[integrations] Chrome extension — Gemini bulk history sync (Phase B/C)#21
Open
alanshurafa wants to merge 25 commits intomainfrom
Open
[integrations] Chrome extension — Gemini bulk history sync (Phase B/C)#21alanshurafa wants to merge 25 commits intomainfrom
alanshurafa wants to merge 25 commits intomainfrom
Conversation
Chrome MV3 extension that captures AI conversations into Open Brain via the REST API. First-run config screen collects API URL and key (stored in chrome.storage.local). All ExoCortex-specific hardcoded Supabase project URLs removed — extension is fully configurable. Runtime host permissions model documented.
…licit flag not identity check
….debugger Adds the Phase B foundation for Gemini bulk backfill. Gemini exposes no public conversation API, so the extension attaches chrome.debugger to gemini.google.com tabs and watches for the one internal RPC that Gemini itself uses to load conversation history (batchexecute rpcids=hNvQHb). On loadingFinished the service worker reads the response body via the debugger protocol, parses Gemini's framed positional-array envelope, and yields one normalized turn per user/assistant exchange. All turns are funneled through the existing processCaptureRequest pipeline so they inherit retry queue, sensitivity filter, fingerprint dedup, and session metrics — no parallel /ingest path. This commit is debugger infrastructure only. Phase C (the Sync All orchestrator that drives per-conversation navigation) lands in a follow-up commit. Live StreamGenerate / ambient capture is NOT ported — the extension's public release deliberately dropped ambient capture, and this port preserves that policy. Manifest adds the minimum permissions needed: `debugger` (attach only to gemini.google.com, observe one RPC pattern) and `scripting` (for the Phase C sidebar enumerator). Version bumps 0.4.0 → 0.5.0. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds the Phase C orchestrator that drives full-history backfill. The
state machine lives in a pure helper (lib/gemini-sync-state.js) so it
can be unit-tested under node --test without a Chrome stub; the
orchestrator (background/gemini-sync.js) owns all chrome.* calls.
Flow:
1. Enumerate the Gemini sidebar via chrome.scripting.executeScript
(scrolls the list until IDs stop growing).
2. Open one background sync tab, drive it through each conversation.
3. Per conversation: register a waiter keyed by the conversation ID,
navigate the tab, wait for Phase B to notify via
notifyHistoryCaptured(id, totals).
4. Fingerprint dedup at the ingest layer guarantees re-runs are safe
(duplicate turns return duplicate_fingerprint / existing).
Resilience:
- Resumable across MV3 SW restarts via chrome.storage.local state.
- User-cancelable at any time; Sync All button relabels to Resume
Sync when a run was paused mid-flight.
- Tab-health check before every navigation catches Google bot
challenges (CAPTCHA / login prompts that redirect off Gemini) and
transitions gracefully to a CANCELED paused state instead of
burning through the queue with silent timeouts.
Anti-bot throttle:
- 4–12 s jittered delay between conversations (sub-millisecond
precision so whole-second clusters don't fingerprint as a bot).
- 20–35 s "reading pause" every 10 conversations to break cadence.
Tuned to stay under Google's challenge threshold (earlier uniform
4 s cadence tripped the challenge around conversation 21).
Incremental path:
- syncIncremental() filters the sidebar against lifetime everSyncedIds
and navigates only the delta. Capped at 20 per run so scheduled
use stays quiet.
- Auto-sync (4-hour cadence, opt-in) drives incremental sync.
Tests (35 cases) cover state transitions, pendingIds deduplication and
cap enforcement, completion/failure bookkeeping, progress summary, and
the waiter registry's resolve/abort/abortAll semantics.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Loads the Phase B debugger and Phase C orchestrator via importScripts, adds GEMINI_SYNC_* message handlers (start, cancel, resume, incremental, status, auto-sync toggle), and drives the Gemini auto-sync alarm on install/startup. Exposes processCaptureRequest on the service-worker global so the Gemini debugger module can funnel history turns through the same ingest pipeline as manual capture. Classic-script function declarations are already global in the SW scope, but the explicit assignment pins the cross-module contract so it doesn't quietly break if the function is ever rewrapped. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the "Gemini bulk sync not supported" hint with the full Sync All History / Sync New / Cancel button group plus a 4-hour auto-sync toggle. The button auto-labels as "Resume Sync" when a run was paused mid-flight (Google bot challenge, SW restart) so the user can pick up where the state machine left off without re-enumerating. Progress surfaces via a 2-second polling loop that reads GEMINI_SYNC_STATUS, renders percent complete and the captured/dedup counters, and stops the poll as soon as the state machine leaves enumerating/syncing. Paused state shows the failure reason and — when the reason looks like a CAPTCHA — prompts the user to solve it first and then click Resume. The existing shared sync-progress-area (used by Claude/ChatGPT full sync) remains in place unchanged; Gemini's progress lives in its own gemini-sync-progress block so the two flows don't clobber each other. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
README: adds the "Gemini bulk history sync (Phase B/C)" section covering how the debugger-based capture works end-to-end, the "Debugging this browser" banner users will see while syncing, the anti-bot throttling strategy, and how to pause the flow (toggle Gemini off or dismiss the debugger banner). Updates the Supported Sites table row for Gemini, the Usage paragraph, and the Chrome Web Store permission justifications for `debugger` and `scripting`. metadata.json: bumps version 1.0.0 → 1.1.0, adds the `gemini-bulk-sync` tag, updates the `updated` date, and drops the `_todo` field so the current (stricter) metadata schema validates. The TODO it referenced (PR NateBJones-Projects#201 slug) is still explained in the README alongside the prerequisite link. License remains FSL-1.1-MIT. No new runtime dependencies, no binary blobs, no telemetry or third-party hosts. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
github-actions
Bot
added
dashboard
documentation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 770594fe1e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if (!over) return; | ||
|
|
||
| const thoughtId = active.id as number; | ||
| const newStatus = over.id as string; |
There was a problem hiding this comment.
over.id
When a card is dropped over another card, dnd-kit sets over.id to that card's numeric thought ID, not a kanban status string; this value is then sent as status, which fails server validation (VALID_STATUSES) and causes every such drag to revert. In practice this breaks moving items into populated columns because drops frequently land on cards rather than empty column space.
Useful? React with 👍 / 👎.
|
|
||
| try { | ||
| const result = await driveConversation(record.syncTabId, conversationId); | ||
| await updateState((rec) => stateMod.recordCompletion(rec, conversationId, result)); |
There was a problem hiding this comment.
This path records every driven conversation as a completion regardless of capture outcome, and recordCompletion adds the ID to lifetime everSyncedIds; statuses like disabled_platform/other non-ingest outcomes still flow through here, so incremental sync later skips those conversations permanently via filterToNewIds even though nothing was actually captured.
Useful? React with 👍 / 👎.
Comment on lines
+973
to
+974
| if (attachedTabs && typeof attachedTabs.has === 'function' && !attachedTabs.has(syncTabId)) { | ||
| return { healthy: false, reason: 'debugger not attached to sync tab' }; |
There was a problem hiding this comment.
The health check treats "debugger not attached to sync tab" as fatal, but immediately after creating/updating the sync tab there is an async attach race (tabs.onUpdated in the debugger module) where _attachedTabs can still be empty. In that window, the loop cancels the run before first capture; this should wait for attach (or defer this specific check) instead of hard-failing.
Useful? React with 👍 / 👎.
… false-negatives on debugger attach race On fresh sync, ensureSyncTab creates the Gemini tab synchronously but Phase B's chrome.debugger.attach runs async via chrome.tabs.onUpdated. mainLoop's first iteration called checkSyncTabHealthy immediately and could false-flag the run as paused with "debugger not attached" before attach won the race. Move attach awareness entirely into driveConversation (which already does it via waitForDebuggerAttach with a tolerant 2s budget), so the health check focuses on the real CAPTCHA/navigation-away signal. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…meIfInterrupted resumeIfInterrupted() is called at module load and was setting syncInFlight=true only AFTER an await loadState(). A popup-triggered startSync/resumeSync or alarm-triggered syncIncremental arriving during that async gap would observe syncInFlight=false, slip past the guard, and double-enter mainLoop against the same persisted queue. Claim the lock synchronously at the function top before any await. Release it explicitly when we short-circuit out (no-op / stale-reset paths); the happy path hands off to the existing finally block that already clears the lock after mainLoop completes. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…etion totals double-count Two pure-helper fixes with test coverage: 1. waiter register() silently overwrote an existing slot, so a reentrant driveConversation or stale slot that outlived its setTimeout left the old promise hanging until abortAll. Worse, the old timeout's abort-by-id call could then reject the NEW waiter. Now reject the prior waiter with a clear reason before replacing. 2. recordCompletion added the result totals every call even when the id was already in completedIds. A duplicate notify (retry, late-resolve, or re-sync of the same conversation) inflated captured/dedup/turn counts. Gate the totals fold on first-time completion only. Adds two tests (36, 37) that guard both behaviors. Full suite 37/37 pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… alarm + popup Four related fixes across the Phase B/C surface: 1. gemini-debugger.js — reject base64Encoded response bodies (the hNvQHb RPC is always text; base64 means wrong content type or URL misidentification) and cap parseable bodies at 8 MB so a malformed payload can't OOM the service worker. 2. gemini-debugger.js — on empty/malformed extractor output, recover the conversation id from the sync tab's /app/<id> URL and resolve the waiter with zero-result counts. Previously the orchestrator blocked on the full 15s capture timeout even when the response was visibly empty. 3. service-worker.js — Gemini auto-sync alarm now reads status first and skips syncIncremental when the prior run is paused on a Google challenge (state=canceled with pendingIds). Prevents alarm-driven re-runs from re-triggering the CAPTCHA until the user resumes. 4. popup.js — resume-vs-start click dispatch keys off a data-mode attribute set by renderGeminiProgress instead of string-matching the button text. Robust to future copy or localization changes. Tests 37/37 pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…execute parsing Adds 13 tests that exercise extractGeminiHistory against synthesized batchexecute payloads. Guards the shape layers codex called out as under-tested: - XSSI prefix handling and whitespace tolerance - empty / non-history / non-wrb.fr frames return null cleanly - single-turn and multi-turn payloads decode in historyOrder - turns with missing ids / empty prompts drop without aborting the batch - historic timestamp decoding from [seconds, nanos] pair - parseAdaptive length-prefix drift tolerance (+/- 5 bytes) - fuzz: extractor never throws on 7 varieties of garbage input Test total now 50 (was 37). All pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… + resumeIfInterrupted lock leak Round 2 review caught two bugs I introduced in the round 1 fix batch. 1. gemini-debugger.js — the empty-payload notifyHistoryCaptured (captured:0, total:0) path marked the conversation as completed and added it to everSyncedIds, silently skipping it forever on future incremental syncs. If the payload was just a transient parse failure or future Gemini format change, the user loses that conversation. Revert to letting the orchestrator's 15s capture timeout fire naturally, which routes the id to failedIds instead — user can retry via Sync All after an extractor fix ships. 2. gemini-sync.js — resumeIfInterrupted's finally block checked the local record.state (still SYNCING/ENUMERATING) to decide whether to release syncInFlight. The stale-heartbeat branch updated persisted state but not the local copy, so the finally wrongly assumed we were handing off to the happy path and never cleared the lock. Result: after one stale-interrupted detection, every subsequent startSync, resumeSync, and syncIncremental returned 'sync already running' until the SW terminated. Replace the state-based heuristic with an explicit handedOffToMainLoop flag set only when we actually proceed. Tests 50/50 still pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Owner
Author
|
Refreshing checks after markdownlint cleanup merged into fork main. |
Owner
Author
|
Refreshing checks after fork markdownlint workflow fix. |
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
1 participant
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.
Stacks on NateBJones-Projects#214 (chrome-capture-extension). Adds batchexecute-based history backfill (Phase B) and Sync All UI (Phase C). Pre-review fork PR; upstream PR will follow after cross-AI review.
Summary
Adds Gemini bulk history sync to the Chrome capture extension. Google does not expose a public conversation API for Gemini, so the implementation observes Gemini's own internal
batchexecutehistory-load RPC viachrome.debugger(Phase B) and adds a Sync All / Sync New orchestrator that walks the sidebar and drives the debugger through each conversation (Phase C).chrome.debuggerattaches togemini.google.comtabs and watches exactly one URL pattern:batchexecuterequests withrpcids=hNvQHb. OnloadingFinishedthe service worker reads the response body, parses Gemini's framed positional-array envelope, and funnels every user+assistant turn through the existingprocessCaptureRequestpipeline. No DOM scraping, no parallel ingest path.lib/gemini-sync-state.js) plus orchestrator (background/gemini-sync.js). Sync All enumerates the sidebar, opens a hidden tab, and drives it through each conversation one at a time while Phase B observes the history-load RPC. Sync New filters against lifetime synced IDs. Auto-sync (4h, opt-in) runs incremental. Resumable across MV3 service-worker restarts viachrome.storage.local.debuggerandscriptingpermissions; version 0.4.0 → 0.5.0. No other permission changes.gemini-bulk-synctag. License unchanged (FSL-1.1-MIT). No new runtime deps, no binary blobs, no telemetry.Test plan
node --test integrations/chrome-capture-extension/lib/__tests__/gemini-sync-state.test.js→ 35/35 passnode --checkon every new/modified JS file — cleanmarkdownlint-cli2on the updated README — 0 errorscheck-jsonschemaagainst.github/metadata.schema.json— validStacks on
This PR stacks on top of NateBJones-Projects#214 (chrome-capture-extension). It adds files under
integrations/chrome-capture-extension/; merging NateBJones-Projects#214 first, then this PR, applies both layers of the extension cleanly.🤖 Generated with Claude Code