Claude Code · Codex · Cursor · Chenxi
Seven skills for orchestrating, reviewing, executing, and understanding agent work.
| When you need to… | Use |
|---|---|
| coordinate a task-specific native agent team | $orchestrate <task> |
| drive long-lived workers across sessions, models, and machines | $herd <task> |
| find out what is actually happening | /whatup |
| execute an already-approved plan | /execute-plan <plan-path> |
| check a bounded fix preserves related behavior | /seesaw-review <target> |
| review, fix, and re-verify an artifact | /review-cycle <target> |
| get an independent cross-model review | /tribunal-review <target> |
cstack is a Claude Code and Codex skills plugin. The seven skills install
together and share source files. orchestrate uses native subagents when
available; installing the plugin does not supply models or agent tools.
The operational lead owns scope, integration, review decisions, and final acceptance. Codex defaults to GPT-6 Sol; Claude Code defaults to Opus 5.5. Fable and Astra are called for bounded difficult decisions or required independent review. A user can explicitly choose a different capable lead.
$orchestrate <task or approved plan>
The operational lead reads the project, chooses only the roles the task needs, and selects each worker's model and effort from the host's available options. Codex defaults to Sol 6 as lead; Claude Code defaults to Opus 5.5. Luna/low fits bounded extraction, Sol/medium fits scoped implementation, and Astra can advise on difficult or high-risk reasoning. These are starting points, not a fixed roster.
Independent work runs through native subagent tools. Each worker gets explicit file ownership and acceptance criteria; the lead integrates and verifies results. Overlapping writes run sequentially, and simple tasks need no workers. A runtime without the requested model or subagent capability reports the limitation. The skill does not change global model defaults or create custom agent files.
Use the same symlink installation as the other skills, or
install the plugin. In a Codex CLI session, select the default lead with
codex -m gpt-6-sol, then invoke $orchestrate. A user may choose Astra as
lead explicitly. The skill does not change an already-running session's model.
The workflow follows the official subagents guide.
Validation on 2026-09-07: native dispatch accepted two bounded read-only workers requesting Sol/medium and Luna/low, and both returned results. Their feedback closed the ownership-transfer and nested-tool gaps. Returned metadata did not independently identify the serving model/effort, so these remain requested settings, not a latency benchmark or a guarantee about every runtime.
$herd <task or approved plan>
The lead picks a transport per worker: a same-provider peer session for
work that belongs to another live repo, a herdr agent for persistent work
on another model or machine, or a native subagent for bounded one-off work.
An explicit $herd request puts the substantive work with workers; the lead
sets scope, handles authorized permission prompts, checks the result, and
integrates it. Persistent assignments use an execution contract and a
per-task review gate. Only the herdr transport requires herdr.
An Opus lead can send harder scoped implementation to Cursor or Devin pinned to Opus 5.5; a different Opus version or moving alias is not a substitute. Those workers can implement but cannot independently review Opus work. When lead and worker use the same canonical model, a different model must review each substantive change before the lead accepts it. Model identity is checked from launch settings and available runtime labels; this does not attest to the provider's serving model. Sol starts at medium effort and increases effort only when the task warrants it. These choices aim to reduce total cost, but the repository does not claim measured token savings.
Plain-language status, grounded in evidence.
/whatup [PR N | a path | a phrase]
Use it when: the agent has been working for a while and you no longer trust the narration enough to know what is done, what failed, or what happens if you stop.
What it does: rereads the goal and checks the branch, dirty files, recent commits, PR state, artifacts, and test evidence. It also compares earlier "done" and "passing" claims with the fresh state.
What you get: one plain-language screen covering purpose, current state, what works, what failed, what remains, and the decision waiting on you. Every claim carries its evidence beside it or is marked unverified.
whatup is read-only. It names the next step; it does not take it.
The skill takes its premise from "AI Agents Push Humans Out of the Loop" (Mitchell, Ghosh & Passi, 2026): "current approaches to the development and deployment of AI agent systems do not support effective human oversight – they contribute to its degradation."
The paper's audit mechanism checks an agent's narrative against the raw record
to find where they diverge. whatup turns that into a concrete rule: before it
writes the status, it checks earlier claims against what it just verified. A
mismatch becomes the opening line instead of a quietly corrected fact.
An approved plan in; a verified branch out.
/execute-plan <plan-path> [--full-cycle]
Use it when: the plan is already written and approved. You want the agent to stop reopening settled questions and carry it through.
What it does: creates an isolated worktree, executes in a straight line, commits at meaningful milestones, runs the plan's checks, and verifies the evidence before reporting completion.
What you get: a branch or PR plus a table of claims, evidence, and commands you can use to re-verify each claim.
Add --full-cycle to review the plan before implementation, review the final
cumulative diff, and run the plan's real end-to-end check when one exists.
Even without --full-cycle, a worker sharing the lead's canonical model
triggers different-model review before the affected milestone is accepted.
seesaw-review [<patch, PR, or task target>]
A focused patch review: prove the original bug is fixed and the related behavior still works. The lead traces callers, sibling paths and relevant retry/runtime conditions, then verifies concrete preservation checks with one independent peer through Herd: Opus 5.5 for a Sol lead, Fable for Astra, or Astra for a Claude Code lead. Corrections keep the original baseline and checks; two unsuccessful correction rounds escalate instead of accumulating more patches.
It reuses task evidence and returns a bounded PASS / FIX / ESCALATE. It does not
replace required review-cycle or tribunal-review gates, authorize standalone
edits or deployment, or turn unrelated old problems into patch scope.
Review that changes the artifact, not just comments beside it.
/review-cycle [quick] [target] [focus …]
Use it when: code, a plan, or prose is finished and needs a disciplined review before it ships.
What it does: runs a fixed review sequence and applies valid fixes between passes. After each apply step, it reruns the repository's own verification before continuing.
self-review → tribunal → adversarial → simplicity
→ cumulative reread → assumptions → acceptance
What you get: the improved artifact and a complete pass ledger. A skipped
or unavailable step remains visible; it cannot disappear while the verdict
still says SHIP.
Ordinary edits use self-review and relevant checks; invoke this cycle when requested or warranted by impact and uncertainty, including high-risk single-file changes. Completion follows acceptance evidence, not subjective confidence scores.
The review adapts to the target. Code is checked for defects, plans against the repository they intend to change, and prose for false or misleading claims.
A second opinion from a different model lineage.
/tribunal-review [target] [quick|deep] [focus: …]
Use it when: one model reviewing its own work is not enough.
What it does: the lead personally reviews and verifies the artifact while
independent reviewers work in persistent herd panes. Codex Sol or Astra
requires Fable; Claude Code Opus or Fable requires Astra. Cursor uses verified
Grok 4.7 or newer. Devin, whether SWE-2 Max or pinned Opus 5.5, is execution-only.
Reviewers retain their own context for debate, rebuttal and versioned fix verification.
The lead validates findings and owns the final acceptance decision.
| Seat | Weight |
|---|---|
| Orchestrator | 1.0 |
| Claude or Codex peer | 1.0 |
| Cursor / Grok | 1.0 |
| Gemini advisor (only when requested) | 0.5 |
What you get: one deduplicated verdict with consensus findings, contested items, dismissed low-confidence claims, and the exact panel that answered.
The actual assignment establishes whether a seat can answer. An installed CLI may be logged out, unlicensed, or unable to reach its credentials. Missing reviewers are named; an unavailable required Fable/Astra peer leaves the gate open. Panes stay readable while needed. Once the lead has all necessary information and no follow-up needs the context, it saves the evidence and closes task-created panes without waiting for the whole project to finish.
The weights represent lineage independence, not measured accuracy.
approved plan ──▶ execute-plan ──▶ verified branch / PR
│
└─ --full-cycle ─▶ review ─▶ execute ─▶ review + e2e
bounded bug fix ──▶ seesaw-review
finished artifact ─▶ review-cycle
│
└─ Pass 2 ─▶ tribunal-review
lost the thread ─────────────────────▶ whatup
execute-plan, review-cycle, and tribunal-review form the delivery chain.
whatup is independent and read-only, so you can use it at any point.
/plugin marketplace add moremeds/cstack
/plugin install cstack@cstack
This installs all seven skills together, plus the shared Claude Code commands and hooks.
Clone the repository, then symlink each skill directory into the runtime roots you use:
| Source | Claude Code | Codex and other Agent Skills runtimes |
|---|---|---|
skills/<name>/ |
~/.claude/skills/<name> |
~/.agents/skills/<name> |
Use symlinks, not copies. Both runtimes then load the same SKILL.md, leaving
one source to review, test, and publish.
See the official-guidance audit for the source mapping, preserved safeguards, validation, and performance limits. These are workflow changes, not a claim of measured model speedup.
The optional auto-commit hook now commits only explicitly staged work; it never stages the workspace. Test-on-edit remains opt-in and is diagnostic, not acceptance evidence. Hook checks are best-effort helpers, not a complete security boundary.
skills/ the plugin orchestrate, herd, execute-plan, seesaw-review, review-cycle, tribunal-review, whatup
rules/ standing rules maintainer defaults for Claude Code and Codex
hooks/ hard guardrails block or rewrite selected Claude Code tool calls
commands/ entry points small shared Claude Code commands
tests/ contracts mutation-checked workflow and publication guards
The skills are the product. Everything else supports the maintainer's own daily setup and is optional. Machine-specific settings, credentials, memories, plugin inventories, and private skills stay outside this public repository.
Run the contract suite with:
python3 -m unittest discover -s testsAssertions are mutation-checked: each one must fail when the contract it claims to protect is broken. The publication guard also catches private markers before they enter public Git history.