Milestones
List view
- No due date•3/23 issues closed
Production exit gate: upgrade, failed-update recovery, rollback, uninstall/data preservation, security/accessibility/E2E gates, support docs, and Windows/Linux compatibility smoke all pass for the macOS v1 release candidate.
No due date•0/3 issues closedBeta exit gate: first-run onboarding and storage adoption work without Terminal; the Apple Silicon app/DMG are hardened, Developer ID signed, notarized, stapled, SBOM/checksum accompanied, and accepted by Gatekeeper on a clean supported Mac.
No due date•0/2 issues closedAlpha exit gate: the shared desktop shell starts exactly one bundled server, renders the existing console, completes an MCP smoke, reports failure states, and leaves no orphan across 30 lifecycle cycles.
No due date•2/3 issues closedFoundation exit gate: installed server runs from an arbitrary read-only location, confines writes under one application-data root, and preserves source, wheel, container, and CLI behavior.
No due date•2/3 issues closedDiscovery exit gate: relocatable Python/native bundle, application-data-root migration, safe supervisor lifecycle, and Apple signing/notarization path are each proven and recorded before implementation assumptions harden.
No due date•3/4 issues closedThe 102 findings from the 2026-07-25 principal-engineer architecture review that are NOT covered by milestone #7 Boundary hardening. **Method.** 8 parallel subsystem readers -> 8 adversarial refuters (refute-by-default, re-open the cited file) -> 3 independent cross-cutting judges. 51 CONFIRMED / 27 PARTIAL / 1 REFUTED across the reader set, plus **32 findings the refuters caught that their own reader had missed** — a single-pass review would have lost roughly a third of this. **How to read a finding.** `CONFIRMED` = the refuter re-read the cited code and the claim held. `PARTIAL` = something real is there but the original claim was overstated or mis-located; the correction in the body governs. Refuted findings were dropped and are not here. **Calibration.** Severity is scaled to a single-user, local-first OSS tool — not to a hosted multi-tenant service. Loopback-only with no auth, a single uvicorn worker, manual cutover and no HA are correct scoping and were deliberately excluded from the review's verdict. **This is a backlog, not a sprint.** Most items are S-sized. Full context and the cross-cutting analysis: `.claude/notes/scans/architecture-review-2026-07-29.md`.
No due date•42/103 issues closed- No due date•18/18 issues closed
- No due date•9/9 issues closed
- No due date•15/15 issues closed
- No due date•0/24 issues closed
Close the verified defects at the boundaries where the served process hands off to something else: cron, the filesystem, the packaging system, and the operator. Source: the 2026-07-25 principal-engineer architecture review (8 subsystem readers -> 8 adversarial refuters -> 3 cross-cutting judges; 51 CONFIRMED / 27 PARTIAL / 1 REFUTED). All three judges returned `fragile`, by three different routes, and the routes converge on the same boundaries. Every item in this milestone was re-verified by hand against the code before filing. **Scope discipline.** This milestone is remediation of verified defects only. It is deliberately NOT the extensibility work (shared kernel, source-type adapter, handler registry) -- that is design work belonging in a roadmap/1 track, not a findings milestone. **Explicitly out of scope, because they are correct scoping for a single-user local-first tool and must not be 'fixed':** loopback-only with no auth, a single uvicorn worker, manual operator-gated cutover, no HA, and shipping the hybrid/rerank pipeline dormant after measuring +0.000 P@10 at 122x latency.
No due date•10/10 issues closedShip a PyPI-installable v0.1.0 that is safe to install and safe to cite, then let a one-quarter adoption verdict decide whether to keep investing in distribution. **Key results** - A clean-environment `pipx install arxmcp` / `uvx arxmcp` boots the server (frontend assets present in the wheel) with 0 standalone-install crashes - The get_paper wrap-helper tripwire test asserts wrapping IS applied to title/abstract and fails the moment wrapping is removed — the inverted-polarity guard (D8-R02) is corrected - 100% of source_kind=textbook chunks carry a non-OA-allowlisted license token by default, and get_chunk visibly truncates a re-ingested copyrighted textbook body under the existing 300-char policy **Entry points** (`/milestone-pipeline <id>`): `trustworthy-release-m1`, `trustworthy-release-m2`, `trustworthy-release-m3`, `trustworthy-release-m4`, `trustworthy-release-m5` Roadmap: [`plans/trustworthy-release/roadmap.yaml`](https://github.com/chris-dare-dev/arXMCP/blob/main/plans/trustworthy-release/roadmap.yaml) · 32 items (0 done) · id-map: `plans/trustworthy-release/github/issue-map.json`
No due date•4/32 issues closedShip a security-gated /ui/ workbench so the operator can search, read typeset math, curate, and grade their own corpus directly, and export it to their own citation and notes tools. **Key results** - A document-preview request on the tier-1 Windows workstation returns 200, not a 404, for a notebook with an ingested paper - The operator runs a corpus search and opens a full chunk body from a browser, making 0 hand-built JSON-RPC calls and touching none of the 3-search / 4-get_chunk lifetime session caps - Every /ui/ surface that shows chunk or document text renders $TeX$ / MathML as typeset math, not raw markup, verified against real corpus chunks rather than fixture pages **Entry points** (`/milestone-pipeline <id>`): `researcher-workbench-m1`, `researcher-workbench-m2`, `researcher-workbench-m3`, `researcher-workbench-m4`, `researcher-workbench-m5` Roadmap: [`plans/researcher-workbench/roadmap.yaml`](https://github.com/chris-dare-dev/arXMCP/blob/main/plans/researcher-workbench/roadmap.yaml) · 37 items (0 done) · id-map: `plans/researcher-workbench/github/issue-map.json`
No due date•0/37 issues closedShip a linearized bulk write path, scale-guard telemetry with documented inflection thresholds, GPU-capable ingest, and a contained, poison-screened ingest pipeline - validated by an instrumented 10K-paper capacity checkpoint that gates Tier-5 entry. **Key results** - A bulk/batch ingest run (bulk_ingest, oai_delta, re_embed, notebook drivers) produces exactly one end-of-run index build instead of one per-paper rebuild, verified by an MVCC version-count assertion in a batch-ingest test - The daily ops report and /status surface all ~8 scale-guard gauges (lancedb version_count, dir_bytes, BM25 bytes, startup seconds, notebook count, backup duration/bytes, per-stage ingest throughput) with D5-R07's inflection table encoded as documented thresholds - /readyz returns 200 with ARXMCP_BM25_PHASE=off as the default even when the BM25 pickle load path is forced to fail, verified by a startup test **Entry points** (`/milestone-pipeline <id>`): `scale-ops-hardening-m1`, `scale-ops-hardening-m2`, `scale-ops-hardening-m3`, `scale-ops-hardening-m10`, `scale-ops-hardening-m11` Roadmap: [`plans/scale-ops-hardening/roadmap.yaml`](https://github.com/chris-dare-dev/arXMCP/blob/main/plans/scale-ops-hardening/roadmap.yaml) · 40 items (0 done) · id-map: `plans/scale-ops-hardening/github/issue-map.json`
No due date•0/46 issues closedShip a truthful, budget-sane, spec-current MCP tool and resource surface for arXMCP, plus the minimal orchestrator loop that puts real agent traffic through it for the first time. **Key results** - A single interactive session completes at least 30 search_papers calls and 100 get_chunk calls with zero RETRIEVAL_CAP_REACHED rejections (today's lifetime ceiling is 3 and 4, with a documented rotate-id bypass that provides no real security) - Zero tool descriptions in tools/list assert functionality that is unshipped, cancelled, or reserved-but-inert, confirmed by a description-vs-handler truth audit run at every batched schema re-pin - get_chunk resolves a batch of at least 20 chunk ids in one call, collapsing the search-then-fetch loop from N+1 round-trips to 2 **Entry points** (`/milestone-pipeline <id>`): `agent-platform-m1`, `agent-platform-m3`, `agent-platform-m8` Roadmap: [`plans/agent-platform/roadmap.yaml`](https://github.com/chris-dare-dev/arXMCP/blob/main/plans/agent-platform/roadmap.yaml) · 28 items (3 done) · id-map: `plans/agent-platform/github/issue-map.json`
No due date•12/28 issues closedShip the retrieval unlocks — proof-chunk retrievability, LaTeX-routed TED equation search, definition filtering, S2-hydrated citation context, and Stacks Project grounding — batched into one agent-platform schema re-pin window, with every quality-sensitive upgrade (prefixes, reranker, embedder) held behind the evidence-engine track's fixture evidence. **Key results** - get_chunk(include_referenced=True) on a stmt-kind chunk returns its paired proof window(s) via (paper_id, theorem_label, kind='proof') linkage, with a documented section_path/adjacency fallback when theorem_label is null - find_equation resolves LaTeX-form queries onto the TED+dense fusion path (retrieval_mode reports something other than dense_only_stmt_fallback) instead of always bypassing to dense-only - search_papers accepts filters={'include_kinds':['proof']} and returns proof-kind rows with honest excluded_kinds/retrieval_mode fields; the dual-column-RRF default stays OFF until the evidence-engine fixture verdict lands **Entry points** (`/milestone-pipeline <id>`): `retrieval-unlocks-m1`, `retrieval-unlocks-m2`, `retrieval-unlocks-m4` Roadmap: [`plans/retrieval-unlocks/roadmap.yaml`](https://github.com/chris-dare-dev/arXMCP/blob/main/plans/retrieval-unlocks/roadmap.yaml) · 34 items (16 done) · id-map: `plans/retrieval-unlocks/github/issue-map.json`
No due date•16/34 issues closedShip arXMCP's first reproducible harness numbers — retrieval nDCG@5/Recall@10, agent-task ergonomics, and a zero-labeling auto-benchmark — behind a pinned regression ledger, ending the project's never-measured era. **Key results** - `make eval` (dense) records a real nDCG@5/Recall@10 number in `.claude/docs/retrieval-quality-report.md` for the first time, instead of SKIPping. - Every eval run appends a manifest-pinned entry (embedder/chunker/tokenizer/reranker versions + fixture content-hash) to `var/arxmcp/ops/eval/ledger.jsonl`, and watchdog AC1 goes live. - An 8-10 task agent-ergonomics eval reports pass^k, tool-call count, and wrong-tool-selection rate for at least one pre/post milestone comparison, published in the README. **Entry points** (`/milestone-pipeline <id>`): `evidence-engine-m1` Roadmap: [`plans/evidence-engine/roadmap.yaml`](https://github.com/chris-dare-dev/arXMCP/blob/main/plans/evidence-engine/roadmap.yaml) · 19 items (0 done) · id-map: `plans/evidence-engine/github/issue-map.json`
No due date•0/19 issues closed