dogkeeper886

Apply a built ollama37 image onto a testbed through CI — recreate the container, never docker compose by hand

Build the ollama37 CI image on a self-hosted GPU runner — never with a local make

Run a decision-grade FA/CUDA performance experiment via CI — verified throughput, self-reverting testbed

Graph-first verify discipline: re-ground in the codebase knowledge graph before AND after any non-trivial code change, and especially after a failed build/test or wrong output. Re-index the changed repo, re-trace the symbols you touched (callers, callees, data flow), and — for a port/adaptation — diff against the reference project, then verify against the live file before editing. Confirm any relationship- or reachability-based claim (who-calls, what-derefs, which-case-runs) with a graph trace, not grep alone. Use when implementing or porting non-trivial code, when an attempt fails, or before iterating on a fix. Complements (does not replace) the codebase-memory tool reference and codebase-graph-navigation IDE-movement skills.

Execute test cases with the simple judge by default, or opt in the agent judge

Generate YAML test cases from requirements or acceptance criteria

Build the certificate parameters for wifi_connect_enterprise (EAP-TLS/PEAP/TTLS) and diagnose enterprise-WiFi cert failures. Use when connecting to an 802.1X network, or when a connect fails with "Not a CA certificate", "unknown_ca" / AUTHENTICATION_FAILURE, or "mandates server certificate but validation is not enabled".

Build the android-wifi-mcp project. Use when the user asks to build the code, compile, or produce a fresh server bundle and/or companion APK. "Build" means BOTH the Node MCP server and the Kotlin companion app unless the user narrows it.

Execute android-wifi-mcp test cases against the attached Android device. Use when the user wants to run tests, execute test cases, or verify MCP tool behavior end-to-end.

Generate YAML test cases from an issue or feature request that test android-wifi-mcp tools against an attached Android device. Use when the user wants to create tests, generate test cases, or add test coverage for a feature.

Reviews the words of a human-read document — the README, the prose in docs/, or any text written for a person rather than an agent — for whether the phrasing fits its reader: leads with the point, stays brief, carries the right tone, and says the true complete thing and only that. The words half of the human-read doc review (reviewing-typography handles the look). Use when such a doc is written or updated and is about to reach a reader. Judgment over checklist; floor, not ceiling.

Reviews how a human-read document looks — the README, the prose and tables in docs/, or any human-read document — for visual hierarchy, Gestalt proximity, restraint, and walls of text, so a reader can find the point at a glance. The look half of the human-read doc review (reviewing-phrasing handles the words). Use when such a doc is written or restructured. Judgment over checklist; floor, not ceiling.

Reviews any workflow artifact — the commands, skills, and project docs that are the tooling (READMEs, stories, CLAUDE.md, and the like) — against a few goal questions: one clear job, complete, a goal not a frozen spec, fits the project, right for its reader, and (for producers) paired with a review. Its pairing question flags any producer shipped without a paired review. It judges whether an artifact does its job; how a human-read doc (the README, docs/ prose) looks and reads goes to the typography + phrasing review. Floor, not ceiling.

Reviews any workflow artifact — the commands, skills, and project docs that are the tooling (READMEs, stories, CLAUDE.md, and the like) — against five goal questions: one clear job, complete, a goal not a frozen spec, fits the project, right for its reader. Also runs a producer→review pairing coverage pass that flags any producer shipped without a paired review. It judges whether an artifact does its job; how a human-read doc (the README, docs/ prose) looks and reads goes to the typography + phrasing review. Floor, not ceiling.

Reviews the words of a human-read document — the README, the prose in docs/, or any text written for a person rather than an agent — for whether the phrasing fits its reader: leads with the point, stays brief, carries the right tone, and says the true complete thing and only that. The words half of the human-read doc review (reviewing-typography handles the look). Use when such a doc is written or updated and is about to reach a reader. Judgment over checklist; floor, not ceiling.

Reviews how a human-read document looks — the README, the prose and tables in docs/, or any markdown written for a person — for visual hierarchy, Gestalt proximity, restraint, and walls of text, so a reader can find the point at a glance. The look half of the human-read doc review (reviewing-phrasing handles the words). Use when such a doc is written or restructured. Judgment over checklist; floor, not ceiling.

Run K80-specific CI workflows for vLLM on the self-hosted runner. Use when triggering a remote build or smoke test, or deciding between local vs CI.

Build vLLM for Tesla K80 (sm_37) using the Docker builder/runtime pattern. Use when you need to produce a `vllm37-local` image, or are considering a host build (don't).