[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7DDFZHTWG3J589038F09S74":3,"skill-same:01M7DDFZHTWG3J589038F09S74":44},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":13,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":29,"install":30,"resources":43},"01M7DDFZHTWG3J589038F09S74","delegating-to-codex",null,"Use when authoring a non-interactive `codex exec` invocation or deciding whether to hand a task from Claude Code to OpenAI Codex CLI. Provides the validated autonomous invocation, the sandbox\u002Fapproval ladder, the JSON event schema for triage, a when-to-delegate framework, failure modes, and one empirical trial. Trigger phrases include \"delegate to codex\", \"codex exec\", \"hand this to codex\", \"run codex\", \"codex --full-auto\", \"ask codex\", \"second opinion from codex\", \"have codex use a skill\", \"run a skill in codex\", \"codex pr-review\". Do NOT use for interactive pair-programming in the Codex TUI, or for trivial CRUD\u002FUI\u002Ffirst-attempt tasks where a model handoff adds no value. For autonomous Claude (not Codex), use `delegating-to-claude-p` instead.","AgentParadise\u002Fagentic-workspace","AgentParadise","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F242330113?v=4","01M6C52MXXSE01NH5GKMC3N7JA","01M7DDFCPP7TW5GGJZYGFPKMK2",0,"MIT","https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace\u002Fblob\u002Fmain\u002Fplugins\u002Fdelegation\u002Fskills\u002Fdelegating-to-codex\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: delegating-to-codex\ndescription: Use when authoring a non-interactive `codex exec` invocation or deciding whether to hand a task from Claude Code to OpenAI Codex CLI. Provides the validated autonomous invocation, the sandbox\u002Fapproval ladder, the JSON event schema for triage, a when-to-delegate framework, failure modes, and one empirical trial. Trigger phrases include \"delegate to codex\", \"codex exec\", \"hand this to codex\", \"run codex\", \"codex --full-auto\", \"ask codex\", \"second opinion from codex\", \"have codex use a skill\", \"run a skill in codex\", \"codex pr-review\". Do NOT use for interactive pair-programming in the Codex TUI, or for trivial CRUD\u002FUI\u002Ffirst-attempt tasks where a model handoff adds no value. For autonomous Claude (not Codex), use `delegating-to-claude-p` instead.\nplacement: \"Domain skill. Lives at `plugins\u002Fdelegation\u002Fskills\u002Fdelegating-to-codex\u002F` in agentic-primitives. NOT in `.claude\u002Fskills\u002F`; that scope is for meta skills.\"\n---\n\n# Delegating to `codex exec`\n\n## Purpose\n\nInvoke this skill when you are about to dispatch a task from Claude Code to the\nOpenAI Codex CLI in non-interactive mode, or when deciding whether a task is\nworth a cross-model handoff at all. It is the Codex peer of\n[`delegating-to-claude-p`](..\u002Fdelegating-to-claude-p\u002FSKILL.md): both dispatch\nwork to an autonomous agent one-shot; the steering surface and the footguns\ndiffer, and this skill documents Codex's.\n\nThe single most important difference from `claude -p`: **Codex has no built-in\ncost cap.** `claude -p` enforces `--max-budget-usd`; `codex exec` has no\nequivalent. The only hard bound is an external one (see the budget section).\n\n## Captured workflow delegation\n\nWhen `AGENTIC_SESSION_STORE_PROVIDER=local`, use the installed structured shim:\n\n```sh\nsyn-delegate codex --prompt=\"$TASK_PROMPT\" --timeout 600\n```\n\nKeep the `=` in `--prompt=\"$TASK_PROMPT\"`: a prompt that starts with `-` is\notherwise read as an option and the shim exits 2 without launching. The shim\npasses the prompt to the harness after `--`, so it can never become a CLI flag.\n\nUse `--model` when selecting a model explicitly. Run from the intended working\ndirectory. The shim always passes Codex an explicit `--sandbox`: `workspace-write`\nby default, or `read-only` via `--sandbox read-only` or\n`AGENTIC_DELEGATE_CODEX_SANDBOX=read-only`. Any other value is refused before\nlaunch. Before every launch it probes the Codex sandbox live, in the same mode,\ndirectory and environment Codex will get, and launches only if the probe\npasses (see the ladder below); otherwise it records `launch_failed` and exits\n69. The shim grants no other permissions. It records intent before launch, binds the delegate's own\nnative session ID, and preserves its actual exit status even when a caller uses\n`|| true` or a pipeline. Claude's shell hook supplies exact parent context;\nCodex supplies its native `CODEX_THREAD_ID`. Missing parent context or durable\nstorage denies launch. Do not fabricate these context values or fall back to a\nraw CLI to bypass a capture failure.\n\nThe raw CLI examples below describe uncaptured use and underlying harness\noptions. In captured workflows, use the shim so the run can discover this launch.\n\n## The validated invocation\n\n```sh\ncodex exec -s workspace-write \\\n  --json \\\n  -o .\u002Fcodex-last.txt \\\n  -C \u002Fpath\u002Fto\u002Fproject \\\n  \"$TASK_PROMPT\" \\\n  \u003C \u002Fdev\u002Fnull\n```\n\n**`\u003C \u002Fdev\u002Fnull` is not optional.** `codex exec` reads stdin *in addition to* the\nprompt argument. Launched from any context whose stdin never reaches EOF (a\nbackground shell, a CI step, an agent harness), it prints\n`Reading additional input from stdin...` and **hangs forever** - no events, no\nerror, no tokens consumed, just a wedged process until something kills it.\nOmitting it is the single most common way these invocations fail.\n\n> **Flag drift.** Older releases spelled the autonomy mode `--full-auto`. That\n> flag was **removed** by `codex-cli 0.147.0` — passing it now aborts the run\n> with `error: unexpected argument '--full-auto'`. Use `-s workspace-write`\n> instead, as above. `codex exec --help | grep full-auto` returning nothing is\n> the quick check. Everything else in this recipe (`--json`, `-o`, `-C`,\n> `--ephemeral`, `--add-dir`, `--output-schema`, `--skip-git-repo-check`) is\n> unchanged and verified present on 0.147.0.\n\n### Per-flag rationale\n\n- **`-s workspace-write` \u002F `--sandbox workspace-write`** — the realistic\n  autonomy mode: Codex may write files and run commands inside the workspace\n  sandbox, and `codex exec` does not prompt. This is the Codex analog of\n  `claude -p --permission-mode bypassPermissions`, but narrower — it does *not*\n  grant network access or writes outside the workspace. (`--ask-for-approval`\n  is no longer an `exec` flag; `--approve-for-me` routes approval requests\n  through automatic review using the same sandbox.)\n- **`--json`** — emit the run as a JSONL event stream on stdout. Without it the\n  transcript is prose and unscoreable, exactly as `claude -p` text mode is. See\n  the event-schema section for what it carries.\n- **`-o \u003Cfile>` \u002F `--output-last-message \u003Cfile>`** — capture the agent's final\n  message to a file, so a wrapper can read the conclusion without parsing the\n  whole stream.\n- **`-C \u003Cdir>` \u002F `--cd \u003Cdir>`** — set the working root explicitly. Safer than\n  relying on the caller's CWD when invoked from a wrapper.\n- **`\u003C \u002Fdev\u002Fnull`** - close stdin so Codex does not block waiting for input it\n  will never receive. See the warning above; this is the highest-frequency\n  failure mode in practice.\n- Useful additions: **`-m \u003Cmodel>`** to pin the model (otherwise the config\n  default is used), **`--output-schema \u003Cfile>`** to force a structured final\n  response, **`--skip-git-repo-check`** to run outside a git repo.\n- **`--ephemeral`** - **omitted deliberately**, for the same reason\n  `--no-session-persistence` is omitted from the `claude -p` recipe. It\n  persists no session, so the delegated run leaves no on-disk rollout for a\n  session-store sweep to collect. Persistence is a property of DELEGATION, not\n  of one CLI: suppressing it on either side loses the same thing.\n  The narrow case where it is right is explicitly throwaway local work, where\n  you intend not to keep a resumable transcript. If a run persists something\n  you did not want kept, fix that in the session-store or `$HOME`\n  configuration; do not reach for this flag to work around it.\n\n## The sandbox \u002F approval ladder\n\nPick the least privilege that lets the task finish. Granularity Codex exposes:\n\n| Setting | Grants | Use when |\n|---|---|---|\n| `-s read-only` | Read files; no writes, no commands | Analysis \u002F review only |\n| `-s workspace-write` | Write + run inside the workspace; no network | The default for fmt\u002Ftest\u002Fedit loops |\n\nNothing above `workspace-write` is on the ladder. Modes and flags that switch\nCodex's sandbox off are not used by these recipes, and `syn-delegate` refuses\nthem (`--sandbox` accepts only `read-only` and `workspace-write`). When writes\nmust reach a directory outside the workspace, add `--add-dir \u003Cdir>` instead.\n\n### Inside a workspace container: keep the sandbox, fix the container\n\n`-s workspace-write` makes Codex sandbox itself with **bubblewrap**, and\nbubblewrap needs an unprivileged user namespace. Docker's default seccomp\nprofile denies that to a container without `CAP_SYS_ADMIN`, so in a workspace\nstarted with the default profile every sandboxed operation fails:\n\n```\nwarning: Codex could not find bubblewrap on PATH ... will use the bundled\n         bubblewrap in the meantime\nbwrap: No permissions to create a new namespace, likely because the kernel does\n       not allow non-privileged user namespaces.\nFailed to write file \u002Fworkspace\u002Fpalindrome.py\n```\n\nRead that stack carefully: the missing-bwrap line is a **warning** and Codex\ncontinues with a bundled copy. The namespace denial underneath it is the real\nfault, and the write failure is only the symptom. Diagnosing from the first\nplausible line sends you to install bubblewrap, which does not help. Worse,\n`codex exec` still exits 0, so a caller that trusts the exit status records a\nrun that did nothing as completed.\n\nThe fix belongs to the workspace, not the invocation. `agentic_isolation`\napplies the Codex sandbox policy to every image labelled\n`agentic.codex_cli_version` (the Codex-capable workspace images are). It adds a\nnarrow seccomp profile (Docker's default plus `clone`\u002F`unshare` for the user,\nmount, pid, net and ipc namespaces, and `mount`, `umount2`, `pivot_root`)\nwhile `--cap-drop=ALL`, `no-new-privileges` and the read-only root stay in\nforce. On AppArmor hosts (Ubuntu 24.04, where the symptom is\n`bwrap: Failed to make \u002F slave: Permission denied`) it also applies the\nAppArmor profile `agentic-codex-sandbox`, which the host administrator loads\nonce with `sudo apparmor_parser -r`; it admits only the mounts bubblewrap\nperforms, so Codex must run from a directory under `\u002Fworkspace`. With both,\n`-s workspace-write` works as designed: writes inside the workspace succeed\nand writes outside it are denied.\n\n**Nested delegation.** A `workspace-write` delegate gets exactly two things\nbeyond Codex's defaults, as `-c sandbox_workspace_write.*` overrides that\nreplace any `CODEX_HOME` value: two extra writable roots, the retained child\njournal's partition directory and the partition's Claude transcript root\n(`$SPOOL\u002F$PARTITION\u002Fclaude`), and `network_access = true` (egress is still the\ncontainer's network policy). That is what lets the delegate run\n`syn-delegate` itself (Claude -> Codex -> Claude) with the Claude grandchild's\nown transcript captured. Nothing else outside the working directory becomes\nwritable, so:\n- A Codex grandchild cannot start its sandbox (its `CODEX_HOME` is read-only),\n  so it is refused with `codex_sandbox_unavailable`.\n- A `read-only` delegate cannot write the journal at all. A `syn-delegate` it\n  runs starts nothing, exits 70, and prints an\n  `agentic-delegate-launch-denied\u002Fv1` line; the enclosing `syn-delegate` reads\n  that line from its child's output and records `launch_failed` with reason\n  `nested_journal_unavailable` under the Codex child, so the denial is a\n  visible gap, not silence.\n\n`syn-delegate codex` probes `codex sandbox` with the same mode and grant\n(writing inside the extra root) live before every launch and refuses when it\nfails: it records `launch_failed`\nwith reason `codex_sandbox_unavailable` and exits 69, rather than running a\nCodex that cannot execute a single command. There is no status file to trust.\nThe entrypoint logs the same probe at startup (`[entrypoint] codex sandbox:`)\nfor diagnosis only. If you hit the refusal, the workspace lacks the policy or\nyou are outside `\u002Fworkspace`; do not route around it with a raw `codex exec`.\n\n## Budget: there is no built-in cap\n\n`codex exec` reports token `usage` only in the final `turn.completed` event —\n*after* the work is done — and surfaces no USD figure under ChatGPT-subscription\nauth. There is no `--max-budget-usd`. Therefore the bound must come from outside\nthe process:\n\n```sh\n# macOS lacks `timeout`; install coreutils for `gtimeout`, or use a CI step limit\ngtimeout 600 codex exec -s workspace-write --json -o .\u002Fcodex-last.txt \"$TASK_PROMPT\" \u003C \u002Fdev\u002Fnull\n```\n\nTreat an external wall-clock bound (`gtimeout`, a CI job timeout, a wrapper\nwatchdog) as mandatory for unattended runs. A wedged Codex run will otherwise\nconsume tokens until it finishes or is killed by hand.\n\n## `codex exec review` — the built-in diff reviewer\n\n0.147.0 ships a `review` subcommand that scopes itself to a diff instead of\nneeding you to describe one:\n\n```sh\ncodex exec review --base main --json -o .\u002Freview.md \u003C \u002Fdev\u002Fnull\ncodex exec review --uncommitted --json -o .\u002Freview.md \u003C \u002Fdev\u002Fnull\ncodex exec review --commit \u003CSHA> --json -o .\u002Freview.md \u003C \u002Fdev\u002Fnull\n```\n\n**The catch:** `--base` \u002F `--uncommitted` \u002F `--commit` are **mutually exclusive\nwith `[PROMPT]`**. Passing both fails:\n\n```\nerror: the argument '--base \u003CBRANCH>' cannot be used with '[PROMPT]'\n```\n\nSo you get either Codex's stock review rubric *or* your own instructions, never\nboth. It also exposes no `-s` or `-C`, so it is inherently read-only and runs\nagainst the current directory.\n\nChoose accordingly:\n\n| Want | Use |\n|---|---|\n| A quick review with Codex's own rubric | `codex exec review --base \u003Cbranch>` |\n| A review steered at specific risks, or one that runs the test suite to verify its own claims | plain `codex exec -s workspace-write` with a prompt telling it to `git diff \u003Cbase>...HEAD` |\n\nFor the steered form, add an explicit **\"do NOT modify any tracked file\"** to the\nprompt — `workspace-write` permits edits, and you want the review read-only while\nstill letting Codex run tests. Confirm with `git status` afterwards; treat a dirty\ntree as a failed review run, not a bonus.\n\n## What `--json` emits (for triage)\n\nThe stream is JSONL, one event per line. Event shapes observed:\n\n- `thread.started` — `{thread_id}` for the run.\n- `turn.started` — a turn begins.\n- `item.started` \u002F `item.completed` — each agent action. `item.type` is one of\n  `agent_message`, `command_execution` (carries `command`, `aggregated_output`,\n  `exit_code`, `status`), or `file_change`.\n- `turn.completed` — carries `usage` (`input_tokens`, `cached_input_tokens`,\n  `output_tokens`, `reasoning_output_tokens`).\n\nTo triage a run: filter `command_execution` items for non-zero `exit_code` to\nsee what Codex tried and where it stumbled; read the final `agent_message` (or\nthe `-o` file) for the conclusion; read `turn.completed.usage` for token spend.\n\n## When to delegate to Codex\n\nDelegate when a cross-model perspective is likely to help:\n\n- A bug survived **2+ of your own attempts** — a different model may break the\n  fixation.\n- Logic- or algorithm-heavy backend work with many edge cases.\n- You want an independent second implementation to compare against.\n\nDo **not** delegate:\n\n- Trivial CRUD, UI scaffolding, config edits, or first attempts — the handoff\n  overhead exceeds the value.\n- Tasks needing genuine multi-turn clarification — `codex exec` is one-shot\n  (use the interactive Codex TUI, or stay in Claude Code).\n- Work that must not run unsandboxed where you cannot grant `workspace-write`.\n\nFor autonomous *Claude* rather than Codex, use `delegating-to-claude-p`. The two\nare interchangeable dispatch targets; pick Codex specifically when you want a\ndifferent model's reasoning.\n\n## Using a Claude skill (e.g. pr-review) with Codex\n\nCodex does **not** auto-dispatch Claude `SKILL.md` plugin skills — it consults no\nskills index. To make Codex follow a specific skill, **inject the skill's content\nand name it in the prompt.** Two mechanisms:\n\n- **stdin (best for one-shot)** — pipe the skill body in; Codex appends it as a\n  `\u003Cstdin>` block. This is the one case where you do NOT redirect from\n  `\u002Fdev\u002Fnull`: the pipe supplies stdin and then closes it, which is exactly what\n  Codex is waiting for. Name the skill and the task in the prompt arg:\n  ```sh\n  cat path\u002Fto\u002Fskills\u002Freview\u002FSKILL.md | codex exec -s workspace-write --json \\\n    -o .\u002Freview.md -C \u002Fpath\u002Fto\u002Fproject \\\n    \"Apply the review skill provided on stdin to review this repo against plan.md. Follow its Report format exactly. Read-only: do not modify files.\"\n  ```\n- **`AGENTS.md` (best when reused across runs)** — drop the skill content into an\n  `AGENTS.md` at the workspace root; Codex auto-loads it. Still name the skill's\n  action in the prompt — like `claude -p`, a one-shot agent binds to explicit\n  prompt verbs, not to passively-present docs.\n\nSkills that mandate a **structured output contract** (a pr-review rubric, a\nverdict format) transfer especially well: the contract is self-describing, so\nCodex reproduces it faithfully. Inject the *whole* skill body, not a paraphrase —\nthe format section is what steers the output. See Trial T2.\n\n## Failure modes\n\n| Symptom | Cause | Fix |\n|---|---|---|\n| Hangs forever, zero events, `Reading additional input from stdin...` | stdin never reaches EOF and Codex waits on it | Add `\u003C \u002Fdev\u002Fnull`. Detect it by the `--json` stream file stalling at ~39 bytes, the length of that banner; a healthy run passes 100KB within a minute |\n| Run consumes tokens indefinitely | No built-in budget cap | Wrap in `gtimeout`\u002FCI timeout; `usage` is only reported post-hoc |\n| Unscoreable prose transcript | Default output is not JSONL | Add `--json`; read `-o` file for the conclusion |\n| `error: unexpected argument '--full-auto'` | Flag removed by 0.147.0 | Use `-s workspace-write`; re-check with `codex exec --help` |\n| `--base \u003CBRANCH> cannot be used with [PROMPT]` | `codex exec review` takes its built-in rubric OR custom instructions, not both | Pick one: `review --base` for the stock rubric, or plain `codex exec` with your own prompt |\n| Cannot write files \u002F run commands | Default sandbox is `read-only` | Set `-s workspace-write` explicitly |\n| Refuses to run | Not inside a git repo | Add `--skip-git-repo-check` |\n| Writes need a dir outside the workspace | `workspace-write` is workspace-scoped | Add `--add-dir \u003Cdir>`; keep the sandbox mode as it is |\n| `bwrap: No permissions to create a new namespace`, every write fails, exit 0 | Workspace runs Docker's default seccomp profile | Use an image labelled `agentic.codex_cli_version` through `agentic_isolation`, which applies the Codex sandbox policy |\n| `bwrap: Failed to make \u002F slave: Permission denied` | AppArmor host, workspace confined by `docker-default` | Load `agentic-codex-sandbox` on the host (`sudo apparmor_parser -r`) |\n| `bwrap: Can't bind mount ... Permission denied` | Codex started outside `\u002Fworkspace` on an AppArmor host | Run from a directory under `\u002Fworkspace` |\n| `syn-delegate` exits 69, `launch_failed` \u002F `codex_sandbox_unavailable` | The live probe found the Codex sandbox unusable | Same fixes; the stderr line names the bwrap cause |\n| `syn-delegate` exits 2, `--prompt: expected one argument` | Prompt starts with `-` | Use `--prompt=\"$TASK_PROMPT\"` |\n| Nested `syn-delegate` exits 70, `agentic-delegate-launch-denied\u002Fv1` on stderr | Run inside a `read-only` delegate (journal not writable) | Delegate with `workspace-write` if the child must delegate further |\n| `syn-delegate claude` exits 69, `claude_nested_auth_unavailable` | Claude below a Claude Bash call with OAuth only | Supply `ANTHROPIC_API_KEY` or a credentials file, or delegate from Codex |\n| Command fails on a quirky local env (e.g. `pytest` exit 127 under pyenv) | Sandbox shell inherits host PATH quirks | Under `-s workspace-write` Codex often self-recovers (Trial T1); for wrappers, pre-set env via `-c shell_environment_policy...` |\n\n## Trial T1 — empirical reference\n\nA real workspace-write run, recorded 2026-06-08 on `codex-cli 0.137.0`, where\nthe mode was still spelled `--full-auto`. The invocation below is left as it was\nactually executed; on 0.147.0+ substitute `-s workspace-write`.\n\n- **Task**: add `is_palindrome(s)` + a 3-case `pytest` file to a fresh repo, run\n  the tests.\n- **Invocation**: `codex exec --full-auto --json -o \u003Cfile> \"\u003Ctask>\"` (model =\n  config default).\n- **Result**: exit 0, **44s** wall-clock, **1 turn**, 18 items\n  (8 `agent_message`, 8 `command_execution`, 2 `file_change`), tests passed.\n- **Usage** (`turn.completed`): input 237,408 tokens (202,112 cached, ~85% cache\n  hit), output 1,139, reasoning 134. No USD figure surfaced under ChatGPT auth.\n- **Autonomous recovery**: the first `pytest` call failed (`exit_code 127` —\n  pyenv shadowing the binary); Codex retried with\n  `PYENV_VERSION=3.12.0 python -m pytest` and got `3 passed`. The `--json`\n  stream captured both the failed and the recovered command, which is exactly\n  the signal you triage on.\n\nThis is a single data point, not a universal claim. Add your own paired trials\nbefore quoting costs or wall-clock for a different task shape.\n\n## Trial T2 — does an injected skill actually steer Codex?\n\nA paired run recorded 2026-06-08, testing the injection mechanism above. Target:\na scratch repo with a 2-milestone `plan.md` whose milestone 2 (`truncate`) was\ndeliberately left unimplemented, so a correct review must flag the omission.\n\n- **Run A (skill on stdin)**: piped `sdlc\u002Fskills\u002Freview\u002FSKILL.md`; the prompt\n  named it and asked for its Report format. Codex reproduced the skill's exact\n  contract — `## Implementation Review`, the Milestone Status table, ⚠️ Omissions\n  (Blockers), and the `❌ BLOCKERS FOUND` verdict — correctly flagged the missing\n  milestone, and obeyed the skill's read-only\u002Fstatic-review process. 35s; 82,277\n  input tokens (87% cached) \u002F 1,406 output.\n- **Run B (baseline, no skill)**: same task, generic \"review against plan.md\".\n  Codex returned generic `**Findings**` prose — it found the same facts, but with\n  **none** of the skill's structure, and it ran the tests on its own (a different\n  process). 45s; 131,776 input \u002F 1,501 output.\n- **Adherence check**: all four signature markers\n  (`## Implementation Review`, `### Milestone Status`, `Omissions (Blockers)`,\n  `BLOCKERS FOUND`) were present in A and absent in B.\n\nTakeaway: injecting a skill and naming it in the prompt **does** steer Codex's\noutput format and process — but the content must be supplied, since Codex won't\nauto-dispatch it. Finding *fidelity* was similar in both runs because the bug was\nobvious; skills earn their keep on **structure and process discipline**, which is\nexactly what a pr-review skill enforces.\n\n## References\n\n- **`delegating-to-claude-p`** (sibling skill) — the autonomous-Claude dispatch\n  recipe and its empirically-validated flag set; read it for the side-by-side on\n  what steers a one-shot agent.\n- **`experiments:running-experiments`** — the discipline for designing a paired\n  trial to extend or contest Trial T1.\n- Codex CLI flags here are described by behavior to stay valid across releases;\n  verified against `codex-cli 0.137.0` on 2026-06-08, re-verified against\n  `0.147.0` on 2026-08-17 (which removed `--full-auto` and added\n  `codex exec review`). Re-run `codex exec --help` to confirm flag availability\n  in your installed version — this surface has already drifted once between\n  documented releases, so treat the check as part of the recipe rather than an\n  optional precaution.\n","# Delegating to `codex exec`\n\n## Purpose\n\nInvoke this skill when you are about to dispatch a task from Claude Code to the\nOpenAI Codex CLI in non-interactive mode, or when deciding whether a task is\nworth a cross-model handoff at all. It is the Codex peer of\n[`delegating-to-claude-p`](..\u002Fdelegating-to-claude-p\u002FSKILL.md): both dispatch\nwork to an autonomous agent one-shot; the steering surface and the footguns\ndiffer, and this skill documents Codex's.\n\nThe single most important difference from `claude -p`: **Codex has no built-in\ncost cap.** `claude -p` enforces `--max-budget-usd`; `codex exec` has no\nequivalent. The only hard bound is an external one (see the budget section).\n\n## Captured workflow delegation\n\nWhen `AGENTIC_SESSION_STORE_PROVIDER=local`, use the installed structured shim:\n\n```sh\nsyn-delegate codex --prompt=\"$TASK_PROMPT\" --timeout 600\n```\n\nKeep the `=` in `--prompt=\"$TASK_PROMPT\"`: a prompt that starts with `-` is\notherwise read as an option and the shim exits 2 without launching. The shim\npasses the prompt to the harness after `--`, so it can never become a CLI flag.\n\nUse `--model` when selecting a model explicitly. Run from the intended working\ndirectory. The shim always passes Codex an explicit `--sandbox`: `workspace-write`\nby default, or `read-only` via `--sandbox read-only` or\n`AGENTIC_DELEGATE_CODEX_SANDBOX=read-only`. Any other value is refused before\nlaunch. Before every launch it probes the Codex sandbox live, in the same mode,\ndirectory and environment Codex will get, and launches only if the probe\npasses (see the ladder below); otherwise it records `launch_failed` and exits\n69. The shim grants no other permissions. It records intent before launch, binds the delegate's own\nnative session ID, and preserves its actual exit status even when a caller uses\n`|| true` or a pipeline. Claude's shell hook supplies exact parent context;\nCodex supplies its native `CODEX_THREAD_ID`. Missing parent context or durable\nstorage denies launch. Do not fabricate these context values or fall back to a\nraw CLI to bypass a capture failure.\n\nThe raw CLI examples below describe uncaptured use and underlying harness\noptions. In captured workflows, use the shim so the run can discover this launch.\n\n## The validated invocation\n\n```sh\ncodex exec -s workspace-write \\\n  --json \\\n  -o .\u002Fcodex-last.txt \\\n  -C \u002Fpath\u002Fto\u002Fproject \\\n  \"$TASK_PROMPT\" \\\n  \u003C \u002Fdev\u002Fnull\n```\n\n**`\u003C \u002Fdev\u002Fnull` is not optional.** `codex exec` reads stdin *in addition to* the\nprompt argument. Launched from any context whose stdin never reaches EOF (a\nbackground shell, a CI step, an agent harness), it prints\n`Reading additional input from stdin...` and **hangs forever** - no events, no\nerror, no tokens consumed, just a wedged process until something kills it.\nOmitting it is the single most common way these invocations fail.\n\n> **Flag drift.** Older releases spelled the autonomy mode `--full-auto`. That\n> flag was **removed** by `codex-cli 0.147.0` — passing it now aborts the run\n> with `error: unexpected argument '--full-auto'`. Use `-s workspace-write`\n> instead, as above. `codex exec --help | grep full-auto` returning nothing is\n> the quick check. Everything else in this recipe (`--json`, `-o`, `-C`,\n> `--ephemeral`, `--add-dir`, `--output-schema`, `--skip-git-repo-check`) is\n> unchanged and verified present on 0.147.0.\n\n### Per-flag rationale\n\n- **`-s workspace-write` \u002F `--sandbox workspace-write`** — the realistic\n  autonomy mode: Codex may write files and run commands inside the workspace\n  sandbox, and `codex exec` does not prompt. This is the Codex analog of\n  `claude -p --permission-mode bypassPermissions`, but narrower — it does *not*\n  grant network access or writes outside the workspace. (`--ask-for-approval`\n  is no longer an `exec` flag; `--approve-for-me` routes approval requests\n  through automatic review using the same sandbox.)\n- **`--json`** — emit the run as a JSONL event stream on stdout. Without it the\n  transcript is prose and unscoreable, exactly as `claude -p` text mode is. See\n  the event-schema section for what it carries.\n- **`-o \u003Cfile>` \u002F `--output-last-message \u003Cfile>`** — capture the agent's final\n  message to a file, so a wrapper can read the conclusion without parsing the\n  whole stream.\n- **`-C \u003Cdir>` \u002F `--cd \u003Cdir>`** — set the working root explicitly. Safer than\n  relying on the caller's CWD when invoked from a wrapper.\n- **`\u003C \u002Fdev\u002Fnull`** - close stdin so Codex does not block waiting for input it\n  will never receive. See the warning above; this is the highest-frequency\n  failure mode in practice.\n- Useful additions: **`-m \u003Cmodel>`** to pin the model (otherwise the config\n  default is used), **`--output-schema \u003Cfile>`** to force a structured final\n  response, **`--skip-git-repo-check`** to run outside a git repo.\n- **`--ephemeral`** - **omitted deliberately**, for the same reason\n  `--no-session-persistence` is omitted from the `claude -p` recipe. It\n  persists no session, so the delegated run leaves no on-disk rollout for a\n  session-store sweep to collect. Persistence is a property of DELEGATION, not\n  of one CLI: suppressing it on either side loses the same thing.\n  The narrow case where it is right is explicitly throwaway local work, where\n  you intend not to keep a resumable transcript. If a run persists something\n  you did not want kept, fix that in the session-store or `$HOME`\n  configuration; do not reach for this flag to work around it.\n\n## The sandbox \u002F approval ladder\n\nPick the least privilege that lets the task finish. Granularity Codex exposes:\n\n| Setting | Grants | Use when |\n|---|---|---|\n| `-s read-only` | Read files; no writes, no commands | Analysis \u002F review only |\n| `-s workspace-write` | Write + run inside the workspace; no network | The default for fmt\u002Ftest\u002Fedit loops |\n\nNothing above `workspace-write` is on the ladder. Modes and flags that switch\nCodex's sandbox off are not used by these recipes, and `syn-delegate` refuses\nthem (`--sandbox` accepts only `read-only` and `workspace-write`). When writes\nmust reach a directory outside the workspace, add `--add-dir \u003Cdir>` instead.\n\n### Inside a workspace container: keep the sandbox, fix the container\n\n`-s workspace-write` makes Codex sandbox itself with **bubblewrap**, and\nbubblewrap needs an unprivileged user namespace. Docker's default seccomp\nprofile denies that to a container without `CAP_SYS_ADMIN`, so in a workspace\nstarted with the default profile every sandboxed operation fails:\n\n```\nwarning: Codex could not find bubblewrap on PATH ... will use the bundled\n         bubblewrap in the meantime\nbwrap: No permissions to create a new namespace, likely because the kernel does\n       not allow non-privileged user namespaces.\nFailed to write file \u002Fworkspace\u002Fpalindrome.py\n```\n\nRead that stack carefully: the missing-bwrap line is a **warning** and Codex\ncontinues with a bundled copy. The namespace denial underneath it is the real\nfault, and the write failure is only the symptom. Diagnosing from the first\nplausible line sends you to install bubblewrap, which does not help. Worse,\n`codex exec` still exits 0, so a caller that trusts the exit status records a\nrun that did nothing as completed.\n\nThe fix belongs to the workspace, not the invocation. `agentic_isolation`\napplies the Codex sandbox policy to every image labelled\n`agentic.codex_cli_version` (the Codex-capable workspace images are). It adds a\nnarrow seccomp profile (Docker's default plus `clone`\u002F`unshare` for the user,\nmount, pid, net and ipc namespaces, and `mount`, `umount2`, `pivot_root`)\nwhile `--cap-drop=ALL`, `no-new-privileges` and the read-only root stay in\nforce. On AppArmor hosts (Ubuntu 24.04, where the symptom is\n`bwrap: Failed to make \u002F slave: Permission denied`) it also applies the\nAppArmor profile `agentic-codex-sandbox`, which the host administrator loads\nonce with `sudo apparmor_parser -r`; it admits only the mounts bubblewrap\nperforms, so Codex must run from a directory under `\u002Fworkspace`. With both,\n`-s workspace-write` works as designed: writes inside the workspace succeed\nand writes outside it are denied.\n\n**Nested delegation.** A `workspace-write` delegate gets exactly two things\nbeyond Codex's defaults, as `-c sandbox_workspace_write.*` overrides that\nreplace any `CODEX_HOME` value: two extra writable roots, the retained child\njournal's partition directory and the partition's Claude transcript root\n(`$SPOOL\u002F$PARTITION\u002Fclaude`), and `network_access = true` (egress is still the\ncontainer's network policy). That is what lets the delegate run\n`syn-delegate` itself (Claude -> Codex -> Claude) with the Claude grandchild's\nown transcript captured. Nothing else outside the working directory becomes\nwritable, so:\n- A Codex grandchild cannot start its sandbox (its `CODEX_HOME` is read-only),\n  so it is refused with `codex_sandbox_unavailable`.\n- A `read-only` delegate cannot write the journal at all. A `syn-delegate` it\n  runs starts nothing, exits 70, and prints an\n  `agentic-delegate-launch-denied\u002Fv1` line; the enclosing `syn-delegate` reads\n  that line from its child's output and records `launch_failed` with reason\n  `nested_journal_unavailable` under the Codex child, so the denial is a\n  visible gap, not silence.\n\n`syn-delegate codex` probes `codex sandbox` with the same mode and grant\n(writing inside the extra root) live before every launch and refuses when it\nfails: it records `launch_failed`\nwith reason `codex_sandbox_unavailable` and exits 69, rather than running a\nCodex that cannot execute a single command. There is no status file to trust.\nThe entrypoint logs the same probe at startup (`[entrypoint] codex sandbox:`)\nfor diagnosis only. If you hit the refusal, the workspace lacks the policy or\nyou are outside `\u002Fworkspace`; do not route around it with a raw `codex exec`.\n\n## Budget: there is no built-in cap\n\n`codex exec` reports token `usage` only in the final `turn.completed` event —\n*after* the work is done — and surfaces no USD figure under ChatGPT-subscription\nauth. There is no `--max-budget-usd`. Therefore the bound must come from outside\nthe process:\n\n```sh\n# macOS lacks `timeout`; install coreutils for `gtimeout`, or use a CI step limit\ngtimeout 600 codex exec -s workspace-write --json -o .\u002Fcodex-last.txt \"$TASK_PROMPT\" \u003C \u002Fdev\u002Fnull\n```\n\nTreat an external wall-clock bound (`gtimeout`, a CI job timeout, a wrapper\nwatchdog) as mandatory for unattended runs. A wedged Codex run will otherwise\nconsume tokens until it finishes or is killed by hand.\n\n## `codex exec review` — the built-in diff reviewer\n\n0.147.0 ships a `review` subcommand that scopes itself to a diff instead of\nneeding you to describe one:\n\n```sh\ncodex exec review --base main --json -o .\u002Freview.md \u003C \u002Fdev\u002Fnull\ncodex exec review --uncommitted --json -o .\u002Freview.md \u003C \u002Fdev\u002Fnull\ncodex exec review --commit \u003CSHA> --json -o .\u002Freview.md \u003C \u002Fdev\u002Fnull\n```\n\n**The catch:** `--base` \u002F `--uncommitted` \u002F `--commit` are **mutually exclusive\nwith `[PROMPT]`**. Passing both fails:\n\n```\nerror: the argument '--base \u003CBRANCH>' cannot be used with '[PROMPT]'\n```\n\nSo you get either Codex's stock review rubric *or* your own instructions, never\nboth. It also exposes no `-s` or `-C`, so it is inherently read-only and runs\nagainst the current directory.\n\nChoose accordingly:\n\n| Want | Use |\n|---|---|\n| A quick review with Codex's own rubric | `codex exec review --base \u003Cbranch>` |\n| A review steered at specific risks, or one that runs the test suite to verify its own claims | plain `codex exec -s workspace-write` with a prompt telling it to `git diff \u003Cbase>...HEAD` |\n\nFor the steered form, add an explicit **\"do NOT modify any tracked file\"** to the\nprompt — `workspace-write` permits edits, and you want the review read-only while\nstill letting Codex run tests. Confirm with `git status` afterwards; treat a dirty\ntree as a failed review run, not a bonus.\n\n## What `--json` emits (for triage)\n\nThe stream is JSONL, one event per line. Event shapes observed:\n\n- `thread.started` — `{thread_id}` for the run.\n- `turn.started` — a turn begins.\n- `item.started` \u002F `item.completed` — each agent action. `item.type` is one of\n  `agent_message`, `command_execution` (carries `command`, `aggregated_output`,\n  `exit_code`, `status`), or `file_change`.\n- `turn.completed` — carries `usage` (`input_tokens`, `cached_input_tokens`,\n  `output_tokens`, `reasoning_output_tokens`).\n\nTo triage a run: filter `command_execution` items for non-zero `exit_code` to\nsee what Codex tried and where it stumbled; read the final `agent_message` (or\nthe `-o` file) for the conclusion; read `turn.completed.usage` for token spend.\n\n## When to delegate to Codex\n\nDelegate when a cross-model perspective is likely to help:\n\n- A bug survived **2+ of your own attempts** — a different model may break the\n  fixation.\n- Logic- or algorithm-heavy backend work with many edge cases.\n- You want an independent second implementation to compare against.\n\nDo **not** delegate:\n\n- Trivial CRUD, UI scaffolding, config edits, or first attempts — the handoff\n  overhead exceeds the value.\n- Tasks needing genuine multi-turn clarification — `codex exec` is one-shot\n  (use the interactive Codex TUI, or stay in Claude Code).\n- Work that must not run unsandboxed where you cannot grant `workspace-write`.\n\nFor autonomous *Claude* rather than Codex, use `delegating-to-claude-p`. The two\nare interchangeable dispatch targets; pick Codex specifically when you want a\ndifferent model's reasoning.\n\n## Using a Claude skill (e.g. pr-review) with Codex\n\nCodex does **not** auto-dispatch Claude `SKILL.md` plugin skills — it consults no\nskills index. To make Codex follow a specific skill, **inject the skill's content\nand name it in the prompt.** Two mechanisms:\n\n- **stdin (best for one-shot)** — pipe the skill body in; Codex appends it as a\n  `\u003Cstdin>` block. This is the one case where you do NOT redirect from\n  `\u002Fdev\u002Fnull`: the pipe supplies stdin and then closes it, which is exactly what\n  Codex is waiting for. Name the skill and the task in the prompt arg:\n  ```sh\n  cat path\u002Fto\u002Fskills\u002Freview\u002FSKILL.md | codex exec -s workspace-write --json \\\n    -o .\u002Freview.md -C \u002Fpath\u002Fto\u002Fproject \\\n    \"Apply the review skill provided on stdin to review this repo against plan.md. Follow its Report format exactly. Read-only: do not modify files.\"\n  ```\n- **`AGENTS.md` (best when reused across runs)** — drop the skill content into an\n  `AGENTS.md` at the workspace root; Codex auto-loads it. Still name the skill's\n  action in the prompt — like `claude -p`, a one-shot agent binds to explicit\n  prompt verbs, not to passively-present docs.\n\nSkills that mandate a **structured output contract** (a pr-review rubric, a\nverdict format) transfer especially well: the contract is self-describing, so\nCodex reproduces it faithfully. Inject the *whole* skill body, not a paraphrase —\nthe format section is what steers the output. See Trial T2.\n\n## Failure modes\n\n| Symptom | Cause | Fix |\n|---|---|---|\n| Hangs forever, zero events, `Reading additional input from stdin...` | stdin never reaches EOF and Codex waits on it | Add `\u003C \u002Fdev\u002Fnull`. Detect it by the `--json` stream file stalling at ~39 bytes, the length of that banner; a healthy run passes 100KB within a minute |\n| Run consumes tokens indefinitely | No built-in budget cap | Wrap in `gtimeout`\u002FCI timeout; `usage` is only reported post-hoc |\n| Unscoreable prose transcript | Default output is not JSONL | Add `--json`; read `-o` file for the conclusion |\n| `error: unexpected argument '--full-auto'` | Flag removed by 0.147.0 | Use `-s workspace-write`; re-check with `codex exec --help` |\n| `--base \u003CBRANCH> cannot be used with [PROMPT]` | `codex exec review` takes its built-in rubric OR custom instructions, not both | Pick one: `review --base` for the stock rubric, or plain `codex exec` with your own prompt |\n| Cannot write files \u002F run commands | Default sandbox is `read-only` | Set `-s workspace-write` explicitly |\n| Refuses to run | Not inside a git repo | Add `--skip-git-repo-check` |\n| Writes need a dir outside the workspace | `workspace-write` is workspace-scoped | Add `--add-dir \u003Cdir>`; keep the sandbox mode as it is |\n| `bwrap: No permissions to create a new namespace`, every write fails, exit 0 | Workspace runs Docker's default seccomp profile | Use an image labelled `agentic.codex_cli_version` through `agentic_isolation`, which applies the Codex sandbox policy |\n| `bwrap: Failed to make \u002F slave: Permission denied` | AppArmor host, workspace confined by `docker-default` | Load `agentic-codex-sandbox` on the host (`sudo apparmor_parser -r`) |\n| `bwrap: Can't bind mount ... Permission denied` | Codex started outside `\u002Fworkspace` on an AppArmor host | Run from a directory under `\u002Fworkspace` |\n| `syn-delegate` exits 69, `launch_failed` \u002F `codex_sandbox_unavailable` | The live probe found the Codex sandbox unusable | Same fixes; the stderr line names the bwrap cause |\n| `syn-delegate` exits 2, `--prompt: expected one argument` | Prompt starts with `-` | Use `--prompt=\"$TASK_PROMPT\"` |\n| Nested `syn-delegate` exits 70, `agentic-delegate-launch-denied\u002Fv1` on stderr | Run inside a `read-only` delegate (journal not writable) | Delegate with `workspace-write` if the child must delegate further |\n| `syn-delegate claude` exits 69, `claude_nested_auth_unavailable` | Claude below a Claude Bash call with OAuth only | Supply `ANTHROPIC_API_KEY` or a credentials file, or delegate from Codex |\n| Command fails on a quirky local env (e.g. `pytest` exit 127 under pyenv) | Sandbox shell inherits host PATH quirks | Under `-s workspace-write` Codex often self-recovers (Trial T1); for wrappers, pre-set env via `-c shell_environment_policy...` |\n\n## Trial T1 — empirical reference\n\nA real workspace-write run, recorded 2026-06-08 on `codex-cli 0.137.0`, where\nthe mode was still spelled `--full-auto`. The invocation below is left as it was\nactually executed; on 0.147.0+ substitute `-s workspace-write`.\n\n- **Task**: add `is_palindrome(s)` + a 3-case `pytest` file to a fresh repo, run\n  the tests.\n- **Invocation**: `codex exec --full-auto --json -o \u003Cfile> \"\u003Ctask>\"` (model =\n  config default).\n- **Result**: exit 0, **44s** wall-clock, **1 turn**, 18 items\n  (8 `agent_message`, 8 `command_execution`, 2 `file_change`), tests passed.\n- **Usage** (`turn.completed`): input 237,408 tokens (202,112 cached, ~85% cache\n  hit), output 1,139, reasoning 134. No USD figure surfaced under ChatGPT auth.\n- **Autonomous recovery**: the first `pytest` call failed (`exit_code 127` —\n  pyenv shadowing the binary); Codex retried with\n  `PYENV_VERSION=3.12.0 python -m pytest` and got `3 passed`. The `--json`\n  stream captured both the failed and the recovered command, which is exactly\n  the signal you triage on.\n\nThis is a single data point, not a universal claim. Add your own paired trials\nbefore quoting costs or wall-clock for a different task shape.\n\n## Trial T2 — does an injected skill actually steer Codex?\n\nA paired run recorded 2026-06-08, testing the injection mechanism above. Target:\na scratch repo with a 2-milestone `plan.md` whose milestone 2 (`truncate`) was\ndeliberately left unimplemented, so a correct review must flag the omission.\n\n- **Run A (skill on stdin)**: piped `sdlc\u002Fskills\u002Freview\u002FSKILL.md`; the prompt\n  named it and asked for its Report format. Codex reproduced the skill's exact\n  contract — `## Implementation Review`, the Milestone Status table, ⚠️ Omissions\n  (Blockers), and the `❌ BLOCKERS FOUND` verdict — correctly flagged the missing\n  milestone, and obeyed the skill's read-only\u002Fstatic-review process. 35s; 82,277\n  input tokens (87% cached) \u002F 1,406 output.\n- **Run B (baseline, no skill)**: same task, generic \"review against plan.md\".\n  Codex returned generic `**Findings**` prose — it found the same facts, but with\n  **none** of the skill's structure, and it ran the tests on its own (a different\n  process). 45s; 131,776 input \u002F 1,501 output.\n- **Adherence check**: all four signature markers\n  (`## Implementation Review`, `### Milestone Status`, `Omissions (Blockers)`,\n  `BLOCKERS FOUND`) were present in A and absent in B.\n\nTakeaway: injecting a skill and naming it in the prompt **does** steer Codex's\noutput format and process — but the content must be supplied, since Codex won't\nauto-dispatch it. Finding *fidelity* was similar in both runs because the bug was\nobvious; skills earn their keep on **structure and process discipline**, which is\nexactly what a pr-review skill enforces.\n\n## References\n\n- **`delegating-to-claude-p`** (sibling skill) — the autonomous-Claude dispatch\n  recipe and its empirically-validated flag set; read it for the side-by-side on\n  what steers a one-shot agent.\n- **`experiments:running-experiments`** — the discipline for designing a paired\n  trial to extend or contest Trial T1.\n- Codex CLI flags here are described by behavior to stay valid across releases;\n  verified against `codex-cli 0.137.0` on 2026-06-08, re-verified against\n  `0.147.0` on 2026-08-17 (which removed `--full-auto` and added\n  `codex exec review`). Re-run `codex exec --help` to confirm flag availability\n  in your installed version — this surface has already drifted once between\n  documented releases, so treat the check as part of the recipe rather than an\n  optional precaution.\n",{"name":5,"placement":23,"description":7},"Domain skill. Lives at `plugins\u002Fdelegation\u002Fskills\u002Fdelegating-to-codex\u002F` in agentic-primitives. NOT in `.claude\u002Fskills\u002F`; that scope is for meta skills.","main","8341ffb3b4128f405d0797b5f3f3affb25de91ea","plugins\u002Fdelegation\u002Fskills\u002Fdelegating-to-codex","https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace","https:\u002F\u002Fgithub.com\u002FAgentParadise","2026-11-09 13:31:42",{"source":8,"skillName":5,"commands":31,"prompt":41,"zipAvailable":42},[32,35,38],{"tool":33,"command":34},"npx","npx skills add AgentParadise\u002Fagentic-workspace --skill delegating-to-codex",{"tool":36,"command":37},"bunx","bunx skills add AgentParadise\u002Fagentic-workspace --skill delegating-to-codex",{"tool":39,"command":40},"pnpm","pnpm dlx skills add AgentParadise\u002Fagentic-workspace --skill delegating-to-codex","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace\u002Fblob\u002Fmain\u002Fplugins\u002Fdelegation\u002Fskills\u002Fdelegating-to-codex\u002FSKILL.md",true,[],[45,52,59,66,73,80],{"skillNo":46,"name":47,"nameZh":6,"description":48,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":49,"categories":50,"keywords":51,"publishedAt":18},"01M7DDJG3VACD4G63HQ3WWDTV5","testing-expert","Expert knowledge for analyzing testing architecture and planning improvements","https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace\u002Fblob\u002Fmain\u002Fplugins\u002Fsdlc\u002Fskills\u002Ftesting-expert\u002FSKILL.md",[],[],{"skillNo":53,"name":54,"nameZh":6,"description":55,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":56,"categories":57,"keywords":58,"publishedAt":18},"01M7DDJ3Z86RTM7YBDFF4QED1A","security-hardening","Audit and harden repository security — supply chain, CI\u002FCD scanners, secrets, credentials","https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace\u002Fblob\u002Fmain\u002Fplugins\u002Fsdlc\u002Fskills\u002Fsecurity-hardening\u002FSKILL.md",[],[],{"skillNo":60,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"publishedAt":18},"01M7DDHZY3KRZ1PM4YTVR8Y5B5","review","Review implementation against project plan and verify completeness","https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace\u002Fblob\u002Fmain\u002Fplugins\u002Fsdlc\u002Fskills\u002Freview\u002FSKILL.md",[],[],{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":18},"01M7DDHVFQK785K8Z7HX5GRN4A","qa-setup","Verify or create standardized QA commands for polyglot projects","https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace\u002Fblob\u002Fmain\u002Fplugins\u002Fsdlc\u002Fskills\u002Fqa-setup\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":18},"01M7DDHSD25FQ20METGDRWAWED","prioritize","Prioritize review comments by severity","https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace\u002Fblob\u002Fmain\u002Fplugins\u002Fsdlc\u002Fskills\u002Fprioritize\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":18},"01M7DDHPHRVKHR1TZBYNSTA0XX","pre-commit-qa","Run standardized QA checks before committing changes","https:\u002F\u002Fgithub.com\u002FAgentParadise\u002Fagentic-workspace\u002Fblob\u002Fmain\u002Fplugins\u002Fsdlc\u002Fskills\u002Fpre-commit-qa\u002FSKILL.md",[],[]]