Free-text approval policies

Write operational review instructions at Settings → Approval policies. The same editor is shared by the terminal console, web console and desktop clients. Select a target from the current inventory; identifiers do not need to be typed.

The snapshot covers the requesting repository and its owning Project group; it does not claim to aggregate policies from every repository linked to a session.

Enabled instructions are additive, in that order. Narrower text does not remove broader constraints. Empty or disabled policies do not authorize an action. When Control Room is configured, a pending manual request can still be reviewed to propose a useful task intent. An action already approved by the existing non-danger setting does not create an unnecessary model review. Do not put credentials or secret values in policy text.

For example, a Project policy may say: “Allow reading public documentation and running tests in isolated development databases. Ask a person before changing production resources or creating a paid service. Deny access to unrelated projects.” This is review guidance; it is not a replacement for the configured network, account, team or OS boundaries.

Review and actual execution

When at least one policy applies, a pending approval keeps its original risk classification and receives a durable policy-review record. The record binds the original request, current source identity, every applicable policy revision and their content hash. The existing automatic approval of non-danger requests also waits for this review.

An isolated automatic Control Room task reads the exact request, bounded controller-owned Project/repository/host/linked-repository facts and policy text. The reviewer uses an explicitly authorized Control Room account and installed model profile; there is no separate classifier service. Authenticated host launch and a native task claim are required before a verdict is accepted. The reviewer’s managed MCP access is limited to reading, claiming and returning a verdict for its own review. Host-command and session-input MCP operations are blocked for that task. The native CLI retains its existing host/container execution boundaries; this feature does not claim a new OS sandbox.

The verdict is one of:

The restricted reviewer currently requires the verified Claude launch profile and an updated execution host. If the ordinary Control Room uses Codex, devctrl may reuse one unambiguous, already authorized Claude human-chat profile on the chosen host. It preserves the ordinary human and fulfillment agent selection. It never borrows a project account or silently chooses among multiple accounts. The configured model, original human grant and actual reviewer account are recorded separately from the requesting agent and fulfillment configuration.

The reviewer must also satisfy the source repository and linked repositories’ current account restrictions, administrator ceilings and account eligibility. Those revisions bind the review, and native reads, claims and verdicts recheck routing. A corporate request cannot be sent to an excluded personal account. Absent, revoked, ambiguous, locked, out-of-quota or unsupported profiles require a person with a visible reason; no unrestricted reviewer is launched. Codex’s current native review adapter remains unverified; this does not mean the CLI can never provide a restricted execution mode.

Missing Control Room configuration, unavailable models or quota, an expired review, a changed policy/request, uncertainty, conflicting guidance and team policy exceptions require a person. No unavailable model is treated as approval. Reviews run on the server without an open console. Reconnection does not replay submitted task prompts or the approved operation.

A policy verdict and technical fulfillment are separate. Allowing a repository network request still uses its normal persistent-policy and gateway application path. Resource and service requests still need their existing authenticated host receipts. “Allowed” never means a route, secret, service or resource was actually installed. Existing owners and team ceilings remain authoritative.

CLI and MCP

devctrl policies list
devctrl policies show system
devctrl policies show project <project-group-id>
devctrl policies show repository <repository-id>
devctrl policies set repository <repository-id> policy.md
devctrl policies disable repository <repository-id>
devctrl policies review <approval-id>

set reads a bounded UTF-8 text file, fetches the current revision and saves using optimistic concurrency. A concurrent edit returns a conflict rather than being silently overwritten.

Control Room uses its existing control_operations and control_request MCP transport. Discover GET /v1/text-policies/scopes, GET /v1/text-policies and PUT /v1/text-policies. Source sessions can inspect only their applicable layers through inspect_current_project; they cannot set trusted policy text in an approval request. The isolated reviewer reads GET /v1/approval-policy-reviews/:id, claims its native identity through /claim, and publishes its structured /decision. These operations do not bypass the normal authentication or ownership checks.

Current limitations

Policy text is interpreted by the configured Control Room model, rather than a deterministic policy language. It can request human review and does not promise that every instruction is machine-verifiable. A separate manual approval remains available, and should be used when the review is unavailable or inconclusive.

Remember a task, rather than approving every step

While reviewing a normal access request, Control Room can propose a concise natural-language policy such as “Allow configuring CI for this Project.” CI is only an example: the same mechanism can describe another operational objective. A proposal may cover routine setup, newly created task resources and subsequent source commits; it is not limited to an existing runner or Service ID.

The approval screen offers the existing Approve once action and a separate Approve and remember this intent action. Read and edit the proposed title, coverage and exclusions, and choose the exact repository or owning Project-group scope. Confirmation is an authenticated user decision. A native reviewer cannot confirm its own proposal or obtain authority by returning “allow” when no policy or confirmed intent authorizes the request.

Remembered decisions persist on the server across process and controller restarts. Their scope binds the owner and repository incarnation or existing Project group, rather than a commit hash or provider process. Future requests in that scope carry the current remembered decisions into the same isolated Control Room review. The model evaluates whether the current action fits the confirmed intent. It can permit routine work without asking the user again, including new resources that fit the purpose. A different task, conflicting policy, uncertainty or a hard team boundary still needs a person. No keyword or regular-expression match substitutes for that contextual review.

Inspect, edit or revoke remembered intents in Settings → Approval policies. Edits use revision checks; revocation invalidates pending model verdicts. It does not delete resources or roll back work already performed. Intent history is paged separately, rather than loaded into every console snapshot.

Saving an intent, approving the current request and completing its technical work are distinct. If the intent is saved but a technical check rejects the current request, the interface reports both facts. An interrupted decision never replays an uncertain operation merely because its approval row is still pending. The server settles a saved verdict from the actual approval record, or returns the uncertainty to a human with an explicit reason.

devctrl policies intents repository <repository-id>
devctrl policies intent <intent-id>
devctrl policies revoke-intent <intent-id> <revision>
devctrl policies confirm-intent <approval-id> confirmation.json
devctrl policies edit-intent <intent-id> change.json

The JSON files use the same strict API schemas as the shared interface. Exact proposal hashes, current intent revisions and selected-action/plan hashes remain required. Confirmation and intent edits are not available to native MCP callers; Control Room can propose and read the user's decision instead.