Specs & Plans

spec Mail pipeline off the Anthropic API — PARTIAL design spec Status: in-progress · Updated: 2026-09-17 · Related: implementation plan. Supersedes the mail half of cheaper classification. Sep 16, 2026 In progress plan Mail pipeline off the Anthropic API — PARTIAL implementation plan Status: in-progress · Updated: 2026-09-17 · Related: design spec. Sep 16, 2026 In progress spec Private work ledger: Review my work — design spec Sep 15, 2026 In progress plan Private work ledger: implementation plan Sep 15, 2026 In progress spec Organization-managed Outlook mailboxes Approved organization-managed Outlook access without new North users. Implementation in progress; new capability uncommitted/unverified. Personal defaults, e… Sep 10, 2026 In progress plan Organization-managed Outlook mailboxes — implementation plan Approved organization-managed Outlook access without new North users. Implementation in progress; new capability uncommitted/unverified. Personal defaults, e… Sep 10, 2026 In progress spec Outlook automatic recovery — design spec Seven slices implemented; final Astra and Opus finders CLEAN, no new actionable regression. In-progress means lane acceptance pending; CI not yet available,… Sep 9, 2026 In progress plan Outlook automatic recovery — implementation plan Seven slices implemented; final Astra and Opus finders CLEAN, no new actionable regression. In-progress means lane acceptance pending; CI not yet available,… Sep 9, 2026 In progress spec MCP tool-list refresh — design spec The first draft proposed a per-session posture-polling watcher and treated the delivery chain as understood. Dual review falsified enough of that to reshape… Aug 27, 2026 In progress spec NetDocuments firm-wide connection + firm/personal connections split — design spec The connection is read-only document access, and the firm has no per-person NetDocuments ACL in play today. The per-user shape therefore buys nothing and cos… Aug 27, 2026 In progress plan MCP tool-list refresh — implementation plan A separate describer and logger. Do not widen describeToolCallEnvelope (audit-boundary.ts:149) or logToolCallEnvelopes (mcp-app.ts:170): they return arrivals… Aug 27, 2026 In progress plan NetDocuments firm-wide connection + firm/personal connections split — implementation plan One additive migration (column with a default + two CHECKs + one partial index) and one additive grant; no data migration. Reverting the product code restore… Aug 27, 2026 In progress plan OneDrive absence as a first-class state — implementation plan North has two document-store connectors (OneDrive, NetDocuments) and must run on either, or neither. Today revoked is treated as a fault everywhere, and one… Aug 26, 2026 In progress spec Meeting capture integrity and end-to-end recording telemetry — design spec On 2026-08-20 one user recorded three consecutive Microsoft Teams meetings — 30, 24 and 43 minutes — on the Windows desktop client. All three produced full-l… Aug 25, 2026 In progress plan Meeting capture integrity and end-to-end recording telemetry — implementation plan Phase 0 comes first because the decisive artefact self-destructs: polaris-win.log is capped at 256 KB and deleted rather than rotated. Every day of normal us… Aug 25, 2026 In progress spec NetDocuments MCP federation — design spec reviewer, both adversarial, both returning revise, converging on the same defects. Following the Aug 17, 2026 In progress plan NetDocuments MCP federation — implementation plan Every outbound NetDocuments HTTP subrequest (each search page, each fetch chunk) is wrapped: on 401 → refresh route → retry once; a second 401 → reauth-requi… Aug 17, 2026 In progress spec Polaris Windows — real embedded app (WebView2, Mac parity) — design spec The Mac client became one real app in July (native-macos-app spec): a main window hosts the authenticated apps/web surface in a WKWebView, and the companion… Jul 31, 2026 In progress plan Polaris Windows — real embedded app (WebView2, Mac parity) — implementation plan WinUI cannot build on Linux, so every phase is one PR through the standard polaris-win loop: Jul 31, 2026 In progress spec Morning preprod autodeploy — design spec Every morning at 10:00 Europe/Paris, deploy origin/main to preprod automatically — but only Jul 28, 2026 In progress spec Async research delegation — design spec Jul 24, 2026 In progress plan Async research delegation — implementation plan Jul 24, 2026 In progress spec Desktop build & release pipeline — design spec Jul 22, 2026 In progress spec Knowledge retrieval precision — design spec Jul 20, 2026 In progress plan fix: Title/entity-aware knowledge retrieval ranking — implementation plan Jul 20, 2026 In progress spec Absent corpus as a first-class result — design spec North safely stops instead of inventing a missing document, but its negative search outputs are bespoke prose. Adjacent files can then look like genuine hits… Jul 18, 2026 In progress spec Preserve Word markup in document reads — design spec The repository already has a deterministic OOXML revision reader in @workspace/document-authoring. It surfaces insertions, deletions, and comments for upload… Jul 18, 2026 In progress spec Mail-matter-document linkage — design spec Matter resolution can correctly identify a tracked matter such as MSQR-103 while matter-scoped status and “what are we waiting on?” questions still return no… Jul 18, 2026 In progress spec Partial state instead of opaque timeouts — design spec A combined matter-document workflow can spend an unbounded interval inside one north__matter_documents_search call. The explicit matter_ids path walks as man… Jul 18, 2026 In progress plan Absent corpus as a first-class result — implementation plan Turn conservative negative paths in knowledge, typed-precedent, and single-matter searches into one stable RESULT: ABSENT contract. It records exact searchab… Jul 18, 2026 In progress plan Preserve Word markup in document reads — implementation plan Ship one opt-in mode: "markup" vertical slice through the deterministic DOCX reader, native read_matter_document, and north__matter_document_read. The mode r… Jul 18, 2026 In progress plan Mail-matter-document linkage — implementation plan Repair the existing north__matter_todos path so one tracked matter id resolves both current tracked-id links and historical Clio-source-id links on classifie… Jul 18, 2026 In progress plan Partial state instead of opaque timeouts — implementation plan Bound the MCP explicit multi-matter document-search path and return truthful, resumable partial state without claiming to repair host-side rendering. Jul 18, 2026 In progress spec Working notes separation — design spec One assistant turn may persist text → tool → text → tool → text. The shared chat renderer currently presents every non-empty text part as answer Markdown, so… Jul 16, 2026 In progress spec North dans Claude Cowork — plugin + first-party MCP server — design spec Cowork est la surface AI que les lawyers connaissent déjà ; il n'a aucun accès au brain North. Symétriquement, NOS-120 Deliverable B a dessiné un first-party… Jul 9, 2026 In progress plan North Cowork plugin — implementation plan Ship order : C1-probe ∥ C0 (le probe fournit l'endpoint que C0.1 monte) → C2 → C3 → C4 → (C5 ∥ C6 si C6 GO) → C7. Security review Opus + Codex avant merge su… Jul 9, 2026 In progress plan Eval Coverage Completion — the whole surface, honestly — roadmap Three of the five workstreams shipped on main on 2026-07-07; the stale statements below are annotated in place. Jul 3, 2026 In progress plan Inbound revision awareness — Implementation Plan Jul 3, 2026 In progress plan LegalWork-parity master roadmap What changed since 2026-07-03, verified on main (0291b774): Jul 3, 2026 In progress plan Share hardening — pre-share secret scanner + read-only viewer tier — implementation plan Access resolution has one choke point: Jul 3, 2026 In progress spec North OS — AWS Production Infrastructure — design spec These four are settled and baked into the sections below — not open for the "review" pass: Jul 1, 2026 In progress plan North OS — AWS Production Infrastructure — implementation plan Move North OS off the single Scaleway box onto the per-client replicable AWS estate the spec defines: Terragrunt-per-tenant IaC, shared EKS + dedicated per-f… Jul 1, 2026 In progress spec Granola connector + meeting-to-matter-memory — design spec A lawyer at HSE runs their meetings through Granola (Business plan). After a call, North should: Jun 29, 2026 In progress spec The North Learning Loop — design spec what happened was it good? safe to ship? apply + compound Jun 27, 2026 In progress plan North learning loop — implementation plan The spec's Capture layer leans on the product_event stream from the analytics+feedback plan. That stream is frozen inside PR #17 (AWS-estate-gated) — grep pr… Jun 27, 2026 In progress spec North recall + meeting capture — feature architecture a recall memory like screenpipe (the agent captures context — audio + screen — during a session and can Jun 22, 2026 In progress spec Matter-aware, firm-wide RAG Mail Intelligence — design spec Turn every connected mailbox into one firm-organized, matter-indexed mail brain. Each lawyer's mail is classified (per-user, as P3 does today); mail that con… Jun 5, 2026 In progress plan Matter-aware, firm-wide RAG Mail Intelligence — implementation plan (rev 2) Ground truth re-verified against main (HEAD 0291b774, 2026-07-06) and the loop branches. What changed since rev 3: Jun 5, 2026 In progress spec WhisperX inference service — design spec The meetings package (Spec B) needs to turn an audio file into a speaker-labelled transcript. We chose to self-host that capability rather than send privileg… May 29, 2026 In progress plan WhisperX Inference Service Implementation Plan You can build and test the HTTP shape + contract on a laptop using a tiny CPU model (tiny / base) and diarize=false, then swap to large-v3 + pyannote on the… May 29, 2026 In progress

spec Actionable matters: Clio team, person filters, personal recency and document access Draft for plan review (NOS-384). /matters shows matter, client, status, the explicit Clio responsible lawyer and responsible staff by stable Clio User id (or… Sep 15, 2026 Planned plan Actionable matters — implementation plan Draft for NOS-384. Plan only, five disjoint proposed workstreams for the configured seats: connector assignment snapshot (clio_assignments metadata key, obse… Sep 15, 2026 Planned spec NetDocuments filing back + folder watch — design spec R1 fixed the first round's write-safety gaps. R2 fixes what the second round found underneath them: the parts where the design named a mechanism that does no… Sep 7, 2026 Planned plan MCP session continuity — recurrence risk and residual work — implementation plan Fifteen failure modes that make a claude.ai or Claude desktop chat lose its North tools, six of them added by review and proven in the source. No recurrence… Sep 7, 2026 Planned plan NetDocuments filing back + folder watch — implementation plan The REST client is no longer written from scratch: PR #216's verified closed package is extracted and gains one upload primitive, and its allow-list guard tr… Sep 7, 2026 Planned spec Second client instance — design spec North's second client needs a usable instance by 2026-09-20; this spec recommends one firm per deployment on its own host, inventories every place the produc… Sep 3, 2026 Planned plan Second client instance — implementation plan Rebaselined plan for standing up the second client's instance: the Clio-gated onboarding step blocks a OneDrive-first firm on either topology, the client's E… Sep 3, 2026 Planned spec MCP Enterprise Managed Auth — design spec deferred non-goal), specs/2026-08-18-mcp-connection-stability-design, PR #251 (3a532052, fixed client identity) Aug 27, 2026 Planned plan MCP Enterprise Managed Auth — implementation plan Companion to specs/2026-08-27-mcp-enterprise-managed-auth-design. Branch Aug 27, 2026 Planned spec Meeting "no speech detected" as a first-class outcome — design spec Worth stating precisely, because a plain reading of NOS-310 suggests the user-facing copy is missing. It is not. Aug 19, 2026 Planned plan Meeting "no speech detected" as a first-class outcome — implementation plan Two of the four phases below need a word from Christophe first, and one of them needs it within about two days before evidence is destroyed by retention. Not… Aug 19, 2026 Planned spec Meeting capture, transcript-first — design spec This replaces the audio-first pipeline, where the desktop client uploaded the recording to POST /api/desktop/meetings, we encrypted it into R2, and a worker… Jul 28, 2026 Planned plan Meeting capture, transcript-first — implementation plan The clients must ship before the audio path is removed. A lawyer on an older build keeps uploading audio, and that must keep working until nobody does it any… Jul 28, 2026 Planned spec The judgment brain — MINE mechanism — design spec The three mechanisms form a pipeline. MINE is the head of it — everything downstream assumes its output exists. Jul 24, 2026 Planned plan The judgment brain — MINE mechanism — implementation plan Execution order: A1 now → A2 productizes → A3 gates. Starting A1 flips the doc status: bun docs/superpowers/status.ts set judgment-brain-mine in-progress. Al… Jul 24, 2026 Planned spec The firm-judgment lever — design spec The learning-loop spec defined Lever #2 as success-trace retrieval: "On a new task, retrieve how this firm successfully did a similar one before; few-shot it… Jul 23, 2026 Planned plan The firm-judgment lever — implementation plan Execution order: B1 now → B2 once B1 has validated units → B3 gates the release. Starting B1 flips the doc status: bun docs/superpowers/status.ts set firm-ju… Jul 23, 2026 Planned plan North — Roadmap to September 2026 Jul 15, 2026 Planned spec North OS AWS GitOps cutover — design spec Jul 8, 2026 Planned spec North OS dev server migration — design spec Jul 7, 2026 Planned spec Failure Mining — trace intelligence for the learning loop — design spec Replit's continual-learning writeup (Catasta et al., 2026-07-06) describes three measurement pillars feeding their improvement loop: offline benchmarks, prod… Jul 7, 2026 Planned plan North OS Dev Server Migration Implementation Plan Jul 7, 2026 Planned plan Failure Mining — implementation plan The one migration. Jul 7, 2026 Planned plan Eval WS-2 — Mail-intelligence quality batteries — plan Jul 6, 2026 Planned spec MCP federation (client + first-party server) — design spec North has no MCP anywhere — no client, no server, no config table, no engine that speaks the protocol. Every integration today is a bespoke connector package… Jul 3, 2026 Planned spec Voice-action execution (voice that operates the app) — design spec North's voice path today has exactly three verbs, and none of them mutate the app. The full path: Polaris (Swift menu-bar, half-duplex push-to-talk; STT via… Jul 3, 2026 Planned plan MCP federation (client + first-party server) — implementation plan Jul 3, 2026 Planned plan Voice-action execution (voice that operates the app) — implementation plan What changed since 2026-07-03 (every claim below verified against main today): Jul 3, 2026 Planned spec MatterBench — Agentic Legal Benchmark — design spec Jul 2, 2026 Planned spec Agent workspace + execution sandbox — design spec Build it as four components, smallest-valuable-first. The through-line: the sandbox is an unprivileged byte factory; delivery stays on the existing privilege… Jun 30, 2026 Planned plan Agent workspace + execution sandbox — implementation plan agent (LLM) Jun 30, 2026 Planned spec On-Demand, Per-Matter Document Ingestion — design spec North runs two distinct retrieval paths, and the feature sits between them: Jun 26, 2026 Planned plan Skill back-testing (v2) — implementation plan v2 adds back-testing: a dry-run of a skill against a trusted fixture, surfaced on the /skills/:id edit page, with a regression diff against the prior body an… Jun 19, 2026 Planned

spec Continuous Clio source ingestion and explicit NetDocuments matter correspondence Agents must not rediscover implicit Clio-to-NetDocuments links each turn. North ingests the populated Clio sources continuously, records explicit NetDocument… Sep 14, 2026 Implemented plan Open-ended working profile — implementation plan Implemented source: optional open-ended questions, B1 exact-pair approval, bounded briefing and best-effort closeout. Native Cowork manual acceptance deferre… Sep 14, 2026 Implemented spec Cheaper classification — design spec Use GLM for thin mail classification and Gemini Flash Lite for document classification. Accept occasional semantic misses, not weaker privacy or exclusion co… Sep 10, 2026 Implemented spec External task tools — scheduler removal cutover North supplies generic tools; external MCP hosts schedule. Implemented under documented waivers and exceptions. Historical nine Actions checks passed on 8e6a… Sep 10, 2026 Implemented plan Cheaper classification — implementation plan Make two local classifier transport changes, preserve safety and existing contracts, then run bounded verification. Implemented in f09db4d1; unmerged, undepl… Sep 10, 2026 Implemented plan External task tools — cutover plan and evidence Remove North scheduling, preserve authorized history, and expose generic immediate tools. Implemented under documented waivers and exceptions. Historical nin… Sep 10, 2026 Implemented plan Clio read-only reactivation — implementation plan Sep 9, 2026 Implemented spec Meeting recording card — the pill — design spec Sep 8, 2026 Implemented spec NetDocuments investigation — design spec Sep 8, 2026 Implemented plan NetDocuments investigation — implementation plan Sep 8, 2026 Implemented plan NetDocuments browse + ids — implementation plan Two read-only tools on the existing netdocs_mcp connection: netdocs_browse (cabinet → workspace → folder → documents, with ids) and netdocs_workspace_lookup… Sep 7, 2026 Implemented spec Handover — Clicky as a North OS surface the north_os repo so it can finish brainstorming → write the docs → plan → Sep 3, 2026 Implemented spec Mail search fusion — design spec Preprod, agent_tool_audit, 30 days to 2026-09-01: Sep 1, 2026 Implemented spec NetDocuments long-document reads — design spec A lawyer asking "what does document 151 say about reimbursable costs" costs minutes, and the time is spent in North, not in NetDocuments. Sep 1, 2026 Implemented plan Mail search fusion — implementation plan Phases serialize on main. Not in this plan: pipeline instrumentation, claude-sonnet-5, the firm_noise reorder, deterministic linking, activity-gated rollups,… Sep 1, 2026 Implemented spec CI speed — design spec (v9, wave 1 shipped) Source: gh api …/actions/runs/<id>/jobs for the 30 most recent completed runs (24 PR, 6 push-to-main), aggregated with a throwaway script. Job durations are… Aug 31, 2026 Implemented plan CI speed — implementation plan (v9 — wave 1 shipped) Branch chore/ci-speed, worktree ~/workspace/north-os-ci-speed. PRs: 0 (#307) · 0.5 (#302) · 0.6 (#318) · 0.7 (#319) · 0.8 (#324) · 1a (#320) · 1b (#322) — al… Aug 31, 2026 Implemented spec North workflow intelligence — design spec North knows a great deal about matters and almost nothing about the people working them. Aug 27, 2026 Implemented plan North workflow intelligence — implementation plan Settled with Christophe on 2026-08-27. Changing one of these is changing the design, not the implementation. Aug 27, 2026 Implemented spec Connector-family tool gates — design spec The v2 draft escalated one question: when a connector is disconnected, does a tool that reads retained local data disappear (the provenance rule, as the know… Aug 26, 2026 Implemented spec Proactive memory — save and use, edit later — design spec Aug 26, 2026 Implemented plan Connector-family tool gates — implementation plan Branch fix/mcp-connector-family-gates, worktree .claude/worktrees/mcp-polish. No schema change; bun db:generate must not run on this branch. MCP surface plus… Aug 26, 2026 Implemented plan Proactive memory — implementation plan Aug 26, 2026 Implemented spec Raw matter mail access — firm roles + workspace policy — design spec The value it receives is the Better Auth organization role (apps/web/lib/chat/router.ts:445-449 → RunChatInput.role, documented as such at packages/chat-runt… Aug 25, 2026 Implemented plan Raw matter mail access — firm roles + workspace policy — implementation plan Aug 25, 2026 Implemented spec MCP connection stability — design spec The omp reviewer (verdict: implement with listed changes) falsified parts of the first draft. Corrections kept visible: Aug 18, 2026 Implemented plan MCP connection stability — implementation plan Two workers in herdr panes under w41 (tab 2), orchestrated per docs/WORKFLOW.md. Backend → omp (--approval-mode yolo); frontend → claude --permission-mode au… Aug 18, 2026 Implemented spec Matter document reads — bounded containment, honest deadlines — design spec Both reviewers falsified parts of the first draft. Corrections, kept visible rather than quietly edited away: Aug 14, 2026 Implemented spec Desktop sign-in via email OTP — design spec A lawyer downloads the desktop app, taps Sign in, receives a magic-link email, clicks it — and lands signed into North in their browser, while the desktop ap… Aug 3, 2026 Implemented plan Desktop sign-in via email OTP — implementation plan Three phases, each independently reviewable and CI-gated. Backend first (it is testable end-to-end on this box and unblocks the app), then the native app, th… Aug 3, 2026 Implemented spec RAG-first retrieval on portfolio questions — design spec A partner asked which deals have Freddie Mac as the LIHTC tax-credit equity investor. North answered "none, across two independent sweeps", then added: "The… Jul 31, 2026 Implemented plan RAG-first retrieval on portfolio questions — implementation plan Two disjoint file sets, built in parallel by two workers, plus the measurement the fix has to move. Jul 31, 2026 Implemented spec MCP host approval — a refused call, and how North says so — design spec premises), docs/superpowers/specs/2026-07-18-mcp-absent-corpus-design.html, Jul 30, 2026 Implemented spec Meetings global search, personal visibility & task aggregation — design spec Four moves, one feature: Jul 29, 2026 Implemented spec Self-serve domain auto-join — design spec Let a lawyer whose email domain belongs to a firm sign in to North without an Jul 29, 2026 Implemented spec Zero-downtime blue/green preprod deploys — design spec Make preprod deploys cut nothing. Today every deploy ends in Jul 29, 2026 Implemented plan Meetings global search, personal visibility & task aggregation — implementation plan Four workstreams on one branch (feat/meetings-global-search): Jul 29, 2026 Implemented plan Self-serve domain auto-join — implementation plan One branch, feat/self-serve-domain-autojoin, tight diff on the auth surface. Jul 29, 2026 Implemented plan Zero-downtime blue/green preprod deploys — implementation plan ships in-repo + the cutover runbook; the deploy bot executes cutover in a deploy window. Jul 29, 2026 Implemented spec MCP write catalogue + proactive Cowork — design spec Two user-facing frustrations with Cowork over the North MCP server: Jul 28, 2026 Implemented plan MCP write catalogue + proactive Cowork — implementation plan server→client requests with relatedRequestId — elicitation cannot work without this). Jul 28, 2026 Implemented spec Downloads page polish — design spec Jul 23, 2026 Implemented spec Knowledge search speed — 6.9s → <1s — design spec The lexical leg hit this exact same planner failure at 184k chunks and was fixed with a MATERIALIZED CTE fence (see the long comment in retrieval/lexical.ts)… Jul 23, 2026 Implemented spec Observability bundle — design spec The fleet already produces telemetry that nobody consumes. Jul 23, 2026 Implemented spec OneDrive matter-document search — bounded-concurrency tree walk — design spec Jul 23, 2026 Implemented spec search_matters 3.3s → tens of ms — design spec Jul 23, 2026 Implemented spec Chat pre-turn parallelization + real tool duration_ms — design spec Jul 23, 2026 Implemented spec Voyage round-trip reduction — design spec The measured cold path is ~1,080ms p50, and after this work it is still ~1,080ms p50. That is the honest result: the remaining time is either an irreducible… Jul 23, 2026 Implemented plan Downloads page polish — implementation plan Jul 23, 2026 Implemented plan Knowledge search speed — implementation plan Concrete work plan for the design spec. Branch perf/nos-209-knowledge-search-speed; the wave's only schema-toucher. Rev 2 incorporates the omp plan review: r… Jul 23, 2026 Implemented plan Observability bundle — implementation plan Work packages are ordered by dependency. WP-1 unblocks the scrape config in WP-2. Jul 23, 2026 Implemented plan OneDrive walk concurrency — implementation plan Jul 23, 2026 Implemented plan search_matters speed — implementation plan Jul 23, 2026 Implemented plan Chat pre-turn parallelization + real tool duration_ms — implementation plan Jul 23, 2026 Implemented plan Voyage round-trip reduction — implementation plan The spec's Decisions 1–4 ship as documentation and gates — they are measured rejections, and the issue's whole value is that they stay rejected. Code lands f… Jul 23, 2026 Implemented spec Chat detail level — design spec Jul 21, 2026 Implemented spec Meeting lifecycle recovery — design spec Jul 21, 2026 Implemented spec Meetings — production readiness — design spec Jul 21, 2026 Implemented spec Meeting → RAG ingestion — design spec Jul 21, 2026 Implemented plan Chat detail level — implementation plan Jul 21, 2026 Implemented plan Meetings — production readiness — roadmap Jul 21, 2026 Implemented spec Clause taxonomy normalization — design spec Jul 20, 2026 Implemented spec MCP latency and tool-surface polish — design spec The Cowork MCP demo is slow for two different reasons. First, several retrieval calls are intrinsically slow because they cross embedding or Microsoft Graph… Jul 20, 2026 Implemented plan Clause taxonomy normalization — implementation plan Jul 20, 2026 Implemented plan MCP latency and tool-surface polish — implementation plan Reduce the context and orchestration cost of successful Cowork MCP calls without changing legal retrieval identity or provenance. The measured root cause is… Jul 20, 2026 Implemented spec Meeting system-audio capture + local retention — design spec The v1 meeting recorder captures the microphone only. Other participants are audible solely through speaker bleed — with headphones they vanish from the reco… Jul 17, 2026 Implemented spec Skill jurisdiction profile + catalogue filters — design spec Every built-in skill defaults to enabled for every org (firm_skill_state: no row = enabled). The catalogue now spans jurisdictions — 202 skills including Ont… Jul 17, 2026 Implemented plan Skill jurisdiction profile + catalogue filters — implementation plan Jul 17, 2026 Implemented spec North as an OAuth 2.0 authorization server for Cowork (MCP) — design spec The C3 read catalogue (PR #77) is live and correct but unreachable from Cowork: Cowork speaks OAuth, our server speaks a static bearer. The captured flow (20… Jul 16, 2026 Implemented spec Housekeeping side channel — design spec A completed bookkeeping tool result is the source record. The tool call and its full result remain in the existing activity chrome or existing first-class re… Jul 16, 2026 Implemented spec North Mac onboarding — design spec The Mac app still ships the Clicky-era onboarding: a creator video streamed from Mux (CompanionManager.setupOnboardingVideo), background music (ff.mp3), a sc… Jul 16, 2026 Implemented plan North OAuth provider for Cowork (MCP) — implementation plan Ship order: P0.5 → P1 → P2 → P3. Security review Opus + Codex before merging P1 and before flipping the P2 guard. One mergeable deliverable (small PRs by pha… Jul 16, 2026 Implemented plan North Mac onboarding — implementation plan import AppKit Jul 16, 2026 Implemented plan Harness-supremacy follow-ups — implementation plan Every cwb-08 attempt was unscored: one malformed judge verdict nulls the whole attempt (eval-cowork-bench.ts:684-703), safeJudgeCall returns "" on throw, and… Jul 15, 2026 Implemented spec Big-document subagent reads — parallel fan-out document reading — design spec Three separate failures compound on a big document, and the pagination fix only solves the first. Jul 9, 2026 Implemented spec Cowork benchmark battery — design spec Three gaps, all confirmed by the 2026-07-09 harness survey: Jul 9, 2026 Implemented spec Dynamic context cleaner — design spec Jul 9, 2026 Implemented spec OpenRouter multi-model access — design spec Today the runtime is single-provider by construction. packages/agent-runtime/src/model.ts is the only model registry: a MODELS record of four Claude definiti… Jul 9, 2026 Implemented spec Audit-grade redline provenance — design spec Anthropic labels its own Claude-for-Word surface "not recommended for final deliverables or litigation filings" and excludes it from Enterprise audit logs an… Jul 9, 2026 Implemented spec Verifiable coverage — document manifest, thread ledger, turn verifier — design spec The 2026-07-09 Gill A201 incident is the concrete case: a 190k-char GC contract arrived carrying 5 tracked changes and 21 margin comments. read_document appe… Jul 9, 2026 Implemented plan Big-document subagent reads — implementation plan The spec (2026-07-09) was drafted against the in-flight pagination fix and before verifiable-coverage phases 1–2 existed. Both are now real, and this plan ad… Jul 9, 2026 Implemented plan Cowork benchmark battery — implementation plan Seven phases (0–6), each independently shippable and tested. Order matters through Phase 3: telemetry before fixtures before scorers/arms before the battery.… Jul 9, 2026 Implemented plan Dynamic context cleaner — implementation plan Jul 9, 2026 Implemented plan OpenRouter multi-model access — implementation plan Ship order: P0 → (P1 ∥ P2) → P3 → P4 ∥ P5. Every phase passes root bun run typecheck + bun run lint; phases touching agent-runtime's export surface also run… Jul 9, 2026 Implemented plan Audit-grade redline provenance — implementation plan Four phases, each independently shippable. Phase 1 is the load-bearing one (from its ship date, every write is recorded — D8's "provenance starts at ship" cl… Jul 9, 2026 Implemented plan Verifiable coverage — implementation plan Manifest → ledger → verifier is the dependency chain, not just a preference: the ledger needs totalChars and the enumerated markupItems (manifest) to compute… Jul 9, 2026 Implemented spec Conversational Trajectory Evals — a user-simulator layer for the whole conversation — design spec Before WS-5, the harness had 19 batteries. Every one tested a capability in one to three turns: mail classification precision/recall (WS-2), the proactive-gl… Jul 7, 2026 Implemented plan Conversational Trajectory Evals — implementation plan WS-5 schema v1.1 and runner are implemented. Phases 0–6 are complete; cheap lint-in-CI (7.1) is complete; MatterBench graduation history/reporting, verifier… Jul 7, 2026 Implemented spec Weekly "What's new" newsletter — design spec Jul 6, 2026 Implemented plan Weekly "What's new" newsletter — implementation plan Jul 6, 2026 Implemented spec Native macOS app — design spec Jul 4, 2026 Implemented plan Electron desktop cleanup — implementation plan Jul 4, 2026 Implemented plan Native macOS app — implementation plan Jul 4, 2026 Implemented spec Diligence upgrade — parallel subagent fan-out + click-to-source tabular verification — design spec LegalWork reference (from docs/reviews/2026-07-03-legalwork-vs-north-os.md): the tabular-review skill makes the orchestrator (1) resolve rows (docs) + column… Jul 3, 2026 Implemented plan Diligence upgrade — parallel subagent fan-out + click-to-source tabular verification — implementation plan The phases are ordered so value ships early and risk is front-loaded onto cheap changes: Jul 3, 2026 Implemented spec Polaris Windows — meeting capture (detection + recorder) — design spec Christophe shipped sub-feature 1 of North meeting capture (PR #24, on main): an encrypted, firm-scoped Jul 2, 2026 Implemented spec Cross-Matter Document Access — design spec The 550 10th Ave deal spans ≥4 matters (GTHM-121 equity/JV, GTHM-131 recast, GTHM-133 WF Mezz, GTHM-134 syndication), all client Gotham. Org docs like the OA… Jul 1, 2026 Implemented plan Cross-Matter Document Access — implementation plan Per standing rule ([[realistic-agent-eval-no-cheating]]): real conversation-style prompts, agent self-discovers everything, then the retrieval and the produc… Jul 1, 2026 Implemented plan Collaborative document writes in shared chats — plan Jun 30, 2026 Implemented plan Realistic Lawyer-Task Evals — implementation plan We do not fabricate fixture docs. Two real corpora already exist, each exercising a different production path — this mirrors the product's own model (no-bulk… Jun 30, 2026 Implemented spec Org-Scoped Chat Sharing with Realtime Collaboration — design spec Two tiers are fully derivable from created_by + row existence — no permission column (that's YAGNI and immutable-migration debt; a future read-only tier is a… Jun 27, 2026 Implemented plan Org-Scoped Chat Sharing with Realtime Collaboration — implementation plan Jun 27, 2026 Implemented spec Meeting auto-detection — design spec Granola's magic isn't transcription — it's that you never have to remember to hit record. You join a call and it's already capturing. North should do the sam… Jun 25, 2026 Implemented spec Product Analytics, Feedback & Session-Replay Posture — design spec Split "track everything + feedback + Hotjar-thing" into three separate jobs and solve each with the cheapest tool that respects attorney–client privilege. Do… Jun 25, 2026 Implemented spec User & Matter Memory + Chat Import — design spec Don't build a new memory subsystem — finish the one that already exists. agent_memory_candidate + save_memory already model user_private/matter/firm_global s… Jun 25, 2026 Implemented plan Product Analytics, Feedback & Session-Replay — implementation plan What changed since the 2026-06-26 reconciliation (all claims verified against the repo on 2026-07-06): Jun 25, 2026 Implemented spec North OS preproduction deploy — design spec Stand up a preproduction environment for North OS on the same Scaleway box that runs dev, with: Jun 22, 2026 Implemented plan Meetings package backend — implementation plan packages/meetings/ Jun 22, 2026 Implemented plan North OS preproduction deploy — implementation plan New directory deploy/preprod/: Jun 22, 2026 Implemented spec Skill forking + chat-based authoring — design spec Firm-managed skills v1 let a firm disable any of the 162 built-ins and author new markdown skills (form-based Add dialog + asset upload). It does not let a f… Jun 19, 2026 Implemented plan Skill forking + chat-based authoring (v1) — implementation plan Every increment ends green: bun run typecheck + bun run lint from root + the touched packages' tests. One conventional commit per increment (commit only when… Jun 19, 2026 Implemented plan Word surface v2 — in-place tracked changes — Implementation Plan resume: z.object({ Jun 19, 2026 Implemented spec Firm-managed skills — design spec Skills today are static, in-code, and firm-agnostic: 160 markdown methodologies under packages/agent-runtime/src/skills/library/, indexed by a generated cata… Jun 18, 2026 Implemented spec Word surface — cross-surface conversation continuity — design spec Make Microsoft Word a thin client of the North conversation spine, so a lawyer can: Jun 18, 2026 Implemented plan Firm-managed skills — implementation plan Every increment ends green: bun run typecheck + bun run lint from root + the touched packages' tests. One conventional commit per increment (commit only when… Jun 18, 2026 Implemented plan Word surface — cross-surface conversation continuity — Implementation Plan GET https://graph.microsoft.com/v1.0/shares/u!{base64url(webUrl)}/driveItem?$select=id,webUrl,parentReference Jun 18, 2026 Implemented spec Artifact figure provenance — enforce "every figure comes from a tool result" — design spec The legal-assistant agent fabricates figures into artifact cells. The eval Jun 17, 2026 Implemented plan Artifact figure provenance - implementation plan Jun 17, 2026 Implemented spec Durable desktop voice path + topic-aware thread router — design spec The desktop voice path (Polaris → clicky-gateway /dispatch → runDesktopChat → the desktop two-phase graph) goes KO after one conversation. Turn 1 (a fresh th… Jun 12, 2026 Implemented plan Durable Desktop Voice Path + Topic-Aware Thread Router Jun 12, 2026 Implemented spec Desktop Two-Phase Ack Interaction — design spec The desktop voice path works end-to-end but feels unresponsive: the lawyer speaks, then waits in silence while the whole pipeline runs (router LLM → screen-c… Jun 11, 2026 Implemented plan Desktop Two-Phase Ack — Implementation Plan Two independent reviews (Claude + Codex) found 3 BROKEN + 5 RISKY items, all verified against the real code. The task bodies below are kept for context, but… Jun 11, 2026 Implemented plan Notch island port + threads panel — Implementation Plan Jun 11, 2026 Implemented plan Phased ingestion — explicit Analyze / Ingest split + discovered sub-folder launch Today "Start ingestion" is one irreversible chain: discovery agent (Sonnet, ~$1–3/folder) → auto-enqueued walk → outbox → per-file LLM commits (the bulk spen… Jun 11, 2026 Implemented spec Clicky Smart Interaction + Reliability — design spec A real lawyer at HSE tests Clicky next week. Today the desktop path works end-to-end but has rough edges that undercut a first impression, and it does not ye… Jun 10, 2026 Implemented spec Agent document editing & creation, scoped to the lawyer's personal space — design spec Let the agent create documents and modify documents (both chat-uploaded docs and firm-synced drive docs) under one non-negotiable rule: Jun 10, 2026 Implemented spec Lawyer toolbelt P1 — PDF packaging, structured extraction, trackers — design spec P1 of the lawyer toolbelt attacks HSE pains #1 (signature packets), #2 (closing binders), #3 (title review) and #7 (trackers). Ten new agent tools plus one a… Jun 10, 2026 Implemented spec Lawyer toolbelt P2 — dates, calculations, term checks, email drafts, workflows — design spec P2 closes the P1 follow-ups and ships the next tool wave, mapped to the proposal's phasing: Jun 10, 2026 Implemented spec Lawyer toolbelt P3 — design spec P1 shipped the PDF/packaging, extraction, and tracker groups. P2 shipped dates, calculation, term-checks, the precedent/playbook tools, and draft_email/revis… Jun 10, 2026 Implemented plan Batch-API ingestion + pause/resume — half-price bulk ingestion you can stop and restart cleanly Bulk ingestion bills every document twice (classify + facts, both Haiku 4.5) at synchronous prices. The Completed CDs folder alone is 7,353 files ≈ $190 sync… Jun 10, 2026 Implemented plan Clicky Smart Interaction + Reliability Implementation Plan import { describe, expect, it } from "vitest" Jun 10, 2026 Implemented plan Agent document editing & creation, scoped to the lawyer's personal space — implementation plan Increment-by-increment; end every increment green (bun run typecheck + bun run lint from root + the touched package's tests) before the next. Stage files by… Jun 10, 2026 Implemented plan ElevenLabs Realtime STT Implementation Plan Locked from the official docs (2026-06): Jun 10, 2026 Implemented plan Lawyer toolbelt P1 — implementation plan Every increment ends green: bun run typecheck + bun run lint from root + the touched packages' tests. One conventional commit per increment. Jun 10, 2026 Implemented plan Lawyer toolbelt P2 — implementation plan Every increment ends green: bun run typecheck + bun run lint from root + the touched packages' tests. One conventional commit per increment. Jun 10, 2026 Implemented plan Lawyer toolbelt P3 — implementation plan Every increment ends green: bun run typecheck + bun run lint from root + the touched packages' tests. One conventional commit per increment. Jun 10, 2026 Implemented plan .msg email parsing — Outlook emails become ingestable knowledge properties. The ingestion pipeline has no parser for it, so parseByMime Jun 10, 2026 Implemented plan RAG eval harness — a testable ingestion+retrieval system with a measured success rate The pipeline now works (328 documents ingested end-to-end), but "works" is Jun 10, 2026 Implemented plan Agent RAG retrieval — wire the chat agent to the ingested knowledge store Increment-by-increment; end every increment green (bun run typecheck + bun run lint from root + the touched package's tests) before the next. Stage files by… Jun 9, 2026 Implemented spec Mail-Intelligence P3 — the Outlook mail write path — design spec Make Outlook mail ingested → classified → stored → readable by the agent. P2 delivered the ingestion plumbing (OAuth, subscriptions, webhook, source.email.re… Jun 4, 2026 Implemented plan Mail-Intelligence P3 — implementation plan (rev 2) Increment-by-increment; end every increment green (bun run typecheck + bun run lint from root + the touched package's tests) before the next. Stage files by… Jun 4, 2026 Implemented spec Clio matter index + instant picker search — design spec Replace the matter picker's live per-keystroke Clio search with a local index for Jun 1, 2026 Implemented spec Meetings package — design spec The prior draft was structurally sound but built on a foundation that does not exist in the current repo. This spec corrects every divergence against the rea… May 29, 2026 Implemented plan Meetings Package Implementation Plan This plan was drafted when 0023 was the migration head and against an older checkout. A planner re-verified every load-bearing fact against the current workt… May 29, 2026 Implemented spec Clicky gateway extraction — design spec Stop bending the apps/web Next.js server to host Polaris's WebSocket. Extract every Clicky-specific surface (HTTP routes + WS + Redis subscriber + dispatch +… May 28, 2026 Implemented plan Clicky gateway extraction — implementation plan Files this plan creates or modifies, by phase. Each file has one clear responsibility — focused units that can be read and edited in isolation. May 28, 2026 Implemented spec Clio matter picker + per-connector manage pages — design spec Make Clio tracking selective. Today connecting Clio means "sync all 4000 matters at HSE" — that is a real blocker for go-live and a UX problem for any firm w… May 27, 2026 Implemented plan Clio matter picker + per-connector manage pages — implementation plan Today's journal lands at 0021_hesitant_maximus.sql. This work owns 0022_clio_matter_picker.sql. The Clio matter subsystem (matter table, connector, worker, c… May 27, 2026 Implemented plan Polaris foundation — manual test plan Walk through the entire foundation end-to-end against real infrastructure May 27, 2026 Implemented spec Clio matter subsystem — design spec Build the Matter subsystem for North OS: a first-class matter entity, sourced from Clio for HSE, that becomes the central organizing unit for chats, document… May 26, 2026 Implemented plan Clicky Foundation Implementation Plan Files this plan creates or modifies, by phase. Each file has one clear responsibility — focused units that can be read and edited in isolation. May 26, 2026 Implemented plan Clio matter subsystem — implementation plan P5.1 is now represented by journal entry 0019_agent_tool_audit_with_check, so this work owns the next migration number: 0020_matter_subsystem.sql. The Clio d… May 26, 2026 Implemented plan Outlook Product Surface — Design + Implementation Plan In scope: May 26, 2026 Implemented plan P5.1 — Outlook Surface Remediation Plan In scope (8 findings — see § 1 for the consolidated mapping): May 26, 2026 Implemented spec Outlook ingestion seam — design spec Close the Outlook ingestion gap identified by the P3.R re-audit (Review E): May 25, 2026 Implemented plan Outlook Ingestion Seam Implementation Plan Run: cat packages/db/src/migrations/_journal.json && echo "---" && ls packages/db/src/migrations/ | tail -8 May 25, 2026 Implemented plan Port hardening — implementation plan The foundation. Other phases depend on the roles existing. May 22, 2026 Implemented spec `ask_user` question tool — design The legal-assistant agent cannot ask the lawyer a question mid-turn. When a May 21, 2026 Implemented spec Chat rebuild on @assistant-ui — design The chat is half-finished. Eight defects, all user-visible or trust-critical: May 21, 2026 Implemented spec Clicky foundation — design spec Build the Clicky launcher end to end: a macOS menu-bar surface where the May 21, 2026 Implemented spec Clicky as a North OS surface — vision The north star for the Clicky workstream. It captures what Clicky becomes May 21, 2026 Implemented spec Connections subsystem — design spec Build the Connections subsystem — the product surface that lets a law firm connect its document store to North OS and ingest its documents into the knowledge… May 21, 2026 Implemented spec Handoff skill — design spec handoff document and launches a fresh Claude in a new herdr space, pre-loaded to May 21, 2026 Implemented spec Per-thread model selector — design A lawyer can choose which Claude model a chat thread runs on — Opus 4.7 (deep analysis), Sonnet 4.6 (balanced — default), or Haiku 4.5 — and see, on every as… May 21, 2026 Implemented spec Spawn worktree agent skill — design spec Let a Claude session running inside herdr spin off a fresh, May 21, 2026 Implemented plan Connections subsystem — implementation plan The backend/contract foundation. Migration 0007, then the corpus_category field flowing from ingest_sources through the source.document.* event into knowledg… May 21, 2026 Implemented plan Per-thread Model Selector Implementation Plan Create packages/agent-runtime/src/model.test.ts: May 21, 2026 Implemented spec Chat Document Upload + `read_document` Tool — Design Let a lawyer attach a document to a chat and have the legal-assistant agent read May 20, 2026 Implemented spec Forma services port — design spec Port the backend services of the Forma MVP into north_os — service by service, each validated independently. Not a monolithic copy: each service is a self-co… May 20, 2026 Implemented spec LangGraph agent architecture — design test). That work is wiped; this is the proper foundation. May 20, 2026 Implemented spec Legal skills system — design runtime this extends). May 20, 2026 Implemented spec Signup & onboarding flow — design spec The groundwork landed magic-link login and the Better Auth organization plugin, but there is no way for a new firm to come on board. Today a user can only si… May 20, 2026 Implemented spec spec-bmad skill — design spec north_os currently produces superpowers-style design specs (docs/superpowers/specs/YYYY-MM-DD-<topic>-design.md) — a single lightweight file written before i… May 20, 2026 Implemented plan Chat Document Upload Implementation Plan Design: docs/superpowers/specs/2026-05-20-chat-document-upload-design.md. May 20, 2026 Implemented plan Forma services port — implementation plan P0 is the prerequisite for every service. No per-service go/no-go — it must complete before P1. May 20, 2026 Implemented plan LangGraph Agent Architecture Implementation Plan In apps/web/app/api/[[...route]]/route.ts, delete the chatRoute import line and the app.route("/", chatRoute) block (and its comment). The file returns to: h… May 20, 2026 Implemented plan Legal Skills System Implementation Plan import { afterEach, describe, expect, it } from "vitest" May 20, 2026 Implemented plan Signup & Onboarding Implementation Plan Create packages/db/src/schema/signup.ts: May 20, 2026 Implemented spec Pre-port cleanup — design spec (Forma side) Prepare the Forma repo so the upcoming port of its backend layer to a new product repo (north_os) is mechanical. Every artifact carried into the new repo mus… May 19, 2026 Implemented spec Pre-port groundwork — design spec (north_os side) Scaffold the north_os repo so the upcoming port of the backend layer from the MVP repo (forma-apps/forma) lands cleanly. Concretely: May 19, 2026 Implemented plan Pre-port cleanup implementation plan (Forma) This phase produces 8 new packages, 4 new apps, deletes ~12 routes + 2 apps, and renames ~50 files. The change surface is large but every change is local: no… May 19, 2026 Implemented plan Pre-port groundwork implementation plan (north_os) This phase creates 12 new workspace packages (mostly as empty slots) + 4 new worker app slots + extracts the DB out of apps/web into its own package + adds H… May 19, 2026 Implemented spec Forma `docs/` Reorganization — Design Spec Reorganize forma/docs/ so it (a) reflects what the MVP code actually shipped, (b) is navigable in seconds without prior context, and (c) serves as the refere… May 18, 2026 Implemented

spec Modular intake loops — design spec (v2) The tips loop works. It is also built as a fixed product object rather than a configuration of Aug 25, 2026 Superseded plan Modular intake loops — implementation plan (v3) Companion to specs/2026-08-25-modular-intake-loops-design (v2). Branch Aug 25, 2026 Superseded spec Tips loop as first-class product tooling — design spec The NetDocuments tips loop should be declarable from a conversation: a lawyer says "watch this mailbox for NetDocs tips and send me a digest", the agent sets… Aug 17, 2026 Superseded plan Tips loop as first-class product tooling — implementation plan Eleven units, ordered so each is independently testable. Backend units go to omp --approval-mode yolo; unit 8 is frontend and goes to claude --permission-mod… Aug 17, 2026 Superseded spec Eval gate split — cheap batteries per PR, agent-quality ratchet at the deploy — design spec Jul 23, 2026 Superseded plan Eval gate split — implementation plan Jul 23, 2026 Superseded spec NetDocs tips loop — design spec Jul 22, 2026 Superseded plan NetDocs tips loop — implementation plan Jul 22, 2026 Superseded spec Eval agent-quality grounding fix — design spec Jul 20, 2026 Superseded plan Eval agent-quality grounding fix — implementation plan Jul 20, 2026 Superseded spec Desktop shell — design spec Build order: C1 → C2 → C3 → C4-human → C4-agent. Total ≈ 3–4 weeks calendar to the full feature set, with most coding parallelizable across a Claude/Codex fl… Jul 3, 2026 Superseded plan Desktop shell — implementation plan Jul 3, 2026 Superseded spec Memory Calibration — the lawyer's trust tool — design spec Build the calibration tool in four composable layers, smallest-valuable-first. The through-line: the lawyer is the judge; the model only triages what's worth… Jun 29, 2026 Superseded plan Memory Calibration v2 — implementation plan Migration slot 0057 and the mine.ts hot loop are contended by several pieces, so the order is fixed: Jun 29, 2026 Superseded spec Smart proactive glance — design spec The proactive glance today is a one-shot guesser. runProactiveGlance (packages/chat-runtime/src/glance.ts) feeds one screenshot + the foreground app name to… Jun 18, 2026 Superseded plan Smart Proactive Glance Implementation Plan Jun 18, 2026 Superseded spec Proactive actionable nudge — design spec The proactive glance today surfaces a silent observation — a one-line message plus an optional deepLink that, when tapped, opens a browser thread. It is pass… Jun 12, 2026 Superseded plan Proactive Actionable Nudge Jun 12, 2026 Superseded spec Island indicator + proactive nudges — design spec Two desktop surfaces on the Polaris macOS app, prototyped fast: Jun 11, 2026 Superseded plan Island indicator + proactive nudges — Implementation Plan Jun 11, 2026 Superseded spec Meetings package — design Lawyers at HSE (and future firms) already record themselves on Loom, Zoom, Microsoft Teams, and ad-hoc calls. They use Granola heavily for internal/team meet… May 28, 2026 Superseded plan Meetings package MVP — implementation plan git checkout main May 28, 2026 Superseded spec Chat-pipeline glue — design doc Define how the chat surface in apps/web (built with the Vercel AI SDK + assistant-ui + ai-elements) invokes pipeline runs as agent tools. The pipeline runner… May 19, 2026 Superseded