Repository navigation
Description
Problem
(measured: source baseline) Verified at 27a32e80998a983affdb244de7d232249963e0ed applies to all current-source claims in this document; historical revisions and GitHub observations are identified separately.
(measured: owner statement, 15 Nov 2026, 07:43 America/Buenos_Aires) “another issue for the review feature is that the model is always asking for permission and it should be fully autonomous.” The owner identifies the action as Source Control → Review uncommitted changes.
(measured: supplied screenshot) A Codex (OpenAI) review is paused at PENDING APPROVAL / Command approval requested, with Cancel turn, Decline, Always allow this session, and Approve once. The displayed command is a single /usr/bin/bash -lc invocation combining file inspection (cat), git diff, a Node check script, and node --test. The screenshot establishes one approval stop; “always” / “on every step” is the reported frequency, not a measured rate.
Two independent static diagnoses ran on the same clone, one with Codex and one with Claude Code. Neither saw the other's output. Their findings are merged here. Points both reached are the main explanation; points where they differ are listed under Open questions.
(inferred; both agree) This is not a launch fallback or a lost setting. Every chat review is hard-coded to Supervised (runtimeMode: "approval-required"). For Codex that maps to approval policy untrusted plus a readOnly sandbox, which can stop even a read-oriented command for approval. The current behavior matches the implemented design (decision D41) and living docs. The requested change is an unattended review that can inspect the selected changes and return findings without per-command permission decisions, while still protecting the checkout.
Triage proposal
- (inferred; both agree) Type: enhancement. The code, tests, living docs and design D41 say a chat review starts Supervised and that Codex may ask before some read commands. Source Control: a Review button starts a read-only agent that reviews the branch and uncommitted changes #293 promised a read-only session, not an autonomous one. One secondary copy defect does not change the label: the Codex guard line says writes ask for approval, but read-only commands can ask too.
- (inferred; agent's call) Priority: P2. The present behavior matches the delivered policy and is not a demonstrated regression. (inferred; Claude preferred P1/medium because the default reviewer is the Codex chat and unattended review is blocked for that path.)
- (inferred; agent's call) Size: L. The recommended direction decouples Codex approval from sandbox, carries a server-owned review policy, and needs Codex/Claude coverage plus safety validation. A one-line Full access substitution is not the proposed solution. (inferred; Claude suggested size:M, and said L if a new
RuntimeModeliteral is chosen.) - (inferred; both agree) Labels: enhancement, needs-triage, size:L (Claude would use
size:M). (measured) These labels exist ingh label list. No trust or profiler label is proposed. - (inferred; both agree) Release view: not a release blocker on the available evidence. Do not ship an autonomy change that removes the review's write protection.
- (inferred; both agree) Diagnosis gate: Inferred, not Confirmed. The source path is traced; no live reproduction or one-variable outcome flip was allowed. The actual running binary, settings and provider request were not inspected.
(measured) The original feature issue, #293, is closed as completed. Its initial proposal recommended inheriting the user's default runtime mode; open question 1 asked whether approval-required should replace that. Design D41 later settled on Supervised. This request changes that delivered tradeoff; it is a policy follow-up, not a duplicate of the already-delivered button. Button size/styling remains the companion issue (draft).
Expected vs Actual
| Surface | Expected | Actual |
|---|---|---|
| Codex chat review | (inferred; proposed) Inspect the selected checkout and return findings without command approval prompts. | (measured: screenshot) One command awaits approval. (measured: owner) Repeated requests interrupt the review. |
| Review access | (inferred; proposed) Autonomy is bounded by enforceable read-only access, independent of the host chat's edit permissions. | (measured: source) Review chat explicitly uses approval-required; Codex maps it to untrusted plus readOnly. |
| Normal chat comparison | (measured: source) Normal chat uses its composer override, then thread mode, then the Full access default. | (measured: source) The review launcher does not inherit that choice; review settings select agent, model and prompt, not a runtime mode. |
| Claude review | (inferred; proposed) Equivalent unattended inspection where an enforceable boundary is available, with an honest unavailable state otherwise. | (measured: source) The shared launcher selects approval-required plus plan, translated to Claude permission mode plan. (measured) No Claude failure was supplied. (inferred; Claude diagnosis) This exact approval card is unlikely for Claude, because BiBCode never opens a Claude permission request. |
Environment and supplied evidence
- (measured: supplied evidence) The review surface belongs to BiBCode; the Activity panel has the OpenAI/ChatGPT mark and a session usage chip. The owner identifies Codex as the provider.
- (measured: screenshot) The command is one
/usr/bin/bash -lcinvocation containingcat,git diff,nodeandnode --test; it is described here by class to avoid exposing workspace details. A plan/session usage footer is also visible; its account-related text is omitted. - (measured) The screenshot does not identify the installed BiBCode version, provider CLI version, actual approval policy, sandbox state, model, or full environment. A POSIX shell path does not prove the desktop OS or whether the execution environment is local or remote.
- (measured) Source baseline:
HEADandorigin/mainboth equal27a32e80998a983affdb244de7d232249963e0ed. This identifies the inspected source, not the running app. (measured: Claude) The feature first shipped in1a6a0987b(11 Nov 2026), first released in v0.16.0, merged in PR Batch: pipeline, Git lock, Markdown, themes, Host dashboard, Review button and terminal fixes (#340 to #348, #290, #292, #293) #349. - (inferred; investigation recipe, not executed) In a disposable fixture with uncommitted changes and Codex selected as the review agent, open an established thread, then click Source Control's Review action. Have the review perform ordinary file/Git inspection and a safe fixture script; capture launch/turn policies and whether a command approval appears. Repeat with the host set to Full access, then with Claude. No deterministic live reproduction is claimed here.
Static trace
(measured: source) Verified at 27a32e80998a983affdb244de7d232249963e0ed for every path:line reference in this section and the source-backed tables below. Test references describe assertions read, not tests run. Both diagnoses traced this path independently and agreed on the mechanism.
- (measured) Entry points and scope.
apps/web/src/components/SourceControlPanel.logic.ts:134-151selects uncommitted scope when the tree is dirty and no branch-ahead scope applies. The button callscodeReview.onStart(scope)atSourceControlPanel.tsx:1162-1191; Review once with supplies a one-off agent at:1218-1238; Review with instructions… returns to the same callback at:1732-1740.apps/web/src/keybindings.ts:347-364resolvessourceControl.reviewor the unclaimed Mod+Alt+Shift+R fallback;ChatView.tsx:5337-5340→:5097-5106invokes the same handler. No separate native-menu review launcher was found in the searched desktop sources. - (measured) Reviewer and prompt.
apps/web/src/components/ChatView.tsx:4937-4950,5002-5025resolves the savedsettings.codeReviewAgentor the menu override.packages/contracts/src/settings.ts:560-563,632-640defaults to the Codex chat and defines agent/model/prompt settings.apps/web/src/codeReview.ts:92-110refuses unavailable reviewers instead of substituting another provider.:26-46resolves project → user → built-in prompt and adds “Do not edit files.” The built-in prompt atsettings.ts:579-588asks for Git/file inspection, not mandatory test execution. - (measured) The policy decision.
apps/web/src/codeReview.ts:49-57explicitly returnsruntimeMode: "approval-required"for every chat driver andinteractionMode: "plan"only for Claude. No setting is read.ChatView.tsx:5069-5091supplies those modes tocenterPanelActions.createPromptedChatPanel, with the selected model,codeReview: true, and label Code review · <agent>. This is an explicit override, not a fallback caused by a missing setting. Pinned bycodeReview.test.ts:122-135. - (measured) Hooks, stores and RPC.
apps/web/src/centerPanelActions.ts:297-304binds the worktree and thread commands.:430-445reserves a new tab inuseCenterPanelStore;:383-397sendsthreadDefaultsand the review flag toworktree.createPanel.:526-551guards duplicate launch and sends the first message;:472-480passes the same modes tostartTurn.packages/client-runtime/src/state/worktrees.ts:469-480routes panel creation to the selected environment.state/threadCommands.ts:167-188→operations/commands.ts:193-202dispatchesthread.turn.start. - (measured) Server authority.
apps/server/src/production/worktree_catalog_rpc.rs:1571-1589creates a panel with the requested modes, review marker and model, but resolves the checkout and branch from its host.apps/server/src/production/provider_runtime.rs:6284-6317,6348-6353,6437-6449carries turn modes, or persisted thread modes when no override is supplied, intoProviderLaunchRequestwith that checkout. The new panel owns a provider session; the review does not reuse the host's runtime policy. - (measured) Codex wire policy.
provider_runtime.rs:14688-14694converts"approval-required"intoCodexRuntimeMode::ApprovalRequired; an unknown runtime-mode string becomes Full access (fail-open).:14826-14836uses that for session options.apps/server/src/provider/codex/runtime.rs:1112-1121,1149-1162sendsapprovalPolicy: "untrusted",sandbox: "read-only"on thread start/resume.apps/server/src/provider/codex/model.rs:726-737sendsapprovalPolicy: "untrusted"andsandboxPolicy: {type: "readOnly"}for the turn. Full access instead pairsneverwithdangerFullAccess; auto-accept-edits pairson-requestwithworkspaceWrite. No existing chat runtime mode expressesneverplusreadOnly. (measured: Claude) Thebuild_turn_start_paramstests (model.rs:1612,:1640) check the model andclientUserMessageIdonly; no test pinsapprovalPolicy/sandboxper mode. - (measured) Approval reaches the card.
codex/runtime.rs:2958-2986retainsitem/commandExecution/requestApprovalas a pending request and emitsrequest.opened.production/provider_runtime.rs:13017maps it toapproval.requested;apps/web/src/session-logic.ts:355-405derives pending approvals.components/chat/ComposerPendingApprovalPanel.tsx:13-24renders the screenshot heading;ComposerPendingApprovalActions.tsx:21-52sends cancel/decline/acceptForSession/accept.codex/runtime.rs:1626-1662forwards the decision to Codex; it does not change the review launch policy. - (measured) Claude is on the shared path. It gets
approval-requiredplusplan, not Codex sandbox fields.production/provider_runtime.rs:20639-20655gives interactionplanprecedence over runtime mode;:18600-18634launches the CLI in stream-json print mode with--permission-mode planand--setting-sources=user,project,local. In ordinary default interaction, runtime modes map todefault,acceptEdits, orbypassPermissions.provider/claude/runtime.rs:2528-2541turnsExitPlanModeoutput into review findings. A search ofapps/server/srcfound only the definition ofopen_permission_request(:2548), not a caller. (inferred) Identical approval-card behavior for Claude is therefore unproven; plan-mode refusals, plan-exit handling and configuration still need their own acceptance evidence. - (measured) Terminal contrast.
apps/web/src/components/chat/providerTerminalActions.ts:226-230,255-304replaces normal terminal bypass flags with Claude--permission-mode planor Codex--sandbox read-only --ask-for-approval never, then appends the prompt as an argument.ChatView.tsx:5027-5066opens that AI Terminal. Codex terminal reviews already request the no-prompt/read-only combination, but that does not prove chat or platform enforcement. - (measured) Normal chats and settings.
packages/contracts/src/orchestration.ts:128-137defines the three runtime modes and defaults to Full access.ChatView.tsx:1464-1466picks composer override → thread value → default. Plain panels useDEFAULT_RUNTIME_MODE(centerPanelActions.ts:114-118); side chats deliberately use approval-required (:120-125). The reviewed settings schema/UI has no global or review-specific default approval-mode setting. The MCP server access setting is a different authority boundary. - (measured) Intent and existing coverage. Living docs say chat reviews start Supervised (
docs/user/workspace-ui.md:2325-2351;docs/architecture/runtime-modes.md:7-25).apps/web/src/codeReview.test.ts:122-135asserts approval everywhere;centerPanelActions.test.ts:947-982asserts the same modes at panel creation and first send;apps/desktop/e2e/specs/source-control-review.e2e.ts:243-261expects the Supervised picker and uses a fixture provider, not live Codex. Historical designdocs/plans/2026-11-11-code-review-button-design.md:207-220(D41) explicitly accepts thatuntrustedmay prompt for read commands, as an agent's call overriding Source Control: a Review button starts a read-only agent that reviews the branch and uncommitted changes #293's initial default-mode recommendation. The helper already returnedapproval-requiredin639bd965f. - (measured: Claude) The
codeReviewflag is not yet a server policy. It exists in orchestration (apps/server/src/orchestration/engine.rs:540,:4929,:4988) butgit grepfinds it nowhere inproduction/provider_runtime.rsorprovider/. A server-enforced review policy must carry it into the provider session request.
Ranked causes
| Rank | Candidate and prediction | Evidence level |
|---|---|---|
| 1 | Explicit Supervised review policy selects Codex untrusted; replacing only approval behavior while retaining read-only enforcement should remove routine review prompts. |
(inferred; both agree) Strong source chain and consistent screenshot. The launch helper and provider mappings are measured; the installed run was not traced or changed. |
| 2 | Structural gap: CodexRuntimeMode ties approval and sandbox together, so chats cannot express never + readOnly, even though the Codex Terminal reviewer already requests that pair. |
(measured; both agree) The three mapping sites pair the existing modes that way. |
| 3 | The mixed shell invocation includes executable Node code that the provider does not consider trusted, or requests a sandbox escalation. Simpler safe reads might pass where this batch asks. | (guess) The screenshot shows command shape, not the provider's classification, script behavior, or escalation metadata. Codex source is not vendored. |
| 4 | acceptForSession may be scoped to the identical command, so “Always allow this session” would not cover the next, different pipeline. Provider version, loaded configuration or a manual picker change could also change the effective policy. |
(guess) No running configuration or Codex trusted-set was read. The explicit source default already explains why approvals are possible. |
(inferred; both agree) An accidental loss of a global Full access preference is not supported by the inspected path: the mode is deliberately assigned, and no such review preference is exposed. A purely stale Activity badge is also a weaker explanation than an actual approval request, because the screenshot contains the command decision card, not just a status label.
Spec
(inferred; agent's scope call; both agree on chat-review scope) This proposal changes chat reviews, including the shared Codex/Claude launch policy. Existing AI Terminal review behavior is comparison evidence and regression coverage only; extending its enforcement is outside this issue unless a regression appears. Ordinary chats, side chats, auto-fixing, Git writes, publishing, Cursor/Grok unsandboxed autonomy, and the companion issue (draft)'s button appearance are out of scope.
(inferred; proposal, not approved implementation design) Goal: one review action produces findings from permitted inspection without the user approving individual commands, while the reviewed checkout stays unmodified. Non-goals: auto-fixing, staging/committing/pushing, PR comments, changing normal-chat or side-chat defaults, replacing the model's judgment, or the companion issue (draft)'s button appearance.
AGENTS.md requires an approved design note before a non-trivial architectural change. This issue is the policy follow-up; an amendment to D41 must be approved before implementation.
R1 — P0: Autonomous Codex review
(inferred; proposed requirement, from owner intent) WHEN a user starts a supported Codex chat review, BiBCode SHALL run permitted inspection and deliver findings without waiting for command approval.
- (inferred; acceptance) GIVEN a dirty fixture checkout and a host in each of the three existing runtime modes, WHEN review starts from the button, shortcut, or one-off menu, THEN file reads and Git inspection complete, findings arrive, zero command approval cards appear, and the host mode stays unchanged. A complete-caller server test from a review panel (
codeReview: true) through session open, resume andturn/startasserts the chosen approval/sandbox payloads. Today there is no such pin. - (inferred; acceptance) GIVEN an unsupported server/provider capability, WHEN review is requested, THEN it is refused before execution with a reason, or it starts in today's Supervised mode; no silent Full access fallback or provider substitution occurs.
R2 — P0: Preserve the review boundary
(inferred; proposed requirement, from the existing read-only goal) WHILE an autonomous review runs, BiBCode SHALL prevent the review's tools and descendant processes from modifying the reviewed checkout or accessing host resources outside its declared inspection boundary.
- (inferred; acceptance) GIVEN a fixture containing tracked, staged and untracked changes, WHEN review attempts file edits, shell redirection, Git index/ref mutation, symlink escape, or a child process writing outside its permitted scratch area, THEN the operation is denied without approval, the protected content/index/refs remain identical, and findings identify any resulting evidence gap.
- (inferred; acceptance) GIVEN a custom prompt, repository instruction, hook or tool request asking for broader access, THEN it cannot disable the boundary or acquire unrestricted host filesystem/network access.
- (inferred; acceptance) GIVEN denied inspection, THEN the review uses other permitted evidence or returns explicitly incomplete findings; it does not repeatedly wait for permission or claim denied checks passed.
R3 — P0: Keep policy through the complete lifecycle
(inferred; proposed requirement) WHEN an autonomous review is retried, resumed or reconnected, BiBCode SHALL retain its review policy and checkout identity until it ends. Non-review threads SHALL keep their current behavior for each existing runtime mode.
- (inferred; acceptance) GIVEN a provider reconnect/restart between two inspection commands, THEN both run with the same no-prompt read-only policy; follow-up turns in the review retain it, and normal chats and side chats retain their chosen policy.
- (inferred; acceptance) GIVEN a double click, delayed panel creation, close-before-send, or failed first send, THEN at most one review starts, a closed panel sends no prompt, and unsent text remains recoverable without launching a more permissive session.
R4 — P1: Claude parity with truthful availability
(inferred; proposed requirement) WHEN Claude is offered for autonomous review, BiBCode SHALL either provide the same no-prompt inspection boundary or refuse the launch with a specific capability reason.
- (inferred; acceptance) GIVEN a supported Claude configuration, WHEN review reads the fixture, THEN findings arrive without command approvals, edits remain blocked and no Implement action appears, including after reload.
- (inferred; acceptance) GIVEN settings/hooks that would approve plan exit, allow edits or enable bypass, THEN those cannot weaken an advertised autonomous boundary. If protection cannot be established, the review action explains its unavailability before starting; merely using
bypassPermissionsdoes not pass.
R5 — P1: Truthful UI, docs and verification
(inferred; proposed requirement) WHEN the review policy changes, BiBCode SHALL describe the effective access and incomplete checks consistently in the reviewer tooltip, settings, composer and findings.
- (inferred; acceptance) Update the Codex guard line (
codeReview.ts:70and the docs/tests that assert it). Today it says writes ask for approval, which understates that read commands can ask too. After the change it should distinguish unattended inspection from Full access (for example, “Runs in Codex's read-only sandbox without asking. Writes are blocked.”). A rejected script is reported as not run rather than successful. - (inferred; acceptance) Update the living runtime/provider/workspace docs, D41, and affected testing runbooks; record
UI.mdandvercel-react-best-practicesreviews for changed React code. - (inferred; acceptance) The focused and broader gates listed below pass on the candidate change; failures and unavailable platform coverage are recorded, not waived by component-only assertions.
R6 — P2: Optional script execution and settings
(inferred; optional proposal) WHERE a review supports executing repository check scripts or tests, BiBCode SHALL run them only within the declared isolated tool boundary and report skipped checks as limitations. A later settings control may offer “Run on its own (read-only)” versus “Ask before commands.”
- (inferred; acceptance) A harmless Node fixture and a
node --testfixture can finish without prompts when supported; a fixture that attempts writes, network access or child-process escape is contained. If this execution capability is omitted, review still completes through R1's file/Git inspection and clearly says scripts/tests were not run.
Safety and alternatives
(inferred; both agree on the meaning of “fully autonomous”, agent's call) Starting review authorizes inspection, reasoning, permitted tool execution and findings without further command approvals. It does not authorize edits to the user's work, commits, publishing, installations, arbitrary credential reads, or unbounded execution on the host. Writes, if required for tool scratch or provider state, must be isolated from the checkout and explicitly bounded in the design. With approval never there is no escalation path, so a blocked write cannot be approved by mistake.
(inferred; risk; both agree) A node script or node --test executes unreviewed repository code; the name does not prove read-only behavior. An arbitrary bash -lc batch can execute substitutions, shell startup behavior and child processes. Auto-approving based on a cat, git diff or node prefix is insufficient. Prompt injection in the reviewed diff can steer that execution. Read-only filesystem access alone also does not settle network exfiltration, reads of secrets, resource exhaustion or mutation via connected tools. (guess) Codex's readOnly sandbox may disable network by default; confirm this for the supported Codex version before relying on it.
- (inferred; recommended direction; both agree) Separate no-prompt approval behavior from write authority. For Codex, request
neverplusreadOnlyconsistently at session start/resume and turn start, with a server-owned review policy and capability checks, matching the Codex Terminal reviewer. Verify the actual provider/platform boundary before advertising it. Claude needs separate enforcement evidence; plan mode alone is not an immutable sandbox. - (inferred; alternative) Offer only structured read operations and reject arbitrary code execution. This limits test-driven evidence but avoids granting host execution merely to remove prompts.
- (inferred; rejected shortcut; both agree) Reuse Full access, or auto-answer approval cards with
acceptForSession: it broadens authority or may run an approved command outside the sandbox, and does not establish the required read-only boundary. - (measured: Claude; fail-closed requirement)
runtime_mode()andclaude_mode()map an unknown string to full access (provider_runtime.rs:14688-14694,:20639-20647). Any new runtime-mode literal must be capability-gated so an older server never receives it. If the sandbox is unavailable (including Windows, listed as unverified in the design note at:153), fall back to today's Supervised mode or refuse; never to unsandboxed autonomy. - (inferred; design gate) Choose and approve the enforcement approach, allowed filesystem/network/tool scope and resource bounds before implementation. This draft specifies the outcome and safe failure behavior; it is not approval of a new sandbox architecture. Cursor and Grok reviews have no sandbox and stay as they are.
Tests to add and repository gates
(measured: source baseline above) Existing review tests encode the current Supervised policy. (inferred; proposed acceptance evidence) Update them deliberately and add complete-caller coverage:
- (inferred) Web tests:
codeReview.test.ts,centerPanelActions.test.ts,components/ChatView.hooks.test.tsx,components/SourceControlPanel.test.tsx, unchanged terminal argument regression tests, and a capability-gate test. Assert all chat entry points, exact panel and first-turn policy, unchanged host settings, duplicate suppression and failure recovery. - (inferred) Rust tests: panel creation/serialization, provider launch, Codex thread start/resume/turn-start payloads for review and non-review threads, Claude launch/plan transitions, unexpected approval handling, denied operation completion, and policy restoration across restart. Cover malformed/older-server capability input without privilege escalation. Today no test pins Codex
approvalPolicy/sandboxper mode. - (inferred) Extend
apps/desktop/e2e/specs/source-control-review.e2e.tsfor thread defaults and the guard line. That spec uses a fixture provider, so it cannot prove live Codex asks nothing. Isolated native-provider sandbox tests must separately prove prevented writes and tool escape. A mockedreadOnlyJSON field alone is not safety proof. - (inferred; future commands, not run here) From
apps/web:vp test run --project unit src/codeReview.test.ts src/centerPanelActions.test.ts src/components/ChatView.hooks.test.tsx src/components/SourceControlPanel.test.tsx src/components/chat/providerTerminalActions.test.ts. - (inferred; future gates, not run here) From the repository root:
cargo fmt --all --check;cargo test -p bibcode-server --lib;cargo clippy --workspace --all-targets -- -D warnings;vp run test;vp check;vp run typecheck; andvp run check:contractsif contracts change. - (inferred; future isolated native gate, not run here) After preparing disposable roots and fixtures per the native runbook:
BIBCODE_E2E_SPEC=./specs/source-control-review.e2e.ts vp run test:ui:desktop, with the required packaged build prepared byvp run test:ui:desktop:build. Never point this validation at production state. - (inferred) Review/update
docs/testing/cross-platform-validation.md:713-729, which currently requires the Supervised picker, and record platform-specific sandbox evidence for every advertised platform. Preserve supervised/attributed process lifecycle, cancellation and bounded cleanup. Record the required profiler coverage line without prompt/command content.
Workarounds today
These are current options, not the proposed fix:
- (measured: source) Choose Codex Terminal as the Code review agent (Settings → Agents → Code review, or the review menu). It already launches with
--sandbox read-only --ask-for-approval never. Actual platform enforcement was not exercised. - (inferred) Choose a Claude chat reviewer: it runs in plan mode, and BiBCode was not found to open a Claude permission-request card.
- (measured: source) Switching the review panel to Full access removes the prompts, but it also removes the sandbox (
dangerFullAccess). Not recommended for reviews.
Related and duplicate check
(measured) Read-only searches of open and closed issues, including not-planned closures. Queries included review, approval, permission, autonomous (0 hits), “full access”, “Review uncommitted”, “Source Control” review, command approval, and label:wontfix (none exist). No direct duplicate was found.
- (measured) Source Control: a Review button starts a read-only agent that reviews the branch and uncommitted changes #293, closed/completed — “Source Control: a Review button starts a read-only agent that reviews the branch and uncommitted changes.” Closest predecessor, not an unresolved duplicate. It introduced the button, reviewer settings, prompt and terminal read-only flags. Open question 1 is this runtime-mode choice; D41 later settled it as Supervised. Reference Source Control: a Review button starts a read-only agent that reviews the branch and uncommitted changes #293; do not reopen it or allege its first proposal is current behavior.
- (measured) Flaky on Windows arm64: review prompt shim tests time out after 10 seconds #367, closed/completed — flaky Windows arm64 review prompt shim tests. Same feature family, different defect.
- (measured) Chat: open a side chat that starts with the thread's context and leaves the thread untouched #265, closed/completed — side chats also start Supervised. Different action; do not change that default incidentally.
- (measured) Left panel: agent activity status (Working / Pending Approval / Awaiting Input / Done) is unreliable — review Orca's tracking and fix #187 and Review Activity (chat and terminal) against current provider docs #144, closed/completed — Activity status / Activity adapter audit, not Source Control review launch policy.
- (measured) Unattended CLI diagnostics: let bibcode pair with a remote environment and download its bundle directly #202, closed/completed — unattended diagnostics downloads. Different approval boundary.
- (measured) [Critical] Codex chat is fully broken once BiBCode's MCP is registered and enabled: mcp_servers.bibcode collision #273, Codex chat: stdout JSON-RPC line over 8 MiB closes the connection and exhausts delivery retries #316 and Codex: turn errors never match codexErrorInfo (camelCase on the wire) #162, closed/completed — Codex startup, transport and error-parsing defects, not a pending command approval.
- (measured) macOS measurements owed for #119 (Source Control rendering) and #96 (reconnect heap) #134 and Terminal re-attach: keep a server-side screen state instead of replaying a raw byte tail #350, closed/not-planned — performance measurements and terminal replay. Neither covers this policy.
- (inferred) No direct duplicate found. The separate button-appearance work remains the companion issue (draft).
Open questions
- (inferred; design owner, before implementation) Which enforceable boundary will provide the required autonomy: constrained provider sandboxing (
never+readOnly), structured read tools, or isolated repository execution? Set filesystem, credential, network, non-shell-tool and resource limits. Confirm whether CodexreadOnlyblocks network and whethernode --testmay write to a temp directory. - (inferred; product/design owner) Is optional script/test execution in R6 wanted in the first increment? The proposed minimum completes review from file/Git reads and reports unrun checks.
- (inferred; product owner; agent's call in this draft is “no”) Does “fully autonomous” also require a way to write temp or build output, for example a workspace-write sandbox on a throwaway copy of the checkout? This draft assumes no: tests that need to write fail and are reported.
- (inferred; product owner) What should the review panel's runtime picker show once the default is “read-only, no prompts”? Should OpenCode reviews become autonomous if its rules can deny edits without asking?
- (inferred; release owner) Decide whether Claude enforcement ships alongside Codex or remains explicitly unavailable for autonomous review until R4 is proved. Do not advertise parity on the basis of plan-mode arguments alone. On Windows, Codex sandbox support is listed as unverified; R1 must fall back to asking there, never to unsandboxed autonomy.
- (guess) Which commands does
untrustedlet through, and how far doesacceptForSessionextend, in the supported Codex version? This decides whether “every step” means every command or every non-trusted one. - (measured; evidence gap) The installed BiBCode and Codex versions for the screenshot are unknown.
- (inferred; triage) Codex preferred P2 and size:L; Claude preferred P1/medium and size:M (L if a new runtime-mode literal is chosen). This draft takes P2 and size:L.
Read-only work record
(measured) Two independent static diagnoses, one Codex (gpt-6-astra, high) and one Claude (claude-opus-5-5, medium), each on a fresh clone at 27a32e80998a983affdb244de7d232249963e0ed. Neither saw the other's output. Used git show/grep/log/diff and rev-parse, read-only gh issue list / gh issue view / gh label list, and worktree checks. No fetch was needed. The historical helper already hardcodes approval-required; no “used to be autonomous” regression is established.
(measured) No app, provider CLI session (beyond local codex --help in the Claude diagnosis), dev server, test, build, install, database, production state, configuration, credential, trace or profiler file was opened or run. No BiBCode process was inspected, contacted, signalled or restarted. CodeGraph refresh was skipped because it would write; no graph existed in this clone.
(measured) Validation commands listed above were not run, by the static-only constraint. No fix or successful runtime outcome is claimed. No issue/PR was filed, commented on or edited; only draft output artifacts were written outside the clone.
(measured) Final git rev-parse HEAD / git rev-parse origin/main still matched the recorded SHA after each diagnosis; the clone working tree stayed clean. The draft was checked for identifying paths, usernames and hostnames; none were included.
Activity
fix: close the open issue batch and the findings its validation showed
mubeda commentedon Nov 19, 2026
The batch pull request delivers the core of this issue and declines the rest. This issue closes with it; the declined parts are listed here with reasons so nothing is left implied.
Delivered
neverin Codex's read-only sandbox, the same policy as the Codex terminal review. The server decides this from the thread's code-review marker, in all three stored runtime modes, and the stored mode is not changed. If Codex still sends a command approval, a file-change approval or a question, BiBCode declines it at once and no card opens. An older server stays Supervised; nothing falls back to Full access.readOnlyCodexReviews, the guard line in the tooltip, the review menu and Settings reads "Runs on its own in Codex's read-only sandbox, like the Codex terminal review." Docs are updated.Declined
Verified with a real Codex (codex-cli 0.169.0), at protocol level only: the review policy is accepted on thread start and turn start, a read command ran with no approval request, and nothing was written. Not verified: a review through the packaged app with a real Codex, and any run on Windows; the automated tests use a fake Codex.