Foxfire1st · Foxfire1st · Jul 12, 2026 · Jul 12, 2026 · Jul 12, 2026 · Jul 12, 2026
diff --git a/.agents/skills/l-01-agent-lifecycles/SKILL.md b/.agents/skills/l-01-agent-lifecycles/SKILL.md
@@ -281,6 +281,25 @@ For ordinary spawned seats, settings are the sole developer-controlled spend sur
 `spawn_agent_session` callers declare role and level, never harness/model/effort or direct
 launch/session spend controls.
 
+### Hosted role dispatch is three explicit states
+
+Every role that dispatches another hosted role uses one exact session id through all three states:
+
+1. Call `spawn_agent_session` with `context` omitted and `submit=false`. Success is
+   `spawned-unbriefed`; it creates and binds the seat but assigns no work.
+2. Call `hosted_session_readiness(session_id=<returned-id>, wait_seconds=<finite-bound>)`.
+   Only `status=ready` advances that exact seat to harness-ready. Spawned-only and not-ready seats
+   are **not active work** and receive no task instructions.
+3. Post exactly one durable `operator_inbox_post` row addressed by that same `agent_id`, with
+   `message_kind="dispatch-brief"` and `deliver_to_hosted=true`. Treat the seat as briefed only when
+   that row reports both `deliveryState=delivered` and
+   `deliveryDetail=harness-log-confirmed`.
+
+If step 3 fails, the original durable row stays pending on the original spawned session for the
+standard injector/supervisor retry path. Never duplicate the row, append another visible draft,
+or respawn merely because delivery is pending. Settings-owned `sessionCommands` remain launch
+configuration; settings-owned `promptKeywords` ride the post-readiness dispatch brief exactly once.
+
 ## settings.json Orchestration Block
 
 Machine/user overrides layer over the role-file defaults, in the **global agentic settings file**
@@ -337,7 +356,8 @@ per-harness via the effective registry (claude `--model`/`--effort`; a mapping-l
 env-only; a session-vocabulary effort like claude's `ultracode` is delivered as a post-launch
 `/effort` paste). Unknown effort values REFUSE at dispatch naming the harness's vocabulary — the
 CLI would warn-and-silently-degrade. The free-form escape hatch (`launchArgs` verbatim argv,
-`promptKeywords` riding the brief paste, `sessionCommands` pasted before the brief) is never
+`promptKeywords` riding the post-readiness brief exactly once, `sessionCommands` applied during
+fresh-session launch) is never
 validated, only recorded in spawn provenance; `orchestration.harnesses` teaches the framework new
 TUIs or pre-customizes builtin launches (manual: `docs/reference/harnesses.md`).
 `orchestration.loops` (the three-party-loop knobs: per-level loop sets, round cap, reviewer

diff --git a/.agents/skills/l-01-agent-lifecycles/roles/architect.md b/.agents/skills/l-01-agent-lifecycles/roles/architect.md
@@ -110,6 +110,16 @@ host harness rules.
 Hat-collapse is allowed here because this is the owner/developer-facing seat. The same collapse is
 not allowed in spawned role seats.
 
+## Hosted Role Dispatch
+
+Every horizontal expansion from this seat follows the shared three-state protocol in `../SKILL.md`:
+`spawn_agent_session(context omitted, submit=false)` must return `spawned-unbriefed`; then
+`hosted_session_readiness` must return `status=ready` for the exact returned session id; only then
+post one exact-agent durable `dispatch-brief`. A spawned-only or not-ready orchestrator is not
+active work. Count it briefed only from `deliveryState=delivered` plus
+`deliveryDetail=harness-log-confirmed`; a failed delivery stays pending on the same row and session,
+never a duplicate brief or automatic respawn.
+
 ## Design And Drawing Board
 
 When the developer is still shaping the work, the architect wears `roles/designer.md` inline:
@@ -247,8 +257,8 @@ developer or the configured distinct decider; the architect does not approve its
 | model   | highest-reasoning | developer-facing architecture and ruling quality need the strongest model |
 | effort  | high              | decision framing is not the place to economize |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste (settings-only; never validated) |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | developer-facing owner surface | `read_ar_files` · onboarding · route indexes · `task_doc` · inbox · gates for developer hand-offs · `spawn_agent_session` |
 
 Settings.json `orchestration.roles.architect` overrides these, and `orchestration.rolesPerLevel.<level>.architect` overrides per dispatch level (role-file defaults < settings < level override; spawn knobs manual: `docs/reference/harnesses.md`).
diff --git a/.agents/skills/l-01-agent-lifecycles/roles/curator.md b/.agents/skills/l-01-agent-lifecycles/roles/curator.md
@@ -100,8 +100,8 @@ verdict.
 | model   | mid-reasoning  | precise onboarding edits and reference checking |
 | effort  | medium         | scales with onboarding blast radius via settings |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste (settings-only; never validated) |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | onboarding surface | native reads/edits in memory worktree · native reads in code worktree · c-05 workflow · local route indexes · shell checks · inbox |
 
 Settings.json `orchestration.roles.curator` overrides these, and `orchestration.rolesPerLevel.<level>.curator` overrides per dispatch level (role-file defaults < settings < level override; spawn knobs manual: `docs/reference/harnesses.md`).
diff --git a/.agents/skills/l-01-agent-lifecycles/roles/designer.md b/.agents/skills/l-01-agent-lifecycles/roles/designer.md
@@ -91,8 +91,8 @@ absorbs architect, orchestrator, manager, worker, strategist, or reviewer work.
 | model   | high-reasoning     | reframe + blast-radius reasoning wants a strong model                 |
 | effort  | high               | design leverage justifies the thinking budget                        |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste (settings-only; never validated) |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | bird's-eye toolkit | route indexes · onboarding · `grepai_search` · `cgc_*` · `read_ar_files` · `task_doc` · inbox |
 
 Settings.json `orchestration.roles.designer` overrides these, and `orchestration.rolesPerLevel.<level>.designer` overrides per dispatch level (role-file defaults < settings < level override; spawn knobs manual: `docs/reference/harnesses.md`).
diff --git a/.agents/skills/l-01-agent-lifecycles/roles/manager.md b/.agents/skills/l-01-agent-lifecycles/roles/manager.md
@@ -37,6 +37,16 @@ Roles expand horizontally into new chats; sub-agents drill vertically inside thi
 bounded analysis or report checks. A spawned manager never absorbs architect, orchestrator,
 strategist, reviewer, curator, or worker briefs.
 
+## Hosted Role Dispatch
+
+Every worker, reviewer, or curator dispatch below uses the shared three-state protocol in
+`../SKILL.md`: call `spawn_agent_session` with `context` omitted and `submit=false`, retain the exact
+`spawned-unbriefed` session id, and require `hosted_session_readiness(...)=status=ready` for that
+same id before posting one exact-agent durable `dispatch-brief`. Spawned-only and not-ready seats
+are not active work. Briefed means both `deliveryState=delivered` and
+`deliveryDetail=harness-log-confirmed`. If delivery fails, keep the original row and session pending
+for standard retry; never duplicate the brief or respawn merely because delivery is pending.
+
 ## Lens
 
 - **Opening move:** on a developer-declared takeover, first run `../SKILL.md`'s
@@ -99,8 +109,9 @@ stops belong to the orchestrator via the system-specialist protocol.
   fix rounds resume the SAME builder); **every round must shrink the finding set** — a
   non-shrinking round escalates to the orchestrator immediately, with the full round history
   attached, regardless of the count.
-- `spawn_agent_session(worker)` — a **fresh session** on the leaf: the brief (compiled from
-  `../templates/worker-brief.md`) is pasted + submitted, with `env={"AR_SPAWN_ROLE": "worker"}` and
+- `spawn_agent_session(worker)` — a **fresh session** on the leaf: after exact-session readiness,
+  one durable `dispatch-brief` compiled from `../templates/worker-brief.md` is delivered and proven,
+  with `env={"AR_SPAWN_ROLE": "worker"}` and
   the **qualified** leaf key `<repository>/<master>/<docId>`; together they claim the worker's
   `(leaf, role)` seat, and the worker edits inside the leaf worktrees the brief names.
 - **Process and ack the worker's signals — passive contract.** A turn-report artifact is expected at
@@ -224,8 +235,8 @@ own lifecycle if you need its state).
 | model   | mid-reasoning  | leaf review + coordination; strong but below the orchestrator    |
 | effort  | medium         | one master's scope, not the portfolio                            |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste (settings-only; never validated) |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | coordination + review + leaf lifecycle | `task_doc` · `read_ar_files` · gates · `spawn_agent_session` · `session_retire` (your own master's worker/reviewer/curator seats only) · worktree lifecycle (start · closeout · integrate · finalize) · C-11/`c-09` · inbox |
 
 Settings.json `orchestration.roles.manager` overrides these, and `orchestration.rolesPerLevel.<level>.manager` overrides per dispatch level (role-file defaults < settings < level override; spawn knobs manual: `docs/reference/harnesses.md`).
diff --git a/.agents/skills/l-01-agent-lifecycles/roles/orchestrator.md b/.agents/skills/l-01-agent-lifecycles/roles/orchestrator.md
@@ -27,6 +27,16 @@ architect or owning seat via the inbox. Roles expand horizontally into new chats
 vertically inside this seat for bounded analysis. A spawned orchestrator never absorbs another
 role brief and never performs architect/developer-facing hat-collapse.
 
+## Hosted Role Dispatch
+
+Every manager, strategist, reviewer, or system-specialist dispatch below means the complete shared
+three-state protocol in `../SKILL.md`: spawn with `context` omitted and `submit=false` and retain the
+exact `spawned-unbriefed` session id; require `hosted_session_readiness(...)=status=ready` for that
+same id; then create one exact-agent durable `dispatch-brief`. Spawned-only and not-ready seats are
+not active work. Briefed means both `deliveryState=delivered` and
+`deliveryDetail=harness-log-confirmed`. A delivery failure remains pending on its original row and
+session for standard retry; never duplicate its brief or respawn it merely for pending delivery.
+
 ## The Event Loop
 
 **Opening move, every session — new or resumed** (resumption is the common case, not the
@@ -56,8 +66,9 @@ exception):
 
 **Profile check (takeover).** Before heavy work in any job: if this session's harness/model/
 effort is wrong for the run (resolved: role file < settings), spawn the right chair —
-`spawn_agent_session` with `AR_SPAWN_ROLE=orchestrator` + a conversation-handover packet
-(`../templates/conversation-handover-packet.md`) — and hand over; the architect still talks to the
+`spawn_agent_session` with `AR_SPAWN_ROLE=orchestrator`, then deliver a conversation-handover packet
+(`../templates/conversation-handover-packet.md`) as the post-readiness `dispatch-brief` — and hand
+over; the architect still talks to the
 developer, and backend orchestrator seats stay behind the relay.
 
 Several jobs can be active across a day; the loop routes per event. The frame's phase axis stays
@@ -234,7 +245,8 @@ acceptable and recorded in durable notes.)
 Dispatch independent ready masters in parallel by default up to
 `orchestration.concurrency.maxParallelMasters`. Sequential execution is the exception and must
 name a gate, a shared-file one-writer dependency, or an explicit ruling. For each ready master
-(dependencies integrated into super), call `spawn_agent_session(manager)` with a brief compiled from
+(dependencies integrated into super), run the three-state hosted-role dispatch for
+`spawn_agent_session(manager)`, compiling its post-readiness `dispatch-brief` from
 `../templates/manager-brief.md` (`env={"AR_SPAWN_ROLE": "manager"}`, the **qualified** leaf key
 `<repository>/<master>/<docId>`; together the environment role and qualified leaf claim the
 manager's `(leaf, role)` seat; the brief carries the load-bearing base fact: master branches off
@@ -444,8 +456,8 @@ task, fill small blanks, escalate real deltas).
 | model   | highest-reasoning | portfolio blast-radius judgment wants the strongest model |
 | effort  | high              | the bird's-eye seat; not the place to economize |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste (settings-only; never validated) |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | full bird's-eye + orchestration | route indexes · onboarding · `grepai_search` · `cgc_*` · `read_ar_files` · `task_doc` · gates · `spawn_agent_session` · `session_retire` (any seat, portfolio-wide) · worktree/C-11 |
 
 Settings.json `orchestration.roles.orchestrator` overrides these, and `orchestration.rolesPerLevel.<level>.orchestrator` overrides per dispatch level (role-file defaults < settings < level override; spawn knobs manual: `docs/reference/harnesses.md`).
diff --git a/.agents/skills/l-01-agent-lifecycles/roles/reviewer.md b/.agents/skills/l-01-agent-lifecycles/roles/reviewer.md
@@ -180,8 +180,8 @@ integrated on super.
 | model   | high-reasoning   | adversarial review wants a strong, skeptical model           |
 | effort  | high             | the last line of defense before a handover; do not economize |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste (settings-only; never validated) |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | review surface   | `read_ar_files` · `memory_quality_check` · `drift_check` · `grepai_search` · `cgc_*` · `system/tools.md` checks · report templates · inbox |
 
 Settings.json `orchestration.roles.reviewer` overrides these, and `orchestration.rolesPerLevel.<level>.reviewer` overrides per dispatch level (role-file defaults < settings < level override; spawn knobs manual: `docs/reference/harnesses.md`).
diff --git a/.agents/skills/l-01-agent-lifecycles/roles/strategist.md b/.agents/skills/l-01-agent-lifecycles/roles/strategist.md
@@ -197,8 +197,8 @@ Everything else — `task_doc`, `worktree_*`, `lifecycle_*`, `gate_*`, `spawn_ag
 | model   | highest-reasoning | whole-portfolio dependency + blast-radius reasoning wants the strongest model |
 | effort  | high              | the sprint plan parameterizes every downstream loop; not the place to economize |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste (settings-only; never validated) |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | read-only analysis surface | `read_ar_files` · `grepai_*` · `cgc_*` · `context_packet` · `drift_check` · notes-draft write · inbox |
 
 Settings.json `orchestration.roles.strategist` overrides these, and `orchestration.rolesPerLevel.<level>.strategist` overrides per dispatch level (role-file defaults < settings < level override; spawn knobs manual: `docs/reference/harnesses.md`).
diff --git a/.agents/skills/l-01-agent-lifecycles/roles/system-specialist.md b/.agents/skills/l-01-agent-lifecycles/roles/system-specialist.md
@@ -93,8 +93,8 @@ The orchestrator owns the final decision: fixable-in-session -> order a targeted
 | model   | fable   | system diagnosis and report synthesis |
 | effort  | high    | degradation triage is high-impact |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | provider diagnostics + native reads + inbox | provider_status · provider_diagnostics · provider_watchers when explicitly ordered · logs/metrics reads · inbox |
 
 Settings.json `orchestration.roles.system-specialist` overrides these, and

diff --git a/.agents/skills/l-01-agent-lifecycles/roles/worker.md b/.agents/skills/l-01-agent-lifecycles/roles/worker.md
@@ -147,8 +147,8 @@ leaf well, ask when the leaf itself is in question" default.
 | model   | mid-reasoning  | competent implementer on a scoped leaf |
 | effort  | medium         | scales with leaf difficulty via settings |
 | launchArgs | — | free-form escape: verbatim harness argv (settings-only; never validated, recorded in spawn provenance) |
-| sessionCommands | — | free-form escape: lines pasted + submitted into the fresh session before the brief (settings-only; never validated) |
-| promptKeywords | — | free-form escape: prepended as the first line of the dispatch brief paste (settings-only; never validated) |
+| sessionCommands | — | settings-owned launch configuration: lines pasted + submitted during fresh-session launch (never validated; not brief delivery) |
+| promptKeywords | — | settings-owned keywords prepended exactly once to the post-readiness dispatch brief (never validated) |
 | tools   | build surface  | native edit · read-only AR retrieval · prescribed checks · inbox |
 
 Settings.json `orchestration.roles.worker` overrides these, and `orchestration.rolesPerLevel.<level>.worker` overrides per dispatch level (role-file defaults < settings < level override; spawn knobs manual: `docs/reference/harnesses.md`).