# AGENTS.md — Build guide for Cotypist

## Project scope
Start with a paused menu-bar assistant for one supported text field. Generate local next-word suggestions, show a non-focus-stealing panel and accept only the current suggestion with Tab; Escape or further typing dismisses it.

Catalogue verdict: kinda. The idea is simple: watch what you type, predict the rest, show it as ghost text, accept with Tab. The prediction part is genuinely easy now, a small local model or even a personal n-gram table over your own writing gets you most of the way. The hard part is everything around it: reading the current text field and caret position through the macOS Accessibility API, drawing an overlay that tracks a moving cursor, and not breaking in Electron apps, Chrome, terminals, password fields, or anything that reimplements text editing. You can build something that works in native AppKit fields in a couple of long sessions and feels magic. Getting it to work everywhere, at typing latency, without eating keystrokes or leaking into your bank login, is where the actual product lives.
Use the implementation prompt below to define the deliverable. Complete each phase's acceptance checks before extending the scope.

## Working agreement
- Inspect the repository and its existing instructions before choosing paths, dependencies or commands. Keep one coherent stack and explain changes to the proposed architecture.
- Plan a vertical slice that accepts a real input and produces the useful output described below. Persist only the state the prompt calls for; respect memory-only and upstream-managed workflows. Use fixtures only when they are clearly labelled.
- After scaffolding, document the actual install, development, check and build commands in README and keep them synchronized with the package or project manifest. Do not report commands as successful unless they ran.
- Work in small steps. At handoff, list implemented flows, checks actually performed, remaining blockers, and any credentials or provider setup the owner must supply.
- Do not publish, spend money, contact customers, delete source data or run irreversible migrations without the project owner's authorization.

## Prerequisites
- macOS, Xcode and a selected deployment target. Ask for each OS permission only when its feature is enabled; code signing and distribution are separate release tasks.
- Implementation components: Swift/SwiftUI and AppKit for a visible menu-bar state and non-focus-stealing suggestion panel. Accessibility context and prediction request/revision tokens held in memory only; a local n-gram predictor first, with an optional local model after cancellation is reliable. Codable settings contain only app allowlists, pause preferences and shortcut configuration. No persistent text-context database or history.
- Scope boundary: Universal inline completion and language-model quality are not promised.

## Stack and architecture
- Swift/SwiftUI and AppKit for a visible menu-bar state and non-focus-stealing suggestion panel.
- Accessibility context and prediction request/revision tokens held in memory only; a local n-gram predictor first, with an optional local model after cancellation is reliable.
- Codable settings contain only app allowlists, pause preferences and shortcut configuration. No persistent text-context database or history.
- Domain model: allowlisted apps, bounded caret contexts, ephemeral prediction requests and document/caret revision tokens

## Security and data integrity
- Start capture or automation paused. Respect denied permissions, exclude secure fields and sensitive applications, and expose a visible pause/stop control.
- Correctness boundary: Never persist raw surrounding text; secure fields are excluded and late predictions cannot insert into a changed caret or application.
- Keep caret context, suggestions and request tokens ephemeral. Cancel them on focus/caret/revision change, clear them on pause or app exit, and accept text only if the original target revision still matches.
- Export/import only allowlists, shortcut settings and preferences. Never serialize caret text, suggestions, prediction traces or surrounding document contents; restart paused with no restored text context.

## Agent implementation rules
- Project rule — data model: allowlisted apps, bounded caret contexts, ephemeral prediction requests and document/caret revision tokens
- Project rule — preserve this invariant: Never persist raw surrounding text; secure fields are excluded and late predictions cannot insert into a changed caret or application.
- Project rule — acceptance evidence: Switch apps while a prediction is pending and insert nothing; denied Accessibility permission leaves typing untouched and the assistant paused.

## Optional agent skills and references
- Optional external skill: [swiftui-expert-skill](https://github.com/AvdLee/SwiftUI-Agent-Skill/blob/main/skills/swiftui-expert-skill/SKILL.md) — Build and review native SwiftUI views, state management, navigation, accessibility and rendering performance. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
- Optional external skill: [swift-concurrency](https://github.com/AvdLee/Swift-Concurrency-Agent-Skill/blob/main/skills/swift-concurrency/SKILL.md) — Design Swift async tasks, actors, isolation, cancellation and safe data sharing, including Swift 6 migration. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
- Optional external skill: [sharp-edges](https://github.com/trailofbits/skills/blob/main/plugins/sharp-edges/skills/sharp-edges/SKILL.md) — Review security-sensitive APIs and configuration for dangerous defaults and easy-to-misuse interfaces. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.

Read the linked SKILL.md and its dependencies before adding a skill. Select only the skills matching this project's runtime and task; their documentation does not supply API access, credentials or approval to perform external actions. Pin the reviewed revision where the tool supports it. Follow the chosen agent's documented project-level installation mechanism.

## Distribution ideas
These are optional planning notes. Obtain the owner's approval before publishing or contacting anyone.
- Demonstrate the actual Cotypist-inspired workflow with owned or clearly labeled sample data: Start with a paused menu-bar assistant for one supported text field. Generate local next-word suggestions, show a non-focus-stealing panel and accept only the current suggestion with Tab; Escape or further typing dismisses it.
- Publish a reproducible walkthrough with this observable result: Switch apps while a prediction is pending and insert nothing; denied Accessibility permission leaves typing untouched and the assistant paused.
- Explain who can operate this scoped tool, its setup and ongoing costs, and these remaining product gaps: Universal inline completion and language-model quality are not promised. Avoid guaranteed savings, performance scores or implied endorsement.

## Engineering roadmap
1. Phase 1 — Scope and fixtures. Implement this bounded workflow: Start with a paused menu-bar assistant for one supported text field. Generate local next-word suggestions, show a non-focus-stealing panel and accept only the current suggestion with Tab; Escape or further typing dismisses it. Record prerequisites, select representative user-owned fixtures and document the unsupported features: Universal inline completion and language-model quality are not promised.
2. Phase 2 — Separate settings from ephemeral input. Persist only the app allowlist, shortcut and pause preferences. Keep bounded caret context, suggestions and request/revision tokens in memory with cancellation. There is no context history to migrate, back up or export; secure fields are excluded before prediction.
3. Phase 3 — Complete the first useful path. Implement the workflow's input, review and output interface, with clear controls and explicit empty/error states. Keep caret context, suggestions and request tokens ephemeral. Cancel them on focus/caret/revision change, clear them on pause or app exit, and accept text only if the original target revision still matches.
4. Phase 4 — Permissions and integration failure. Start capture or automation paused. Respect denied permissions, exclude secure fields and sensitive applications, and expose a visible pause/stop control. Request integration credentials and permissions only for the enabled feature; show a disconnected state instead of mock results.
5. Phase 5 — Portable handoff. Export/import only allowlists, shortcut settings and preferences. Never serialize caret text, suggestions, prediction traces or surrounding document contents; restart paused with no restored text context. Include setup, operating limits, fixture walkthrough and shutdown/restart instructions in the README.
6. Phase 6 — Acceptance scenarios. Switch apps while a prediction is pending and insert nothing; denied Accessibility permission leaves typing untouched and the assistant paused. Repeat the workflow after restart and with a denied permission or unavailable dependency; show recoverable failure rather than a success placeholder.

## Paid-product capabilities outside this build
- Coverage: your build will work in native text fields and fail or misbehave in Electron apps, browsers, and terminals
- Latency polish: suggestions that arrive after you already typed the words are worse than nothing
- Safety rails: skipping password fields, secure input mode, and sensitive apps takes deliberate work
- Personalization that actually improves over months of your writing rather than a one-off corpus dump
- Signed, notarized, auto-updating distribution and the permission onboarding that makes it survive OS updates

## Implementation prompt
WORKING SLICE
Start with a paused menu-bar assistant for one supported text field. Generate local next-word suggestions, show a non-focus-stealing panel and accept only the current suggestion with Tab; Escape or further typing dismisses it.

Build this scoped Cotypist-inspired workflow with a documented data model and visible failure states.

Architecture
- Swift/SwiftUI and AppKit for a visible menu-bar state and non-focus-stealing suggestion panel.
- Accessibility context and prediction request/revision tokens held in memory only; a local n-gram predictor first, with an optional local model after cancellation is reliable.
- Codable settings contain only app allowlists, pause preferences and shortcut configuration. No persistent text-context database or history.

Prerequisites and limits
macOS, Xcode and a selected deployment target. Ask for each OS permission only when its feature is enabled; code signing and distribution are separate release tasks.
Outside this release: Universal inline completion and language-model quality are not promised.

Data model and correctness
allowlisted apps, bounded caret contexts, ephemeral prediction requests and document/caret revision tokens
Invariant: Never persist raw surrounding text; secure fields are excluded and late predictions cannot insert into a changed caret or application.
Keep caret context, suggestions and request tokens ephemeral. Cancel them on focus/caret/revision change, clear them on pause or app exit, and accept text only if the original target revision still matches.

Security and privacy
Start capture or automation paused. Respect denied permissions, exclude secure fields and sensitive applications, and expose a visible pause/stop control.

Recovery and export
Export/import only allowlists, shortcut settings and preferences. Never serialize caret text, suggestions, prediction traces or surrounding document contents; restart paused with no restored text context.

Implementation order
1. Phase 1 — Scope and fixtures. Implement this bounded workflow: Start with a paused menu-bar assistant for one supported text field. Generate local next-word suggestions, show a non-focus-stealing panel and accept only the current suggestion with Tab; Escape or further typing dismisses it. Record prerequisites, select representative user-owned fixtures and document the unsupported features: Universal inline completion and language-model quality are not promised.
2. Phase 2 — Separate settings from ephemeral input. Persist only the app allowlist, shortcut and pause preferences. Keep bounded caret context, suggestions and request/revision tokens in memory with cancellation. There is no context history to migrate, back up or export; secure fields are excluded before prediction.
3. Phase 3 — Complete the first useful path. Implement the workflow's input, review and output interface, with clear controls and explicit empty/error states. Keep caret context, suggestions and request tokens ephemeral. Cancel them on focus/caret/revision change, clear them on pause or app exit, and accept text only if the original target revision still matches.
4. Phase 4 — Permissions and integration failure. Start capture or automation paused. Respect denied permissions, exclude secure fields and sensitive applications, and expose a visible pause/stop control. Request integration credentials and permissions only for the enabled feature; show a disconnected state instead of mock results.
5. Phase 5 — Portable handoff. Export/import only allowlists, shortcut settings and preferences. Never serialize caret text, suggestions, prediction traces or surrounding document contents; restart paused with no restored text context. Include setup, operating limits, fixture walkthrough and shutdown/restart instructions in the README.
6. Phase 6 — Acceptance scenarios. Switch apps while a prediction is pending and insert nothing; denied Accessibility permission leaves typing untouched and the assistant paused. Repeat the workflow after restart and with a denied permission or unavailable dependency; show recoverable failure rather than a success placeholder.

Acceptance
Switch apps while a prediction is pending and insert nothing; denied Accessibility permission leaves typing untouched and the assistant paused.
Use real source data or clearly labeled fixtures. Explain unsupported input and provider failures; do not fabricate analytics, delivery receipts, accuracy claims or security guarantees.

Optional agent guidance
Optional external skill: [swiftui-expert-skill](https://github.com/AvdLee/SwiftUI-Agent-Skill/blob/main/skills/swiftui-expert-skill/SKILL.md) — Build and review native SwiftUI views, state management, navigation, accessibility and rendering performance. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
Optional external skill: [swift-concurrency](https://github.com/AvdLee/Swift-Concurrency-Agent-Skill/blob/main/skills/swift-concurrency/SKILL.md) — Design Swift async tasks, actors, isolation, cancellation and safe data sharing, including Swift 6 migration. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
Optional external skill: [sharp-edges](https://github.com/trailofbits/skills/blob/main/plugins/sharp-edges/skills/sharp-edges/SKILL.md) — Review security-sensitive APIs and configuration for dangerous defaults and easy-to-misuse interfaces. Review its instructions and compatibility before use; it does not grant deployment, data-access or publication permission.
Project rule — data model: allowlisted apps, bounded caret contexts, ephemeral prediction requests and document/caret revision tokens
Project rule — preserve this invariant: Never persist raw surrounding text; secure fields are excluded and late predictions cannot insert into a changed caret or application.
Project rule — acceptance evidence: Switch apps while a prediction is pending and insert nothing; denied Accessibility permission leaves typing untouched and the assistant paused.

## Completion evidence
Demonstrate the prompt's acceptance scenarios against the scoped workflow. Include setup from a clean checkout and failure recovery. Check persistence across restart and export/restore only for the state the prompt says to store; for memory-only tools, confirm that temporary content is discarded as specified. Record actual results and remaining limitations. A detailed plan alone does not establish a working replacement.