# Orchestration DevRites defines workflow semantics; Codex and Claude provide native orchestration. ```text AGENTS.md / CLAUDE.md / skills ↓ host interprets the requested workflow ↓ host runs .codex/agents/.toml or .claude/agents/.md ↓ host waits, follows up, and delivers results ↓ root reconciles semantics; engine checks deterministic structure/state/safety ``` ## Authority - The public `rite-*` skill is the root orchestrator. It owns scope, human questions, decisions, result reconciliation, `.devrites/**` artifacts, phase transitions, and explicit irreversible-action approval. - Build maintains the strict candidate manifest. Prove, Review, and Seal bind their artifacts to the engine-computed digest; Polish owns final durable rollups and affected re-proof; Ship is candidate-read-only. - The root never writes source or tests. Both hosts enforce that as workflow policy (ADR-0035): Claude's root is not in plan mode because it must write `.devrites/**`, and Codex's parent must be workspace-capable for its writer child to execute. - The role catalog has one source/test writer, `devrites-slice-wright`. Claude and Codex make only that exact specialist writable. - Leaves never invoke leaves, ask the human, change phase, push, deploy, migrate live data, or write canonical workflow state. Only an eligible isolated-worktree wright may create one local unpushed transfer commit; same-worktree leaves never commit. Skills name the exact role and bounded task in normal language. The host owns internal spawn fields, scheduling, waiting, follow-ups, and result delivery. DevRites does not parse rollout logs or maintain dispatch receipts. On OMP, each `task` item selects its specialist through `agent`, including every item in a `tasks[]` batch: `devrites-code-reviewer`, `devrites-test-analyst`, or `devrites-slice-wright` for those roles. Writing “you are…” in a task or naming it after a specialist does not select that profile. Lane suffixes stay on engine role labels, not native profile names. Check the actual host selector before admitting results: the engine records declared roles and host handles, not authenticated native profile identity. A missing profile remains a gap, not permission to select a generic/default child. Quick, Standard, and Full select evidence/review depth only; they are not agent API versions or execution tiers. The shared contract lives in [`orchestration-profiles.md`](../pack/.claude/skills/devrites-lib/reference/orchestration-profiles.md). Persisted state schemas remain explicitly versioned because readers need stable data contracts; no compatibility protocol is inserted between a skill and the host's native agent call. ## Native permission boundary Claude project settings set no `defaultMode` (ADR-0035); its slice-wright profile uses `permissionMode: acceptEdits`. Codex uses a workspace-capable root because children cannot elevate above the parent permission ceiling. Its exact slice-wright uses `default_permissions = ":workspace"` and the other 16 specialists use `default_permissions = ":read-only"`. The eight `overhaul-*` agents of the explicit `/overhaul` skill sit outside this split: Claude gives them every tool and Codex gives them `:workspace` ([ADR-0031](adr/0031-overhaul-agents-keep-every-tool.md)). The four `fast-*` agents of `/rite-fast` also sit outside it: `fast-builder` uses `acceptEdits` on Claude and `:workspace` on Codex, and the other three are read-only ([ADR-0033](adr/0033-rite-fast-own-agents.md)). If the exact role is unavailable, the root stops instead of substituting a generic agent or doing specialist/writer work inline. ## Result closure Each leaf's final response is its complete admission packet. It repeats every finding, outcome, evidence citation, limitation, and gap that the root must reconcile; an earlier message followed by a bare `done` or a link is malformed. The root owns collection and accounts for every required or applicable role. Timeout, absence, failure, or malformed output remains a gap, and a successful sibling cannot erase it. An evidenced clean result is valid. Reviewers must account for their rubric and inspected scope, but they never manufacture a finding to fill a quota. ## Read-only Claude workflow pilot Claude installs `.claude/workflows/devrites-readonly-review.js` as an optional adapter for immutable-candidate discovery, four independent reviewer roles, one adversarial verification pass, and a completeness check. Its script owns only transient fan-out and intermediate results. It cannot write source, tests, `.devrites/**`, Git, proof, lifecycle state, or shared services; it never invokes `devrites-slice-wright`. The caller still admits every returned role result under the normal result-closure contract and performs final reconciliation. A workflow timeout, missing role, malformed result, unread input, or incomplete verification is a gap. The adapter does not replace `/rite-review`, alter required rosters, or create another durable state plane. Codex keeps the same portable skill/agent semantics through native agent dispatch and has no fake workflow mirror. Provider parity applies to review meaning and evidence, not to this optional Claude orchestration optimization. Use the pilot only when an immutable candidate already exists; lifecycle writes, proof execution, human gates, and final decisions remain with the rite root. Host capabilities are admitted independently. Codex CLI agent threads and inherited sandboxes establish context/permission separation, not filesystem worktree isolation; without an explicit named-agent worktree and reconciliation interface, Codex uses the serial same-worktree writer. Likewise, goals or hooks do not prove native time/event activation. Unsupported activation remains `unavailable` rather than being emulated by a DevRites scheduler or shell loop. The claim-bounded acceptance rows live in [`codex-acceptance.json`](../evals/native-host/codex-acceptance.json). ## Slice-wright lifecycle For Claude or Codex, the root states the smallest exact project-relative source and test paths directly in the dispatch task. Directories, globs, traversal, symlink escapes, duplicates, and `.devrites/**` are invalid. 1. The root records the pre-dispatch `git diff --name-only`. 2. The host runs the exact `devrites-slice-wright` and waits for its result. 3. The root compares the returned file list and `git diff --name-only` with the task contract and rejects any extra path. 4. The root inspects the test diff for deletion, skipping, focus markers, or loosened assertions. 5. Accepted changes receive repository/CI proof and applicable read-only review. 6. The root records exact candidate rows and canonical evidence/state. Proof must name a positive, discriminating assertion and decisive observed signal; skipped/zero/assertion-free/tautological/unexecuted/exit-only results cannot establish behavior. Exact-path scope is instruction-backed on both hosts; native sandboxes provide the broader writer/read-only split. An unauthorized delta is rejected and the same bounded writer must restore it before work continues. The root never widens the contract or rewrites source. A serial native-worktree pilot may isolate `devrites-slice-wright` when the candidate baseline is committed and clean, the repository is not a submodule child, baseline proof is green, and the host exposes explicit result reconciliation. The wright returns one local unpushed transfer commit; root proves its exact paths and unchanged base before native transfer, then re-proves the reconciled candidate. Missing transfer, conflict, extra paths, moved base, or cleanup failure stops with worktree evidence preserved. No ad hoc copy/cherry-pick/merge occurs from the read-only root. Parallel writers use the `/rite-build` worktree lifecycle: independently vetted contracts, disjoint declared/hidden write surfaces, one retained controller lease, and per-sibling review/proof before serialized integration. `--parallel N` on Autocomplete is one global active-writer cap across authorized sequence workspaces; without the flag it uses AFK `max_parallel`, retained lease capacity on resume, else 10. No additional user flag is needed. Standalone Build remains scoped to its requested workspace. While a writer runs, Autocomplete prepares or re-Vets recorded continuations and refills from the whole eligible pool. A future slice needs an explicit, current Plan/Vet `Workspace dependencies` declaration; absent or stale declarations never become independence merely because files differ. Only durable integrated producer checkpoints release dependent work. Human/access/safety gates retain their scope. Each worker keeps its source workspace identity, isolated cwd, candidate, proof, state and `WIP():` checkpoint. Even a lone initial writer is isolated when using a parallel-capable controller, so later work can refill without creating same-tree co-writers. Explicit `--parallel 1` remains one writer total. Unsupported host isolation falls back truthfully to serial execution. The controller drains and cleans its lease before non-lease product corrections, shared Polish writes or checkpoints; unrelated control commits are never blessed as a new expected tip. Final per-workspace proof/Seal and exact release-union Ship approval remain mandatory. See the [controller workflow](../pack/.claude/skills/devrites-lib/reference/parallel-workspaces.md) and [worktree lifecycle](../pack/.claude/skills/rite-build/reference/parallel-batch.md). ## Engine boundary The engine owns deterministic: - local managed install, update, and uninstall against caller-supplied candidates; - strict project-candidate validation/digesting, content-bound readiness, and final structural plus exact-binding checks; - atomic answer/drop/batch resolve and transactional close mutations; - secret scanning and version reporting. Native hosts, skills, exact agents, repository tools, or CI own: - agent discovery, dispatch, scheduling, waiting, follow-up, and results; - semantic readiness, traceability, acceptance/evidence quality, doubt, review reconciliation, test-quality assessment, capability interpretation, semantic upgrade, and recovery routing; - normative spec grammar re-read, qid allocation, Clarify cursor edits, AFK budget accounting, recovery attempt accounting, and read-only install/host diagnosis; - exact-release bundle/binary acquisition before invoking local engine install operations; - repository inspection, search, and test/build/lint/release execution; - repository JSON/schema/generated-artifact validation; - session history, compaction, status prose, progress, plugins, and memory. The engine has no agent bridge, semantic readiness protocol/digests, heuristic prose parser, capability-ledger interpreter, compatibility telemetry, or old aliases. Its `migrate` command performs deterministic schema normalization; artifact stubs do not establish semantic readiness or historical proof ([ADR-0029](adr/0029-v5-workspace-schema-and-native-migration.md)). Provider/consumer changes use the existing plan and traceability artifacts: one canonical contract plus provider- and consumer-side asserting tests that both consume it. Spec capability impact and lossless MODIFIED folding remain with Spec/Polish. These rules add no dispatch protocol, schema registry, or phase. Any harness adapter must expose the same read-only `check candidate` command, preserve exact binding lines and the Build/Prove/Polish/Review/Seal/Ship role boundaries, and stop rather than substitute another hash or inline specialist. See [ADR-0015](adr/0015-read-only-root-native-orchestration.md), [ADR-0017](adr/0017-native-codex-writer-agent.md), [ADR-0018](adr/0018-native-sandbox-instruction-writer-boundary.md), [ADR-0020](adr/0020-thin-engine-native-orchestration-boundary.md), [ADR-0022](adr/0022-native-orchestration-thin-engine.md), and [`standards/agents.md`](../pack/.claude/skills/devrites-lib/reference/standards/agents.md).