt Log — ver § Debug mode.
- VSIX:
npm run vsix para validar localmente; ver CHANGELOG.md sección [0.5.0].
0.4.0
- Project memory: per-workspace-folder JSON under extension global storage (
ghostPrompt/projectMemory/v1/): bootstrap excerpts, optional editor-ingest with LRU/quotas, reconcile before LM requests when contextMode: project and projectMemoryEnabled.
- Hygiene: TTL GC for unused stores, optional file watchers on indexed paths only, command to clear memory for this workspace.
- Docs: README privacy section,
ARCHITECTURE.md storage update, roadmap v0.4 phases A–F complete.
0.3.0
- Architecture:
src/ reorganized into layered domains (core/, system/, engines/, destinations/, host/, session/, etc.) and a documented completion pipeline (CompletionProvider, provider registry).
- Copilot LM: existing
vscode.lm path unchanged as default (ghostPrompt.completionProvider: copilot).
- OpenCode (optional):
ghostPrompt.completionProvider: opencode, API client connecting to http://127.0.0.1:4096, CLI detection, webview catalog + exclusions, @opencode-ai/sdk bundled in the VSIX.
- Docs & tests: README provider section and Developing and tests; CHANGELOG; unit tests mock the OpenCode runtime (no network in CI); optional live OpenCode integration test behind
GHOST_PROMPT_OPENCODE_INTEGRATION=1 (npm run test:integration).
- Release: build a VSIX anytime with
npm run vsix; git tag v0.3.0 optional—see CHANGELOG.md.
0.2.5
- Unified GhostPrompt session — single host session state for Activity Bar and Panel: shared draft, settings, loading/suggestion UI, and one active completion request across views.
- Tighter style calibration — completion instructions use explicit
STYLE_CONCISE / STYLE_BALANCED / STYLE_DETAILED rules for repeatable length behavior.
- Narrow layout parity — control strip wraps instead of horizontal-only scroll; hints wrap on small widths; optional
__ghostPromptCapabilities hook for future layout flags.
- Regression tests — multi-view store notifications and style injection in completion requests (
npm run test).
0.2.4
- Pricing-aware model tiers: model classification now uses passive runtime metadata (
pricing, e.g. 0x, 0.33x, 1x) without extra model generation requests.
- No hidden quota usage for tier sync: GhostPrompt does not probe models for pricing; it only reads model catalog metadata.
- UI tier clarity: model labels now use
Included / Premium / Unknown and display pricing multiplier when available.
- Model list deduplication: repeated model entries (same visible label/tier/pricing) are collapsed so the selector does not show duplicates like multiple
GPT-4o.
- Provider grouping: model selector is grouped and ordered by inferred provider (OpenAI, Anthropic, Google, xAI, GitHub, Other) for quicker navigation.
- Final token UX polish: selector rows now show tier/cost with compact textual chips (
[INCLUDED 0x], [PREMIUM 1x], [UNKNOWN]) for faster visual scan.
- Loading feedback polish: status line now shows a spinner while suggestions are being generated (
Buscando sugerencia...).
- Normalization hardening: host-side normalization is now the only source of truth for spacing boundaries, reducing split-word artifacts in inline preview/accept.
- Better leading-space intent: completion instruction now differentiates between "continue current word" (no leading space) and "start new word" (single leading space).
- Terminology and reason cleanup: empty reason updated to
no-included-model and user-facing messages aligned to included-model policy.
- Dev tooling isolation: pricing audit helper remains under
Scripts/ and is not part of production extension packaging.
- Known runtime limitation: in some environments
gpt-5-mini and raptor can timeout without returning suggestion chunks; GhostPrompt now recovers gracefully and prompts retry/model switch.
0.2.3
- Punctuation boundary polish: better spacing when suggestions continue after punctuation (e.g.
:, ;, ,) with no extra duplicated separators.
- Model selector redesign: model dropdown with explicit tier labels and persisted preferred model selection.
- Model transparency: effective model metadata is now traceable in pipeline/debug and visible in webview runtime status.
- Scoped cache keys: cache now separates suggestions by language/style/context/model dimensions to avoid cross-configuration collisions.
- Language stability: auto language detection now uses confidence threshold + hysteresis + fallback for short/mixed inputs.
- Validation and packaging:
npm run check and VSIX packaging verified for 0.2.3.
0.2.2
- Request governor retune: more generous defaults (
requestCooldownMs=500, rateLimitMaxRequests=90, sessionRequestBudget=300) with clearer blocked-state guidance in UI.
- Context quality: new
contextMode=project, adding lightweight workspace/file/language/selection signals plus recent prompts.
- Ghost text robustness: improved normalization for overlap and incomplete trailing-word duplication; safer boundary insertion on
Tab acceptance.
- Suggestion language control:
auto detection from user input plus manual ES/EN override in the webview controls.
- Validation:
npm run check green.
For complete details and historical versions, see CHANGELOG.md.
0.2.0
- Inline ghost-text in the composer, including scroll/height behavior for long suggestions.
- Suggestion pipeline: typed results (
suggestion, empty, error, loading), no silent failures.
- Request governor (dedupe, cache, cooldown, rate limit, session budget) to limit accidental over-calling.
- Model policy: default included-model selection with optional
anyModel override; commands and settings for policy/debug.
- Quality controls: suggestion style (
concise / balanced / detailed), maxSuggestionChars, optional session context (contextMode).
- Webview UX: compact chip controls, keyboard/a11y polish, layout fixes for narrow Activity Bar views.
- Tests: Vitest suite (
npm run test / npm run check).
- Docs: debug flow guide, README screenshots (collapsible), repository metadata for marketplace.
0.0.1
Initial release: dual-panel registration (Activity Bar + bottom Panel), ghost-text completions via vscode.lm, Tab-to-accept, Enter-to-send.
License
This project is released under the MIT License.
Badges are informational only (versions, stack, and links). They do not imply endorsement by Microsoft, GitHub, or the VS Code team. GitHub, GitHub Copilot, and Visual Studio Code are trademarks of their respective owners.
Forks, clones, and renaming
The MIT License allows others to copy, modify, and redistribute the code, including under a different product name, as long as they include the original copyright notice and a copy of the MIT license in the substantial portions they distribute. That means someone could fork the repo and publish “AutoCompletion”-style marketing or chase SEO: that is legally permitted for the code under MIT, provided the license terms are honored.
What MIT does not automatically grant:
- exclusive use of the name GhostPrompt (that is trademark territory; a registered mark offers stronger protection than OSS license alone);
- any rights to use third-party trademarks (Copilot, VS Code logos, etc.) beyond what those companies allow in their brand guidelines.
If you want stronger naming protection, consider registering a trademark for the product name in your jurisdiction and publishing clear branding in the Marketplace listing first.
