@sporhq/spor 0.9.0 → 0.11.0

package/.claude-plugin/plugin.json

@@ -2,6 +2,6 @@
   "name": "spor",
   "displayName": "Spor Context Compiler",
   "description": "Maintains a typed, versioned knowledge graph and compiles compact briefings from it: session-start injection, per-prompt relevance digests, capture at discovery, end-of-session distillation, decision queue.",
-  "version": "0.9.0",
+  "version": "0.11.0",
   "author": { "name": "losthammer" }
 }

package/API.md

@@ -117,12 +117,21 @@ changes the description without a deploy.
 ### `add_edge`
 
 Micro-mutation. Input `{ "id": "<node>", "type": "<edge type>",
-"to": "<target>" }` — accepts canonical, alias, and inverse forms; inverse
-forms are flipped onto the target before writing. No revision echo is needed.
-Output `{ "status": "updated|skipped", "id": <node actually modified>,
-"revision", "warnings" }` (`skipped` = edge already present — the call is
-idempotent). Both nodes must exist. The tool description carries the same
-registry-generated vocabulary as `put_node`.
+"to": "<target>", "attrs"?: { "<k>": "<v>" } }` — accepts canonical, alias, and
+inverse forms; inverse forms are flipped onto the target before writing. No
+revision echo is needed. Output `{ "status": "updated|skipped",
+"id": <node actually modified>, "revision", "warnings" }` (`skipped` = edge
+already present — the call is idempotent). Both nodes must exist. The tool
+description carries the same registry-generated vocabulary as `put_node`.
+
+The optional `attrs` carries trailing flat edge attributes — the per-assignment
+`profile:` override on an `assigned → agent` edge is the motivating case. Values
+round-trip only simple `[A-Za-z0-9_-]` tokens (the frontmatter edge grammar);
+richer values need `put_node`, and `type`/`to` are reserved. With `attrs`, a
+duplicate `(type, to)` becomes an **upsert**: same attributes → still `skipped`;
+different attributes → the edge's attribute set is replaced (not merged) in
+place. Omitting `attrs` never touches an existing edge's attributes, so a bare
+`add_edge` is unchanged.
 
 ### `set_status`
 
@@ -227,11 +236,18 @@ Omitting both (or passing empty arrays) filters nothing.
 ### `ask_question`
 
 File a question the graph could not answer. Input `{ "text": "<the
-question>", "title"?: "<short title>", "mentions"?: ["<node id>", ...] }`
-(routing considers `mentions` first). The question becomes a durable node,
-deterministically routed to the steward of the closest relevant node
-(unrouted if none matches), and joins the decision queue until answered.
-Answer by writing a node with an `answers` edge to the question.
+question>", "title"?: "<short title>", "mentions"?: ["<node id>", ...],
+"project"?: "<slug>" }` (routing considers `mentions` first). The question
+becomes a durable node, deterministically routed to the steward of the closest
+relevant node (unrouted if none matches), and joins the decision queue until
+answered. Answer by writing a node with an `answers` edge to the question.
+
+By default the question's project is derived from its relevance neighborhood
+(its `mentions`, then the compiler's picks), falling back to the asker's home
+project. An explicit `project` overrides that derivation — pass it for a
+mention-less question whose neighborhood would otherwise yield nothing (a
+dispatched background agent injects its session project here, since the
+launcher environment never reaches it).
 
 ### `run_workflow`
 
@@ -301,9 +317,9 @@ endpoint is the REST twin of a core call:
 | `GET /v1/me` | `spor whoami`/`status`, onboarding | identity echo for the bearer token → `{person, name, email, bound, is_admin}`. `bound:false` means the token authenticates but maps to **no person node** (legacy/OAuth, or minted before the node existed), so routed questions and the personal queue will be empty — the client warns on it (the silent identity-degradation signal). `is_admin` reflects the `stewards→root` edge that gates the token-admin surface |
 | `GET /v1/briefing/{project}` | session-start | read the `brief-<project>` node → `{found, version, body, project_brief?, graph_status}`. The slug resolves through project-node aliases (GRAPH.md "Project identity nodes") before lookup. A BARE repo slug also rides up to its home-project grouping: the grouping's `brief-<grouping>` node returns alongside as `project_brief` (the product context spanning sibling repos), matching the shared up-resolution (dec-spor-queue-slug-resolves-to-grouping); passing the repo NODE id (`repo-<slug>`) is the escape hatch that returns only the repo brief, no `project_brief`. Optional `?fp=root:<sha>,remote:<host/path>,...` carries the repo's fingerprints: the server learns them onto the owning project node, and an unknown slug with a known fingerprint files an alias proposal in the queue |
 | `POST /v1/digest` `{query, root?, project?, min_sim?}` | prompt-context, /spor:brief | digest-mode compile → `{found, text}`; `found: false` is a successful empty result. `root` is the structural-walk twin of `query` (the two are mutually exclusive; `root` wins, an unknown id is `422`). Optional `project` is the session slug: the server scopes the compile to it — the same-project relevance boost, the grouping union, and the `always_on` norm `applies_to_*` ride-along — resolving the slug through project-node aliases/groupings inside compile (dec-spor-queue-slug-resolves-to-grouping), exactly as `/v1/queue` does. A bad slug is `422`; **omitting `project` runs the digest project-blind (byte-identical to before)**, so older clients that send only `{query}` are unaffected |
-| `GET /v1/nodes/{id}` | /spor:brief | `get_node` semantics; when a live inbound resolves/answers edge contradicts a still-open status the response carries `resolution`, and open gardener findings about the node ride along as `open_findings` |
+| `GET /v1/nodes/{id}` | /spor:brief | `get_node` semantics; a live inbound resolves/answers edge rides along as `resolution` (carrying the resolver's `summary`/`title` and a `lagging` flag — set when it contradicts a still-open status, clear when the node is already terminal, e.g. an answered question pointing at its answer), and open gardener findings about the node ride along as `open_findings` |
 | `POST /v1/nodes` | drain-outbox, mechanical writers | `put_node` semantics, batch: `{nodes: [...], if_exists: "skip"}` (entries may be raw strings or `{node, if_exists, revision}`) → `{results: [...]}`, 207 when any entry failed |
-| `POST /v1/nodes/{id}/edges` `{type, to}` | scripts, mechanical writers | `add_edge` semantics (§1): normalize/flip, dedupe, append — no revision echo |
+| `POST /v1/nodes/{id}/edges` `{type, to, attrs?}` | scripts, mechanical writers | `add_edge` semantics (§1): normalize/flip, dedupe, append — no revision echo. Optional `attrs` adds trailing flat edge attributes (e.g. a per-assignment `profile:` override); re-adding the same edge with different attrs upserts the set |
 | `POST /v1/nodes/{id}/status` `{status}` | scripts, mechanical writers | `set_status` semantics (§1): one-scalar update through the `transitions()` gate. Setting a work node to an in-progress status also CLAIMS it (same lease as `/claim` below) |
 | `POST /v1/nodes/{id}/claim` `{session?}` | `claim`/`set_status` MCP tools, `spor dispatch` | take the heartbeat-renewed lease (dec-cc-task-claim-lease): writes the durable `assigned` edge once, attributes to `$viewer` from the token (never an argument), and creates the ephemeral lease → `{ok, status, lease: {node_id, by, expires, expires_at, session, claimed_at}, edge}`. A live lease held by ANOTHER person is `409 conflict` naming the holder + expiry (re-claiming your OWN live claim just renews it). `session` scopes the heartbeat (omit to leave it person-scoped, so any of the claimer's sessions may renew — what `spor dispatch` does at the PRE-launch claim, since `claude --bg` self-allocates the run session only at launch; dispatch then renews with the real session once it has read it from `claude agents --json`, dec-spor-dispatch-bg-session-late-bind) |
 | `POST /v1/nodes/{id}/renew` `{session?}` | post-tool heartbeat, `renew` MCP tool, `spor dispatch` | bump the live lease's expiry only — no commit; the heartbeat that keeps a claim from lapsing. A lapsed/stolen lease is `409` (names the current holder). Person-scoped: any of the claimer's sessions may renew; a `session` binds the lease to that run (`spor dispatch` uses this to bind the captured `claude --bg` session post-launch) |
@@ -315,7 +331,7 @@ endpoint is the REST twin of a core call:
 | `POST /v1/distill/report` | distill | sweep telemetry, journal-only (no store mutation): `{facts, captured?, spooled?, rejected?, project?, session?}` → `{status: "reported"}`; zero-fact sweeps report too |
 | `POST /v1/corrections` | /spor:correct | `propose_correction` semantics → 201 `{status, id, revision, warnings}` |
 | `GET /v1/queue?project=&assignee=&type=&exclude_type=&limit=&offset=` | /spor:next, session-start | the ranked decision queue: `{items, count, offset, returned_count, total_count, truncated, next_offset, counts_by_type, counts_by_project, counts_by_suggest, muted?, dormant?, questions, asked, findings, pending, reviews, policy?, generated_at}` — items retired by a live resolves/answers edge are excluded; items hidden by the viewer's `queue_mute` or parked by a future `wake:` date (QUEUE.md §4) are counted, never silently dropped; `questions`/`findings`/`pending` are the routed-to-me-plus-unrouted views for the authenticated identity, `asked` is the questions you filed, and `reviews` is the nodes whose review is requested of you (an open `review-requested` edge to your person node — explicitly targeted, no unrouted fallback). `limit` is the page size (default 20, **max 100**, clamped not rejected) and `offset` skips that many items in the ranked order (default 0); the `counts_*`/`total_count` aggregates always cover the FULL ranked set regardless of the page, so one call answers "how many issues vs tasks" without paging, while `truncated`/`next_offset` let a client walk the rest by re-requesting with `offset=next_offset` until `next_offset` is null. Pagination is offset over a point-in-time ranked slice (the queue re-ranks every call), not a cursor — it resumes the same slice only across an unchanged ranking. `project` resolves through the shared up-resolution (dec-spor-queue-slug-resolves-to-grouping): a bare repo slug unions its home-project grouping's member queues, the repo NODE id (`repo-<slug>`) pins one repo, a grouping id (`proj-<slug>`) is used directly; **omitting `project` is the cross-project firehose** (every repo's queue at once). `assignee=<person-id>` scopes to the work that person carries (their `assigned`/`stewards` edges) — a manager's "who is carrying what"; `assignee=me` binds to the caller (empty if the token maps to no person node). `type=`/`exclude_type=` (comma-separated, repeatable) whitelist/blacklist node types from the ranking (exclude wins on overlap) — a hard scope filter applied before scoring, so the aggregates describe the filtered queue (task-cc-queue-filtering-enhancements) |
-| `POST /v1/questions` `{text, title?, mentions?}` | ask_question's REST twin | file a question node; deterministically routed to the steward of the closest relevance-neighborhood node, unrouted if none → 201 `{status, id, routed_to, via, asker, revision, warnings}` |
+| `POST /v1/questions` `{text, title?, mentions?, project?}` | ask_question's REST twin | file a question node; deterministically routed to the steward of the closest relevance-neighborhood node, unrouted if none → 201 `{status, id, project, routed_to, via, asker, revision, warnings}`. `project` is derived from the relevance neighborhood (then the asker's home project) unless an explicit `project` slug overrides it — pass that for a mention-less question (a dispatched agent injects its session project); a malformed slug → 400 |
 | `POST /v1/gardener` | ops cron / on demand | run a gardener sweep now; findings filed as queue items → `{filed, resolved, ..., generated_at}` |
 | `GET /v1/lens/{id}/render?format=html\|text\|json` | browsers, teammates without a checkout | run a lens OR workspace node and render its view tree (html default, plain text, or the raw tree as json). Read-only — no action forms; writes stay with `/v1/nodes` and the MCP tools. Auth is the caller's bearer header OR a signed read-only **render ticket** for shared links (browser links can't carry an Authorization header): `?ticket=<blob>` is accepted once and exchanged via a 302 for an HttpOnly `spor_render_ticket` cookie (kept out of URLs, logs, and view-to-view hrefs). The ticket binds `$viewer` to the recorded sharer and the render shows a "Viewing as &lt;sharer&gt;" banner. The former `?token=<PAT>` sharing path is **removed** — a shared link can never carry a write-capable credential |
 | `POST /v1/lens/{id}/ticket` `{expires?}` | sharing a view | mint a signed, expiring, read-only render ticket for the lens/workspace, recording the authenticated caller as the sharer → `{ticket, url, lens_id, sharer_person_id, exp}`. `expires` is `<N>d` or an ISO date (default `7d`, max `30d`); the caller must be bound to a person node (else `422 no_person`). The ticket carries no write scope and is honored only on the render route |
@@ -327,6 +343,10 @@ endpoint is the REST twin of a core call:
 | `POST /v1/admin/agents` `{label, owner?, id?, pubkey?}` | `spor agent create`, onboarding | create a person-owned `agent` node + its `owned-by` edge (`owner` defaults to the caller's person; `id` derives from `label`) → 201 `{id, owner, spiffe, pubkey, status, revision}`. 409 dup id / 422 invalid / 403 non-admin. Admin-only, same `stewards→root` gate as `/v1/admin/tokens` |
 | `POST /v1/agents/{id}/token` `{session?, audience?, expires?}` | `spor dispatch` | **self-serve** (NOT admin): mint a short-TTL, per-session token scoped to agent `{id}` → 201 `{token, expires_at, agent, session}` (`session: null` when deferred). Authorized iff the caller's person **owns** the agent (its `owned-by` edge) — else `403`; `404` unknown agent; `422` a SUPPLIED `session` that is malformed. `session` is now **OPTIONAL** (deferred binding, dec-spor-dispatch-bg-session-late-bind): `spor dispatch` mints it deferred because `claude --bg` self-allocates the run session only at launch, then binds the real one via `POST /v1/agents/session` below. The token carries `{agent, session?}` (the person is derived from the `owned-by` edge at verify time); a write under it is stamped agent-on-behalf-of-person (§1). A caller `expires` may only shorten the default TTL, never extend it |
 | `POST /v1/agents/session` `{session}` | `spor dispatch` (post-launch) | **late session binding** for a session-deferred agent token (dec-spor-dispatch-bg-session-late-bind). Authenticated by the **agent token itself** (the bearer hash identifies its own record — no agent id in the path, no ownership re-check), so only an agent-scoped token may call it (`403` otherwise). Sets that token's `session` → `200 {ok, agent, session}`. **Write-once**: idempotent on the same value (`{unchanged: true}`), `409 conflict` on a different one (a token's session is provenance, not a mutable field); `422` missing/malformed `session`. Every subsequent write under the token then stamps the bound session |
+| `POST /v1/agents/{id}/capabilities` `{harnesses?, reachable_mcp?, skills?, plugins?, deny?}` | `spor capabilities publish`, session-start auto-publish | **publish** this box's machine capabilities to the fleet scheduler (task-spor-remote-fleet-scheduler, dec-spor-machine-profile-satisfiability). The remote twin of the machine-local `dispatch.capabilities` map: the server collapses the body with the SAME `effectiveCapabilities()` the client runs (so a raw `{probed,declared,deny}` map or the already-flat axes both work; also accepts a `{capabilities: {...}}` envelope) and stores it BESIDE the agent node (operational store, not the durable git-tracked node — capabilities are machine-local, probe-refreshed, never committed) → `200 {agent, capabilities, published_at, last_seen, published_by, session?, changed}`. Authorized iff the caller **owns** the agent (its `owned-by` edge) OR is the agent itself (a self-publish under an agent token) — else `403`; `404` unknown agent; `422` a malformed map. A publish stamps both `published_at` (when the CAPS last changed) and `last_seen` (last contact); `last_seen` ALSO advances on the cheap `POST .../heartbeat` below, and the host-match keys staleness off `last_seen` not `published_at`. Beyond the manual verb, `session-start` AUTO-publishes here in remote mode whenever a `dispatch.agent` is configured (task-spor-fleet-capabilities-autopublish-session-start) — bounded + fail-open, so every session refreshes this box's caps and last-contact without a manual call; disable with `SPOR_CAPABILITIES_PUBLISH=0` |
+| `GET /v1/agents/{id}/capabilities` | steward fleet view, debugging | read back an agent's published capabilities → `200 {agent, capabilities, published_at, last_seen, published_by, session?}`; `404` if none published. Readable by the **owner**, the **agent itself**, or an **admin** (a stewards→root fleet-capacity view) — else `403` |
+| `POST /v1/agents/{id}/heartbeat` | post-tool mid-session liveness tick | **liveness ping** (task-spor-fleet-scheduler-hardening): refresh this box's `last_seen` WITHOUT re-uploading capabilities — the cheap "still here" signal, decoupled from a caps re-publish, so a box that published once and runs for hours stays a live fleet host. The host-match keys staleness off `last_seen`, so a box that keeps heartbeating is never demoted while a genuinely dead one ages out under `max_age` → `200 {agent, capabilities, published_at, last_seen, …}` (the refreshed record). Same owner/self gate as publish — else `403`; `404` unknown agent OR nothing published yet (publish before heartbeat — liveness without caps is meaningless to the scheduler); `422` a malformed agent id. The CLIENT caller (task-spor-fleet-scheduler-client-heartbeat-tick): the `post-tool` hook ticks this in REMOTE mode whenever a `dispatch.agent` is configured (the SAME opt-in as the session-start auto-publish), piggybacking on write-activity but THROTTLED to one ping per `dispatch.heartbeatIntervalMs` (default 5min) — so a long session keeps `last_seen` fresh between session-starts (which today refresh it the expensive way, via a full re-publish) without re-probing. Bounded + fail-open; disable with `SPOR_HEARTBEAT=0` |
+| `GET /v1/profiles/{id}/hosts` `?owner=me\|person-X&max_age=<dur>` | `spor capabilities hosts <profile>`; `spor dispatch` (auto on a FORK B refusal) | **host-match** a `type: profile` against every agent's published capabilities using the SAME pure `satisfies()` matcher the client runs locally → `200 {profile, satisfiable: [{agent, owner, published_at, last_seen, age_seconds}], unsatisfiable: [{agent, owner, published_at, last_seen, age_seconds, reasons}], counts}`. Satisfiable hosts are freshest-first (by `last_seen`); the unsatisfiable carry the matcher's own reasons (the failing atoms), enabling **substitution-free re-routing** — pick a box that satisfies the profile, NEVER substitute a different one (dec-spor-machine-profile-satisfiability FORK B). The CLIENT consumer (task-spor-fleet-scheduler-autoroute-dispatch): `spor capabilities hosts` lists the re-route targets directly, and when `spor dispatch` refuses because THIS box can't satisfy the resolved profile it calls this endpoint and names the satisfiable hosts to re-route to — or, when none satisfy it, escalates to the owner (fail-soft: an unreachable scheduler degrades to a generic hint). **Visibility is steward-scoped** (task-spor-fleet-scheduler-hardening): the whole-fleet view (every member's boxes + caps) is a multi-tenant cross-member disclosure, so an **admin** (stewards→root) sees the whole fleet and may scope to any `owner=person-X`, while an ordinary **member** is scoped to THEIR OWN boxes (default `owner` = the caller's person; an agent token resolves to its owner; `owner=me` is the explicit form) and a member asking for a colleague's `owner=person-X` is `403`. `max_age` (`30m`/`12h`/`7d`/ms) demotes hosts whose `last_seen` is older than it to unsatisfiable (the liveness filter). `404` unknown/non-profile id; `422` bad `max_age`/`owner` |
 
 Path parameters (node ids, project slugs) must match
 `^[a-z0-9][a-z0-9-]*$`. Request bodies are capped at 1MB

package/GRAPH.md

@@ -58,7 +58,10 @@ Rules:
   omit them.
 - Edges may point at ids that don't exist yet; the compiler skips them. Don't
   delete an edge just because the target is missing — it marks a node worth
-  creating.
+  creating. An edge may also carry extra flat attributes after `to:` —
+  `- {type: assigned, to: agent-X, profile: profile-Y}` — preserved on the edge
+  object (the per-assignment profile override; see "The agent orchestration
+  layer"). Plain `{type, to}` edges are unchanged.
 - `commits` is an optional inline list of repo-qualified git shas
   (`commits: [wf@1a2b3c4d, ...]`, kebab-case repo slug + 7–40 hex) linking
   the node to the code commits that implement it (task-cc-commit-linking).
@@ -95,6 +98,8 @@ Rules:
 | question   | `question-` | a routed ask the graph could not answer (queueable; status `open`/`answered`, gated) |
 | person     | `person-` | a member of the org — the identity anchor for `$viewer` binding and Tier-2 question routing (team mode; see "People, routing, and onboarding") |
 | agent      | `agent-`  | a person-owned automation principal — a dispatched session's durable identity, owned by a person via an `owned-by` edge; its writes attribute "agent on behalf of person" (see "Agents") |
+| profile    | `profile-`| a reusable runtime+capability bundle an agent runs under: `harness`, `model`, `skills`/`plugins`/`mcp`. Its runtime fields ARE the dispatch satisfiability spec; `capturable: false` (see "The agent orchestration layer") |
+| routine    | `routine-`| owner-scoped trigger→action automation (`owned-by` a person): declarative `when → do` rules over graph events that dispatch only the owner's agents, AND-ed with org policy; `capturable: false` (see "The agent orchestration layer") |
 | capture-pending | `cap-` | raw captured text that fit no schema; filed by the server for later triage (QUEUE.md §2.3); born status-less, closed only as `merged` (content now in proper node(s)) or `rejected` (no durable fact) — a `transitions()` gate rejects other statuses at write time |
 | finding    | `find-`   | a gardener observation about another node, filed as a queue item (QUEUE.md §6) |
 | repo       | `repo-`   | durable git-repo identity: slug aliases + repo fingerprints; heals renames at read time (below) |
@@ -606,6 +611,118 @@ edges:
   mirroring `person`, `repo`, and `workflow-run`. (So the distiller's emit
   vocabulary deliberately omits `agent`/`owned-by`.)
 
+## The agent orchestration layer
+
+The layer ABOVE agent identity (dec-spor-agent-orchestration-layer): how work is
+routed to agents, and how per-person automation fires on graph events. The node
+model is `person ──owns──▶ agent ──uses-profile──▶ profile`, with owner-scoped
+`routine` nodes driving automation. It composes with — never bypasses — the org
+policy layer. The schemas ship in the seed pack, and so do the machine-local
+`dispatch.capabilities` map + `satisfies()` matcher + fail-soft dispatch
+(task-spor-dispatch-capabilities-satisfiability); the routine engine and the
+remote fleet scheduler are still deferred.
+
+### profile — the reusable runtime+capability bundle
+
+A `profile` node (prefix `profile-`) is "the HOW" an agent dispatches under,
+factored out of the agent node so a toolset is reusable across agents and people.
+
+```markdown
+---
+id: profile-docs-writer
+type: profile
+title: Docs-writer profile
+summary: Claude-code on Opus with the writing + spor skills and the spor MCP server.
+harness: claude-code
+model: opus
+skills: [writing, brief]
+plugins: [spor]
+mcp: [spor]
+status: active
+date: 2026-06-18
+---
+```
+
+- `harness:` (`claude-code` | `codex` | `opencode` | …) selects the launcher
+  (dec-cc-portable-core-adapters: claude-code → `claude --bg`, others → their
+  CLIs); `model:` → the harness `--model`; `skills`/`plugins` are preloaded;
+  `mcp` is merged into the strict `--mcp-config` dispatch writes, so the agent's
+  toolset is exactly the profile plus the agent-spor server, nothing ambient
+  (dec-spor-session-identity-active-record).
+- **The runtime fields ARE the satisfiability spec** — there is no separate
+  requirements block (dec-spor-machine-profile-satisfiability). A machine
+  declares ATOMIC capabilities in a machine-local `dispatch.capabilities` map
+  (built like `dispatch.repos`, never committed), and `satisfies(machine,
+  profile)` checks `profile.harness ∈ machine.harnesses ∧ profile.mcp ⊆
+  machine.reachable_mcp ∧ profile.skills ⊆ machine.skills ∧ profile.plugins ⊆
+  machine.plugins ∧ profile ∉ machine.deny`. No satisfying machine → dispatch
+  fails soft and LOUD, leaves the assignment intact, NEVER substitutes a
+  different profile. The probe seeds `reachable_mcp: [spor]` from CONFIGURED-ness
+  (a bound Spor server/connector, remote mode) rather than a network ping — the
+  agent-spor server is part of every dispatched session's toolset by construction
+  (above), so an `mcp: [spor]` profile satisfies on a fresh box with no manual
+  `allow-mcp` (task-spor-mcp-reachability-deterministic-seed). Forward-compatible
+  with the deferred remote fleet scheduler (each agent publishes its capabilities;
+  same vocabulary).
+- **Reusable + both-scoped, with override.** Profiles are PERSONAL and
+  ORG-PUBLISHED (a curated, vetted toolset), with personal override. Org-published
+  profiles are where this meets policy: a policy can require that work of a risk
+  class go to an agent whose profile is org-approved (curated-toolset-as-governance).
+- `status:` is declarative (`active` default); `capturable: false` (created
+  deliberately, never distilled).
+
+### routine — owner-scoped trigger→action automation
+
+A `routine` node (prefix `routine-`), `owned-by` a person, holds declarative
+`when → do` rules. Triggers are graph events (a status change — e.g. to a
+resolving state, or to `done` — or an edge like `reviewed-by`/
+`changes-requested-by`) with `where:` filters; actions are the bounded verb set
+`create-node` / `assign` (to a SPECIFIC agent, optionally with a `profile:`
+override) / `dispatch` / `set-status` / `reassign` / `escalate`. Because the
+regex frontmatter parser is flat, an instance carries its `rules` as a fenced
+` ```json ` block in its BODY, parsed by the deferred routine engine.
+
+Two invariants (dec-spor-agent-orchestration-layer): (1) **only the owner's
+agents are ever dispatched** — the `person → routine → agent` RFC 8693 act-chain
+is the audit trail; (2) **personal routines accelerate, org policy gates — they
+AND, never bypass** — agent-on-behalf-of-X counts as X for the self-approval ban
+and the definition-of-done quorum, so an owner's reviewer-agent cannot
+auto-approve their implementer-agent's work past the org bar. Per
+dec-spor-orchestration-routine-requires-threads (thread 1) the bounded
+declarative register ships FIRST; attached sandboxed code is a later
+schema-gated escape hatch. The routine ENGINE and its vetting (dry-run, mandatory
+retry budget, self-approval-floor activation review, fire-time org-policy gate —
+thread 2) are deferred, so this seed schema is declarative and UNGATED;
+`capturable: false`.
+
+### Routing is explicit assignment; the profile cascade
+
+Routing is EXPLICIT `assigned` to a SPECIFIC target, not an eligibility match:
+`assigned → person` is human work (the default human pool); `assigned → agent-X`
+makes the task a dispatch candidate (X's profile says how and what it may touch).
+An agent-targeted `assigned` edge may carry an optional `profile:` attribute —
+the durable per-assignment override (`{type: assigned, to: agent-X, profile:
+profile-Y}`, thread 3). Profile precedence, explicit wins: `--profile` dispatch
+flag > assignment-edge `profile:` attribute > the agent's default `uses-profile`
+edge.
+
+### The `requires:` risk-class register
+
+`requires:` is a flat list on a WORK node naming the risk/permission classes the
+work may touch (`requires: [shell, prod-creds]`). It is a **registry-declared
+extensible enum** — a `type: schema` node with `kind: register` and `register:
+requires` (seed: `schema-requires`), so an org grows the vocabulary by editing a
+schema node, never a code change. The kernel exposes it as a partition
+(`graph.registry.requiresClasses()`); it stays policy-free. The register is
+**DISTINCT from machine-satisfiability**: satisfiability asks "can this box LAUNCH
+the profile"; `requires:` asks "what may this work touch", validated against the
+assigned profile (a task's `requires:` must be ⊆ the profile's granted classes,
+else warn/refuse) and gated by org policy via the same governs-traversal the
+definition-of-done quorum uses (dec-spor-orchestration-routine-requires-threads
+thread 4). The seed set is small — `shell`, `prod-creds`, `browser`, `network`,
+`human`, `filesystem-write`, `paid-api`. `human` is unsatisfiable by any agent:
+assign that work to a person.
+
 ## Edge types and traversal weights
 
 | edge             | weight | meaning                                          |
@@ -618,7 +735,7 @@ edges:
 | `resolves`       | 0.9    | this node fixes/closes the target                |
 | `blocks`         | 0.7    | target cannot proceed until this node does       |
 | `answers`        | 0.7    | this node answers that question (inverse `answered-by`); pulls the answer through the asker's next compile |
-| `assigned`       | 0.5    | work is assigned to this person                  |
+| `assigned`       | 0.5    | work is assigned to this person OR agent (the explicit-routing edge; an agent target may carry a `profile:` per-assignment override) |
 | `reviewed-by`    | 0.5    | this person reviewed and approved the node — counts toward a policy quorum |
 | `changes-requested-by` | 0.5 | this person reviewed the node and requested changes — not an approval |
 | `relates-to`     | 0.5    | weak association                                 |
@@ -626,6 +743,7 @@ edges:
 | `stewards`       | 0.4    | this person stewards an area/spec/norm — the Tier-2 question-routing key |
 | `grouped-under`  | 0.3    | this repo's home project grouping (inverse `groups`); structural membership, not work dependency |
 | `owned-by`       | 0.3    | this agent is owned by that person (inverse `owns`); structural identity binding, not work dependency |
+| `uses-profile`   | 0.3    | this agent's default profile (the runtime+capability bundle it dispatches under); structural config binding, overridable per assignment/dispatch |
 | `routed-to`      | 0.3    | a question routed to this person for answering   |
 | `review-requested` | 0.3  | a review of this node is requested of this person (pending) — surfaces in their queue |
 | `compiled-for`   | —      | briefing → its task/query (provenance only)      |

package/QUEUE.md

@@ -142,6 +142,15 @@ is the default and graph-resident schema nodes override/extend it. The
 `schema` node type itself is recognized natively by the core — no
 schema-for-schemas regress.
 
+Beyond `node-schema`/`edge-schema`, a schema node may carry a non-type `kind`:
+`policy` (the org-defined policy layer, §below), `queue-policy` (the ranking
+override singleton), and `register` — a registry-declared extensible **enum**
+the kernel exposes as a partition (`graph.registry.register(name)` /
+`requiresClasses()`) instead of a hardcoded table. The seed `requires` register
+(`schema-requires`, the work-node risk/permission axis) is the first; an org
+grows the vocabulary by writing a resident `kind: register` schema with the same
+`register:` name (graph beats seed, higher CalVer wins — like every other kind).
+
 Resolution is **graph beats seed wholesale, regardless of version**: a
 resident override replaces the seed entry for that type entirely, so a seed
 behavior change (a new `transitions()` gate, a default, a prefix) does *not*