next-prompt — next-prompt suggestions for pi and Oh My Pi

A pi / Oh My Pi (OMP) coding-agent extension that, after an agent turn fully settles and the input editor is empty, computes the single most logical next instruction you'd type and shows it. Three render modes:

• widget (default) — a colored below-editor line: ↳ next: <suggestion> (Alt-/ to accept)
• ghost — inline greyed ghost text in the input box (renders even when the editor is unfocused, e.g. after switching tabs/apps)
• both — inline ghost AND the below-editor widget simultaneously

OMP editor-coexistence note: ghost/both work on OMP too. OMP has no public editor-owner getter, so next-prompt cannot detect another custom-editor extension that installed first in the same session — the last installer wins there (Pi captures and can restore the prior owner). If ghost rendering ever fails, the default editor is restored and the mode falls back to widget.

Accept with Alt-/ (default; configurable) to fill the input box. Any other key dismisses; backspace down to empty re-arms the last suggestion after a short delay (no new model call). No suggestion while streaming; the suggestion is cleared and any in-flight model call aborted the instant you submit, start a turn, or the agent starts.

What a suggestion is — and is not

A suggestion is extension output only: one line next-prompt computes with its own model call and renders in the input area (ghost/widget). It is never part of the coding agent's reply, and the extension never injects it into the conversation — only accepting it (Alt-/) puts the text in the editor, and only the user submits it.

The required shape — the bare imperative the user would type next, not narration or a status report — describes this extension's rendered line only. It is not a style instruction for a coding agent's own messages, and reading it must never lead an agent to append prompt-shaped "suggestion" lines to its replies.

Manual trigger (autoTrigger: false)

By default suggestions are computed automatically after every settled turn. Set autoTrigger to false to switch to manual-only mode: nothing is computed automatically, and the accept key doubles as the trigger —

1. press Alt-/ once → compute the next-prompt suggestion;
2. press Alt-/ again (once it is shown) → accept it into the input box;
3. press Alt-/ while a suggestion is being generated → ignored (no-op; never two concurrent requests).

{ "autoTrigger": false }

When autoTrigger is true (the default), the accept key still triggers a fresh computation whenever no suggestion is currently showing.

Install

Pi and OMP auto-discover extensions from standard locations.

From npm / the pi package gallery

The published package is @gamaraan/next-prompt, published under the npm account gamaraan:

pi install npm:@gamaraan/next-prompt

A specific release can be pinned with:

pi install npm:@gamaraan/next-prompt@0.1.0

OMP

Install through the OMP plugin manager (observed from omp plugin --help / omp plugin install --dry-run):

omp plugin install npm:@gamaraan/next-prompt

Pin a specific release the same way:

omp plugin install npm:@gamaraan/next-prompt@0.1.0

After installing, run omp plugin doctor and confirm zero plugin errors. The package manifest uses the pi.extensions form, which OMP accepts directly and loads with its legacy @earendil-works/pi-* import remapping — there is no separate OMP package.

From GitHub

The source repository is gamaraan/next-prompt-extension:

pi install git:github.com/gamaraan/next-prompt-extension

To pin a GitHub release or commit, append the tag or commit reference:

pi install git:github.com/gamaraan/next-prompt-extension@v0.1.0

Manual — copy the file

Copy next-prompt.ts into your global pi extensions directory:

cp next-prompt.ts ~/.pi/agent/extensions/next-prompt.ts

Or, for a single project only, place it in the project-local extensions directory (loads after the project is trusted):

cp next-prompt.ts .pi/extensions/next-prompt.ts

Manual — reference from settings

Add the path to the extensions array in ~/.pi/agent/settings.json:

{
  "extensions": [
    "/absolute/or/relative/path/to/next-prompt.ts"
  ]
}

Restart pi (or start a new session) after installing.

Configure

Interactive: /next-prompt-config

Run the /next-prompt-config slash command for a guided walkthrough of every configurable option except systemPrompt (that one is config-file-only) — a model picker (lists all available models), render mode, thinking level, accept key, re-arm delay, transcript/recent-turn/suggestion caps, and cross-provider disclosure. Changes are saved to the host agent dir (~/.pi/agent/next-prompt.json on Pi, ~/.omp/agent/next-prompt.json on OMP) and the host reloads so they take effect immediately.

Config file

All fields optional. Merged global + project (project overrides global per top-level key; the nested model block is replaced wholesale, not merged). The config root comes from the host's CONFIG_DIR_NAME:

Host	Global	Project
Pi	~/.pi/agent/next-prompt.json	<cwd>/.pi/next-prompt.json
OMP	~/.omp/agent/next-prompt.json	<cwd>/.omp/next-prompt.json

{
  "model": { "provider": "ollama", "model": "deepseek-v4-flash:0731-cloud" },
  "thinking": "low",
  "acceptKey": "alt+/",
  "autoTrigger": false,
  "renderMode": "both",
  "rearmDelayMs": 2000,
  "maxTranscriptChars": 12000,
  "maxRecentTurns": 10,
  "maxSuggestionChars": 240,
  "allowCrossProvider": false
}

Field	Default	Notes
model	current model (ctx.model)	{ provider, model, sessionId? }. If the configured model isn't found, pi notifies once (warning) and falls back to the current model. For opencode/opencode-go the wizard also stores a sessionId, sent as the x-opencode-session header that gateway requires (it answers 400 MissingSessionID without one); it is minted once per model and reused, so suggestions keep a stable route across sessions. Other providers ignore it.
thinking	unset	Reasoning level for the suggestion model: "minimal"/"low"/"medium"/"high"/"xhigh"/"max". Set "low" for faster suggestions. Passed as reasoning to the model call.
acceptKey	"alt+/"	Any pi-tui KeyId (e.g. "alt+/", "ctrl+space", "shift+enter"). Intercepted before the base editor, so keys like ctrl+space (\x00) won't pollute the box. Accept only fires when a suggestion is showing and the autocomplete dropdown is closed.
autoTrigger	true	When true (default), suggestions are computed automatically after every settled turn. When false, manual-only: the accept key doubles as the manual trigger (first press generates, second press accepts, in-flight press is a no-op).
renderMode	"widget"	"widget" (below-editor line), "ghost" (inline greyed text in the box), or "both" (inline ghost + below-editor widget). On OMP, ghost/both work too (see the editor-coexistence note above).
rearmDelayMs	2000	Delay (ms) before re-arming the last suggestion after the user deletes back to empty. No new model call.
systemPrompt	built-in extractor	Config-file only (not prompted by /next-prompt-config). See SYSTEM_PROMPT in next-prompt.ts.
maxTranscriptChars	12000	Tail-truncation of the conversation transcript sent to the model.
maxRecentTurns	all	Disclosure minimization: only the last N user/assistant turns are sent (tool results are never sent regardless). Invalid values fail closed — suggestions are disabled.
maxSuggestionChars	240	Cap on the returned suggestion length (visible width; a hard code-point bound of 4× this value also applies, so zero-width payloads cannot bypass the cap).
allowCrossProvider	false	When true, a configured suggestion model on a different destination (provider + endpoint + model route) than the active model may be used — but only after explicit per-project consent (see Security). When false, fall back to the active model silently. Project config can never loosen a global false.
allowCrossProviderPairs	[]	Directional provider pairs that skip the consent dialog: [["activeProvider", "suggestionProvider"]] (e.g. [["opencode-go", "openai"]]). Set via the dialog's "Always allow for this provider pair" option (saved to the global config) or by hand. Case-insensitive; the reverse direction is NOT implied. Invalid entries fail closed — suggestions are disabled.
debug	false (absent)	When true, appends one JSON line per decision to <agent dir>/next-prompt-debug.log: event name, model, transcript/response sizes, stop reason, token counts — never transcript or suggestion text. Absent or false means no file is written at all. Toggled by the last step of /next-prompt-config.

Why alt+/ is the default accept key

• tab — conflicts with pi's path-autocomplete and /template dropdown.
• ctrl+tab — many terminals send it as plain tab/\t or swallow it (window/tab switcher), so it's unreliable.
• ctrl+space — works (sends \x00, which pi-tui maps to ctrl+space); the extension intercepts it before the base editor so it no longer pollutes, but some terminals remap Ctrl-Space to IME toggle.
• alt+/ — sends an unambiguous \x1b/ sequence, not bound by pi or most terminals, and is memorable ("accept the suggested next command"). Recommended.

Override with any KeyId, e.g. "acceptKey": "ctrl+space".

Security: cross-provider transcript disclosure

The extension sends the conversation transcript (user + assistant text only — tool results, thinking blocks, and tool-call arguments are skipped) to the suggestion model. This is no more than what the active model already saw if the suggestion model is on the same destination as the active model. A destination is the provider label plus the endpoint origin plus the resolved model's routing id: two different downstream models behind one gateway (e.g. openai/gpt and openai/claude at the same https://gateway.example/v1) are different destinations. If you configure a suggestion model on a different destination, the transcript is sent to that second destination, which may have different data-handling terms.

Cross-destination disclosure is opt-in and fail-closed:

• allowCrossProvider defaults to false. With false, a configured model on a different destination is never used; the extension silently falls back to the active model (and, when there is no active model, computes nothing).
• With true, the first time a different destination would receive the transcript, the host shows a dialog naming the destination and the transcript size, with three choices: Allow once (this project), Always allow for this provider pair, and Decline. "Allow once" persists consent per project + destination (provider + endpoint + model route) in <agent dir>/next-prompt-consent.json (a 0600 file outside the repository; ~/.pi/agent on Pi, ~/.omp/agent on OMP); declining blocks that destination for the rest of the session without re-prompting. "Always allow" additionally saves the directional provider pair ([active provider, suggestion provider]) to the global config (via the same atomic 0600 write), so that exact direction never prompts again in any project — the per-destination consent record is kept too, so the dialog also stays silent when the config write is refused.
• Consent is keyed by the full destination identity: changing the endpoint or the model route invalidates a stored grant and prompts again. Records written by older versions (without a model route) never match and also re-prompt — fail closed.
• Project config (<cwd>/.pi/next-prompt.json on Pi, <cwd>/.omp/next-prompt.json on OMP) is only honored for trusted projects, and can never loosen a global allowCrossProvider: false or increase a global maxTranscriptChars cap — repository content cannot silently redirect your transcript. Pi gates project config on its isProjectTrusted() API. OMP exposes no project-trust API, so OMP follows the configuration loader's default (trusted) and enforces the same global privacy floors and consent flow; there is simply no host project-trust signal to consult. An existing-but-unreadable or syntactically invalid global/project config disables suggestions entirely rather than falling back to defaults (both hosts).

Mitigations:

• buildTranscript redacts obvious high-entropy secrets (AWS AKIA…, OpenAI sk-…/sk-proj-…/sk-ant-…, GitHub ghp_…/github_pat_…, GitLab glpat-…, Google AIza…, Slack xoxb-…, JWTs, PEM private key blocks, and key=value assignment forms) from both user and assistant text before sending. This is defense-in-depth, not comprehensive secret detection — destination consent and least disclosure are the primary controls.
• maxRecentTurns minimizes what is sent by limiting the transcript to the last N turns.
• Suggestion output is sanitized before rendering: terminal control sequences (OSC/CSI/DCS/APC/PM/SOS — both ESC-prefixed and 8-bit forms), C0/C1 controls, DEL, carriage returns, bidi overrides, unpaired surrogates, and oversized zero-width payloads are stripped or bounded, so model text can never execute terminal commands (e.g. OSC 52 clipboard writes).

How it works

1. On session_start (interactive mode only — Pi TUI ctx.mode === "tui", OMP ctx.hasUI === true; headless/RPC/JSON sessions never compute), a global ctx.ui.onTerminalInput listener is registered to detect the accept key editor-independently. ghost/both install a render-only GhostEditor via setEditorComponent — never re-installed on settle. If another extension owns the editor on Pi, the ghost is still attempted (with a warning); only if the ghost install or its render pass actually fails does the extension restore the prior owner and fall back to widget mode. On OMP there is no editor-owner getter, so a failed ghost restores the default editor instead; OMP also has no host-side extension-editor teardown, so next-prompt resets its editor to default at the next session_start. The host clears extension listeners when the UI is reset; each fresh session_start (reload/new/resume/fork) re-registers exactly one listener and re-installs the editor once.
2. On completion:
   • Pi: agent_settled (its fully-settled contract) — if the editor is empty, the controller calls ctx.modelRegistry.complete(model, { systemPrompt, messages }, { signal, reasoning }) with the resolved model, the configured thinking level, and the (redacted, tail-truncated) transcript.
   • OMP: only a terminal agent_end (never when event.willContinue is true — continuations, automatic retries, and pending continuation turns produce no suggestion) — the controller lazily imports completeSimple from the remapped legacy pi-ai module and calls completeSimple(model, { systemPrompt: [prompt], messages }, { apiKey: modelRegistry.resolver(model), signal, reasoning }). OMP emits the extension agent_end before the session fully unwinds (ctx.isIdle() is still false at handler time), so the terminal event itself is treated as the settle signal on OMP; Pi keeps its agent_settled + real-time idle gates. Stale-output protection is identical on both hosts (input-generation bumps, aborts, and render-time guards). Pi never loads the OMP module; OMP never calls modelRegistry.complete.
3. The returned text is sanitized (terminal controls stripped, trimmed, de-quoted, de-fenced, collapsed to one line, capped) and shown — via setWidget (widget/both), via the inline ghost overlay (ghost/both), or both.
4. The accept key is intercepted before the base editor: if a suggestion is showing, it fills the editor via ctx.ui.setEditorText and swallows the key. Any other key dismisses the suggestion immediately (widget and ghost) and delegates to the base editor; deleting back to empty re-arms the last suggestion after rearmDelayMs (no new model call). Dismissing via Escape/arrows never re-arms, and typing invalidates any in-flight suggestion request.

Develop

Clone and run the checks with Bun:

bun install
bun run typecheck        # Pi API types (default tsconfig.json)
bun run typecheck:omp    # OMP 17.2.13 API types (tsconfig.omp.json)
bun test
bun run verify:package

The extension imports @earendil-works/pi-coding-agent, @earendil-works/pi-tui, and @earendil-works/pi-ai. These are provided by your pi installation — list them in peerDependencies with "*" (do not bundle). If your editor's TypeScript LSP can't resolve them, link them from your pi install's node_modules (git-ignored here; nothing is downloaded). The pinned @oh-my-pi/* packages are dev-only test dependencies for tsconfig.omp.json — they never appear in peerDependencies or runtime imports; on OMP the legacy @earendil-works/pi-* imports are remapped to the host's bundled packages at load time.

Edit next-prompt.ts in place and restart the host (pi or OMP) to pick up changes.

Compatibility

Pi: supported range 0.84.0 – current (0.84.0 minimum; 0.85.1 validated live on 2026-09-09). ModelRegistry.complete() — which the extension calls directly — was added in pi 0.84.0, so older 0.80–0.83 releases are not supported. CI runs the unit suite and typecheck against both the oldest supported (0.84.0) and the current (0.85.1) @earendil-works/pi-* packages (.github/workflows/verify.yml — compat matrix) on every PR and before every tag publish.

OMP: supported range 17.2.12 – 17.2.13 (17.2.12 was the researched API version; 17.2.13 validated live on 2026-09-09). OMP runs the extension through its legacy pi.extensions manifest and @earendil-works/pi-* import remapping — the same published package works on both hosts. OMP-specific behavior:

• Lifecycle: suggestions compute on a terminal agent_end only; willContinue: true events (tool-loop continuations, automatic retries) never compute. Pi keeps its agent_settled contract.
• Transport: OMP completes via completeSimple + the model registry's auth resolver; Pi keeps modelRegistry.complete.
• Rendering: OMP supports widget, ghost, and both; because OMP has no editor-owner getter, a ghost failure restores the default editor (Pi restores the captured prior owner) and another custom-editor extension installed first in the same session is not detected.
• Trust: OMP has no project-trust API; project config follows the loader default (global privacy floors and consent are unchanged).

CI runs a minimum/current matrix for BOTH hosts (.github/workflows/verify.yml, shared by PRs and tag publishing): pi 0.84.0/0.85.1 typecheck + unit suite; OMP 17.2.12/17.2.13 typecheck:omp + unit suite; the packed artifact is installed on both hosts (isolated OMP profile + omp plugin doctor must report zero errors; pi install + pi list must register the extension). Releases additionally require a dated manual TUI smoke record for the exact version in TUI_SMOKE_TEST.md — the tag cannot publish without it.

Manual TUI smoke tests

Terminal-rendering and live-interaction behavior cannot be fully verified in unit tests. Before a release, run the checklist in TUI_SMOKE_TEST.md in a real pi session AND a real OMP session, and record the results.

Design & development

Source and design discussion live in the GitHub repository.

License

MIT — see LICENSE.