ThreadBack AI v1.3.1 — Reliable Continuous Reconciliation

Fixes task-reconciliation starvation after Manifest V3 service-worker restart with persisted fair revisit scheduling and bounded follow-up reconciliation. Semantic thresholds are unchanged.

The BGE-small-en-v1.5 INT8 semantic engine remains fully local and offline: no remote AI, API key, backend or telemetry.

Controlled cross-source fixtures and affected real Chromium workflows passed. The original screenshot records were unavailable; this release does not claim to reproduce those exact records. Existing embedding limitations remain.

ThreadBack AI v1.2.0 — Local AI Intent Engine

Continuous local AI intent management with bundled BGE-small-en-v1.5 INT8: page-to-task assignment and task-to-task semantic reconciliation, automatic duplicate-task merging, conservative splitting, latent one-off handling and expiry, stale-task archival and archived-task reactivation. Grounded workflow renaming and bounded action history with safe Undo are included.

Fully offline semantic inference: no remote AI, API key, backend or telemetry. Durable MV3 pending-event recovery preserves browsing events through worker termination.

Verified: 197 tests passed; typecheck, lint, build, packaging and security checks passed. Actual unpacked Chromium: 31/31 checks, including worker recovery with 0 lost and 0 duplicate events. Ten same-purpose pairs, ten close-but-different pairs and twenty curated generalization flows passed; these are engineering fixtures, not universal accuracy estimates.

Known limitations: broad/acronym research may remain split. Automatic splitting currently requires fully represented tasks of at most five pages. Embedding similarities are not calibrated probabilities; semantic mistakes remain possible. Installed Chrome/Edge distributions and exact RAM usage were not independently qualified.

ThreadBack AI v1.2.0 — Local AI Intent Engine

AI-first local semantic intent clustering with bundled BGE-small-en-v1.5 INT8. Every new semantic assignment is evaluated on-device from bounded metadata. No API key, backend, remote AI or telemetry.

Uniform candidate ranking, bounded page/centroid support, stable purpose anchors and multi-page workflow reconciliation improve continuity. Singleton and uncertain pages remain latent until supported. A durable pending-event queue preserves intent memory through MV3 worker termination.

Verified: final automated suite, packaging/security checks and actual unpacked Chromium workflows, including offline model loading, arbitrary-topic grouping, Resume, manual controls and browser restart. Worker interruption lost 0 events and duplicated 0.

Limitations: embedding similarity is not calibrated probability or proof of user purpose. Closely related goals may merge and ambiguous pages may split/remain latent. Installed Chrome/Edge distributions and exact RAM usage were not separately qualified.

ThreadBack AI v1.1.1 — Local Semantic Intent Upgrade

TinySemanticClassifier now resolves ambiguous intent candidates locally alongside the released stage-one pipeline.

• Ambiguity-only invocation: 0.25 <= score < 0.50; valid semantic decisions require confidence >= 0.65.
• No API key, remote AI, telemetry, backend, extra permissions or additional model fetch.
• Conservative cross-domain grouping with word-boundary concepts, generic/domain filtering and bounded subword features.
• Malformed metadata, classifier exceptions and unresolved fallback preserve the original heuristic result.
• 145 tests pass; typecheck, lint, build, extension verification and security scan pass. Actual isolated Chromium: all 13 candidate checks passed; the affected resolver path and final UI smoke passed again after narrowing the software category. See HANDOFF.md for exact staged evidence.

Download and extract the ZIP, then Load unpacked the folder containing manifest.json in Chrome/Edge Developer mode.

The 50-case benchmark is a curated development regression set, not general accuracy evidence. Exact RAM was not independently measured. Heuristic/topic ambiguity and existing development-toolchain advisories remain. Installed Chrome/Edge distributions were not independently qualified.

v1.1.0 and all earlier releases are preserved; this audited follow-up is v1.1.1.

ThreadBack AI v1.1.0 — EurekaDev 2026

Local-first MV3 extension with a tiny bundled metadata concept classifier. No neural runtime, backend, model download, API key or telemetry. Heuristics remain primary and survive model failure.

Validation: 111 tests passed, typecheck/lint/build/packaging/secret scan passed; Actual Chromium: 10 existing core checks passed; the affected local model failure/recovery and final UI/privacy smoke passed on the final build. See HANDOFF.md for staged verification details.

Install: download the ZIP, extract it, enable Developer mode in Chrome/Edge, and Load unpacked the extracted folder containing manifest.json. No development server or API key is needed.

Installed Chrome/Edge distributions were not independently qualified. Remote AI/page extraction and native side-panel registration are not implemented. Local storage is not application-encrypted. Current audit: zero production advisories; 10 development-toolchain advisories remain (3 moderate, 5 high, 2 critical). See HANDOFF.md, SUBMISSION.md and DEMO_SCRIPT.md.

ThreadBack AI v1.0.1 — EurekaDev 2026

Submission-ready local-first MV3 extension.

Validation: 96 tests passed, typecheck/lint/build/packaging/secret scan passed; expanded isolated actual Chromium search-context, worker-restart and persistence smoke passed.

Install: download the ZIP, extract it, enable Developer mode in Chrome/Edge, and Load unpacked the extracted folder containing manifest.json. No development server or API key is needed.

Installed Chrome/Edge distributions were not independently qualified. Remote AI/page extraction and native side-panel registration are not implemented. Local storage is not application-encrypted. Current audit: zero production advisories; 10 development-toolchain advisories remain (3 moderate, 5 high, 2 critical). See HANDOFF.md, SUBMISSION.md and DEMO_SCRIPT.md.

ThreadBack AI v1.0.0 — EurekaDev 2026

Submission-ready local-first MV3 extension.

Validation: 79 tests passed, typecheck/lint/build/packaging/secret scan passed; isolated actual Chromium smoke passed.

Install: download the ZIP, extract it, enable Developer mode in Chrome/Edge, and Load unpacked the extracted folder containing manifest.json. No development server or API key is needed.

Installed Chrome/Edge distributions were not independently qualified. Remote AI/page extraction and native side-panel registration are not implemented. Local storage is not application-encrypted. Current audit: zero production advisories; 10 development-toolchain advisories remain (3 moderate, 5 high, 2 critical). See HANDOFF.md, SUBMISSION.md and DEMO_SCRIPT.md.