Repository navigation
Lane D: agent isolation delivery and handoff #28
Description
Owner: Codex Lane D implementation task. Baseline: 5eec4b5 (remote main). Build step 5 foundation; T1, T2, isolation portion of T9. Ordered PRs: D1 invocation contract and isolated clone; D2 restricted pinned container and startup probes; D3 vendor egress and phase enforcement; D4 adapters and bounded settlement; D5 combined real-Docker gate. No runner/store/UI rewiring.
D1 branch: codex/agent-isolation-d1. Exact files: agents/contract.ts, git/clone.ts, test/agent-contract.test.ts, test/agent-clone.test.ts. No overlap with PR #23. Done when: npx vitest run test/agent-contract.test.ts test/agent-clone.test.ts and npm run typecheck pass, then full unit validation on the pushed head.
Interface: immutable clone identity and trusted committed head; invocation includes phase, exact approved argv, finite deadline, attempt ID and captured context/version. Planning/questions cannot grant commands. Cancellation requests and process termination are distinct; terminal results preserve original reason and exit status. F owns current-context admission/publication and persisted compare-and-swap. D1 creates a staging clone only; D2 must relocate metadata onto a separate runner-owned filesystem and allocate bounded work storage before invocation. D1 never launches agents.
D1 dependencies: none. D2-D5 follow validated predecessor tasks on documented prerequisite branches if not yet merged. F and G production invocation remain blocked until D5 merges. Detailed evidence and review rounds will be recorded in each PR.
Activity
mchwang commentedon Sep 24, 2026
D1 implemented in PR #31 at validated head 489066e. Targeted tests: 24 passed; full unit suite: 214 passed; typecheck passed. Both push and PR CI are green, including browser tests. Copilot review is pending; no merge authorization has been given.
Next-step preflight: Docker Desktop server 27.4.0 is running (overlay2/extfs). Local CLIs report codex-cli 0.153.4 and Claude Code 2.1.281; these are observations, not selected image pins. Codex auth file is present. CLAUDE_CODE_OAUTH_TOKEN and ANTHROPIC_API_KEY are absent; no credential contents were read or published. The planned authenticated Claude container startup probe requires a locally configured setup-token. User input requested for that credential dependency. D2-D5 are not marked complete; no mock probe or skipped authentication check will count as acceptance.
mchwang commentedon Sep 24, 2026
D1 review round 1 complete: fixed both inline findings and reproduced/fixed summary validation and deadline concerns. All four have failing-before/passing-after regressions. Current validated head: d110d81. Current results: 28 targeted tests, 218 unit tests, typecheck, and both CI runs pass; review round 2 is pending. Both inline threads are resolved. PR #31 records the review-lesson audit. No merge performed. Authenticated Claude container-probe credential remains unavailable, so D2-D5 remain pending rather than being represented as validated.
mchwang commentedon Sep 24, 2026
D1 review loop converged in two Copilot rounds. Round 1 fixes have failing-before/passing-after regressions; round 2 returned Findings: None and confirmed resolved inline findings. Its unsupported-Windows overview concern was reproduced with win32.relative and deferred to #36, consistent with the plan scope. Current head d110d81: 28 targeted / 218 unit / 39 browser tests pass, typecheck passes, both CI runs green, mergeable, zero unresolved threads. Review-lesson audit is in PR #31. No merge performed.
Stopped at the authenticated container-probe credential dependency: no Claude setup-token is available to this task. D2-D5 remain unimplemented and no claims of container/network/adapter isolation are made. Configure CLAUDE_CODE_OAUTH_TOKEN locally (do not paste a credential into an issue/chat), then resume D2 on a documented prerequisite branch from #31 or merged main. Original checkout is untouched; D1 worktree is /Users/maxhwang/github/codeboost-lane-d1 on codex/agent-isolation-d1.
mchwang commentedon Sep 24, 2026
D2 owner: Codex Lane D implementation task. Branch: codex/agent-isolation-d2, stacked on D1 validated head d110d81 because D1 is not merge-authorized. Baseline remote main is now b181d15; D2 will remain based on D1 and will revalidate after prerequisite integration. Exact proposed owned files before edit: agents/container/Dockerfile, agents/container/image.ts, agents/container/profile.ts, agents/container/probe.sh, agents/container/run.ts, test/agent-container.test.ts, test/fixtures/agent-container/**, .github/workflows/agent-isolation.yml. No runner/store/UI/common-CI/package-manifest changes. Interface: consumes D1 captured InvocationInput and prepared task-clone staging; produces a lockdown-validated container profile. D2 does not implement vendor egress (D3), adapters/cancellation/bounded output (D4), or combined gate (D5). Done when: real-Docker tests prove pinned image/startup lockdown, separated read-only Git metadata, phase-correct /work access, non-root/no capabilities/read-only root, ephemeral bounded scratch/HOME, absent host paths, Git status, and authenticated Codex/Claude startup probes. If Docker Desktop cannot enforce the planned persistent byte+inode task filesystem, startup must fail closed and the limitation will be recorded rather than mocked.
mchwang commentedon Sep 24, 2026
D2 opened as stacked PR #40 on D1 at validated head 17d71d0. Exact-head local evidence: authenticated real-Docker 9/9 (both providers read the mounted hidden schema value), typecheck passed, full unit suite 225/225 across 14 files. PR is mergeable with zero unresolved threads; four CI jobs are running. Full current-diff self-review found no remaining issues, and Copilot round 1 was requested. D3-D5 remain explicitly deferred in the PR; normal D2 profiles default to no network.
mchwang commentedon Sep 24, 2026
D2 is review-converged in PR #40 at a4bd301. Exact-head validation: authenticated Docker 15/15, full suite 231/231, typecheck and all four CI checks pass; nine review rounds ended with Findings: None and no unresolved threads. No merge is authorized. Starting D3 as a separate stacked PR.
mchwang commentedon Sep 24, 2026
D3 is implemented in PR #44, stacked on D2 / #40. Exact pushed head 353efb0 passed typecheck, the full suite (245/245), and authenticated Codex/Claude container probes (20/20) through vendor-only egress. CI and the required review loop are now in progress; no merge authorization has been given.
mchwang commentedon Sep 24, 2026
D3 is implemented in #44 and review-converged on validated head
981e831637c1d19c3b4c759e18cc5ec06e8a5f33.Proceeding to D4 as a separate stacked PR.
mchwang commentedon Sep 28, 2026
New lane D follow-up: #75. E4 (#45) recorded real planning output through the D4 adapters, and none of it passes E2 validation. The planning commands do not pass the answer schema (
--json-schema/--output-schema), and withfeatures.shell_tool=falseCodex has no way to read files. #75 blocks E4 and asks for a decision on how Codex reads code in read-only phases.🤖 Generated with Claude Code
mchwang commentedon Oct 20, 2026
Lane D is done. D1–D5 merged (#31, #40, #44, #47, #50). Follow-ups merged since then: #66 (#85, #90), #76, #81/#99 (#100), #93 (#106), #95, #97, and the closed #51 runner-contract items (#109). Recorded in docs/architecture.md, lane D row.