Codex Harness Runtime: Boundaries, Hooks, Tools, Permissions

Runtime contract for Codex harness turns. For setup and routing, see Codex harness. For config fields, see Codex harness reference.

Overview

Codex manages the native model loop, native thread resume, native tool continuation, and native compaction. OpenClaw handles channel routing, session files, visible message delivery, OpenClaw dynamic tools, approvals, media delivery, and a transcript mirror around that boundary.

For native connected apps, Codex also owns the final per-thread app and tool policy. OpenClaw caches a runtime-and-workspace-scoped plugin/installed snapshot, reads exact configured plugin details, provisionally admits only explicitly allowed, ownership-proven apps, and creates a deny-by-default native thread. One app/installed request verifies the actual thread ID without forcing an inventory refresh. Native app execution begins only after Codex confirms the app is enabled and callable for that thread.

This check finishes before OpenClaw injects history, starts a turn, or commits a thread binding. Failed persistent provisional threads are deleted; ephemeral threads are unsubscribed. OpenClaw retires the app-server connection when safe cleanup cannot be confirmed. Supervised branches also clean up their temporary probe and preserve recovery state if cleanup fails.

Account-wide app access cannot override an explicitly disabled configured workspace plugin. OpenClaw uses its installed snapshot and reads only that exact plugin's details to identify and deny its apps; it never scans unrelated marketplaces or activates the plugin.

Prompt routing follows the selected runtime, not just the provider string. A native Codex turn gets Codex app-server developer instructions; an explicit OpenClaw compatibility route keeps the normal OpenClaw system prompt even when it uses Codex-flavored OpenAI auth or transport.

OpenClaw starts and resumes native Codex threads with Codex's built-in personality disabled (personality: "none") so workspace personality files and OpenClaw agent identity stay authoritative. Native Codex keeps Codex-owned base/model instructions and project-doc loading otherwise. An ordinary policy-restricted turn has no native filesystem environment, so OpenClaw carries the bounded workspace AGENTS.md snapshot as thread-level developer instructions instead. Lightweight, ring-zero, message-only, and tool-disabled internal turns suppress project-doc loading and that fallback carrier.

OpenClaw developer instructions cover OpenClaw runtime concerns: source-channel delivery, OpenClaw dynamic tools, ACP delegation, adapter context, and the active agent workspace profile files. Skill catalogs and tool-routed MEMORY.md pointers are projected as turn-scoped collaboration developer instructions. When memory tools are unavailable, active BOOTSTRAP.md content and full MEMORY.md fall back to plain turn input context instead.

When openclaw_direct.sessions_yield is available, those instructions also tell a native Codex parent to end the current turn when a child's result should arrive in a later turn. Native wait_agent remains for an intentional same-turn wait when the immediate next step is blocked on the child; completion polling loops are not a substitute.

Most OpenClaw dynamic tools use the searchable openclaw namespace. Tools marked catalogMode: "direct-only" use openclaw_direct, which Codex keeps directly model-visible as DirectModelOnly instead of exposing it to nested Code Mode execution.

Thread bindings and model changes

When an OpenClaw session is attached to an existing Codex thread, the next turn resends the currently selected model, approval policy, sandbox, approvals reviewer, and service tier to app-server. Switching from openai/gpt-5.5 to openai/gpt-5.2 keeps the thread binding but asks Codex to continue with the newly selected model.

Supervised bindings are the exception. The OpenClaw model picker stays locked, and resumes omit model and provider overrides so Codex restores the canonical thread's persisted model and provider. A separate native Codex control can change that persisted pair, and the initial snapshot can produce Codex's normal model-difference warning; the outer OpenClaw model and fallback chain never substitute for either.

Supervision and safe continuation

Codex supervision is an opt-in capability of the same codex plugin. It discovers native threads through a separate connection and projects only non-archived sessions into the Gateway catalog. Without explicit appServer connection settings, that connection uses managed user-home stdio while the ordinary harness remains agent-scoped. Listing and metadata reads are passive: they do not resume a thread, subscribe OpenClaw to its live events, or answer its approvals.

For a stored or idle session on the Gateway computer, Continue as branch creates a normal, model-locked Chat and mirrors bounded user and assistant history through the source's last terminal persisted turn. The first normal Chat turn installs the real approval handlers and uses a temporary native fork to pin the snapshot without a model or provider override. Codex App Server uses its current native configuration and returns the selected pair; it emits its normal warning if that model differs from the source's last recorded model. On the same supervision connection, OpenClaw starts the canonical appServer-source Codex harness thread under its cwd and runtime policy with exactly the returned model and provider for that initial start, injects the bounded visible history, and archives the temporary fork. The source is never resumed. The canonical thread has the full OpenClaw harness tool surface; reasoning, tool calls, and tool results from the source are not cloned into it. The private connection scope survives pending and committed binding states, so every later turn remains on that connection with native auth and provider configuration. Disabled supervision or binding/connection drift fails closed rather than switching to the ordinary agent-home harness.

The original CLI, VS Code, Atlas, or ChatGPT source remains eligible for both catalogs. The canonical branch is a native Codex thread, but its source kind is appServer; native clients may filter that source kind, so its appearance in Codex Desktop is not guaranteed.

Active sources cannot start a new branch or be archived; an existing supervised Chat can still be opened. notLoaded means activity is unknown, not idle; OpenClaw allows archive for a local idle or notLoaded row only after explicit no-other-runner confirmation and a fresh process-local status read. Codex serializes thread mutations within one App Server process but does not provide an exclusive cross-process runner or approval-owner lease, so that read cannot prove that another process is not using the thread. OpenClaw blocks a known active binding owner for the exact target or any non-archived spawned descendant returned by Codex's paginated descendant query. Enumeration errors, cycles, and safety-limit exhaustion fail closed. Native archive can still race a new turn in another process, so confirmation covers unknown clients and the gap between status read and archive. A supervised model-locked Chat cannot be deleted while it protects the native binding.

Paired-node catalogs stay metadata-only in the initial release. The current node invoke boundary is request/response and cannot carry the long-lived turn events, approval requests, or streaming output required by a real Codex harness binding. Remote Continue and Archive therefore remain unavailable even when the row is idle.

See Codex supervision for operator setup and the visible Control UI behavior.

Visible replies and heartbeats

Direct/source chat turns through the Codex harness default to automatic final assistant delivery for internal WebChat surfaces, matching the Pi harness contract: the agent replies normally and OpenClaw posts the final text to the source conversation. Set messages.visibleReplies: "message_tool" to keep final assistant text private unless the agent calls message(action="send").

Codex heartbeat turns get heartbeat_respond in the searchable OpenClaw tool catalog by default so the agent can record whether the wake should stay quiet or notify. Heartbeat initiative guidance is sent as a Codex collaboration-mode developer instruction scoped to the heartbeat turn; ordinary chat turns stay in Codex Default mode. The heartbeat monitor's cron scratch is appended to the heartbeat prompt when present.

Hook boundaries

OpenClaw never relies on project-level or global Codex hooks.json files to determine plugin behavior. Instead, for the native tool and permission bridge, OpenClaw injects per-thread Codex configuration covering PreToolUse, PostToolUse, PermissionRequest, and Stop.

When Codex app-server approvals are turned on, meaning approvalPolicy is not set to "never", the default injected native hook configuration leaves out PermissionRequest. This lets Codex's app-server reviewer and OpenClaw's approval bridge manage real escalations that come after review. To force the compatibility relay regardless, add permission_request to nativeHookRelay.events. Other Codex hooks, including SessionStart and UserPromptSubmit, stay as Codex-level controls and are not surfaced as OpenClaw plugin hooks under the v1 contract.

For OpenClaw dynamic tools, execution happens after Codex requests the call, so plugin and middleware logic runs inside the harness adapter. Codex Code Mode receives generic dynamic results as text and serializes nested dynamic calls; callers must parse JSON-like output and cannot depend on Promise.all for concurrent submission. With Codex-native tools, Codex holds the canonical tool record. OpenClaw can mirror selected events but cannot modify the native thread unless Codex offers that through app-server or native hook callbacks.

Codex app-server report-mode PreToolUse events postpone plugin approval until the corresponding app-server approval arrives. If an OpenClaw before_tool_call hook returns requireApproval while the native payload sets openclaw_approval_mode: "report", the native hook relay records the plugin approval requirement and returns no native decision. When Codex later sends the app-server approval request for that same tool use, OpenClaw presents the plugin approval prompt and maps the decision back to Codex. Codex PermissionRequest events form a separate approval path and can still pass through OpenClaw approvals when that bridge is configured.

Codex app-server item notifications also supply async after_tool_call observations for native tool completions not already handled by the native PostToolUse relay. These serve telemetry and compatibility only; they cannot block, delay, or alter the native tool call.

Compaction and LLM lifecycle projections derive from Codex app-server notifications and OpenClaw adapter state, not from native Codex hook commands. before_compaction, after_compaction, llm_input, and llm_output are adapter-level observations, not exact copies of Codex's internal request or compaction payloads.

Codex native hook/started and hook/completed app-server notifications are projected as codex_app_server.hook agent events for trajectory and debugging. They do not trigger OpenClaw plugin hooks.

Experimental sandbox process streaming

Native sandbox execution stays opt-in through appServer.experimental.sandboxExecServer. When enabled for an active OpenClaw sandbox, sandboxed processes stream ordered stdout, stderr, or PTY output notifications. OpenClaw keeps only a bounded recent-output buffer for polling and replay, so long-running processes cannot grow the app-server bridge without limit. Process exit and cleanup remain tied to the sandbox-owned process. Failed environment registration never falls back to host execution.

See Sandboxed native execution for configuration and local-only transport restrictions.

Paired-device remote-exec is distinct from the experimental local sandbox flag: Codex app-server and model auth stay on the Gateway, while an explicitly authorized managed exec-server on the node owns process, filesystem, capability, and credential-free HTTP operations. The Gateway rejects authentication, cookie, API-key, and other sensitive HTTP headers before they reach the node; authenticated HTTP must run on the Gateway. The existing duplex node channel carries the Codex JSON-RPC stream without starting an OpenClaw worker child or consuming a worker slot. Each attempt owns an isolated Gateway app-server client so its remote environment registration retires with that attempt. Disconnect ends the active attempt and its remote processes; reconnect allows only a fresh attempt. Normal Codex turns work, but /btw side questions fail closed because they are not yet placement-bound. The placement workspace does not confine execution: process and filesystem access remain bounded only by the node's operating system account.

V1 support contract

Supported in Codex runtime v1:

Not supported in Codex runtime v1:

Native permissions and MCP elicitations

For PermissionRequest, OpenClaw only returns explicit allow or deny decisions when policy decides. A no-decision result is not an allow: Codex treats it as no hook decision and falls through to its own guardian or user approval path.

Codex app-server approval modes do not include this native hook by default. That behavior holds unless permission_request is explicitly present in nativeHookRelay.events or a compatibility runtime provides it.

When an operator selects allow-always for a Codex native permission request, OpenClaw records that specific provider/session/tool input/cwd fingerprint for a limited session window. The stored decision matches only on exact values: a different command, arguments, tool payload, or cwd triggers a new approval.

Codex MCP tool approval prompts go through OpenClaw's plugin approval flow when Codex sets _meta.codex_approval_kind to "mcp_tool_call". Plugin, account, Computer Use, and MCP approval classification happens before standard input handling. A denied policy or an unmappable approval schema yields an explicit decline and never becomes a general-purpose form.

OpenClaw supports app-server MCP elicitation modes form, openai/form, and url. Standard and extended forms hold at most 12 fields. OpenClaw maps field names to Gateway-safe question IDs, keeps the original names in accepted content, and shows fields in sequential groups of up to three. Each field can offer at most four choices; fields and choices beyond those limits are declined rather than cut off. Supported fields are free-form strings, string enum or oneOf choices, booleans, numbers and integers, and multi-select string arrays. Free-form string values cap at 4,096 characters. String length, email, uri, date, and date-time constraints and numeric or array bounds get validated before an accepted response is returned. Optional fields, required fields, and valid defaults keep their schema meaning.

openai/form also supports a single-select openai/imagePicker field with up to four bounded item IDs and titles. OpenClaw uses only those IDs and titles; it does not fetch or render item images. An unknown extended field type produces a visible operator message and an explicit decline. This visible fallback is part of the openai/form capability contract.

URL elicitations appear as literal text with explicit Continue and Decline choices. OpenClaw does not fetch or open the URL. URLs cap at 2,048 characters, must use HTTP or HTTPS, cannot include credentials, and cannot contain control or invisible characters. Invalid URLs produce a visible explanation and an explicit decline.

Codex request_user_input and ordinary MCP elicitations share one per-turn interactive queue. The Control UI renders each non-secret Gateway question, and a single choice uses typed channel buttons when the channel supports them. Button taps, Control UI answers, and the next queued plain-text reply resolve the same exact app-server request. serverRequest/resolved selects a request by its outer string-or-integer JSON-RPC ID; attempt abort, timeout, and cleanup cancel the current owner. Late answers cannot resolve a queued replacement.

Only an explicit field isSecret: true or Codex question isSecret: true enables secret handling. Secret form fields are requested one at a time through the warned ephemeral text-reply path and never create durable Gateway question records. OpenClaw does not infer secrecy from field names.

For the general plugin approval flow that carries these prompts, see Plugin permission requests.

Queue steering

Active-run queue steering maps onto Codex app-server turn/steer. With the default messages.queue.mode: "steer", OpenClaw batches steer-mode chat messages for the configured quiet window and sends them as one turn/steer request in arrival order.

Codex review and manual compaction turns can reject same-turn steering. In that case, OpenClaw waits for the active run to finish before starting the prompt. Use /queue followup or /queue collect when messages should queue by default instead of steering. See Steering queue.

Codex feedback upload

When /diagnostics [note] is approved for a session on the native Codex harness, OpenClaw also calls Codex app-server feedback/upload for relevant Codex threads, including logs for each listed thread and spawned Codex subthreads when available.

The upload goes through Codex's normal feedback path to OpenAI servers. If Codex feedback is disabled in that app-server, the command returns the app-server error. The completed diagnostics reply lists the channels, OpenClaw session ids, Codex thread ids, and local codex resume <thread-id> commands for the threads that were sent.

If you deny or ignore the approval, OpenClaw does not print those Codex ids and does not send Codex feedback. The upload does not replace the local Gateway diagnostics export. See Diagnostics export for the approval, privacy, local bundle, and group-chat behavior.

Use /codex diagnostics [note] only when you want the Codex feedback upload for the currently attached thread without the full Gateway diagnostics bundle.

Compaction and transcript mirror

When the selected model uses the Codex harness, native thread compaction belongs to Codex app-server. OpenClaw does not run preflight compaction for Codex turns, replace Codex compaction with context-engine compaction, or fall back to OpenClaw or public OpenAI summarization when native compaction cannot be started. OpenClaw keeps a transcript mirror for channel history, search, /new, /reset, and future model or harness switching.

Explicit compaction requests, such as /compact or a plugin-requested manual compact operation, start native Codex compaction with thread/compact/start. OpenClaw keeps the request and shared-client lease open until Codex emits the matching contextCompaction completion item and then reports the compaction turn as completed. If that terminal turn exceeds the configured compaction timeout, OpenClaw requests a native turn interrupt. The lease and per-thread compaction fence remain held until Codex reports terminal state or confirms the interrupt RPC. If Codex does not confirm within the interrupt grace period, OpenClaw retires the connection before releasing the fence. Remote connections also detach the matching thread binding so later work cannot overlap an unconfirmed remote turn. Other turns on a retired connection fail and can retry on a fresh client. Client closure, request cancellation, or a failed compaction turn returns a failed operation. Automatic context-pressure compaction is Codex's job; OpenClaw only starts native compaction for manually requested triggers.

When a context engine requests Codex thread-bootstrap projection, OpenClaw projects tool-call names and ids, input shapes, and redacted tool-result content into the fresh Codex thread. It does not copy raw tool-call argument values into that projection.

The mirror includes the user prompt, final assistant text, and lightweight Codex reasoning or plan records when the app-server emits them. OpenClaw records the native compaction start and terminal status, but it does not expose a human-readable compaction summary or an auditable list of which entries Codex kept after compaction.

Because Codex owns the canonical native thread, tool_result_persist does not rewrite Codex-native tool result records. It only applies when OpenClaw writes an OpenClaw-owned session transcript tool result.

Media and delivery

OpenClaw continues to own media delivery and media provider selection. Image, video, music, PDF, TTS, and media understanding use matching provider/model settings such as agents.defaults.mediaModels.image, agents.defaults.mediaModels.video, pdfModel, and tts.

Text, images, video, music, TTS, approvals, and messaging-tool output continue through the normal OpenClaw delivery path; media generation does not require the legacy runtime. When Codex emits a native image-generation item with a savedPath, OpenClaw forwards that exact file through the normal reply-media path even if the Codex turn has no assistant text.

Related

• Codex harness
• Codex harness reference
• Codex supervision
• Native Codex plugins
• Plugin hooks
• Agent harness plugins
• Diagnostics export
• Trajectory export