Devin

CORRECTED 2026-09-12 (hello-18 serving wave W-c): the reason this row carried said canon carries no attachment entity AT ALL, and hello-15 made that FALSE - canon.Attachment landed with the support-desk domain (packages/canon/src/support-desk.ts:962), the artifact holds 35 attachment rows with real bodies, and renderer-zendesk resolves them by id today. The empty answer still stands on the NARROWER fact, which is this row's own: every attachment in this universe hangs off a record some OTHER product owns, and NO generation attaches one to an AGENT SESSION - a session here starts from repository work rather than from a file somebody dragged in, so nobody ever uploaded one, and Devin's own answer for a session with none is the empty list this renderer's sessionAttachments() has returned since the host opened. THE ASK is an attachment whose PARENT is an agent session

the org-scoped twin of the enterprise attachments row, held to the same answer by test. CORRECTED 2026-09-12 (hello-18 serving wave W-c): the reason this row carried said canon carries no attachment entity AT ALL, and hello-15 made that FALSE - canon.Attachment landed with the support-desk domain (packages/canon/src/support-desk.ts:962), the artifact holds 35 attachment rows with real bodies, and renderer-zendesk resolves them by id today. The empty answer still stands on the NARROWER fact, which is this row's own: every attachment in this universe hangs off a record some OTHER product owns, and NO generation attaches one to an AGENT SESSION - a session here starts from repository work rather than from a file somebody dragged in, so nobody ever uploaded one, and Devin's own answer for a session with none is the empty list this renderer's sessionAttachments() has returned since the host opened. THE ASK is an attachment whose PARENT is an agent session

generate — new canon entity AgentSessionInsight

generate — new canon entity AgentSessionInsight

refuse — conversational content: the session transcript is what Devin and the engineer said to each other, and this universe does not invent it (the standing founder-adopted default; revisit if a client asks)

refuse — a snapshot blueprint and its builds exist only because a write created them, and this universe accepts no writes (invariant #3) — the async-artifact shape ruled in coverage/reviews/2026-09-01-jira-modes.md

refuse — a snapshot blueprint and its builds exist only because a write created them, and this universe accepts no writes (invariant #3) — the async-artifact shape ruled in coverage/reviews/2026-09-01-jira-modes.md

refuse — a snapshot blueprint and its builds exist only because a write created them, and this universe accepts no writes (invariant #3) — the async-artifact shape ruled in coverage/reviews/2026-09-01-jira-modes.md

refuse — a snapshot blueprint and its builds exist only because a write created them, and this universe accepts no writes (invariant #3) — the async-artifact shape ruled in coverage/reviews/2026-09-01-jira-modes.md

refuse — a snapshot blueprint and its builds exist only because a write created them, and this universe accepts no writes (invariant #3) — the async-artifact shape ruled in coverage/reviews/2026-09-01-jira-modes.md

refuse — a snapshot blueprint and its builds exist only because a write created them, and this universe accepts no writes (invariant #3) — the async-artifact shape ruled in coverage/reviews/2026-09-01-jira-modes.md

refuse — Devin's code-scanning add-on is not enabled for this enterprise, and Devin answers 403 to every code-scan read for an enterprise without it — no scan, finding, metric or profile can exist here

refuse — Devin's code-scanning add-on is not enabled for this enterprise, and Devin answers 403 to every code-scan read for an enterprise without it — no scan, finding, metric or profile can exist here

refuse — Devin's code-scanning add-on is not enabled for this enterprise, and Devin answers 403 to every code-scan read for an enterprise without it — no scan, finding, metric or profile can exist here

refuse — Devin's code-scanning add-on is not enabled for this enterprise, and Devin answers 403 to every code-scan read for an enterprise without it — no scan, finding, metric or profile can exist here

refuse — Devin's code-scanning add-on is not enabled for this enterprise, and Devin answers 403 to every code-scan read for an enterprise without it — no scan, finding, metric or profile can exist here

refuse — the id half of the service-user pair: this universe's only non-human identity is the CI bot and it holds no Devin seat, so the service-user list is empty, so no id can resolve and Devin's own 404 is the only possible answer (the by-id shape ruled in DECISIONS 2026-09-01 [coverage/A0])

CORRECTED 2026-09-12 (hello-18 serving wave W-c). The 2026-09-10 correction already replaced this row's original reason - "this universe has no identity provider" - because hello-15's canon.IdentityProviderConfig and canon.IdentityGroup (packages/canon/src/ci-config.ts:359, :401) made it false: a SAML provider with SCIM enabled, the administrator who switched it on, the date it was switched on, and four groups derived from ACCESS facts, which renderer-circleci serves today (/organizations/{org_id}/groups, ruled 2026-09-07). Two of the three supporting clauses that correction left have since gone false in their turn, and this is the second correction this row has needed for the same reason: a ruling resting on an absence canon later filled. canon DOES attach a role to an identity group (hello-17's grants - a membership row with scope_kind 'group', reader.groupRole(), which renderer-circleci serves), and canon DOES bind a person to an external uid (hello-18's externalUid - person.external_uid, which GitLab's saml/{uid} and scim/{uid} resolve). The empty answer still stands, and it now stands on the FIRST clause alone - which is the one the vendored document actually asks about: `IdpGroupUser` is "a user whose membership is derived from IDP group assignments", and NO Devin membership here is derived that way. Every Devin seat is assigned DIRECTLY to a person by the organization, which is the same absence every member detail already reports as idp_role_assignments: []. THE ASK is now one fact rather than two: a Devin SEAT a person holds BECAUSE an identity group confers it. A directory identity on a person is not that, and a role on a group is not that either. NOTE for the next wave: the five idp-GROUPS rows in this fence still carry the falsified "canon attaches no role to an identity group" clause, and they are deliberately untouched here because on those rows the correction may be a re-SERVE rather than a re-word - `IdpGroup.role_assignments` is RoleAssignment[] and hello-17's groupRole() may now answer it - which is a ruling this wave cannot make alone

refuse — knowledge notes are authored free-text prose a human wrote, the same class of content as session messages, and this universe does not invent authored prose (the standing founder-adopted default; revisit if a client asks)

refuse — knowledge notes are authored free-text prose a human wrote, the same class of content as session messages, and this universe does not invent authored prose (the standing founder-adopted default; revisit if a client asks)

refuse — knowledge notes are authored free-text prose a human wrote, the same class of content as session messages, and this universe does not invent authored prose (the standing founder-adopted default; revisit if a client asks)

SERVED 2026-09-12 (post-hello-18 serving pass). The 2026-09-10 correction on this row named its own expiry: the empty stood because "canon attaches no role to an identity group and binds no person to an external uid", and THE ASK was exactly those two facts. Both have landed - hello-17's features.grants writes one membership row per identity group with scope_kind 'group' (reader.groupRole(), decisions/2026-09-11-2304-hello-17-grants.md), and hello-18's features.externalUid puts person.external_uid on every human (which GitLab's saml/{uid} and scim/{uid} resolve) - and the 2026-09-12 wave that re-worded the idp-USERS rows left the note this answers: "on those rows the correction may be a re-SERVE rather than a re-word". A GROUP IS REGISTERED WITH THIS ENTERPRISE WHEN CANON BINDS IT TO A ROLE: that binding is what registering one MEANS, canon records no Devin-side registration object, and gating all five rows on it is what keeps the three lists answering the same set. This is the ACCOUNT-scoped registry and IdpGroupResponse declares ONE field, the name - no role, no member count, no created instant - so it publishes the same groups the two IdpGroup lists do rather than every directory group: three lists that could disagree about which groups exist would be invariant #5 broken across one family. EVERY MEMBER STILL REPORTS idp_role_assignments: [] AND THAT IS NOT A CONTRADICTION - it is the vendored document's own distinction: IdpRoleAssignment is a role "inherited via IDP group membership" and IdpGroupUser is a user whose "membership is derived from IDP group assignments", both facts about a SEAT, and every Devin seat here is assigned DIRECTLY by the organization (canon's ai_seat), which is the fact the idp-USERS rows were re-ruled onto on 2026-09-12. IdpGroup.role_assignments is what the GROUP confers, not what any member has received. Gated on groupRole(), so on every artifact without features.grants - all 127 registered pins - the lists are Devin's empty envelope and the by-name reads are its 404, byte for byte what these rows served before; conformance asserts that on the hello-7 harness those pins ride.

SERVED 2026-09-12 (post-hello-18 serving pass). The 2026-09-10 correction on this row named its own expiry: the empty stood because "canon attaches no role to an identity group and binds no person to an external uid", and THE ASK was exactly those two facts. Both have landed - hello-17's features.grants writes one membership row per identity group with scope_kind 'group' (reader.groupRole(), decisions/2026-09-11-2304-hello-17-grants.md), and hello-18's features.externalUid puts person.external_uid on every human (which GitLab's saml/{uid} and scim/{uid} resolve) - and the 2026-09-12 wave that re-worded the idp-USERS rows left the note this answers: "on those rows the correction may be a re-SERVE rather than a re-word". A GROUP IS REGISTERED WITH THIS ENTERPRISE WHEN CANON BINDS IT TO A ROLE: that binding is what registering one MEANS, canon records no Devin-side registration object, and gating all five rows on it is what keeps the three lists answering the same set. The by-name half, and its refusal was DISSOLVED rather than overruled: the row was refuse because "the IdpGroup list is empty, no name can resolve, and Devin's own 404 is the only possible answer". The list is not empty now, a name it carries answers 200 with the list's own item, and a name it does not carries Devin's ordinary resource 404 - the same status the ruling wanted, produced by the served route rather than asserted by a table, and WITHOUT x-sandboxapis-refusal, which conformance asserts is absent. EVERY MEMBER STILL REPORTS idp_role_assignments: [] AND THAT IS NOT A CONTRADICTION - it is the vendored document's own distinction: IdpRoleAssignment is a role "inherited via IDP group membership" and IdpGroupUser is a user whose "membership is derived from IDP group assignments", both facts about a SEAT, and every Devin seat here is assigned DIRECTLY by the organization (canon's ai_seat), which is the fact the idp-USERS rows were re-ruled onto on 2026-09-12. IdpGroup.role_assignments is what the GROUP confers, not what any member has received. Gated on groupRole(), so on every artifact without features.grants - all 127 registered pins - the lists are Devin's empty envelope and the by-name reads are its 404, byte for byte what these rows served before; conformance asserts that on the hello-7 harness those pins ride.

SERVED 2026-09-12 (post-hello-18 serving pass). The 2026-09-10 correction on this row named its own expiry: the empty stood because "canon attaches no role to an identity group and binds no person to an external uid", and THE ASK was exactly those two facts. Both have landed - hello-17's features.grants writes one membership row per identity group with scope_kind 'group' (reader.groupRole(), decisions/2026-09-11-2304-hello-17-grants.md), and hello-18's features.externalUid puts person.external_uid on every human (which GitLab's saml/{uid} and scim/{uid} resolve) - and the 2026-09-12 wave that re-worded the idp-USERS rows left the note this answers: "on those rows the correction may be a re-SERVE rather than a re-word". A GROUP IS REGISTERED WITH THIS ENTERPRISE WHEN CANON BINDS IT TO A ROLE: that binding is what registering one MEANS, canon records no Devin-side registration object, and gating all five rows on it is what keeps the three lists answering the same set. The org-scoped by-name twin, on the same terms as its enterprise twin: a name the registry carries answers 200 with the list's own item, a name it does not is Devin's ordinary resource 404 with no refusal header, and an {org_id} nobody has is the org wrapper's 404. EVERY MEMBER STILL REPORTS idp_role_assignments: [] AND THAT IS NOT A CONTRADICTION - it is the vendored document's own distinction: IdpRoleAssignment is a role "inherited via IDP group membership" and IdpGroupUser is a user whose "membership is derived from IDP group assignments", both facts about a SEAT, and every Devin seat here is assigned DIRECTLY by the organization (canon's ai_seat), which is the fact the idp-USERS rows were re-ruled onto on 2026-09-12. IdpGroup.role_assignments is what the GROUP confers, not what any member has received. Gated on groupRole(), so on every artifact without features.grants - all 127 registered pins - the lists are Devin's empty envelope and the by-name reads are its 404, byte for byte what these rows served before; conformance asserts that on the hello-7 harness those pins ride.

refuse — an automation INSTANCE exists only because somebody configured one, and this universe accepts no writes (invariant #3); the vendor-fixed half of the surface — schemas and templates — is derived instead

derive

nobody here ever wrote a playbook, so the playbook list is empty

refuse — the id half of the playbook pair: nobody here ever wrote a playbook, so the playbook list is empty, so no id can resolve and Devin's own 404 is the only possible answer (the by-id shape ruled in DECISIONS 2026-09-01 [coverage/A0])

this universe's only non-human identity is the CI bot and it holds no Devin seat, so the service-user list is empty

refuse — the id half of the service-user pair: this universe's only non-human identity is the CI bot and it holds no Devin seat, so the service-user list is empty, so no id can resolve and Devin's own 404 is the only possible answer (the by-id shape ruled in DECISIONS 2026-09-01 [coverage/A0])

refuse — the git-provider connection is an install-time OAuth object on Devin's side, and this universe's repositories are served by our own GitHub/GitLab mirrors rather than through a Devin-held provider grant

refuse — the git-provider connection is an install-time OAuth object on Devin's side, and this universe's repositories are served by our own GitHub/GitLab mirrors rather than through a Devin-held provider grant

SERVED 2026-09-12: RepositoryResponse over repo/commits/git_tree_entry. provider_repository_id is githubId(repo.id) - the very id the GitHub replica publishes for the same repository - and the connection host is github.com, which every Devin session's own pr_url already names. The connection id is required with no column and is MINTED deterministically from the ORGANIZATION, because this enterprise has exactly one connection and every repository is read through it; a per-repository id would claim as many connections as repositories. indexing_status is nullable and is null: canon carries no Devin index

derive

generate — A WIDENING, NOT A BUILD: canon.AdminAuditEvent exists (packages/canon/src/support-desk.ts:690) and its own doc comment calls it "GENERIC by construction". The action vocabulary (AuditAction, :602), the change vocabulary (AuditChangeSlot, :633), the target vocabulary (AuditTargetKind, :611) and the RFC 5737 source-address rule (AuditSourceBlock, :652) all shipped, and Zendesk serves the table today (reader.listAuditEvents(), ruled 2026-09-07). TWO things are missing and neither is the entity: git-host members on AuditTargetKind, and a generation that WRITES personActor — the field is declared for exactly this wave and "this generation writes none" (:695-697)

refuse — guardrails are an enterprise-security beta this org never turned on, and Devin answers 403 to guardrail-violation reads for an enterprise with no guardrail policy configured

SERVED 2026-09-12: OrganizationResponse over the canonical `org` row this host already resolves {org_id} against. created_at is reader.orgCreatedEpoch(), updated_at the newest canonical event in any of the organization's repositories, and BOTH ACU limits are declared integer|null and are null - this enterprise buys ACUs and spends them (the consumption family reports what was spent) and nobody here capped a session or a cycle, which is the same fact /v3/enterprise/consumption/acu-limits/devin already answers empty

refuse — no Devin Review exists in this universe - a review record is created by a trigger through the API and this universe accepts no writes (invariant #3), and canon's ai_pull rows attribute AUTHORSHIP rather than review (the two pulls Devin produced were reviewed by people, person-14 and person-04); Devin's own answer when no review exists for the resolved commit is its 404, which the operation's own description states. RULED by the founder 2026-09-12

SERVED 2026-09-10 (hello-16): reader.listMembershipRoles() (packages/artifact/src/reader.ts:3042) in Devin's PaginatedResponse[Role] envelope. `role_type` is REQUIRED with a closed enterprise|org enum canon does not carry, so it is READ OFF the permission list canon does: a role holding `org:admin` administers the whole organization, everything below it acts inside one org — conformance asserts the derivation rather than the current five answers. THIS ROW IS WHAT MAKES THE MEMBERS SURFACES RESOLVE: every RoleAssignment.role they publish carries a role_id this operation must answer, so from hello-16 on both read this one catalog and conformance asserts that every role a member holds is in it

refuse — a schedule exists only because somebody created one, and this universe accepts no writes (invariant #3) — nobody here scheduled a recurring Devin run

this org never adopted tagging, which is the same fact the already-ruled session `tags` rows carry (DECISIONS 2026-08-20 [devin])

refuse — an enterprise infrastructure/security administration read, and the sandbox token is an enterprise-member token rather than a superadmin one; Devin answers 403 (the ruled instance-admin precedent, coverage/reviews/2026-09-01-jira-modes.md)

refuse — CORRECTED 2026-09-12 (hello-18 serving wave W-c): the reason this row carried said the artifact carries no file data, and hello-15 made that FALSE - attachment.body holds real bytes for 35 rows and renderer-zendesk serves them. The refusal still stands, on two facts that ARE this row's own. FIRST, no uuid this route could resolve is ever advertised: no generation attaches an attachment to an agent session, .../sessions/{devin_id}/attachments answers [] on every session, so this organization has no attachment with that uuid and Devin's own 404 is the whole answer. SECOND, the SUCCESS answer the document declares is a 307 to an IP-enforcing proxy or a presigned S3 URL - a link into storage this replica has no equivalent of - so even with bytes to serve this row could not answer 307, which is the ruled avatar-bytes precedent (coverage/reviews/2026-09-01-jira-modes.md row 6) applied to a redirect rather than to a body

this enterprise configures no per-seat ACU cap; any figure printed here would be a placeholder copied off the spec example, which invariant #4 forbids (the ruled OpenAI rate_limits shape, DECISIONS 2026-08-20 [P31/openai])

refuse — an enterprise infrastructure/security administration read, and the sandbox token is an enterprise-member token rather than a superadmin one; Devin answers 403 (the ruled instance-admin precedent, coverage/reviews/2026-09-01-jira-modes.md)

refuse — an enterprise infrastructure/security administration read, and the sandbox token is an enterprise-member token rather than a superadmin one; Devin answers 403 (the ruled instance-admin precedent, coverage/reviews/2026-09-01-jira-modes.md)

SERVED 2026-09-12: QueueResponse MEASURED off the fleet rather than asserted - queue_size is the sessions this artifact left unfinished, read through the same status map /v3/enterprise/sessions publishes on every session, and it is 0 at `normal` because every Devin session in canon is terminal (hello-17: 65 completed, 2 failed, no running row). It goes non-zero on its own the day the generator leaves a session in flight

SERVED 2026-09-12 (hello-19 `ciBindings`): the org-level secrets this enterprise hands its CODING AGENT, off reader.orgCiBindings(org, 'coding-agent'). The 2026-09-01 reason asked for a new canon entity ActionsSecret, and hello-19's `ci_binding` IS that entity canonically — one named thing handed to something that runs, with one boolean saying whether its value may be published. THE `coding-agent` AUDIENCE AND NOTHING ELSE: Devin IS the coding agent, while `pipeline` is the build system's audience (GitLab's CI/CD variables, GitHub's Actions secrets) and `dependency-bot` is the updater's — every ci_binding accessor names a scope AND an audience precisely so a renderer has no call it could forget to filter, and conformance asserts that no other audience's key appears here. BOTH ROWS ARE SERVED RATHER THAN THE MASKED HALF, because SecretResponse requires `is_sensitive` as a boolean: Devin's store holds the agent's model key and the workspace it operates in alike, so is_sensitive IS the canonical `masked` column and both values appear. No value is published because SecretResponse declares no value field — the same shape canon holds for a masked binding, which is the argument the whole domain rests on. `secret_id` is minted from the canonical row id (the devinApiKeyId rule); `key` is the name the theme wrote; `note` is null because canon records none; `created_by` is the enterprise administrator's Devin user id, which /v3/enterprise/members/users/{user_id} resolves (invariant #5), and the list is empty rather than naming nobody on an artifact with no Devin seat; `created_at` is the binding's own instant; `secret_type` is key-value and `access_type` is org; `updated_at`/`updated_by` are null exactly when `updated_epoch` equals `created_epoch`, which is canon's own record of never rotated. Gated on hasCiBindings(); an older artifact answers the named hello-19 CI-binding gap, because [] there would say this enterprise hands its agent nothing while the same artifact publishes the sessions that agent ran

SERVED 2026-09-12 on hello-17: PatUserSelf for the founding member - the same `athena` GET /user and GitLab's currentUser resolve to. api_key_id and api_key_name read off a REAL person_credential row (kind personal-token, state active), whose title IS the key's name, and user_id is a Devin member id /v3/enterprise/members/users/{user_id} resolves. Both halves need hello-17 (callerCredentials for the token, adminSeat for the owner's Devin seat); an older generation is NAMED rather than answered with a minted key