claudecodex

ammo

Hand-ported PAVE workflow: run a staged, multi-agent campaign that profiles and optimizes Pumice inference latency on your own NVIDIA GPU host, inside a local Pumice checkout, on Claude Code or Codex (see its README).

1 on GitHubby jinhuang12v1.0.0updated 9 days ago
/

Plugin ammo from marketplace jinhuang12-plugins in jinhuang12/pave. Commands read from the repo's manifest — 8 days ago.

This plugin runs code on your machine — its hooks fire automatically on events like tool use. The exact commands are listed below; read them before installing.

What it installs

Skills (2)

  • ammo-report

    Generate a standalone technical report from completed AMMO campaign artifacts, including measured results, charts, implementation details, failures, and reusable lessons. Read by the ammo-report-writer role at Stage 7b of an ammo:ammo campaign; not for direct use.

  • ammo

    Profile and optimize Pumice inference latency on NVIDIA GPUs, inside a local Pumice checkout. Use when targeting a specific model, GPU, dtype, parallelism, and workload. Runs a staged multi-agent campaign (baseline, profiling, debate, isolated implementation tracks, integration, report) that takes h

Subagents (8)

  • ammo-auditor

    Independent AMMO gate auditor for T_AUDIT_S1, T_AUDIT_S2, T_AUDIT_S45, and T_AUDIT_S67 that reconstructs the stage from primary evidence in Phase 1, applies the hook-delivered checklist in Phase 2, and returns PASS, BLOCKED, or NEEDS_INVESTIGATION.

  • ammo-champion

    Argues for a specific GPU kernel optimization candidate in adversarial debate, runs micro-experiments to gather evidence, and critiques competing candidates.

  • ammo-delegate

    Runs one bounded evidence task (trace parsing, source tracing, roofline math, prior-art search, an assigned experiment, a bounded profiler run, or the report fact-check) for any AMMO agent that spawns it: ammo-champion, ammo-impl-champion, ammo-auditor, pave-init:adversarial-investigator, or the lea

  • ammo-impl-champion

    GPU kernel implementation champion for AMMO optimization tracks. Implements kernel optimizations, then writes and runs its kernel correctness & speedup checks before the E2E sweep.

  • ammo-report-writer

    Generate publication-quality GPU kernel optimization reports from completed AMMO campaign artifacts. Reads profiling data, candidate evaluations, and implementation results to produce REPORT.md with matplotlib charts, code listings, and lessons learned.

  • ammo-researcher

    AMMO Stage 1 baseline and profiling, and Stage 2 bottleneck mining, for Pumice campaigns.

  • ammo-resolver

    Resolve merge conflicts and semantic interactions between accepted tracks (PASS or GATED_PASS) in AMMO Stage 6. The lead spawns it when a composition needs a code resolution.

  • ammo-transcript-monitor

    Periodic adversarial reviewer that monitors impl-champion agents (Stages 4-5) through their harness transcript logs. Detects methodology errors, scope drift, and reward-hacking during implementation. Not used for debate-stage champions.

Hooks (19)

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-team-spawn-guard.sh"

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-pretool-guard.sh"

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-skill-files-guard.sh"

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-env-default-guard.sh"

  • PreToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/worktree-create-with-build.sh"

  • PreCompact

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-precompact.sh"

  • SessionStart

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-postcompact.sh"

  • Stop

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-stop-guard.sh"

  • SubagentStop

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-subagent-release.sh"

  • SubagentStop

    "${CLAUDE_PLUGIN_ROOT}/hooks/audit-phase2-guard.sh"

  • WorktreeCreate

    "${CLAUDE_PLUGIN_ROOT}/hooks/worktree-create-with-build.sh"

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-gpu-release.sh"

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-monitor-reminder.sh"

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-audit-start-stamp.sh"

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-state-validate.sh"

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-next-step-reminder.sh"

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/ammo-artifact-layout-warn.sh"

  • PostToolUse

    "${CLAUDE_PLUGIN_ROOT}/hooks/inject-audit-phase2.sh"

  • WorktreeRemove

    "${CLAUDE_PLUGIN_ROOT}/hooks/worktree-remove-cleanup.sh"

View source on GitHub

More from jinhuang12/pave

All 3 plugins →

pave-init-generated workflow: bring the vLLM-Neuron platform plugin fork to parity with upstream GPU vLLM through user-gated, evidence-backed campaigns. Ships the fifteen Neuron skills its campaign stages load inside the plugin (see its README).

16 skills · 6 agents · hooks

1

pave-init

claudecodex

Create a PAVE-based Claude Code workflow plugin (a lead skill plus registered role agents) from a stated goal, with adversarial review gates.

3 skills · 10 agents · hooks

1

Maintain ammo?

This page updates itself from your repo. Add the badge to your README, or get it Featured to pin it above the organic results with a followed link to your site.

claude & codex plugins3 plugins