ai-dynamo

Checks a Dynamo pull request for design fit, testing and validation, code quality, and the pull request description. Use when making a pull request to Dynamo or reviewing one.

Checks a Dynamo change for performance problems in code that runs per output token, KV block, KV event, request, connection or worker. Compares it against the cost mechanisms behind Dynamo's merged performance fixes, guards external semantics, and checks the…

Ensures every piece of work in ai-dynamo/dynamo starts from a tracked issue and every pull request references one. Use when starting new work, creating or sorting its Linear issue into a feature or workstream project, opening or editing a pull request, or…

Describes how a fix reaches a Dynamo release branch after code freeze - cherry-picking a merged main fix into release/X.Y.Z, landing a release-only fix, or sponsoring an external contributor's Cherry-Pick Request - with the cherry-pick Linear labels, the…

Adds, updates, moves, or removes content on the Dynamo Fern docs site — standard docs pages, catalog-driven recipe and feature-benchmark pages, examples, recipes, and translations — keeping everything in line with the documentation style guide. Use for any…

Reports a defect in the optimization skillpack itself (the skills under .agents/skills/, the documents under agent-docs/, the role contracts under agents/, and the repository AGENTS.md files) as a GitHub issue on ai-dynamo/dynamo, using the repository's…

Drafts a candidate baseline DynamoGraphDeployment from interview requirements when no catalog recipe matches the user's model, hardware, and backend, presenting per-decision evidence for the user's confirmation. Use only from user-interviewer at interview…

Consults the repository performance rules and applicable Dynamo and engine guides to select one evidence-backed optimization proposal, then writes the generator's knowledge-consult.md reasoning record. Use after perf-analyzer completes a valid AIPerf analysis…

Bump the AIPerf package version (e.g. 0.10.0 -> 0.11.0) across pyproject.toml, the mock server, and the version strings shown in docs, then open a PR to main

Use when running, deploying, submitting, attaching to, or retrieving results from a single AIPerf benchmark on a Kubernetes cluster with the aiperf kube CLI (AIPerfJob CR, operator mode or direct mode).

Use when preparing a real multi-node Kubernetes cluster for AIPerf - installing or upgrading the aiperf-operator Helm chart, installing JobSet, wiring a private registry pull secret, choosing node placement and a storage class for the results PVC, or scoping…

Use when running a parameter sweep, multi-run confidence trial, or adaptive/Bayesian search on Kubernetes with aiperf kube sweep and the AIPerfSweep CR - authoring sweep parameter axes, addressing child variations, or collecting sweep aggregates.

Use when an AIPerf benchmark on Kubernetes is stuck, failing, crash-looping, OOM-killed, pending, queued, stalled, showing a high error rate, or its results cannot be retrieved - diagnosing an AIPerfJob or AIPerfSweep from CR status, pod state, and logs.

Review the current branch against origin/main, capture findings in artifacts/code-review.md as a living document, validate every finding against the actual code, reproduce confirmed issues with the aiperf CLI against the in-repo mock server, and draft inline…

This skill should be used when the user asks to "create a linear issue", "add a linear issue", "track this in linear", "open a linear ticket", "log this in linear", "create a ticket", or mentions creating an issue for tracking work in Linear.

Use when reviewing AIPerf code for the SEMANTIC qualities that mechanical linters cannot catch — error-message informativeness, type-hint descriptiveness, docstring example usefulness, naming disambiguation, convention explicitness, comment density, and…

Use when creating a git commit to require Conventional Commits formatting, DCO sign-off, and cryptographic signing

Use when a user asks to tune, optimize, accelerate, or deploy a PyTorch model or inference pipeline for GPU performance.

Use when measuring inference performance of a PyTorch model before and after AITune optimization — capturing baseline, compilation time, tuned throughput, and speedup.

Run full documentation sync (not limited to recent git changes)

Use when inspecting a PyTorch model or pipeline to identify tunable submodules, detect dynamic shapes, and determine the recommended tuning mode before optimization.

Use when creating a new release tag for the project

Use when verifying that a tuned model's outputs match the baseline eager model within numerical tolerance — run after each backend compilation before accepting it as a deployment candidate.

Diagnoses AgentX GPU-versus-AISimulate replay gaps in TTFT, per-request TPOT, throughput, and latency distributions. Use to investigate workload, FPM, frontend, scheduler, or KV-cache fidelity and validate scoped corrections. Not for optimizing simulator host…

Design, implement, and review efficient AISimulate engine-only simulation execution. Use for replay features, correctness fixes, and refactors that affect scheduling, state transitions, event processing, or costs that scale with requests, tokens, workers, or…

Use when running long AIC/aiconfigurator GPU perf auto-collection for a specific GPU/framework/framework_version, including draft-PR checkpoints, resumable collect_xx.py runs, framework-version/kernel preflight, collector fixes/skips, special runtime images,…

Design, implement, and review efficient AISimulate numerical prediction and data lookup. Use for FPE features, correctness fixes, and refactors that affect prediction algorithms, preparation, data representation, caches, or caller cost. Resolve substantial…

Convert InferenceX DB config/benchmark pairs, Dynamo recipe YAML, or confirmed custom serving configs into validated aiconfigurator estimate requests. Use when adapting server configuration, checking whether a recipe can be estimated, generating canonical…

Runs deterministic byte-parity and golden-point qualification campaigns for AI Simulate engine-only offline replay across round-robin vLLM, SGLang, and TRT-LLM configurations, including forced preemption, retraction, handoff, and liveness coverage. Use when…

Use when adding perf data for a new framework version, deciding which backend_version to query or pin in tests, bumping the maintained (current/previous) versions, retiring old version directories, or hitting "is not a queryable version" errors from the…

Convert InferenceX DB config/benchmark pairs, Dynamo recipe YAML, or confirmed custom serving configs into validated aiconfigurator estimate requests. Use when adapting server configuration, checking whether a recipe can be estimated, generating canonical…

Use when validating AIC predictions against real GPU serving measurements and root-causing fidelity gaps (prediction-vs-measured issues, memory/concurrency mismatches, per-op latency divergence). Covers dummy-weight methodology, measurement parity traps,…

Use when running long AIC/aiconfigurator GPU perf auto-collection for a specific GPU/framework/framework_version, including draft-PR checkpoints, resumable collect_xx.py runs, framework-version/kernel preflight, collector fixes/skips, special runtime images,…

Design, add, review, or modify AIC Collector operations and their case population. Use for new Collector ops, backend registry entries, collector/cases YAML, case generators/getters, pruning or deduplication, persisted perf keys, framework-version routing,…

Bump the ModelExpress workspace, Python, and Helm chart version, and update the pinned source-id hashes. Use for any release version bump, and for the separate public-image tag bump on release branches.

Add or change a ModelExpress client CLI argument or environment variable. Use whenever touching ClientArgs, ClientConfig::load(), the Cli struct, or envs.rs/envs.py.

Add a new gRPC service to ModelExpress. Use whenever creating a .proto file, wiring a tonic service into the server, or extending modelexpress_common's generated grpc module.

Keep ModelExpress contributions compliant with the Developer Certificate of Origin. Use whenever creating, amending, rebasing, cherry-picking, or pushing commits, opening or updating a pull request, or diagnosing a failed DCO check in this repository.

Interactive guide for authoring or updating a Grove user guide. Leads a structured dialogue with the user, section by section, and writes the documentation incrementally as each section is confirmed. Also supports continuing work on an existing user guide.…

Interactive guide for authoring a Grove Enhancement Proposal (GREP). Leads a structured dialogue with the user, section by section, and writes the proposal file incrementally as each section is confirmed. Also supports continuing work on an existing GREP.

Audit async-openai dependency upgrades in ai-dynamo/frontend-crates and Dynamo consumers. Use when bumping async-openai, reviewing an upgrade PR, checking whether locally owned or vendored OpenAI protocol types can be deleted or narrowed, detecting drift in…

Generate optimized tool call parsers for the dynamo-parsers crate from HuggingFace model chat templates. Use this when you need to add support for a new model's tool calling format. Takes a HuggingFace model name, analyzes its chat template, compares with…

Performance review of a velo diff, or a performance audit of one subsystem. Maps the change onto velo's hot paths (per message, frame and record), hunts the regression patterns that have shipped here (per-frame cancellation futures, per-record timers, copies,…

Maximum-intensity multi-agent review of a velo diff/branch before it becomes a PR. Fans out repo-specific finder dimensions (correctness, concurrency, FFI/unsafe, wire-compat, semver boundary, shutdown contract, and performance when the diff touches a hot…

Use for source-matched NIXL Python API help on agents, descriptors, metadata, transfers, polling, or cleanup. Do NOT use for install/framework setup.

Guide users through creating or revising a Snapshot Enhancement Proposal (SNEP), including an interactive section-by-section draft. Use for significant Snapshot design changes, not ordinary implementation or small documentation changes.

Drive the real Pi CLI headlessly against a Dynamo or OpenAI-compatible endpoint for pi-dynamo-provider validation. Use when testing Pi provider installs, session header tracing, Pi subagent runs, saved traces, or parent/child session behavior without manually…