poseidon-implement

poseidon-implement

Local, in-session port of the Poseidon issue-implement agent. It does not ship a copy of the Poseidon instructions — it fetches them live from a8cteam51/poseidon-actions@trunk every run, so you always match the deployed version. It then follows those instructions with a thin set of local overrides, running on the model you've selected with no turn cap.

This ports the plain issue-implement, not implement-v2 (the Pressable clone + Playwright verification flow), which is tied to OpsOasis provisioning.

Inputs

$ARGUMENTS: 45 or a8cteam51/some-repo 45 (bare N resolves the repo from the current directory).

Requires: the issue already has an approved ### Poseidon plan comment (run /poseidon-plan first); gh with access to a8cteam51/poseidon-actions (fetch) and write access to the target repo (branches + PRs); and the current directory must be a clone of the target repo (this agent creates branches, commits, and pushes).

How it runs

1. Resolve OWNER/REPO and issue number N from $ARGUMENTS. Confirm you are inside a clone of the target repo; if not, STOP and ask the user where their clone lives (or whether to clone it, and where) before continuing.

2. Fetch the live upstream instructions:

   mkdir -p /tmp/poseidon
   gh api repos/a8cteam51/poseidon-actions/contents/issue-implement/action.yml \
     -H "Accept: application/vnd.github.raw" > /tmp/poseidon/issue-implement.action.yml
   

   If this fails (e.g. 404 / no access), STOP and tell the user — do not fall back to remembered or guessed instructions.

3. Read /tmp/poseidon/issue-implement.action.yml. The Run Claude Code step carries TWO instruction blocks — extract and follow BOTH:

   • the prompt: input — the task itself (implement the plan, branch, PR), including the rule that newer post-plan human comments beat stale details in the plan;
   • the --append-system-prompt "<PROMPT>" value inside claude_args — your authoritative operating contract, including its untrusted-input and production-read-only safety rules. Read both in full and follow them exactly, with the local overrides below. You are NOT executing the YAML's steps; you only extract and follow those two embedded prompts. If the fetched file no longer matches this layout, find the embedded prompts wherever they now live, follow them, and tell the user the upstream layout changed.