Claude Skills by agentydragon

Debug props agent traces. Read LLM request/response history, parse tool calls, and "speak with dead" — resurrect a past agent conversation to ask follow-up questions about its decisions.

Propose narrowing of grader match_file_restriction for unrestricted TP/FP occurrences in a specimen. Produces verifiable, link-rich output that lets the user confirm each restriction is correct.

Operate the live props cluster as evaluator — fetch credentials from k8s, call the API at props.allegedly.works, trigger critic/grader runs, and inspect results.

Manual live props deployment testing — sets up Podman infrastructure (postgres, registry, backend) and runs real agent containers. NOT for standard Bazel tests (use `bazel test //props/...` for those).

Update the Claude Code web container reverse engineering effort. Detects changed binaries, captures new references, runs parallel RE subagents for bindiff/decompilation, updates container snapshot/diff, and refreshes all documentation.

Scan cluster health — Flux kustomizations, pod status, recurring crashes, node conditions, active Alertmanager alerts, CNPG databases, certificate expiry — and output an actionable summary with fix plan. Use when user asks "how's the cluster", "cluster health", "what's broken", "check the cluster", or similar.

Migrate a single-instance CNPG PostgreSQL cluster between regions or region-pinned storage classes via streaming replication with sub-second downtime

Audit every Flux Kustomization and HelmRelease over a window (default 7d): classify broken / slow / miswired / propagating / suspended / healthy with per-item culprit, failure counts, and p99 reconcile duration. Use for "what's slow / stuck / been broken this week".

Take and view screenshots of Hetzner Cloud servers via WebSocket VNC console. Use to diagnose issues when text commands fail — boot problems, unresponsive servers, kernel panics, stuck boot screens, graphical output inspection.

Interact with Proxmox VMs - screenshots, keystrokes, network info (user)

Analyze CPAP sleep therapy data from the user's ResMed AirSense 11. Read daily summaries (AHI, leak, pressure, compliance) and per-session waveforms from EDF files in the private cpap-data Forgejo git repo. Use when user asks about sleep quality, CPAP data, AHI, therapy compliance, or sleep analysis.

Reference for querying the BuildBuddy API: build/test logs, invocations, RBE execution details, cache stats, undeclared test outputs, target pass/fail history for bisecting, workflow re-runs. Use for any why-did-CI-fail, show-me-the-log, or when-did-this-test-break task.

Check CI/CD pipeline health: devel status, per-commit checks, release artifacts, pin staleness, image publishing, scheduled jobs. Fixes trivial issues as PRs, proposes diagnosis plans for deeper ones. Use when CI seems broken, red, lagging, or stuck.

Diagnose a Claude Code session's health against the hook daemon SPEC's acceptance criteria, plus setup/pin/bbr-runner/git-hook diagnostics. Use for "did setup go ok", "why isn't bbr working", "why do my commits fail", "selfcheck".

Creating and developing startup hooks for Claude Code on the web. Use when the user wants to set up a repository for Claude Code on the web, create a SessionStart hook to ensure their project can run tests and linters during web sessions.

Creating and developing startup hooks for Claude Code on the web. Use when the user wants to set up a repository for Claude Code on the web, create a SessionStart hook to ensure their project can run tests and linters during web sessions.

Inspect and safely clean stale local dev state: unused Git worktrees, merged branches, orphaned Bazel output bases. Use for pruning worktrees/branches or reclaiming Bazel disk — drive the `workspace-gc` tool and apply PR judgment to REVIEW items.

Audit a debundle spec's named modules for idiomatic JavaScript structure, infer project conventions from source behavior, and maintain current-state architecture notes and reorganization recommendations. Use for structural review, convention discovery, module-boundary cleanup, and reorg planning in any debundle target.

Turn `debundle modules propose` output into named, dispatchable seed clusters for debundle lane workers. Use for reading factorizer proposals, lightly grounding source meaning, choosing tentative destinations, and producing seeds.json without editing the spec or running gates.

Integrate multiple debundle lane-worker branches through a validated merge train. Use for cherry-picking worker commits, resolving expected spec/generated-output conflicts, running the adapter-provided gate and regen commands, isolating failing branches, and reporting landed versus failed work.

Apply one scoped debundle module-assignment or reorganization task in a worktree. Use for confirming atomic-DAG unit coverage, reading binding context, choosing honest module boundaries, editing debundle YAML (directly or via `debundle bindings assign` / `bindings rename` / `modules merge`), running the adapter-provided gate and regen commands, and committing one reviewable worker branch.

Assign descriptive names to unrenamed symbols in any debundle spec. Reads the emitted JS, understands each symbol's implementation and call sites, and updates the YAML spec `name:` field. Does NOT move modules or edit taxonomy. Trigger: user wants to name/rename minified symbols in a debundle RE spec.

Coordinate a generic AI-driven debundling loop across intake, lane worker, architect, integrator, planning, and naming skills. Use for multi-agent debundle rounds, work routing, graph refreshes, gate/regen command broadcast, progress tracking, and adapter-specific workflow control.

Plan and inspect generic JS debundle spec work using read-only `debundle` queries. Use when an agent needs to turn owner_graph.json plus a modules tree into dispatchable module extraction work, query atomic-DAG and coverage status, inspect graph/source context, or decide what debundle spec edits should be made. Generic to any debundle target.

Author forward-compatible `source_match` selectors for a debundle spec, replacing fragile minified-name pins with structural anchors tied to each entity's purpose. Not for renaming symbols (debundle_mint_names) or moving module boundaries (debundle_architect).

Deep forensic debugging that never stops until root cause is found or visibility limit is proven. Use when user wants to understand exactly why something is broken, not work around it. Activates on "why is this happening", "dig deeper", "don't work around it", "I want to understand", "find the root cause", "this seems suspicious", or when a problem suggests deeper breakage.

Take and view screenshots of Hetzner Cloud servers via WebSocket VNC console

Interact with Proxmox VMs - screenshots, keystrokes, network info (user)

Discover and analyze Claude Code session logs from ~/.claude/projects, including finding the current session and extracting tool calls, user messages, and conversation history

Make well-calibrated probability estimates using superforecasting methodology. Use when user asks about probability, likelihood, chance, odds, "will X happen", "when will X happen", "how much will X cost", "what could go wrong", failure modes, risk assessment, forecasting, or any question involving uncertainty and estimation.

Debug props agent traces. Read LLM request/response history, parse tool calls, and "speak with dead" — resurrect a past agent conversation to ask follow-up questions about its decisions.

Propose narrowing of grader match_file_restriction for unrestricted TP/FP occurrences in a specimen. Produces verifiable, link-rich output that lets the user confirm each restriction is correct.

Operate the live props cluster as evaluator — fetch credentials from k8s, call the API at props.allegedly.works, trigger critic/grader runs, and inspect results.

Manual live props deployment testing — sets up Podman infrastructure (postgres, registry, backend) and runs real agent containers. NOT for standard Bazel tests (use `bazel test //props/...` for those).

Update the Claude Code web container reverse engineering effort. Detects changed binaries, captures new references, runs parallel RE subagents for bindiff/decompilation, updates container snapshot/diff, and refreshes all documentation.

You have access to privileged tools which require user approval. Access to the MCP servers providing these tools is gated by an airlock MCP server. The exec environment provides `AIRLOCK_URL` and `OPENCLAW_SESSION_ID` automatically. You can interact with the Airlock using any MCP client; `mcporter` is one such tool available in the exec environment.

Creating and developing startup hooks for Claude Code on the web. Use when the user wants to set up a repository for Claude Code on the web, create a SessionStart hook to ensure their project can run tests and linters during web sessions.

Creating and developing startup hooks for Claude Code on the web. Use when the user wants to set up a repository for Claude Code on the web, create a SessionStart hook to ensure their project can run tests and linters during web sessions.

Audit a debundle spec's named modules for idiomatic JavaScript structure, infer project conventions from source behavior, and maintain current-state architecture notes and reorganization recommendations. Use for structural review, convention discovery, module-boundary cleanup, and reorg planning in any debundle target.

Turn `debundle peel plan-work` output into named, dispatchable seed clusters for debundle lane workers. Use for reading factorizer proposals, lightly grounding source meaning, choosing tentative destinations, and producing seeds.json without editing the spec or running gates.

Integrate multiple debundle lane-worker branches through a validated merge train. Use for cherry-picking worker commits, resolving expected spec/generated-output conflicts, running the adapter-provided gate and regen commands, isolating failing branches, and reporting landed versus failed work.

Apply one scoped debundle peel or reorganization assignment in a worktree. Use for confirming graph peelability, reading binding context, choosing honest module boundaries, editing debundle YAML, running the adapter-provided gate and regen commands, and committing one reviewable worker branch.

Assign descriptive names to unrenamed symbols in any debundle spec. Reads the emitted JS, understands each symbol's implementation and call sites, and updates the YAML spec `name:` field. Does NOT move modules or edit taxonomy. Trigger: user wants to name/rename minified symbols in a debundle RE spec.

Coordinate a generic AI-driven debundling loop across intake, lane worker, architect, integrator, planning, and naming skills. Use for multi-agent debundle rounds, work routing, graph refreshes, gate/regen command broadcast, progress tracking, and adapter-specific workflow control.

Plan and inspect generic JS debundle spec work using `debundle peel`. Use when an agent needs to turn owner_graph.json plus a modules tree into dispatchable module extraction work, query binding-patch status, inspect graph/source context, or decide what debundle spec edits should be made. Generic to any debundle target.

Show the current task stack and context. Use when user says "bt", "backtrace", "stack", "where are we", or asks about current progress on a multi-step task.

Split a large branch with many changes into independent, reviewable PRs. Use when preparing a messy development branch for code review, when asked to "split this into PRs", "make this reviewable", "break this up", or when a branch does too many unrelated things. Produces a DAG of branches/PRs that can be reviewed and merged independently.

Reference for querying BuildBuddy API. Use when investigating failed or slow CI builds, inspecting invocations by commit or branch, reading build or test logs, checking remote execution (RBE) details (exit codes, stderr, worker logs), analyzing cache hit/miss rates, downloading undeclared test outputs from RBE workers, bisecting test failures using target history to find the culprit commit range, viewing aggregated build statistics, triggering workflow re-runs, or getting AI analysis of build...

Check the health of the CI/CD pipeline — devel branch status, per-commit checks, release artifacts, artifact pin staleness, image publishing, scheduled jobs, repinning cadence. Autonomously creates fix PRs for trivial issues (formatting, etc.) or proposes a diagnosis plan for deeper problems. Use when you suspect CI is broken, main is red, releases are lagging, repinning is stuck, or images are falling behind.

Scan cluster health — Flux kustomizations, pod status, recurring crashes, node conditions, CNPG databases, certificate expiry — and output an actionable summary with fix plan. Use when user asks "how's the cluster", "cluster health", "what's broken", "check the cluster", or similar.