chore(release): bump version to 2.0.0 - #6928

Set src/local_deep_research/__version__.py, package.json and the npm lockfile root to 2.0.0. The FastAPI/Uvicorn migration changes HTTP, authentication, Socket.IO, deployment and database behavior, so this is a SemVer major release. The code diff remains limited to those three version files.

Merge this PR last. The release is not yet qualified. Merging the version change starts the automated release pipeline. The release owner performs the individual PR merges.

Each selected fix must land through its focused canonical PR targeting main. #6929 is reference material only; do not accumulate or publish further fixes there. Its historical component tests do not qualify current main or the final 2.0.0 artifacts. Preserve contributor ownership and compare current main and existing PRs before carrying any still-missing change forward.

Current evidence — October 15–16

The October 15, 22:09 UTC (October 16, 00:09 CEST) inventory snapshot contains 222 open PRs and 887 open issues. It includes new #7441–#7450 and the additional main merges listed below. No new issues were opened after the preceding 21:48 UTC snapshot. Related task conversations, PR discussions and the #7199 backup acceptance were read. This is scope triage and remote-state reconciliation, not a claim that every current diff has passed adversarial review.

• Main advanced to 863cabea459e7ef55b8d31ab5ca6ad33b0d295a6. Its fresh Docker CI run is pending at this snapshot; intervening runs cancelled by newer pushes do not qualify main. The preceding 6e88a356fd59 Docker run passed All Pytest Tests + Coverage and amd64/arm64 production-image smoke tests, but skipped browser/LLM/infrastructure suites. That earlier result does not qualify the new main or final release.
• The earlier green Create Release run ran only its version check. Release-gate, package/platform, browser and publishing jobs were skipped. It is not release qualification.
• fix(ci): repair dependency installs, scans, and test failures #6761, fix(docker): pin Python 3.14.8, retire the zipfile backport, and enforce interpreter and expat floors #7077, fix(ci): close scanner coverage and precision gaps #7311, fix(ci): reject incomplete OSV scans #7318, test: clean up settings and chat fixtures before teardown #7324, fix(security): update pip's vendored urllib3 in Docker images #7328, fix(metrics): preserve caller transactions across telemetry #7227 and fix(downloads): host-level attempt budget stops the fallback chain after a rate-limited static fetch #7352 are merged into main. Since the 19:44 snapshot, fix(db): retry the tracker's reset/cleanup deletes on SQLite lock blips #6950, fix(settings): let search reach non-selected providers' settings #6899, fix(security): enforce realpath containment in validate_safe_path (GHSA-gv87-qcw7-gjr8) #6883, fix(database): wrap SQLCipher driver errors #6916, fix(deploy): allow time for Compose shutdown cleanup #6918, fix(security): port trusted proxy rate-limit identity to main #7143, fix(ui): preserve polling after an early socket disconnect #6959, fix: stop model discovery after request cancellation #6926, fix(strategies): give the empty-synthesis placeholder its own message #6948, fix(latex): accept a space before the comma in group citation labels #6954 and fix(benchmarks): reject num_runs < 1 in measure_execution_time #7364 also merged. The later snapshots additionally confirm test(langgraph-agent): pin snippets-only search tools #7189, ci(dependabot): guard first-match-wins group ordering - specific groups before wildcards #6914, fix(docker): start generated Ollama independently #6912, fix(rag): cancel the requested collection index task #6906, ci: verify release decisions and expiring exceptions #6887, fix(library-deletion): log short ids without a misleading ellipsis on deletion paths #6896, chore(deps): bump python from 89fb7d3 to a2b82f3 #7380 and test: resolve ForwardRef dependencies in isolation census #7323 merged. Release-decision and expiring-exception CI from ci: verify release decisions and expiring exceptions #6887 is delivered. The OSV completeness correction is delivered; it is no longer a pending source PR.
• Thirteen code-scanning alerts remain open: six Scorecard pinning findings, the five findings addressed by fix(security): resolve SQLCipher Bearer alerts and validate scan uploads #7377/fix(security): baseline a Gitleaks example and annotate two DevSkim checksums #7400, the Python image finding and the unstructured OSV finding. Two new medium Dependabot alerts cover the MCP SDK in Semgrep's input requirements and locked requirements. Focused fix(release): hash-lock Towncrier installs #7450 now addresses the two Towncrier Scorecard call sites and awaits full hosted CI, merge and fresh indexing; the other four Scorecard findings remain separately unresolved. The source and overlap audit below confirms the MCP findings are not cleared by existing dependency bumps. The unstructured investigation includes existing 🤖 Update dependencies #7379; its failing full pytest does not qualify it as a resolution. chore(deps): bump python from 89fb7d3 to a2b82f3 #7380 changes both Python image digests while retaining Python 3.14.8, so the runtime/image gate still needs fresh qualification; a digest change alone does not establish a patched runtime.
• The latest published release remains v1.10.7. No v2.0.0 release has been published.

Tracker #7137 retains the earlier scope decisions but its checkboxes are stale. At the earlier reconciliation, 123 of 167 referenced PR entries were merged, 41 were open and 3 were closed without merging; 95 unchecked entries were already merged. Those open entries included old sources, alternatives and holds, so they did not represent 41 independent required merges. The source map below reconciles those distinctions without dropping selected behavior.

Release behavior and upgrade requirements

The release moves the web application to FastAPI/Uvicorn and carries database revisions that v1.10.7 does not understand. It includes library identity/retry/indexing corrections, recoverable encrypted backups, stronger settings contracts, safer frontend rendering, search/report improvements and metrics that preserve research ownership and caller transactions. Some selected follow-ups below remain open; the final release notes must describe only what actually lands.

Already-landed behavior includes:

• Setting-key validation with legacy-row compatibility and typed API-toggle writes (fix(settings): reject control characters in setting keys #7062, fix(security): enforce typed API access toggle #7002); atomic migrations, paired database/salt backups, binary Windows salt handling, retention rules and operator-only library paths (fix(database): make SQLite migration schema changes atomic #7099, fix: preserve salts for recoverable database backups #7032, fix(database): write salt files in binary mode on Windows #7257, fix(backup): keep user-retained backups when taking the pre-migration backup #7151, fix(settings): make library storage path operator-only #7152, docs: clarify v2 backup retention and library path upgrades #7224).
• Canonical documents, PDF snapshots, retained text and indexing invalidation across retries and sync (fix(library): preserve canonical documents and PDF snapshots on retry #7202, fix(library): preserve canonical text on duplicate downloads #7229, fix(library): keep documents that still hold content during library sync #7246, fix(rag): preserve text invalidation through indexing and reconciliation #7307), shared-document lifetime handling (fix(database): preserve shared documents and download history on deletion #7068), restored TOML/Evernote loaders (fix(loaders): restore TOML and Evernote extraction #7059), bounded PDF extraction (fix(pdf): bound upload, RAG, and Zotero text extraction #6472, fix(security): bound upload document conversion and zip-container parsing #6515, fix(pdf): follow-ups to #6459 — pypdf version refs, warn when stream release is disabled, tighter memory estimate #6877, fix(pdf): preserve download text at the character ceiling #7230) and explicit owner-verified legacy-library recovery (fix(library)!: remove ambiguous legacy PDF read fallback #7268).
• Provider-aware cost summaries and incomplete-price indicators, search outcome/error metadata, separate provider usage and response-owned Ollama metrics (fix(metrics): honor providers and mark partial cost estimates #7283–fix(metrics): keep raw Ollama metrics with each response #7286). Transport timeouts retain their type; genuine HTTP 429 handling remains intact (fix(llm): preserve timeouts in rate-limit detection #7287). Streaming failure diagnostics preserve original exceptions, chunks, cleanup and research callbacks (fix(llm): log streaming failure types #7289, fix(metrics): isolate callbacks on shared LLMs #7053).
• Correct PDF range responses and ASCII/Unicode download filenames (fix(web): restore download filename fallbacks #7290, fix(library): restore PDF byte-range responses #7291), error-response security headers and explicit cookie test mode (fix(web): apply response security headers consistently #7065, fix(web): make cookie test mode independent of pytest import order #7293), citation handling (fix(citations): recognize bracketed IPv6 URL citations #7046, fix(export): land bounded citations and Quarto safeguards on main #7139), Federal Register extraction (fix(search): clean Federal Register full-text content #7280) and uploaded-filename search (fix(search): find uploaded documents by their listed filename #7054).
• Chat sequencing/pagination and crashed-worker cleanup (fix(chat): allocate message sequences after attempt deletion #6907, fix(chat): release crashed research slots on session deletion #7218, fix(chat): stabilize pagination and validate cursors and titles #7225, fix(web): bound history and library pagination offsets #7064), queue metadata and claim recovery (fix(queue): preserve JSON-string research metadata #7219, fix(queue): recover retryable claims and discard finished rows #7042), persisted Stop entries (fix(research): persist Stop entries in progress logs #7031), long-research worker isolation (fix(api): isolate long research from auth workers #7090), and contributor fixes for recovery guidance, news model selection and PubMed quotes (fix: restore quick-mode error report guidance #7239, fix: preserve worker recovery suggestions in research status #7244, fix: honor selected provider and model in news Run Once #6694, fix(pubmed): stop quote repair from injecting quotes after balanced phrases #7279).
• Dependency/lockfile repairs and scanner execution/coverage corrections (fix(ci): repair dependency installs, scans, and test failures #6761, fix(deps): enforce patched v2 package and toolchain versions #7131, fix(deps): retain basic-ftp 6.2.2 and redact research CLI errors #7200, fix(deps): update source-map-js to patched 1.2.2 #7213, fix(deps): remove sprintf-js from infrastructure tests #7223, fix(deps): remove sprintf-js from accessibility tests #7245, fix(ci): close scanner coverage and precision gaps #7311, fix(ci): reject incomplete OSV scans #7318). Preserve the existing reviewed exception policy; a successful upload alone does not establish a complete scan.

Before upgrading from v1.10.7, follow the upgrade/rollback guide and reverse-proxy guide:

1. Stop the application and take a verified offline backup of data, configuration and external library/index locations. Preserve matching database/salt material. Automatic retention is not a substitute. Rollback requires the old application and pre-upgrade data; never start v1.10.7 against upgraded databases.
2. Preserve existing library locations, including the shipped ~/Documents/LocalDeepResearch/Library default. The path is operator-managed. Changing its deployment override does not move files and applies across accounts; docs(settings): name the library path variable the code reads #7391 corrects the configuration reference and upgrade wording.
3. Run one Uvicorn process. For TLS termination, enable TRUST_PROXY_HEADERS, restrict LDR_SERVER_FORWARDED_ALLOW_IPS to actual trusted proxies and isolate the backend. Update custom Socket.IO clients/proxies to /ws/socket.io.
4. Expect users to sign in again. Queued background work waits when its session expires; already-running research may finish. Missing or unreadable persistent session keys stop startup.
5. Update clients for CSRF headers and 403 failures, ISO dates, 422 malformed numeric path parameters and the documented news/context-overflow contracts. Model-cache refresh and indexing actions use POST with CSRF protection.
6. Correct invalid saved settings and environment overrides before upgrading to strict settings reads. Preserve optional defaults and explicit inheritance of per-engine time periods. Supported news refresh intervals are 1–10,080 minutes.
7. Automatic reads from the ambiguous old shared PDF root are removed. Existing files remain; use the guide's owner-verified recovery process. Assisted migration remains future work (Library resource_id collides on disk: a per-user helper exists but ~8 call sites bypass it #5521).

Verify login, saved settings, research completion, reconnects, library search/indexing and restart behavior before retiring backups.

Selected work still requiring resolution

These are release requirements or selected changes, not blanket merge recommendations. Preserve each canonical PR, review its current diff and discussion, resolve overlaps, and validate against current main. Draft PRs remain held. A green check does not replace review of a newer revision.

Library acceptance must also retain the qualified retry distinctions: completed stored PDFs still repair tracker/queue/collection state while reusing canonical storage; the missing-PDF refill exception must coexist with rejection of conflicting regenerated bytes before mutating either document or tracker. Uploads without provenance must extract their own PDF, and a refused new snapshot must not borrow an unrelated snapshot's text. Accepted text improvements must remain pending across restarts until indexed, including bulk, parallel, direct-batch and scheduled selection. Preserve collection/source/model/provider selection, decorated-URL history without rewriting stored hashes, and rejection of unsupported remote-vector rebuilds before database work.

At the 22:09 UTC snapshot and follow-up mergeability checks, #5632 ea96af0296f8, #7201 c5ba09836579, #7083 b9e6b8066937, #7322 14c1d1ba156c, #7327 6cddd4d4bb0f, #7378 81e5f081eb90, #7391 abcf846f75a1, #7397 eb4a6db53e27, #7400 547eef66e984 and #7363 45a57de485a3 have successful recorded full pytest and no failed/pending checks on those heads, with no conflicts on the latest checks. Approval remains necessary. These recorded passes do not resolve newer dependency/pinning alerts or qualify final main. #6752 ef924d24e628 also has recorded passing full pytest but retains its draft hold and overlap/dependency review. #7288 1f6575dd344b and #7133 80ef0d625c24 have had their conflicts corrected and await fresh full validation. #7288's local checks include 1,952 settings/API/configuration tests with 13 existing xfails, 327 persistence/security tests with one existing xfail, eight new HTTP/SQLite regressions, final diagnostic controls and 91 frontend tests across 17 suites. Independent review found no remaining blocker after the integer composition and bulk-save correction; local checks do not substitute for hosted full pytest.

Full pytest is queued for #6917, #7313, #7089, #7106, #7310 and #7441; #7377 is now running full pytest. #6919 2b8312a1a876, #7444 378c1b5f6bcd, #7443 91a70ae7c7cb, #7317 6ad643e24c74, #7446 82b7394c41a0, #7450 1d32fd40184b and the other selected pushes still await complete current validation. OSV checks now fail on #7228, #7310, #7317, #7330, #7441, #6662, #7443 and #7444. #7443 also fails two migration-quality guards on the backup-status response contract; passing native Windows SQLCipher jobs do not clear those failures. #7263 has both a recorded full pass and later image/full-check failures with a replacement run pending; it is held until the resulting required checks succeed. #7318, #7227 and #7323 are merged, not pending PR validation. The #6986 task also reports its main conflicts resolved at 0c8614792320, with current CI pending; that synchronization supplies no runtime-image waiver.

The new scanner findings have the following source dispositions at main 15a4b89aebd0:

• Scorecard alerts 9735 and 9734 are the two unhashed Towncrier tool installs in release.yml. The other four are intentional fresh-resolution installs of the wheel, extras, source distribution and published package in release-gate.yml/publish.yml. These still match current source; older indexing is not their resolution. Any correction must preserve the ability to detect broken consumer dependency resolution. Existing fix(ci): verify tool hashes and SQL scanner data flow #7082's tool-lock changes do not cover these call sites. New canonical fix(release): hash-lock Towncrier installs #7450 adds the complete binary-only Towncrier lock, isolated tool environments and the sparse-checkout input. Six real-pip offline regression tests pass and all six fail with the original installers. The existing CI suite passes 1,227 tests and 921 subtests, with seven native Semgrep CLI checks skipped. All 153 hashes match official metadata, and both actual render/cleanup paths produce identical notes from all 415 current fragments. Full independent seven-file review and changed-file hooks pass; full hosted CI is requested with test:pytest and is incomplete. The current release classifier treats all six as blocking medium findings, with no active exceptions; older Scorecard documentation does not waive that gate.
• Both Dependabot alerts are the same public MCP SDK advisory, GHSA-84m7-p3x7-pcfv (fixed in 1.30.0), in Semgrep's tool environment. The application lockfile already contains MCP 1.30.0. Official package metadata for Semgrep 1.179.0, both proposed 1.180.0 bumps and the current 1.181.0 release requires exactly MCP 1.29.0. Adding mcp>=1.30.0 would make those requirements unsatisfiable. Require an upstream-compatible scanner release or a separately reviewed remediation strategy; do not use an unchecked transitive override or create a new exception implicitly. 🤖 Update dependencies #7379 is not a fix for these tool requirements.

New inventory decisions

The refreshed review read all 42 newer open issue discussions and 34 newer non-routine open PRs, checked related conversations and compared existing overlapping work. A further pass read the five new issues #7436–#7440 and checked their relationship to current main and existing PRs. These are scope decisions; current code review and CI remain separate.

Nonblocking PR follow-ups remain #7322/#7327/#7329/#7330, #7336/#7338/#7339, #7355/#7356/#7357, #7369/#7370, #7441 and contributor #7415. #7352 and #7364 are now merged. Contributor #7442 is a useful, nonblocking test-only follow-up to #7422/#7340: it makes the unindexed-collection regression discriminate the skip path and updates stale test documentation. Preserve its ownership and require its own review/CI; it adds no runtime fix. Registration defaults (#7329) and unconditional full-CI scheduling (#7372) are separate policy changes; neither silently replaces the current release criteria. #7338's dormant cancellation hook must not be advertised as live end-to-end MCP cancellation. #7330 adds diagnostics; the transaction repair is delivered by merged #7227. #7441 is desirable but nonblocking: it completes the deferred #6848 diagnostics follow-up to merged #6576 without changing refusal decisions. It complements #7322 and is independent of #7321/#7330. Its scope triage is complete; independent merge review and successful current full CI remain required.

Record the unresolved #7360/#7371 dropdown-layout overlap without taking over or closing either contributor PR. #6954 is now merged; compare the remaining #7367 citation-whitespace diff against main before treating it as still-missing behavior, preserving contributor ownership. #7368's log-preview floor is already covered by merged #7354; compare residual differences before treating it as additional work. New provider #7403 and broad ingestion #7416 remain optional.

The #7437/#7438 whitespace and provider-registry follow-ups are implemented in the updated #7228 decoder patch and await its current full CI and merge; they are not separate release gates. #7439/#7440 remain nonblocking test-caption and dormant-fallback follow-ups.

Most #7421–#7435 are explicitly extracted test, documentation or defense-in-depth follow-ups. In particular, #7431 asks a separate ignored-manifest scan-scope policy decision; merged #7318 guarantees completion of the selected invocation, not universal discovery of ignored files. #7429's legacy NULL-mode checks must be considered with the merged migration/defaults. Preserve the existing post-v2 treatment of #7361/#7365/#7373/#7374/#7390/#7392 and their contributor claims. Do not claim universal parser isolation, immediate revocation of every active socket, or complete sanitization parity from narrower selected fixes.

Source choices, overlaps and holds

Preserve already-landed behavior and contributor attribution from the earlier tracker, including #7066/#7086/#7166/#7172/#7165, #7074/#6381, #6434, #7144 and the release/test helpers #6897/#6921/#6932/#7084/#7210/#7212/#7221/#7234/#7255/#7265/#7267. Closed source candidates #6153/#6981/#7136/#7168/#7155/#7158 do not imply missing standalone fixes. Routine dependency updates remain subject to their own compatibility and scan evidence; the previously pending #7296–#7306 updates are now merged.

The broader #4448 Auto Research, #6251 collection-management/provider feature, #6759 shared cache, #7312 configurable reasoning/temperature/sampling and #7416 automatic ingestion/persistence are not release prerequisites. Preserve any already-selected shared correctness fixes through their existing canonical PRs; broader features and overlapping schema changes require their own decision and validation.

Existing deferrals remain: #7236/#7238/#7241/#7254/#7264, #7292/#7294, #7215/#7216/#7220, #7270/#7271/#7272 and #7273/#7274/#7275. Preserve contributor discussions. #5521 assisted legacy migration remains future work; #6356 requires a separate threat-model assessment; #5810/#6593 are not newly promoted release blockers. #7118's benchmark CLI override gap means those runs cannot stand in for qualification of the requested configuration. Related acceptance/follow-up issues such as #7097, #5681, #6847/#7093 and the broader #5835 work remain tracked; existing uncertainty is not evidence of a new missing fix.

Final release gates

• Resolve selected behavior through canonical main-targeted PRs, including the new acceptance checks above. Review current diffs, full conversations and affected callers; preserve contributor ownership. A selection is not a merge recommendation, and a historical integration result is not evidence of current-main delivery.
• For each PR prepared for merge, retain test:pytest, require successful All Pytest Tests + Coverage on the current PR test-merge revision, all other required checks and reviewer approval. Requalify stacked work on main after its parents land. Focused tests, skipped/cancelled runs and results from earlier heads do not satisfy this gate.
• Complete source, dependency and image scanning on the final revision. Verify scan completeness, artifact provenance and indexed finding dispositions; do not treat a warning upload or old candidate's clean report as final qualification. The five findings covered by fix(security): resolve SQLCipher Bearer alerts and validate scan uploads #7377/fix(security): baseline a Gitleaks example and annotate two DevSkim checksums #7400, six Scorecard pinning findings (the two Towncrier fixes are in fix(release): hash-lock Towncrier installs #7450), the Python image finding, the unstructured dependency finding and two MCP SDK tool-requirement alerts still need the appropriate source/reachability disposition or fix and fresh evidence.
• Wait for a patched official Python image; add no new exception. Verify the shipped runtime contains the required fixes, then qualify its pinned image and platform builds. Merged fix(docker): pin Python 3.14.8, retire the zipfile backport, and enforce interpreter and expat floors #7077's 3.14.8 image and proposed fix(security): accept CPython CVE-2026-12345 as an expiring release exception until 3.14.9 #7383 suppression do not close this gate.
• Qualify clean installed wheels and source distributions, dependency resolution, strict package metadata, MCP entry points, Vite assets/manifests, read-only package operation and container startup/lifecycle. Require native amd64/arm64 coverage and the selected Windows SQLCipher Python 3.12/3.14 checks. Historical emulation or Linux separator tests do not substitute for those platform results.
• Run verified v1.10.7-to-v2 upgrade and rollback rehearsals with representative encrypted databases, salts, saved settings and external library/index locations. Verify data/content identity, retention, operator-only paths, migration atomicity, session renewal, restart behavior and failed-upgrade recovery.
• Complete the required browser and end-to-end authentication, research completion/reconnect/polling, news, library/indexing and supported Markdown/math flows on the final source/artifacts. A successful ordinary main Docker run does not supply the skipped suites.
• Regenerate the v2 narrative, changelog, actual merged-PR list and complete notes attachment from the eventual release commit. Preserve Towncrier issue/PR links (fix(release): make Towncrier links work for issues and PRs #7132), complete attachment bytes (fix(release): preserve original release-note attachment bytes #7255), dispatch tracking (fix(release): bind PyPI monitor to its dispatch on main #7134) and cleanup ownership (fix(release): preserve Docker tags and pending release runs #7138). Old preview sizes, source appendices and artifact digests apply only to their original revisions.
• Prepare the advisory release mapping and publication sequence through the separately authorized process. Verify the intended fixed commits and version ranges before release; keep undisclosed material outside this public plan.
• Only after the pre-release gates above and the release owner's final review, merge this version-only PR last and verify the resulting release pipeline and published artifacts.
• After the first fixed release is tagged, complete the separately authorized advisory publication. If that release is 2.0.0, verify the actual affected/patched ranges against the shipped fix. An untagged main fix is not a patched release; this post-tag step is not a prerequisite for merging the version bump.