host-runtime.md

Host runtime contract

The Bureau separates two decisions:

• runtime routing chooses provider model IDs and reasoning effort in RUN_DIR/model-routing.json;
• agent host transport creates, resumes, and waits for agent contexts.

The first-class pairings are:

model-routing.json#runtime	Agent host	Status
claude	Claude Code Agent tool	supported
openai	Codex collaboration tools	supported
codex	alias for openai at startup/reviewer boundaries	supported
grok	Grok Bot Task/executor (GROK.md)	supported on Grok Bot; run-start.sh --runtime grok
openrouter, hermes	no native transport adapter yet	routing-only; fail closed for a run

Do not infer host behavior from a model name. Read runtime once, select the matching column below, and keep that transport for the run.

Spawn, resume, wait, and human forks

Operation	Claude Code host	Codex host	Grok Bot host
Fresh specialist	Agent tool, subagent_type: general-purpose, explicit model	Codex multi-agent spawn, currently exposed as multi_agent_v1.spawn_agent, fork_context: false, explicit model and reasoning_effort	Task executor, blank prompt, no parent transcript; log --plan payload first
Qualified fast Mage pass	normal fresh specialist	one-shot scripts/run-codex-spark-specialist.sh; Spark/high; execute-plan only; no resume	role default; no grok-build helper yet
Resumable Conductor	Agent tool; retain returned agent id	Codex multi-agent spawn, currently exposed as multi_agent_v1.spawn_agent; retain returned agent id	Task executor; retain agent id in delegate-state.json
Resume an idle agent	SendMessage to retained id	multi_agent_v1.send_input to retained id	Task resume or MessageSubagent
Pass a note to a running agent	SendMessage	multi_agent_v1.send_input	MessageSubagent
Wait for liveness/completion	host Agent/SendMessage result	multi_agent_v1.wait_agent	background completion; CheckSubagent if stuck
Genuine human fork	top-level AskUserQuestion	Delegate persists state and asks Robin in its top-level final response; after Robin replies, multi_agent_v1.send_input resumes the Conductor	Delegate persists, asks Robin in this chat, then resumes the Conductor

Codex fresh-context spawns MUST pass no parent transcript. In the current tool surface that means fork_context: false; in older/future surfaces this may be called fork_turns: "none". Passing the parent conversation (fork_context: true, fork_turns: "all", or equivalent) defeats the Bureau's cold-review and specialist-isolation design.

Every native spawn uses RUN_DIR/model-routing.json#roles.<role>:

• pass model explicitly;
• on Codex, also pass reasoningEffort as reasoning_effort;
• never silently inherit the manager's model;
• keep the role's existing first-message identity, input, workflow, attempt, and nonce lines unchanged.

Codex Spark execution profile

The Codex host's native collaboration endpoint currently accepts the configured Terra/Sol child-agent set but rejects Spark as an unknown spawn model. The Codex CLI can start Spark, so the Bureau exposes it as a separate, explicit execution profile rather than weakening the native spawn allowlist:

• policy: model-policy.v2.json#execution_profiles.granular-ui-fast;
• resolved route: model-routing.json#roles.mage.executionProfiles.granular-ui-fast;
• transport: scripts/run-codex-spark-specialist.sh → ephemeral, one-shot codex exec;
• scope: first attempt at a vetted execute-plan Mage prompt for one existing component/style boundary, text-only inputs, and a known local validation command;
• fallback: the resolved Mage role default (normally Sol/high), only when the helper proves HEAD and the worktree are unchanged.

The prompt-folder tag Execution-profile: granular-ui-fast is necessary but not sufficient: the Conductor and helper recheck policy eligibility at dispatch. New features, architecture, state, API/navigation/contracts, dependencies/generated files, cross-coder work, sensitive/external effects, visual/image judgment, release steps, retries, and review fixes stay role-default. Spark never enters multi_agent_v1.spawn_agent, cannot be resumed, and cannot spawn its own subagents. See workflows/execute-plan/build-tail.md § OpenAI fast profile for event and fallback handling.

Grok host transport (Grok Bot)

Grok Bot is a first-class Bureau host. Entrypoint: GROK.md. Start a run with --runtime grok. Claude Code and Codex do not use this path.

Live specialist spawn is the Grok Bot Task executor. It starts blank: no parent transcript, no Grok Bot chat history, no standing-agent memory. That is the isolation proof for this host. CreateAgent fleets, named teammates, and group rooms remain forbidden as specialists.

scripts/run-grok-specialist.sh does not launch a model. Bash cannot spawn Grok Bot Tasks. Before each Task, the Delegate or Conductor runs --plan and logs the JSON payload (role, model, attempt, isolation flags). The helper still exits 2. Then the host issues the Task. A runtime=claude or runtime=openai routing file is rejected by the helper.

Isolation contract

1. model-routing.json#runtime is grok.
2. fork_context is false: Task prompt contains persona text and paths only.
3. No standing agent as a specialist.
4. Routing records grok-4.3 or grok-4.6 from roles.<role>. As of 2026-08-25 this host's Task tool only accepts sand-default; log the resolved Grok model on the SPAWN-EVENT and in the prompt. That is a named model-id gap, not a skip-routing license. grok-build-0.1 stays exec-only.
5. Handoff is on-disk RUN_DIR artifacts, not executor prose. Same producer artifact rule as Codex.

config/runtimes/grok.json sets supports_fresh_context_subagents true because Task executors start blank. If a future host cannot guarantee that, flip it false and fail closed again.

Starting a Grok run

scripts/run-start.sh "$RUN_DIR" --target "$TARGET_REPO" \
  --workflow "$WORKFLOW" --slug "$SLUG" \
  --runtime grok --no-pointer-echo

Codex producer artifact handoff

Codex subagents can complete or stall without leaving the required files on disk. For any producer role whose persona says it writes artifacts (Analyst, Architect, Designer, Spellwright, and build-party coders), the Conductor must treat the on-disk artifacts as the handoff, not the live prose.

After each producer spawn returns or times out:

1. verify the required output paths exist and contain the required marker sections named by the persona;
2. if missing, send one resume/nudge to the same agent: "write the required files now; if filesystem writes are unavailable, return an ARTIFACT PACKET with one fenced block per required file";
3. if the agent returns an ARTIFACT PACKET, the Conductor may mechanically transcribe that exact packet into the named files. This is not Conductor authorship: no paraphrase, no synthesis, no omitted sections. Log it as mechanical-transcription-from-specialist;
4. if neither on-disk artifacts nor a complete artifact packet appear after the nudge, close that agent, log no-handoff, and retry with a shorter, artifact-only prompt or escalate.

Never build an artifact-only retry prompt through shell interpolation of Markdown. Put the prompt in the tool call message directly, or read it from a plain file without command substitution. Markdown fences and backticks are data, not shell syntax.

The Bureau's checked-in AGENTS.md, Delegate persona, and Conductor persona explicitly authorize these Codex subagents when Robin starts a Bureau run. This is framework execution, not ad-hoc delegation.

Integrated Delegate topology

The topology is identical on Claude, Codex, and Grok Bot:

1. The Delegate creates the run and spawns one resumable Conductor.
2. The Conductor spawns fresh specialists and returns a CONDUCTOR-RETURN block at each checkpoint.
3. The Delegate stages a cold checkpoint packet and calls scripts/run-cold-reviewer.sh.
4. The Delegate validates the verdict, then resumes the same Conductor.

Host wording such as “Agent tool” or “SendMessage” is transport syntax, not a different workflow. Use the mapping above.

On Codex, start a new integrated run with:

scripts/run-start.sh "$RUN_DIR" --target "$TARGET_REPO" \
  --workflow "$WORKFLOW" --slug "$SLUG" \
  --runtime openai --no-pointer-echo

On Grok Bot, use --runtime grok (see GROK.md). Claude keeps the existing command without --runtime (or with --runtime claude).

Cold reviewer boundary

scripts/run-cold-reviewer.sh is the single entrypoint for both hosts. The caller stages only the v2 manifest under NN-context; the helper rejects symlinks, log.md, and transcript-like files. Before each provider call it appends the exact task prompt to the warm run log for audit; that log is never part of the reviewer packet.

• Claude uses the existing Read-only, --add-dir "$CTX", --setting-sources "", no-session-persistence invocation.
• Codex copies the staged packet to a temporary directory and runs an ephemeral codex exec with user config and exec rules ignored, network tools disabled, the snapshot read-only, and explicit filesystem denies for the live run, target repository, framework checkout, home directory, and Codex/Claude session stores. The prompt contains snapshot paths only. Codex uses config/delegate-verdict.codex.schema.json, a closed structured-output projection of the unchanged Claude verdict contract.

Both paths emit a normalized verdict file and a Claude-shaped one-shot usage envelope. Callers append exactly one REVIEWER-TOKEN-EVENT per returned envelope.

Token-accounting guarantees

The three JSONL per-leg rows below apply to integrated Delegate-topology Claude runs. A direct-Conductor Claude run has no delegate_session_id with which to locate the top-session transcript tree, so its Conductor/specialist figures remain an explicit legacy gap.

Rail	Claude Code host	Codex host
Cold reviewer	exact one-shot envelope	exact stable codex exec --json turn.completed.usage, normalized by the helper
Top-level Delegate	post-hoc Claude JSONL window keyed by delegate_session_id; exact or explicitly degraded	unavailable
Resumable Conductor	post-hoc Claude JSONL sum across recorded/discovered Conductor legs; exact or explicitly degraded	unavailable
Specialists	post-hoc Claude JSONL joined by SPAWN-EVENT attempt_id and the run-scope nonce; exact or explicitly degraded	unavailable

On Claude, scripts/account-run.sh invokes scripts/aggregate-transcripts.sh at terminal close-out; the retired Stop/SubagentStop scripts are permanent exit-0 compatibility stubs and are not an accounting fallback. “Unavailable” is a named accounting gap, not zero usage. Codex does not expose the same stable JSONL transcript and per-specialist ownership interface, so the aggregator returns a named _runtime_gap before transcript lookup. Until a stable session-usage API exists, do not scrape Codex transcripts or fabricate zero-token specialist events. Run artifacts, checkpoints, and decisions remain fully tracked. See docs/run-accounting.md § B2. The Spark one-shot helper does preserve its exact turn.completed.usage in a normalized envelope, but that per-attempt evidence is not silently promoted into the currently unavailable aggregate specialist rail.

Resume

Resume from RUN_DIR/state.json, RUN_DIR/log.md, and RUN_DIR/delegate-state.json as usual. Read model-routing.json#runtime before selecting a transport. If the retained Codex agent id is still available, use multi_agent_v1.send_input; otherwise spawn a fresh Conductor with fork_context: false, the same RUN_DIR, topology: integrated, and a recovery instruction naming the durable artifacts. Log the recovery because transcript continuity was lost. Spark profile attempts are never resumed: a clean failed first attempt falls back to a new native Mage at its role-default model; an unsafe/dirty attempt requires adjudication.