prompt-enhancement <- Scaffolding architecture - #109
Open
hi0001234d wants to merge 97 commits into
Open
hi0001234d wants to merge 97 commits into
hi0001234d wants to merge 97 commits into
Conversation
…adapters Extracts the classification/session pipeline into src/core/ with zero Node deps so the browser extension service worker can run the same pipeline in B2+. - 7 port interfaces in src/core/ports/ (StoragePort, LLMPort, LogPort, etc.) - Core copies in src/core/classifier/ (all 9 classifier files, port-adapted) - src/core/session-state.ts (SessionStateManager via StoragePort) - src/core/stage2.ts + src/core/decision/pinch.ts (LLMPort-based) - src/core/decision/options.ts (re-export bridge, no file movement) - 6 re-export bridges in src/classifier/ (zero regression for all callers) - 3 CLI adapters: SqlJsStorageAdapter, OpenAILLMAdapter, loggerAdapter - auto.ts + stop.ts wired to core via adapters (same external signatures) - Frozen UI contract + mock harness with 6 fixtures in src/ext-browser/ui/ Zero regression: 118 test files, 4666 tests — all green. Claude/Cursor/Windsurf/Devin pipeline flows all intact.
8 new test files covering every new module introduced in B1: - core/session-state.test.ts — SessionStateManager with mock StoragePort - core/stage2.test.ts — runStage2 with mock LLMPort (+ pure helpers) - core/classifier/LLMProfileClassifier.test.ts — classifyUserProfileLLM with LLMPort - core/classifier/StreamBPresenceClassifier.test.ts — classifyStreamBPresence with LLMPort - core/classifier/PromptClassifier.test.ts — tidfClassifier optionality + tier cascade - core/decision/pinch.test.ts — generatePinchLabel + validatePinchLabel with LLMPort - cli/adapters/llm.adapter.test.ts — OpenAILLMAdapter param forwarding - cli/adapters/storage.adapter.test.ts — SqlJsStorageAdapter load/save/language All tests use mock LLMPort/StoragePort — zero regression on existing 4666 tests. 126 test files, 4866 tests, all green.
…g core pipeline - 7 browser adapters: IDB storage, fetch LLM, console log, browser clock, chrome.storage key store, offscreen embedding, sync-over-async memory bridge - MV3 service worker: full prompt→classify→stage2→advisory pipeline with SW ephemerality-safe IDB flush before every await - Content scripts: ISOLATED-world injector drops main-world.js <script>, forwards postMessage↔chrome.runtime; inject.ts mounts stub panel - MAIN-world script: patches window.fetch; exposes emit helpers for B3/B4/B5 - Offscreen document stub (B5 wires real Transformers.js) - Shadow-DOM stub panel with XSS escaping (B5b replaces with real UI) - Options page: saves OpenAI API key to chrome.storage.local - Chrome + Firefox manifests; esbuild script → dist/ext-chrome/ + dist/ext-firefox/ - 9 unit test files, 72 tests — 4938/4938 total green, zero regressions
Adds 7 test files for modules B2 shipped without coverage (service-worker, offscreen, embedding-offscreen, inject, main-world-injector, panel-adapter, options). Fixes found in the process: postMessage was broadcasting to '*' instead of location.origin, offscreen doc creation used an in-memory flag instead of hasDocument() (breaks on SW restart), manifests were missing content/inject.js so the panel could never mount, and options page gained a real key-validation Test button + self-check panel.
Adds build:ui-harness (esbuild) script and the standalone harness page, completing the B1 UI-developer handoff package. Verified live in Chrome: all 6 real fixtures load and fire panel events with zero console errors.
…on, idempotent-injection guards for stale content-script instances
…on, idempotent-injection guards, dedup fix
…on, live-confirmed bug fixes Capture side confirmed live on real Replit sessions (prompt-submit + response-stop each fire exactly once per turn). Fixes: missing icons, MV3 content-script bundle format, hidden console.debug logs, page-load history replay, duplicate bundling, stale content-script instances on extension reload, duplicate capture from Replit's own DOM re-renders, and a wrong response-stop selector (disabled-toggle never fired; Replit swaps in a separate stop-button element instead).
panel-adapter.ts awaited a content-script reply that main-world-injector.ts never sent, so every advisory silently resolved as dismiss regardless of what the user clicked. Added a symmetric nexpath:panel-event bridge.
…ash) + 4 other real bugs surfaced by finally running tsconfig.ext-browser.json Root tsc --noEmit excludes src/ext-browser entirely — never caught these. Added npm run typecheck:ext so this can't silently regress again.
…ngine gating Ports cli/commands/auto.ts's entire frequency-gating pipeline (warmup, dedup, cooldown, session cap, confidence overrides) into service-worker.ts, previously entirely absent from the browser. Adds frequency + role selectors to the options page, matching the CLI installer's exact value sets/labels/defaults, wired to browser.storage.local under the same key names the engine now reads.
…the mock harness The frozen ui-contract.ts was a stale, pre-brief draft that never matched either the brief document or the real engine's Stage enum. Rewrote it to match the brief exactly, with the stage field corrected to the real 8-value taxonomy (confirmed live via a real 'review_testing' advisory). Also fixed ui-mock-harness/ (6 fixtures + bootstrap script), which imports from this file and was broken the same way — the UI developer never received a working harness despite brief §7 promising one.
…owing it Confirmed real bug: capture silently stopped mid-session with zero error anywhere, matching an empty .catch() on the SW-bound sendMessage calls. MV3 service workers can be asleep/mid-restart when a message arrives — previously any failure there vanished without trace. Now retries once and always logs loudly on final failure.
… toggles Previously only watched childList mutations (element add/remove) — a longer, multi-action response may toggle the stop button's visibility via attribute change on a persistent element instead, which this silently missed. Also logs the detection moment to the page console, independent of message delivery, closing an observability gap.
…mechanism guesses Two prior fixes assumed specific mutation types (childList, then childList+attributes) and both still failed live. Adds a 1.5s interval poll that checks the stop button's actual presence directly, so detection no longer depends on knowing which DOM change Replit uses.
…ctor Three stop-button-based strategies all failed live on longer responses. Rather than a fourth guess at that same mechanism, adds a completely separate signal confirmed present in every test screenshot this session, both working and failing cases — Replit's own turn-completion label.
…ages + reconciliation sweep Root cause of 'first prompt captured, later prompts silent' (seen 2026-07-02 Publish flow and 2026-07-03 prompt 2): elements were marked seen before their text was read, permanently consuming shell-first renders. Not a regression — capture logic was untouched by today's commits; sendMessage theory disproven by the retry logging staying silent.
…; add source-side composer capture at submit
…B5 reuse; replit.ts becomes thin config
…line fallback; adapter capture tiers corrected to include dom-events
…at/v2), TipTap composer capture, stop/version response-stop detectors, per-agent inject dispatch
…t) + stage-2 verdict observability (silent LLM decline)
…<id> persist POST replayed historical prompts on load (spurious advisory)
…erged all projects into one session; landing pages skip capture
…ive MV3 teardown; options Recent Activity section + debug-channel access
…jected feedback clears funnel; auto-reload agent tabs on update; Replit-flow regression tests
…rectly after navigation into a project (Replit's new UI has no re-capture channel)
…in rejection, Recent Activity 20-row cap, fire-path verdict persistence
…pathEndsWith guard), TipTap composer, #usr message observer, sessionStorage-backed stash for hard-nav landing flow, inject dispatch
… (Lovable aria-label relabel) Lovable renamed its composer aria-label 'Chat input' -> 'Ask Lovable...', so the exact-label selector matched nothing and 'Send to your agent' fell back to clipboard instead of pasting + auto-submitting. injectViaSimulatedPaste now accepts a prioritised selector list (original label kept as top priority) and prefers the first rendered match. Purely additive; no existing selector or code path removed. Live-verified on lovable.dev.
…lassifier-parity fix(ext-browser): harden agent composer lookup against site DOM drift…
…he public repo Relocated docs/capture-recon (third-party API/DOM reverse-engineering) and docs/architecture (internal classifier design) to the private planning submodule. Cleaned the now-dangling recon-doc citations in code comments (comment-only). CHANGELOG keeps the Lovable composer aria-label drift-fix entry.
…lassifier-parity chore: remove internal reverse-engineering + architecture docs from t…
… merge/sub-11-into-scaffolding-architecture # Conflicts: # src/classifier/AbsenceDetector.ts # src/classifier/signals.ts
…rchitecture Merge/sub 11 into scaffolding architecture
…surface Injection is entirely declarative (content_scripts + web_accessible_resources), so the `scripting` permission was never used by any code path. Remove it from both manifests (now storage + tabs only) to minimise the store-review permission surface, and add a manifest guard test so an unused permission can't reappear. No runtime behaviour change: the permission was inert. Full ext suite green (7464/7464).
…ermission fix(ext-browser): drop unused `scripting` permission; pin permission …
…ub-11 User experience improvements sub 11 -> scaffolding
Chrome derives the store summary from the manifest `description` (not editable in the dashboard). Update both manifests to the VS Code extension's proven pattern — "AI coding assistant for Replit, Bolt, StackBlitz, and Lovable — a behaviour-guidance layer for vibe coders." — for brand consistency across CLI/VS Code/browser and store discoverability, within Chrome's 132-char limit. Guard test now asserts description parity + length. Inert metadata; no runtime change.
…ead decision) Drop the StackBlitz brand word from all user-facing text — manifest description (store summary), options-page supported-sites badge, and privacy page — per team-lead branding decision. Also removes a stale `scripting` mention from the privacy page (that permission was dropped in 0.1.5). No regression: the stackblitz.com host support (host_permissions, agent-hosts routing, service-worker patterns, and its tests) is intentionally KEPT — bolt.new / stackblitz behaviour is unchanged. typecheck:ext + ext suite (380/380) + build:ext green.
…gnment Fix/ext store summary brand alignment
…ub-11 User experience improvements sub 11 -> scaffolding
…wport Bug #1 — the advisory popup showed with NO selectable options (only "Send to your agent") on a short terminal/popup window. computeLayout's sticky-header tiers could fill the viewport down to avail === 0, so the option window slice came out empty. Two additive guards, no change to the avail/windowing formulas the budget tests lock: (1) reserve MIN_OPTION_BAND_ROWS (3) when dropping Tier 2 header rows so the header stops one step before it leaves zero room for options; (2) a hard floor in the option windowing loop that always keeps at least the focused option's first line, so options.length > 0 never renders zero options. +6 regression tests (rows 6/10/12/14 tall-header repro + roomy baseline) that fail on the pre-fix code. Full suite 7472/7472; no existing render-loop test changed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…viewport On a short browser window the fixed-size header squeezed .np-scroll to 0px, so no selectable option rendered and a blind Enter jumped straight to the send-confirm. Let the header shrink/clip (why-help drops, pinch+question stay) and floor the options band at min-height:56px. Inert on normal windows (byte-identical render). Reproduced live across 6 fixtures at vp<=230px (blank 19->0); +regression test.
…-viewport fix(ext-browser): never collapse the options band to zero on a short …
fix(render-loop): never render zero selectable options on a short vie…
Scaffolding architecture -> main
- drop the removed `scripting` permission from the store permission justifications (manifest + guard test are storage/tabs only); note injection is declarative so no scripting perm is needed - align the AMO reproducible-build Node version to 20 (what the publish workflow actually builds the uploaded zip with); 24 was the outlier
docs(ext-browser): correct PUBLISH.md permission + Node-version drift
Fix/ext publish doc corrections
…t/Lovable/Bolt order - move the Scope disclaimer from the top to the end (kept — clarifies this policy covers the browser extension only, not the CLI/VS Code) - 'four supported AI coding sites' -> 'three' (StackBlitz not user-facing) - reorder site lists to Replit, Lovable, Bolt.new (site-access + affiliation line + meta description); bump last-updated date
Reorder the site-name list in the manifest description (store summary), README, and CHANGELOG to match the privacy page. Display/metadata only — no code, routing, or functionality touched (routing keys off hostnames in agent-hosts.ts). Full suite 7474/7474, manifest guard 11/11, typecheck:ext clean, description 95/132 chars.
docs(privacy): team-lead edits — Scope box to end, three sites, Repli…
Fix/privacy page teamlead edits
Inline source maps inflated service-worker.js to ~5.3 MB, which Firefox AMO's
linter rejects ('File is too large to parse'). Store builds must not ship maps.
SW 5.31 MB -> 1.76 MB; store zips 1.6 MB -> 0.51 MB. Opt in for local debugging
with NEXPATH_EXT_SOURCEMAP=1. No functional change; typecheck:ext clean, suite 7474/7474.
Fix/ext drop inline sourcemaps
fix(ext-browser): drop inline source maps from the store build
The prior AMO add-on (gecko id {1c62eba0-...}) is being deleted to publish a
LISTED (public) submission. A deleted AMO add-on id is permanently unusable, so
a new gecko id is required. Version stays 0.1.5 — no functional change.
…o-id Fix/ext firefox relist new gecko
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.