feat: add advanced context actions and optional translation

This commit is contained in:
golem
2026-10-05 15:38:16 -06:00
parent 33c4dad00f
commit 5308932bf1
34 changed files with 1579 additions and 159 deletions
@@ -0,0 +1,166 @@
# Advanced Context Menu Implementation Plan
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
**Goal:** Add optional contextual prompt presets and opt-in translation while preserving the existing simple action.
**Architecture:** Pure modules define settings, actions, and launch preparation. Firefox adapters register menus and open tabs/windows; a packaged dialog handles language entry, truncation consent, and errors. Direct and dialog launches share preparation and opening code.
**Tech Stack:** Firefox Desktop MV2 WebExtensions, ES modules, vanilla HTML/CSS, node:test, web-ext; Node.js 22+.
**Spec:** `docs/superpowers/specs/2026-10-05-advanced-context-menu-design.md`
## Global Constraints
- Firefox Desktop remains the only target.
- Preserve the current `storage`, `menus`, and `activeTab` permissions.
- No new host permissions, telemetry, intermediary service, runtime dependency, or API key is part of this design.
- Retain the current 8,000-character serialized URL budget.
- Users still review and submit in ChatGPT; no auto-submit or internal ChatGPT request modification.
- Options remain inline in `about:addons` and use existing visual styles.
- Preserve existing uncommitted changes; they are the working baseline, not disposable noise. Review the working tree before creating an isolated execution checkout; do not branch from HEAD alone and lose that baseline.
- Stage task-specific changes only. Many target files already contain prior work: inspect each staged diff and avoid committing unrelated hunks.
## Review Focus
- A selection inside a link must expose selection actions only (Task 2/5).
- A repeated `{query}` template must remain intact while budgeting repeated source text (Task 3).
- A language containing `&` or non-BMP characters must display literally and obey the 100-code-point limit (Task 1/5/6).
- Background suspension or two simultaneous translation dialogs must not mix or lose an acknowledged request (Task 4).
- Cancel, duplicate Enter, and a failed launch must never send twice or report success prematurely (Task 4).
## File map and contracts
- Modify `lib/defaults.js`, `lib/settings.js`: new defaults and legacy migration.
- Create `lib/actions.js`: context selection, action IDs, prompt rendering, language labels.
- Modify `lib/build-url.js`: retain URL encoding, expose untruncated prompt serialization.
- Create `lib/prepare-action.js`: pure validation and length decisions.
- Create `background/open-chat.js`: tab/window opening with fallback.
- Create `background/dialogs.js`: initial request handoff to packaged dialog.
- Create `background/menus.js`: menu registration and visibility updates.
- Modify `background/background.js`: compose modules and route entry points.
- Create `dialog/action.html`, `dialog/action.css`, `dialog/action.js`, `dialog/controller.js`: accessible dialog shell and controller.
- Modify `options/options.html`, `options/options.js`; create `options/menu-settings.js`: new controls, dependencies, validation.
- Tests mirror these responsibilities under `test/`; docs changes remain in README/CHANGELOG.
Shared data types (plain JavaScript objects, documented with JSDoc):
`ActionRequest = { actionId, source, settings, language? }`, where settings is a
sanitized snapshot, source is a string, and actionId is one of `ask`, `explain`,
`summarize`, `debug`, `translate`, `ask-link`, `review-link`, `ask-page`,
`summarize-page`. Only `ask` permits an empty source (blank-chat shortcut).
`Preparation = {kind:'ready',url} | {kind:'language'} |
{kind:'shortened',url} | {kind:'error',message}`.
## Task 1: Settings schema and migration
**Files:** Modify `lib/defaults.js`, `lib/settings.js`, `test/settings.test.js`.
**Interfaces:** Preserve `mergeSettings(stored)`, `load()`, `save(partial)`.
New keys: contextMenuMode (`off|simple|advanced`, default simple), showTranslate
(false), translationMode (`ask|preferred`, default ask), translationLanguage
(empty). Export `normalizeLanguage(value) -> string` from settings: trim,
return empty on invalid type or more than 100 code points.
- [ ] Add failing assertions for `mergeSettings({showContextMenu:false}).contextMenuMode === 'off'`, absent legacy => simple, valid new mode wins over legacy, invalid new mode falls back through legacy, and unknown keys are removed.
- [ ] Add assertions that 100 emoji are accepted, 101 rejected, whitespace trims, and wrong language/mode/boolean types default. Update existing expected setting shapes without dropping their coverage.
- [ ] Run `node --test test/settings.test.js`; confirm new assertions fail for the intended missing behavior.
- [ ] Implement defaults and migration. `load()` requests both new keys and legacy `showContextMenu`; save writes normalized new keys, not a legacy boolean. Do not delete old data as an initialization side effect.
- [ ] Run `node --test test/settings.test.js`; require nonzero passing tests. Verify load against a fake storage containing only the legacy key.
- [ ] Review/stage only this task and commit `feat: migrate context menu and translation preferences`.
## Task 2: Context and dedicated prompt definitions
**Files:** Create `lib/actions.js`, `test/actions.test.js`.
**Interfaces:** `getContext(info) -> {kind:'selection'|'link'|'page',source}|null`;
`getActions(context,settings) -> [{id,title}]`; `renderPrompt(request,sourceOverride=request.source) -> string`.
`getContext` accepts menu event data, prioritizes nonblank selection then link,
then page URL, and excludes password inputs and media-only contexts. Empty
editable fields get no page group. Link/page URL validation belongs to Task 3.
- [ ] Add tests asserting linked selection chooses selection, empty selection chooses link, media-only and password contexts return null, and page background uses pageUrl. Selection order must be ask/explain/summarize/debug, with translate appended only when enabled.
- [ ] Add prompt assertions: default action applies the saved template; Explain/Debug ignore it; Translate includes the language; link-safety prompt asks for evidence/uncertainty and never claims completed checks. Preserve literal `$&`, `$$`, backticks, Unicode and braces in source text. Escape `&` in menu language labels.
- [ ] Run `node --test test/actions.test.js`; confirm intended failures.
- [ ] Implement fixed action definitions and dedicated prompts matching the spec table. Extract/reuse existing literal template replacement instead of implementing conflicting replacement semantics. Blank `ask` renders empty even with a nonempty template.
- [ ] Run `node --test test/actions.test.js test/build-url.test.js`; expect all passing.
- [ ] Commit reviewed task hunks as `feat: define contextual prompt actions`.
## Task 3: Safe action preparation and length decisions
**Files:** Modify `lib/build-url.js`; create `lib/prepare-action.js`, `test/prepare-action.test.js`; extend `test/build-url.test.js`.
**Interfaces:** `serializePrompt(settings,prompt) -> string` encodes an already
rendered prompt without truncating. Keep legacy `buildChatGptUrl` compatibility
for existing tests until routing is replaced. `prepareAction(request) -> Preparation`
uses Task 2 rendering, Task 1 normalization, and `MAX_URL_CHARS`.
- [ ] Add assertions for ready URLs <=8000, translation needing a language, invalid/credential-bearing/non-HTTP URLs producing error, and oversized URLs producing error rather than shortened. Accepted URL source must survive unchanged within decoded prompt.
- [ ] Add budget tests for repeated `{query}`, emoji, whitespace boundaries, hint overhead, oversized instructions, and the shortening marker. Every shortened result retains all fixed instructions and substitutes only a shortened source; a small source remains byte-identical after decoding.
- [ ] Run `node --test test/prepare-action.test.js`; confirm failures.
- [ ] Implement validation and binary search over source code points, measuring the fully rendered/encoded prompt each time. Reserve the marker `[Selected text shortened]` only when shortening. Prefer a late whitespace boundary as in existing builder. Reject unknown action IDs and missing required sources.
- [ ] Run `node --test test/prepare-action.test.js test/build-url.test.js test/actions.test.js`; require passing assertions and nonzero discovery.
- [ ] Commit reviewed changes as `feat: prepare actions with explicit truncation decisions`.
## Task 4: Launch helper and reliable extension dialogs
**Files:** Create `background/open-chat.js`, `background/dialogs.js`, `dialog/action.html`, `dialog/action.css`, `dialog/action.js`, `dialog/controller.js`, `test/open-chat.test.js`, `test/dialog.test.js`.
**Interfaces:** `openChat(browser,url,openIn) -> Promise<void>` rejects after
final failure. `showActionDialog(browser,payload) -> Promise<void>` transfers
`payload={request,preparation}` to the dialog. `initializeDialog({browser,document,window})`
owns UI and launch after handoff, using Tasks 1–3 and openChat.
- [ ] Write boundary tests: new-window fallback opens a tab, background-tab uses active:false, final failure rejects; language whitespace blocks submit, Cancel/Escape never launch, double submit launches once, rejected launch permits retry, shortening requires explicit consent.
- [ ] Write handoff tests with two opaque IDs and distinct sources: wrong sender/ID cannot claim data; after acknowledgment clearing background state does not affect dialog completion; closed/uninitialized dialogs release requests and show retry rather than processing missing data.
- [ ] Run `node --test test/open-chat.test.js test/dialog.test.js`; confirm failures.
- [ ] Implement tab/window helper and dialog shell. Use runtime extension-page messaging plus a random request ID, verify sender extension URL and associated tab ID, transfer a structured snapshot, acknowledge receipt, then release background payload. Keep launch handling active through handshake; register listeners before opening. Apply a finite 15-second initialization timeout, with clear retry copy; never time out user language entry after handoff. No selected content in URLs, logs, or storage.
- [ ] Implement dialog controller: validate language, rerun preparation, show shortening confirmation if needed, disable launch controls while pending, show errors/retry, close only after success. Use textContent for all source-derived UI. Window creation falls back to an extension tab; closing that surface cancels. Use the same shell for direct-launch error messages.
- [ ] Run boundary tests and manually probe actual Firefox event-page handoff/idle behavior before accepting this task. If lifecycle assumptions fail, correct the handoff design within the approved no-persistence/no-permission constraints before wiring menus.
- [ ] Commit task-specific files as `feat: add action dialogs and shared launch handling`.
## Task 5: Register adaptive menus and route the shortcut
**Files:** Create `background/menus.js`, `test/menus.test.js`; modify `background/background.js`.
**Interfaces:** `initializeMenus({browser,load,onAction})` registers menus,
storage listener, onShown/onClicked, and serializes rebuilding. onAction receives
an ActionRequest. `runAction(request)` in background uses prepareAction then
openChat for ready or showActionDialog otherwise; errors use dialog feedback.
- [ ] Add tests: Off has no root, Simple has one selection-only direct action, Advanced has one root with correct children; onShown reveals exactly one context group and refreshes; selection over link wins; live translation/mode changes rebuild without races.
- [ ] Add assertions that unexpected/stale item IDs or mismatched click contexts do not launch, preferred-language labels escape ampersands, and click creates one settings/source snapshot. Verify shortcut invokes `ask` independently of menu mode, including empty selection.
- [ ] Run `node --test test/menus.test.js`; confirm failures.
- [ ] Implement adapter using existing serialized removeAll/create pattern, namespaced stable IDs and visibility updates. Verify current Firefox menu APIs from Mozilla primary docs at execution time. Preserve the shortcut's focused-frame selection contract and existing fallback behavior; do not expand page-reading scope.
- [ ] Route both entry points through shared preparation, including truncation consent for Simple and shortcut. Remove obsolete menu listeners to prevent duplicates.
- [ ] Run all unit tests. Manually check AGENTS.md frame cases; if the existing ancestor-focus bug is reproduced, make the focused-frame guard correction with a regression check before claiming shortcut preservation.
- [ ] Commit reviewed changes as `feat: wire simple and advanced context menus`.
## Task 6: Options UI, release cleanup, and complete verification
**Files:** Modify `options/options.html`, `options/options.js`, `options/options.css` if needed; create `options/menu-settings.js`, `test/menu-settings.test.js`; modify `README.md`, `CHANGELOG.md`.
**Interfaces:** `initializeMenuSettings({document,onChange}) -> {render(settings),collect()}`
owns context/translation controls. Existing options renderer delegates these
keys rather than assuming every preference maps to a checkbox/text input.
- [ ] Write UI boundary tests for Off/Simple/Advanced dependent visibility, opt-in translation, preferred-language validation, retaining hidden preferences, migration display, and Restore defaults. Assert language length uses code points rather than HTML UTF-16 maxlength alone.
- [ ] Run `node --test test/menu-settings.test.js`; confirm failures.
- [ ] Add controls with approved exact labels, reuse existing styles and debounce/save behavior. Invalid preferred language displays inline guidance; launch behavior remains dialog fallback. Preserve active input against storage echoes. Update restore/shortcut copy that assumes five settings or a checkbox.
- [ ] Remove temporary URL-test HTML and handler. Update README/CHANGELOG and privacy section to describe URL sharing, menu modes, opt-in translation, length consent, and unchanged address-bar behavior.
- [ ] Run `npm test`; require lint with zero warnings, nonzero passing unit tests, and a successful ZIP build. Inspect ZIP entries for new dialog assets and absence of any `docs/` or `test/` entries; do not alter packaging exclusions unless inspection finds a defect.
- [ ] Run `npm run dev` and execute every manual check from the approved spec and AGENTS.md, including light/dark/narrow settings/dialogs, keyboard focus/cancellation, concurrent requests, idle/wakeup, long selection, URL rejection, menu precedence, restart persistence, and all shortcut frame cases. Record actual observed results; ask for user assistance if native interaction is unavailable rather than counting a headless launch as behavioral coverage.
- [ ] Run `git diff --check`, review permissions/search URL/package changes, and commit only reviewed task hunks as `feat: expose advanced menu preferences and usage guidance`.
## Completion and review
Execution must begin with a baseline snapshot/status review and the appropriate
isolation skill; unresolved existing work must not be overwritten. No release,
AMO upload, push, or merge is authorized by this plan. Finish with a review of
the complete change and a report of automated and actual Firefox checks,
including any checks still requiring user interaction.
Plan self-review: spec sections map to Tasks 1–6; all five Review Focus cases
are assigned tests. Shared signatures use the ActionRequest/Preparation types
above. Lifecycle handoff is intentionally verified before entry-point wiring.