Cogis — AI Search

Single-purpose Chrome Manifest V3 extension that searches the user's own conversation history and returns clickable pointers into each AI lab. A launcher, not a knowledge base.

V1 distribution: dev-unpacked only. Load the extension/ folder in Chrome — there is no hosted app server.

Milestone status

Milestone	Scope	Status
M1	ChatGPT end-to-end (including Projects via search)	Merged to dev
M2	Perplexity end-to-end (Library + gated Spaces)	Merged to dev
M3	Claude end-to-end (Recents + Projects)	Merged to dev
M4	Gemini	Merged to dev
M5	Selector hotfix manifest (data-only)	Merged to dev
M6	Debug panel + optional ping	In progress

Docs

• docs/agent_blueprint.md — binding Phase 1 blueprint: architecture, behavioral acceptance criteria, escalation list
• docs/spikes/ — S1–S6 lab contract findings
• docs/handoff-prompts.md — official source of truth for milestone hand-off prompts: implementer (Cursor) and dual-review (Perplexity Computer). Attached as agent context; the operator pastes only the short per-milestone block
• docs/backlog.md — deferred follow-ups (BL-0xx)

Load unpacked

1. Open chrome://extensions
2. Enable Developer mode
3. Load unpacked → select the extension/ directory in this repo
4. Pin Cogis — AI Search and open the popup
5. Stay logged into ChatGPT, Perplexity, Claude, and/or Gemini in the same browser profile
6. After code changes, click Reload on the extension card before re-testing

Content scripts dynamically import shared extension/lib/* modules. Those paths are declared under web_accessible_resources per lab host (required for MV3 content-script import()).

Develop

npm ci
npm test
npm run lint
npm run format:check

DevDependencies only (ESLint 9, Prettier 3, Vitest 3). Runtime is plain HTML/CSS/JS — no bundler.

CI

GitHub Actions (.github/workflows/ci.yml) runs lint, format check, and unit tests on PRs and pushes to dev / main. The job name — and branch-protection check context — is build-and-test.

After the first green run: Settings → Rules → require status check build-and-test on dev and main.

Capability matrix (M1–M4)

Platform	Capability	Strategy	Deep link	Prefill
ChatGPT	full-text	Session search endpoint	https://chatgpt.com/c/{id}	—
Perplexity	title-match	Session list_ask_threads	https://www.perplexity.ai/search/{slug}	https://www.perplexity.ai/search?q=
Claude	title-match	Org session APIs	https://claude.ai/chat/{uuid}	—
Gemini	title-match	DOM-first history rail scan	https://gemini.google.com/app/{id}	—

Popup footnote: Some AIs do not support full-text search.

• ChatGPT: session-authenticated GET /backend-api/conversations/search (Projects included).
• Perplexity: session-cookie POST /rest/thread/list_ask_threads with search_term (Library). Spaces-only recovery via unproven per-Space routes is gated off until a live-proven path lands; C may still return Space-tagged threads when the lab includes them.
• Claude: session-cookie org APIs (GET /api/organizations → paginated chat_conversations + Projects enumeration). Client-side title filter; no stable prefill URL. Reach ceilings (soft caps): root ≈ 5 × pageSize (~100 conversations); per-project ≈ 3 × pageSize (~60). Soft caps still report empty by design when the scanned window has no title match (unread older history is a documented platform limit — BL-024). Incomplete scans are different: HTTP failure, deadline/fetch truncation, or unattempted/failed Projects → unavailable/timeout, not empty (see BL-022). Projects: routes inferred / not yet live Network-tab confirmed — Project-only findability unverified until operator smoke; ladder uses breadth-first page-1 across projects. DOM Recents search fallback deferred (BL-023; M1/M2 endpoint-only precedent).
• Gemini: DOM-first history rail (S4) — no stable first-party history search endpoint at desk. Title-match on /app/{id} links; cascade deep link → https://gemini.google.com/app (no prefill). Reach ceilings (soft caps): ≈ 8 scroll rounds / ~120 distinct history items still authorize empty by design when the history rail (or empty-history state) is proven and the scanned window has no title match (S6 — unread older history is a platform limit). A proven rail without the full Sign-in upsell is an S5 owner/authenticated signal (chip not required). The full S5 login shell still wins over a rail-only heuristic (logged-out empty-history copy must not become empty). Incomplete scans are different: budget exhausted mid-scroll → timeout (history_budget_exhausted); missing/unproven history rail → unavailable (history_rail_missing); never empty. Account chip alone does not prove the rail. Selectors are stub-level pending live polish (highest churn; M5). Flat history + any visible Gems/folder links exposing /app/ ids; no invented Projects clone.

Selector pack (M5)

Bundled extension/lib/selectors/local-pack.json (mirrored by local-pack.js) is always enough to search. On startup (and when a content script first loads its adapter), the extension may HTTPS-fetch an optional data-only hotfix from https://cogis.ai/packs/selectors.json in the background with a short fetch timeout. The remote document may overlay CSS selectors, wait-predicate strings, and same-host URL / relative endpoint path strings only — absolute off-host endpoints are rejected. Fetch or parse failure, timeout, and any executable-looking payload fail closed to the local pack; content scripts hydrate from local immediately so search is never blocked on the remote fetch. No eval, new Function, or remote code import. Pack version / source are exposed via getSelectorPackStatus() for the M6 debug panel.

Debug panel (M6)

Open Debug panel from the popup footer (or navigate to chrome-extension://<id>/debug/panel.html). The panel shows per-platform last latency, hit count, last status / errorCode, and the active selector pack version. Optional anonymous selector-failure ping is off by default; when opted in, the payload is only { platformId, selectorPackVersion, errorClass } (+ optional extension version). This build ships with no ping endpoint URL — pointing a build at a production ping endpoint is a human escalation.

Privacy & ToS

See NOTICE. No API keys. No stored auth. No message-body storage. Selector-manifest fetch carries no query text. Optional M6 ping never includes query text and stays off unless opted in.

License

MIT