Requirements Register

Canonical source: docs/02-business-product/requirements-register.md — this page is rendered from it on every publish; edit the source.

Requirements Register

Product and business requirements that are not owned by a single feature spec: positioning, cross-cutting commitments, and requirements found missing. A feature's own requirements stay in its spec under this folder or in docs/features/<feature>/02-product-requirements.md; this register links to them rather than copying them.

Rules. An id is never reused. Entries are never deleted — a dropped requirement keeps its row with why and who decided. The history column is append-only. A requirement inferred rather than stated by the owner is marked inferred until the owner confirms it.

Status	Means
current	In force as written
changed	In force, reworded or reprioritised; the history says from what
missing	Needed, not yet specified or built; a question for the next intake
dropped	No longer wanted; kept so it is not re-proposed

---

Positioning (ADR-126)

REQ-POS-001

• Statement: The product's public slogan is "Every AI, one workspace." and its description is "Every frontier AI model in one workspace that sees, hears, researches and builds. Pay as you go, bring your team, or run it on your own hardware."
• Status: current · Priority: high
• Source: owner decision, 2026-09-26
• Acceptance: the README, the product vision and the marketing site carry the same wording; no surface still leads with "One Subscription".
• Touches: product-vision.md, root README.md, apps/claw-frontend marketing copy (separate batch).
• History: 2026-09-26 created (DRIFT-001).

REQ-POS-002

• Statement: Every feature named in public copy or the README traces to wired, shipped code, recorded in the flagship catalog with an evidence path. Partial features carry their limit; unbuilt ones are not named.
• Status: current · Priority: high
• Source: owner instruction, 2026-09-26 ("never invent facts")
• Acceptance: each flagship row has an evidence path and a status; the catalog's gap table lists what copy must not claim.
• Mechanism: the audit step in skills/change-the-product-positioning.md; no automated check exists yet.
• History: 2026-09-26 created.

REQ-POS-003

• Statement: Pay-as-you-go is the lead commercial story: usage is metered as connector credit and shown in the visitor's currency.
• Status: current · Priority: high
• Source: owner decision, 2026-09-26
• Acceptance: met by shipped code (flagship 9). Numbers stay in docs/business/.
• History: 2026-09-26 created.

REQ-POS-004

• Statement: The whole product runs on the customer's own hardware, with local models available, as a supported deployment.
• Status: current · Priority: high
• Source: owner decision, 2026-09-26 (pillar "local-first and privacy")
• Acceptance: met by shipped code (flagship 14); local AI is opt-in at install.
• History: 2026-09-26 created.

REQ-POS-005

• Statement: A team or business customer can hold one account with several members, shared billing and pooled usage.
• Status: missing · Priority: Unknown - ask the owner and record the answer.
• Source: implied by the owner's "bring your team" pillar (2026-09-26) — inferred. Also implied by the Team plan's seeded description, "Shared workspaces and a large pooled allowance" (apps/claw-auth-service/prisma/seeders/plan-catalog.json).
• Current state: not built. Users are admin-managed on one install, each with their own plan and quota. The only organisation model is the coding-agent fleet's, which is API-only.
• Until built: copy says "bring your team" and describes admin-managed users, roles, plan grants and usage statistics; it does not promise seats, shared billing or pooling. The Team plan description no longer overclaims (2026-09-29, text only: "A large monthly allowance for heavy daily use.").
• History: 2026-09-26 created; 2026-09-29 Team plan description corrected (owner approved a text-only change).

REQ-POS-006

• Statement: Business customers can sign in with SSO (SAML or OIDC).
• Status: missing · Priority: Unknown - ask the owner and record the answer.
• Source: inferred from the teams/business pillar, 2026-09-26.
• Current state: not built for the web app. A SAML callback exists in apps/claw-agent-service/src/modules/fleet/controllers/saml.controller.ts but creates no session and has no UI.
• History: 2026-09-26 created.

Found missing during the audit

REQ-SEC-001

• Statement: Only a member (with a sufficient role) of a coding-agent organisation may list its members or devices, add a member, or set its SAML metadata.
• Status: done · Priority: high (security)
• Source: found by the ADR-126 flagship audit, 2026-09-26.
• Current state: fixed in commit cbe566511. OrganizationAccessService (apps/claw-agent-service/src/modules/fleet/services/organization-access.service.ts) checks the caller's own membership in the service layer. Members may read the member list; OWNER/ADMIN may add members, change the policy, set SSO metadata and read the device matrix; only an OWNER may grant OWNER; re-adding an existing member is 409. Outsiders, including platform admins, get 404. Production had 0 organisations when fixed, so nothing was exploited.
• Governing rule: rules/16 (IDOR).
• History: 2026-09-26 created; fixed the same day.

Chat supremacy program (2026-09-29)

Intake audit: chat-capability-audit-2026-09.md · benchmark: chat-competitive-benchmark-2026-09.md.

REQ-CHAT-001

• Statement: A user can branch a conversation from any message, get back to the source from the branch, and see the branches cut from a conversation. A branch inherits the conversation up to the fork point (text, attachments, settings, privacy switches) and never anything said after it — including via cross-thread retrieval.
• Status: done · Priority: high
• Source: Chat Supremacy prompt pack, Batch A (owner, 2026-09-29).
• Current state: ADR-130 — lineage columns, full-field copy, family exclusion, GET /chat-threads/:id/lineage, lineage strip + thread-list icon. Merge-back is not built (no product decision on what a merged message is).
• History: 2026-09-29 created and delivered (Batch 1).

REQ-CHAT-002

• Statement: Privacy defects found by the intake audit are fixed before further memory-dependent features: memory extraction must honour a chat's useMemory=false, and a chat with useCrossThreadContext=false must not be retrieved INTO other chats.
• Status: done · Priority: high (privacy)
• Source: intake audit defects 3–4, 2026-09-29 (inferred — owner to confirm the intended semantics of the per-chat switch).
• Current state: fixed 2026-09-30 (Batch 3). message.completed carries useMemory; memory-service skips extraction when it is false. Cross-thread candidates require BOTH switches on. The settings copy in 13 locales now says both directions (and no longer claims cross-thread is off by default, which was untrue). Owner answer (2026-09-30): no — "Use memory" off does NOT stop the chat reading your other chats; only "Use relevant previous chats" controls reading. The two-way reading is confirmed; the inferred mark is removed.
• History: 2026-09-29 created; 2026-09-30 delivered.

REQ-CHAT-003

• Statement: A user can select text in any earlier message, quote it into the composer (up to 3), and ask about it; the answer knows exactly which text was meant, and the quote stays visible on the sent turn.
• Status: done (messages) · Priority: high
• Source: Chat Supremacy prompt pack, Batch B (owner, 2026-09-29).
• Current state: ADR-131. Quoting from citations, file previews, research reports and generated documents waits for batches M and D.
• History: 2026-09-30 created and delivered (Batch 2).

REQ-CHAT-004

• Statement: A user can answer the same question again with a model of their choice (or AUTO), and can edit a question without deleting the conversation after it.
• Status: done · Priority: high
• Source: Chat Supremacy prompt pack, Batch E (owner, 2026-09-29).
• Current state: ADR-132. Regenerate also gained the plan/quota check it was missing. Still open in batch E: send an answer to Repair/Verify/Compare from the message, and "continue" a truncated answer.
• History: 2026-09-30 created and delivered (Batch 4).

REQ-CHAT-005

• Statement: An answer's inline [n] opens the exact source the model was given under that number — and never a source it was not given.
• Status: done (main chat) · Priority: high
• Source: Chat Supremacy prompt pack, Batch M (owner, 2026-09-29).
• Current state: ADR-133, rule 41 §16. Compare/lab lanes, file-page provenance and a grounded-vs-model-knowledge marker are not built.
• History: 2026-09-30 created and delivered (Batch 5).

REQ-CHAT-006

• Statement: When the user asks, in their own words, to remember something or add it to their context, a model decides memory, context pack or both, saves the right text with the right memory type, asks which pack when the user has packs and did not name one, shows a saved card with links to the exact item, keeps it in the chat history, and the AI confirms it is done.
• Status: done · Priority: high
• Source: owner, in chat, 2026-09-30 (with the four design choices recorded in ADR-134).
• Current state: ADR-134. Changing a memory's type happens on the Memory page (the card's link opens the editor), not on the card.
• History: 2026-09-30 created and delivered (Batch 6).

REQ-CONNECTOR-001

• Statement: An administrator can register an OpenAI-compatible provider and configure encrypted connector credentials; synced model identity remains distinct through model catalog and routing.
• Status: partial · Priority: high
• Source: owner prompt packs, 2026-10-03.
• Current state: ADR-157, docs/features/runtime-managed-connector-providers/. Custom definitions and protected built-in status controls are implemented. Live migration, RBAC and product QA remain open. NVIDIA NIM's protocol path is mocked; no live call is claimed.
• Who: owner; implemented and evidence-recorded by the coding agent.

ClawAI Threads

REQ-THREADS-001

• Statement: A user can turn a private chat into a sourced publication through bounded multi-model review, then approve it before it becomes public; approved work supports moderated community participation and independent discovery.
• Status: current · Priority: high
• Source: owner-approved prompt pack decisions, 2026-10-04.
• Acceptance: see the Threads product spec for content types, disclosure, approval, citations, spend cap, moderation, deletion, and indexing requirements.
• Current state: the baseline publication workflow is live. A production review on 2026-10-25 found an already-published thread and a separate run that failed before Judge arbitration; local reliability fixes are recorded in REQ-THREADS-002. This register is not deployment evidence.
• History: 2026-10-04 created for the ClawAI Threads delivery plan.

REQ-THREADS-002

• Statement: Thread generation treats author votes as preferences among candidate drafts. In every round the Judge receives the candidates and their agreement percentage, selects and reviews one candidate, and the Critic reviews that selected draft. Dissent alone does not fail a round. Votes and reviews are checkpointed so a retry resumes saved work instead of repeating completed model calls.
• Status: current · Priority: high
• Source: owner request, 2026-10-25; see ADR-190.
• Acceptance: 50 deterministic local replay cases with split, tie and majority preferences pass the Judge and Critic quality gates (target at least 45/50); no live-provider reliability rate is inferred from the mock replay.
• Current state: all 50 local replay cases pass. The Judge and Critic score, citation and safety checks remain required. Owner-edited single drafts report author agreement as information only.
• History: 2026-10-25 created and implemented; no database migration.