pmwl0128/pokemon_champion_agent

Pokémon Champions UI Session

The local web panel (pcui serve) and you collaborate through ONE shared session store: the panel is the session's eyes and hands, you are the decision participant. This skill adds exactly one thing to your normal workflow — reading session context and submitting gate artifacts through pcui. The team-building discipline itself is unchanged: follow pokemon-champions-team's assisted workflow (UEP) exactly as you always do.

CLI

The pcui command for this machine: {{PCUI_CMD}} (bridge directory: {{BRIDGE_DIR}})

{{PCUI_CMD}} session list                              # newest first
{{PCUI_CMD}} session show --id <S>                     # meta.intent + heads + ledger
{{PCUI_CMD}} artifact get --session <S> --kind <kind>  # latest artifact of that kind (verbatim)
{{PCUI_CMD}} artifact put --session <S> --kind <kind> --file out.json   # or pipe via stdin
{{PCUI_CMD}} run <op> [--file team.json] [--context ctx.json] [--args '<json>']
                                                       # one whitelisted team operator in the daemon

Workflow

1. Find the session: session list, pick the one the user means (usually the newest, or the id they name). session show --id <S> gives meta.intent (the user's build intent typed in the panel), heads (latest artifact per kind), and the ledger.
2. Read what exists: artifact get every kind in heads that matters — context, decision if present (the user's checkpoint answer made in the panel: chooseCandidate + candidateIndex into your slate's teams, requestChanges + note, or proceed), and tune-request if present — see the tune loop below.

The tune loop (tune-request → tune-result)

A tune-request artifact is the user's PLAIN-LANGUAGE tuning ask, written on the calc page against a recommended team: {note, member?, member_set?, team?, form_benchmarks?}. The note may name things the form cannot express — several members or the whole team, environment-filtered goals ("survive every top-50 Ground hit": resolve via meta ranking + dex reverse/type lookups, then benchmark against the concrete threats), or fuzzy targets. TRANSLATING it is your job (the skill never parses free text): turn the note into concrete build-context.benchmarks[] per the team skill's adjustment-intent rule, auto-fill opponent sets from repset/meta modal where unspecified, run team.py tune --context on the team (take it from the draft artifact; if the session has no draft — a request from the standalone tune page — use the request's own team snapshot, and member_set for the member's exact edited item/ability/nature/SP), and artifact put the operator's FULL output as kind tune-result — the panel renders its cards as cliff cards. form_benchmarks are the explicit rows the user already ran in the form; reuse or extend them, don't silently drop them. If the ask implies changing the team (not just SP), say so to the user instead of silently editing members. 3. Work the UEP flow normally (context-audit → grounding → frame → assemble → slate-evaluate → checkpoint → answer-audit), using the four skills as always. Prefer the daemon path: run <op> executes one team operator inside the daemon's resident workers, and a session spec runs several ops in one process — both amortize the per-op sibling-subprocess spawn (dex/meta/calc) that dominates the wall-clock of a COLD team.py <op> call (the Python import itself is cheap; the repeated worker spawns are the tax). Batch related reads (ranking/landscape/ repset/validate) into one session instead of many shell-outs; fall back to a raw team.py only when the daemon is down. When you run an intake question batch (intake --next), also artifact put its output as kind intake so the panel shows the user your current questions live (answers still come through conversation or the panel's decision/intent fields). 4. Draft prose: narrate in the USER'S language, but name every ENTITY by English canonical. In tradeoffs, assumptions[].note, confidence_notes, convergence_rationale, context_summary: species, moves, items and abilities are written EXACTLY as they appear in operator output ("Whimsicott", "Last Respects") — never hand-translated. A hand-written localized name is a hallucination vector nothing downstream can catch; the panel localizes canonical names with dex authority (portrait chips, official names) at render time. 5. Submit each gate artifact as you produce it with artifact put. Kind vocabulary (the panel labels and orders these): context (the build-context you assembled), audit (context-audit receipt), frame, slate (slate-evaluate output), checkpoint (checkpoint output), draft, answer-audit. Submit the artifact's raw JSON exactly as produced — the store keeps it verbatim. decision is PANEL-authored; you only ever read it. Artifact size: the daemon rejects a body over ~2 MiB. Only one artifact risks that — slate (the full per-cell battery + 6-pick-N combos can be several MB). Emit the compact summary for the panel: slate-evaluate … --slate-view summary (or run slate-evaluate with view:"summary"). It keeps survivors, each candidate's legality/mega/structural facts, the CHECK grade counts the card renders, and slate_receipt, while dropping the grid, per-opponent evidence lists and combos the panel never reads. Keep the FULL slate output for your own reasoning and for answer-audit (its claim recompute re-hashes the untrimmed matchup_risk) — only the panel artifact is the summary. 6. Checkpoint pause: when the checkpoint output says pause:true (and the context did not encode direct_final/skip_checkpoint), put the checkpoint artifact, tell the user to decide in the panel, and stop this turn. Next time you are invoked, read the decision artifact first and continue from it. 7. Conflicts: artifact put re-reads the current revision before writing; if it still reports a 409 the panel wrote concurrently — re-run the same put once.

The panel renders every artifact you put in real time (SSE). Do not wait to submit them in a batch at the end; put each gate's output when the gate completes.