From c78bd9cf2358662f9e9731ec0d0156636b60765e Mon Sep 17 00:00:00 2001 From: Claude Date: Sat, 26 Sep 2026 15:00:37 +0000 Subject: [PATCH] docs: say which skills the plugin carries and what they decide MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The README never named the internal skills — not the prompting one this fork just replaced, nor the two beside it. They are not user-invocable, but they decide how a request reaches Codex and how its answer comes back, so their absence left the reader with no way to know why a rescue behaves as it does. The prompting skill gets the notes that matter in practice here: its launch lines use full model slugs because this fork maps only `spark`, `xhigh` is accepted despite the skill's advice to stay at or below `high`, and a test pins both against the plugin's own validators. The result-handling row names the rule most likely to surprise someone: after presenting review findings Claude stops and asks, rather than applying them. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01UXfvnjSC72HsM6EEPVt2Tg --- README.md | 23 +++++++++++++++++++++++ 1 file changed, 23 insertions(+) diff --git a/README.md b/README.md index 9a2fa5e86..a1b02adea 100644 --- a/README.md +++ b/README.md @@ -399,6 +399,29 @@ These are advanced knobs for tuning resource usage in long-running or resource-c One broker serves every session in a workspace. It is shut down by the last session out, never by a session that still has another's work in flight, and it releases its Codex thread subscriptions as clients disconnect, so a departed session's notifications never reach the next one. +### Skills The Plugin Carries + +Three internal skills ship with the plugin. None is user-invocable: they are contracts the rescue +subagent and the result presentation follow, and they are listed here because they decide how your +request reaches Codex. + +| Skill | What it governs | +| --- | --- | +| `codex-cli-runtime` | The forwarding contract: the subagent's only job is to call `task` once and return its stdout unchanged. It also fixes what the subagent may *not* do — no `setup`, `review`, `status`, `result` or `cancel`, no reasoning through the problem itself. | +| `gpt-6-prompting` | How a request is turned into a Codex brief: which model to pick (Luna for bulk and well-specified work, Sol for multi-file implementation and review, Astra for architecture or a second opinion), which reasoning effort, and per-model recipes for diagnosis, implementation, review, research and vision. | +| `codex-result-handling` | How Codex's output comes back to you: its verdict, findings and file paths preserved as reported, findings ordered by severity, and the distinction between fact, inference and open question kept. It also forbids acting on a review by itself — after presenting findings Claude stops and asks which ones to fix, and a failed Codex run is reported rather than replaced by a Claude-side attempt. | + +Two notes on the prompting skill, since it is the one that changes what you get: + +- its launch lines use full model slugs (`gpt-6-luna`, `gpt-6-sol`, `gpt-6-astra`). This fork maps + only one short alias, `spark` → `gpt-5.3-codex-spark`; every other `--model` value is forwarded to + Codex verbatim, so the slug is what works +- it advises staying at `high` effort or below. That is advice about cost, not a limit: `--effort + xhigh` is accepted. `max` and `ultra` are not +- a test reads this fork's own model and effort validators and checks every `--model`/`--effort` the + skill names against them, so the guidance cannot drift into naming something the plugin would + reject or pass through as a literal word + ### Moving The Work Over To Codex Delegated tasks and any [stop gate](#what-does-the-review-gate-do) run can also be directly resumed inside Codex by running `codex resume` either with the specific session ID you received from running `/codex:result` or `/codex:status` or by selecting it from the list.