cloudzero-claude-marketplace

Use when proactively scanning for cost anomalies, unusual spending, unexpected charges, or irregular patterns — during weekly reviews, after incidents, or when something looks off

Use when comparing costs between time periods, environments, accounts, regions, or teams to understand spending differences and identify inefficiencies

Project the monthly cost of an infrastructure definition (Terraform, CDK, CloudFormation, SAM) using CloudZero spend data. Reads IaC files, enumerates resources, and produces a line-item cost breakdown.

Use when a cost spike or unexpected increase has already been identified and you need to find which service, account, or resource is responsible

Use when analyzing whether costs are growing, declining, or stable over time — for forecasting, budget planning, or understanding spending velocity

Use when analyzing costs by organization-specific dimensions like teams, products, business units, or applications for showback, chargeback, or business-aligned cost reporting

Analyze code diffs for infrastructure cost impact using CloudZero spend data. Detects Terraform, CDK, CloudFormation, SAM, K8s, scaling, and application code changes that affect cloud spending.

One-shot, PER-CALL-SITE audit of a repo's EXISTING LLM calls — every place code invokes a model (an SDK/API call, a sub-agent dispatch, an agent frontmatter definition), decomposed by INTENT into the distinct jobs it does, never a grep hit on a model name and never one candidate per file. A skill pinned by one static `model:` key still gets its steps split by intent when severable under Claude Code's per-turn model binding. DELEGATES each call to `model-right-sizer-dryrun` (never re-implements its scoring) and merges results into ONE schema-conformant JSON blueprint committed at the TARGET REPO'S ROOT via a PR — never a markdown table. Distinct from `model-right-sizer-dryrun` (invoke directly for one hypothetical task) and `model-right-sizer-install` (the standing before/after mandate). Use when someone says "audit this repo's model calls", "right-size the models in <repo> per call site", "find every LLM call and right-size it", or "commit a model right-sizing blueprint for <repo>".

The while-work-is-in-flight companion to `model-right-sizer-dryrun`: once `work_routing_map[]` is real and being dispatched as sub-agents, the runbook for keeping two things honest — the status ledger (`status`/`status_updated_at`/`status_note` per row, flipped at every real transition) and the token-budget guard (checking real spend against `budget.token_ceiling`; once `budget_threshold.py`'s `threshold_crossed()` trips at `warning_threshold_pct`, sending `format_budget_warning()`'s string into that unit's next turn). Checks at turn boundaries with whatever usage the dispatch mechanism reports — never a fabricated live ticker. Only applies to rows actually dispatched, never design-time-only `blueprint_rows[]`, and never replaces Pass B's usage report. Use when someone says "dispatch the work-routing map", "run the budget guard while units are in flight", "update the status ledger for unit X", or "did unit X cross its warning threshold".

Preview the model-routing MAP for an intent without building anything. You give it a free-text INTENT — "build a CLI that…", "add a feature that…", "refactor X into Y" — and it invokes the `model-right-sizer` agent in BLUEPRINT-ONLY mode: decompose the work, score each piece, and emit a single schema-conformant JSON blueprint (task→model→effort→budget→ schema→confidence, per `schemas/blueprint.schema.json`), then STOP. No build, no file edits, no after-the-fact usage report — it is the what-would-this-cost / how-would-this-route preview lever, safe to run against any idea, and the JSON it returns is what an orchestrator parses to route dispatch. Read-only. Use when someone says "dry-run the right-sizer on …", "what's the map for …", "how would you route …", or "show me the blueprint for … before I build it".

Tune model-right-sizer's wording knobs (`eval/tuning/knobs.py`) against a REAL, already-measured build's actual token spend — not the synthetic benchmark `model-right-sizer-prompt-tuning` searches, nor a fresh build per candidate. Picks a task from `overfitting_guard.py`'s `HOLDOUT_TASKS` registry (real actual/budgeted pairs already recorded), dispatches 3 INDEPENDENT BLIND dry-runs per candidate (no calibration-ledger access), averages the budget across draws (single-draw noise can flip within/over-budget calls), maps to real actuals, scores via `classify_budget_adherence` + `score_candidate`, diagnoses the miss pattern, proposes ONE wording change, and re-runs to check improvement. n stays fixed per task — flags rather than silently continues once squeezing looks like overfitting. Use when someone says "tune the knobs against this blueprint/build", "iterate the dry run with no prior context against the real actuals", "keep tuning until N%", or "how close does a blind estimate get to the actual cost".

Stamp a model-right-sizer mandate onto the CURRENT repo's agent-instructions file — `CLAUDE.md`, `AGENTS.md`, or both, whichever the repo has. Idempotent and append-only: an existing file is never overwritten, only the marker-delimited mandate block is inserted or refreshed. The mandate's "before" hook runs `model-right-sizer-dryrun` to produce a schema-conformant JSON blueprint for the orchestrator to route by. Also checks whether the `model-right-sizer` agent and `model-right-sizer-dryrun` skill are discoverable and, if either is missing, installs the `model-right-sizer` Claude Code plugin to fix that (falling back to manual instructions otherwise). Deliberately narrow and organization-agnostic — installs only the right-sizing mandate (plus its own dependencies, if absent), not any broader development process. Use when someone says "install model-right-sizer in this repo", "init this repo for model-right-sizer", "add the right-sizer mandate here", or "make this repo consult model-right-sizer every turn".

Empirically measure what each of model-right-sizer's four research-grounded citation layers (Token Economics, IBPO, BudgetThinker, Speculative Decoding) actually does to its blueprints — instead of trusting the citations alone. Renders layer-ablated variants (any of the 16 layer subsets), runs a fixed six-task benchmark through each variant's Pass A blueprint, and for a scoped subset actually executes the recommended build and scores whether real effort stayed within the predicted budget (wrapping `classify_budget_adherence`). Reports each layer's effect in ISOLATION vs. a zero-layer baseline, and every COMBINATION across the full 16-subset grid, so synergy or redundancy is visible, not assumed away. Read-mostly: writes only a scratch directory and a final report, never `agents/model-right-sizer.md`. Use when someone says "does the Token Economics layer actually change anything", "ablate the research layers", "run the layer-ablation study", or "audit model-right-sizer's citations empirically".

Tune the exact WORDING of model-right-sizer's four already-shipped research-grounded layers to maximize real-execution accuracy (effort stayed within budget, per `classify_budget_adherence`) — not whether to include a layer (see `model-right-sizer-layer-ablation`), but how an included layer should be phrased. Runs a discrete coordinate-ascent search (the finite-difference analog of gradient descent for prose) over four wording knobs in `eval/tuning/knobs.py`, each anchored at one spot in the shipped agent text plausibly moving the accuracy ratio: `token_ceiling` margin, how hard the effort dial leans down under difficulty-uncertainty, and the calibration/adherence knobs. Read-mostly: never edits the agent file directly, only proposes the winning wording as a diff to review. Use when someone says "tune model-right-sizer's wording for accuracy", "optimize the budget-ceiling wording", "run a gradient descent / hill-climbing search on the agent prompt", or "which wording maximizes budget-adherence accuracy".

Publish a per-version release report for `eval/token_ceiling_formula.py` every time `FORMULA_VERSION` bumps — the exact configuration (formula, signal weights, calibration constants), a ranked gap list for the next contributor, and what's settled and not worth re-relitigating. Why this is its own skill rather than folding into `model-right-sizer-research-report`: every claim must interweave WHY it matters, in the same breath as WHAT changed — it stays decision-support only if a reader can tell, without a second file, what breaks (wasted spend, false alarms, undetected overruns, a re-biased fleet of budgets) if a number or gap is wrong. Never a new-finding surface — synthesizes only from committed results files. Also the tool for BACKFILLING a report for a past version that shipped before this skill existed. Use when someone says "write the release report for this version", "version the token ceiling formula", "backfill a release report for v0.x", or after any `FORMULA_VERSION` bump.

Package up every result this plugin's tuning/validation research has produced — the layer-ablation study, the prompt-tuning coordinate-ascent passes and the dispatch-floor-awareness/held-out-task work, the averaged-vs-additive `token_ceiling_formula.py` pivot, and the real-work-signal validation experiments — into one condensed, research-paper-style EXECUTIVE report with real charts, built entirely from numbers already recorded in this repo's own dated results files (never invented or rounded up). Publishes a self-contained HTML report (loads the `dataviz` and `artifact-design` skills first) with an abstract, a key-findings table, a handful of figures, limitations stated as prominently as wins, and a reproducibility appendix pointing at the companion skills that can re-run each experiment. Use when someone says "write up all the tuning results", "executive summary of the research", "package the findings into a report", "research report with charts", or "summarize everything we've found so far for leadership".

Prescribe a minimal output schema for ONE agent's handoff to its controller — the `model-right-sizer` agent's "Agent-to-agent message-schema design" lever, scoped to a single seam instead of a whole flow's blueprint. Given a target agent (a path to an existing agent `.md` file, or a description of one not yet written), returns a schema-conformant JSON prescription (`schemas/agent-schema.schema.json`) naming the reusable family the agent's reply fits, typed `in`/`out` fields, an exclusion list, and a ready-to-insert `## Agent-to-agent schema` markdown stamp — reproduced here in portable, organization-agnostic form. Offers to stamp that block directly into the target agent's file, idempotently. The point: an agent that used to hand its controller unscoped prose now hands it typed fields plus one bounded prose slot. Use when someone says "give this agent an output schema", "prescribe a schema for …", "minimize what this agent returns", or "stamp an agent-to-agent contract on …".

Test whether a candidate real-work signal in `eval/token_ceiling_formula.py` (e.g. `context_ingestion_volume`, `investigative_uncertainty`, or new) deserves a nonzero default weight, via the blind multi-draw rating + correlation methodology this repo's research converged on. Ratings must come from genuinely independent sub-agent dispatches seeing ONLY a forward-looking task spec and the signal definitions — never a context holding the real actuals or this repo's retired write-ups, which would turn "blind rating" into transcribing the answer key. Dispatches 3+ independent draws per held-out task, computes per-signal CV and Pearson correlation (alone, and added to the signal sum — dilution, not weak correlation, is the dominant failure found twice), requiring replication on a SECOND held-out task before proposing a nonzero weight. Use when someone says "test this new signal", "does [signal] deserve a nonzero weight", "re-run the signal validation experiment", or "validate real-work signals against real data".

Fetch top unaddressed CloudZero Optimize recommendations, dispatch parallel research agents per item, apply SRE critique, and surface actionable findings with confidence verdicts and per-resource report files. Read-only research only.

Cut a release of this repository (the CloudZero plugin marketplace itself — not a customer-facing plugin). Promotes every touched plugin's own `Unreleased` changelog section to a dated entry, synthesizes those changes into a new SemVer section at the top of the root `CHANGELOG.md`, bumps `.claude-plugin/marketplace.json`'s version, runs the full CI validation suite locally, tags the release, publishes a GitHub Release, and opens the companion docs PR against `Cloudzero/cloudzero-documentation`'s `v2.0` branch so customer-facing docs land in step with the code. This is a maintainer-only, repo-local skill (lives in `.claude/skills/`, not inside a `plugins/*/skills/` directory) — it is never installed by marketplace consumers. Use when told "cut a release", "release the marketplace", "ship a new version", "promote the changelog", or after merging PR(s) that need a version bump and a docs update — mirroring how v1.2.0 (the Model Right Sizer launch) shipped.