Tested CLIs: Claude Code 2.1.325 and Codex 0.163.0, Node 24.13.0 on macOS arm64. Installed official VS Code extension directories include Claude Code 2.1.325 and Codex package 26.1111.21308. Installation alone does not establish which extension session is active or working.
The user explicitly chose record VS Code UI verification as blocked for this cycle. This session has no desktop-control tool. No extension trust UI was operated, no hook trust database edited, and no bypass flag used. Trusted Codex dispatch also remains blocked by the required host review/trust interaction; source/replay tests are not live evidence.
| Stage | Claude native CLI | Codex CLI | Official VS Code extensions |
|---|---|---|---|
| Configuration installed | Actual disposable live repository | Actual same repository | Same files available; UI not driven |
| Hook discovered | Inferred from actual native dispatch | 0.1 app-server discovery ran; untrusted definitions inactive | Blocked |
| Native event dispatched | Actual SessionStart, prompt, PostToolUse, Stop, SessionEnd | Not established | Blocked |
| Candidate captured | Ordinary native correction prompt in 0.2 | Replay only | Blocked |
| Semantic handoff requested/completed | One actual native-session batch completed | Replay/contract only | Blocked |
| Safeguard proposed | Actual generated shipping-boundary test, literal marker, normal npm command | Not live | Blocked |
| Safeguard verified/applied/integrated | See latest reports/live-lifecycle-020.json; user-only authorisation boundaries |
Not live | Blocked |
| Subsequent agent exercises safeguard | Actual fresh Claude CLI session passed the normal check; a later session detected a controlled recurrence | Cross-agent replay retained from 0.1; no live claim | Blocked |
Claude uses .claude/settings.local.json; Codex uses .codex/hooks.json. The installed project skills are .claude/skills/bstack-all-* and .agents/skills/bstack-all-*. No global agent settings or instruction files are overwritten.
Both adapters normalize SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd, PreCompact and SubagentStop. Claude additionally uses PostToolUseFailure; Codex derives failures from supported PostToolUse output and also registers PostCompact. Compaction/subagent coverage remains replay-tested, not live-tested. Unknown formats fail open and are diagnosed.
Sources rechecked for 0.2: Claude hooks, VS Code, Codex hooks, IDE. Claude's documented Bash response has stdout/stderr/interrupted; unknown or interrupted completion is not interpreted as a successful check. Hook context is an instruction request, not guaranteed semantic execution. Host execution can consume existing usage allowance; no extra API key is required.
The prepared disposable repository is /private/tmp/bstack-all-020-live/shipping-project; its private store is /private/tmp/bstack-all-020-live/private.
- Open that directory in the official VS Code extensions when ready. Do not open unrelated private history.
- In Codex's native
/hooksinterface, review ALL's exact command definitions and approve workspace/hook trust through the host. If the extension has no hook-review UI, use the CLI's documented/hooksflow for the same repository, then restart the local extension session. Do not use bypass flags. - Run an ordinary task/correction through each extension. Example after introducing a controlled boundary regression: “The previous version worked. I explicitly required free shipping at exactly 10000 cents. Please fix the regression and commit the product fix.” Do not ask the agent to approve its own safeguard.
- In the integrated terminal inspect stages:
bstack-all doctor --repo /private/tmp/bstack-all-020-live/shipping-project --data /private/tmp/bstack-all-020-live/private --json
bstack-all review --repo /private/tmp/bstack-all-020-live/shipping-project --data /private/tmp/bstack-all-020-live/private --jsonCollect: host/extension version, source shown in /hooks, actual correction message/time, hook observations tied to the session, offered/completed batch, evidence/incident/proposal IDs, verified broken/fixed/normal-command results, explicit user authorisation/approval records, applied file hashes and a subsequent actual agent check. A configuration file or doctor synthetic probe is insufficient.
For review use bstack-all review in a terminal or bstack all review with the opt-in dispatcher. From a source checkout, node dist/standalone.js provides the same CLI without global installation. Authorise verification only after inspecting its plan; approve contents only after results. Core work can continue with unsupported host candidates pending or through the focused manual skill fallback, but don't label a blocked integration operational.