grade-iterate
Phase 3 of building a Claude Managed Agent — the bounded grade→iterate loop. Define a CMA outcome (a required markdown rubric graded by an isolated grader), read each verdict, decide the next move (sharpen / re-run / promote to schedule), and once a version passes, run held-back eval cases in parallel. Use when the user says "grade my agent", "make it pass the rubric", "iterate until it's good", "is it good enough", or when the orchestrator routes phase=grade-iterate. outcome_builder.py builds the user.define_outcome payload (rubric required, max_iterations clamped 1..20 — never unbounded); verdict_reader.py reads the grader result and recommends the next move; eval_scaffold.py generates held-back cases + a parallel run plan (capped at the 25-thread CMA ceiling). Distinct from stage-launch (first launch) and run-without-you (scheduling).
- License 许可
- MIT
- 最近更新
- 2026-08-25
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- grade-iterate
- tags
- [ "cma", "outcome", "rubric", "grader", "grade-iterate", "loop", "max-iterations", "eval", "held-back" ]
- author
- Alireza Rezvani
- license
- MIT
- version
- 2.11.2
- description
- Phase 3 of building a Claude Managed Agent — the bounded grade→iterate loop. Define a CMA outcome (a required markdown rubric graded by an isolated grader), read each verdict, decide the next move (sharpen / re-run / promote to schedule), and once a version passes, run held-back eval cases in parallel. Use when the user says "grade my agent", "make it pass the rubric", "iterate until it's good", "is it good enough", or when the orchestrator routes phase=grade-iterate. outcome_builder.py builds the user.define_outcome payload (rubric required, max_iterations clamped 1..20 — never unbounded); verdict_reader.py reads the grader result and recommends the next move; eval_scaffold.py generates held-back cases + a parallel run plan (capped at the 25-thread CMA ceiling). Distinct from stage-launch (first launch) and run-without-you (scheduling).
- compatible_tools
- [ "claude-code", "codex-cli", "cursor", "antigravity", "opencode", "gemini-cli" ]
03 SKILL.md 正文
Phase 3 — Grade → Iterate (the bounded loop)
This is the plugin's loop: CMA's outcome primitive self-grades the agent's
work in an isolated context and feeds failing verdicts back for the next attempt.
It is always bounded by max_iterations (1..20) — never "improve forever".
See ../../references/loops-and-workflows.md
and the outcome section of
../../references/cma-primitives.md.
Workflow
- Define the outcome.
The rubric is required;python3 scripts/outcome_builder.py \ --sheet ./my-agent/build-sheet.json --max-iterations 5 \ --out ./my-agent/payloads/outcome.jsonmax_iterationsis clamped to 1..20. Send the payload as auser.define_outcomeevent (append to the running session). - Read every verdict first.
Tables the rubric outcome and recommends: SHIP (python3 scripts/verdict_reader.py --result ./my-agent/last-verdict.jsonsatisfied), SHARPEN then re-run (needs_revision), ESCALATE (max_iterations_reached/failed), RESUME (interrupted). With ≤1 iteration left it flips to "make the single highest-value fix or escalate now". - Loop invariant. Each iteration must move ≥1 rubric line fail→pass, or the run halts at the cap and escalates. Don't burn the budget on cosmetic edits.
- Once a version passes, run held-back eval.
Held-back cases (never seen during iteration) run in parallel, capped at the 25-thread CMA ceiling, each graded against the same rubric.python3 scripts/eval_scaffold.py \ --sheet ./my-agent/build-sheet.json --out ./my-agent/eval.json --concurrency 5 - Decide. SHIP as v0, or promote to a scheduled deployment (Phase 4). Record
the verdict on the goal:
goal_state.py set --phase run-without-you.
Hard rules
- Bounded, always. No outcome without a
max_iterationscap. - Read the verdict before acting. The grader's explanation drives the next move.
- Held-back cases are held back. Never grade generalization on cases the agent already iterated against.
Forcing-question library (recommend + cite)
- "What are the 3–5 rubric lines?" Recommend: grounded, checkable criteria. Cite: cma-primitives.md (rubric required).
- "How many iterations before you'd rather look yourself?" Recommend: 3–5. Cite: loops-and-workflows.md (bounded loop).
- "On a fail, sharpen the prompt or the tools?" Recommend: whichever rubric line failed points to. Cite: verdict_reader next-move table.
- "Which cases did the agent NOT see?" Recommend: hold back ≥3 for generalization. Cite: this SKILL (held-back eval).
Tools
scripts/outcome_builder.py— user.define_outcome payload (rubric required, cap 1..20).scripts/verdict_reader.py— grader result → next move.scripts/eval_scaffold.py— held-back cases + parallel run plan (≤25 threads).
04 同仓库更多 Skills
webmcpify
Make a web app agent-ready — propose a WebMCP tool manifest, integrate, verify in a real browser, heal; unrelated code stays untouched. Use for "webmcpify", "add WebMCP", or "expose app actions to AI agents".
vardoger-analyze
Use when the user asks to personalize the GitHub Copilot CLI assistant, adapt Copilot to their style, use vardoger, or analyze their Copilot CLI conversation history. Reads the local session directory at `~/.copilot/session-state/`, extracts recurring preferences and conventions, and writes a fenced personalization block into `~/.copilot/copilot-instructions.md`. Runs entirely on the user's machine via the local `vardoger` CLI (`pipx install vardoger`); no network calls and no uploads. Triggers: 'personalize my copilot', 'analyze my copilot history', 'tailor copilot to me', 'run vardoger', 'update my copilot instructions from history', 'make copilot learn my style'.
transloadit-media-processing
Process media files (video, audio, images, documents) using Transloadit. Use when asked to encode video to HLS/MP4, generate thumbnails, resize or watermark images, extract audio, concatenate clips, add subtitles, OCR documents, or run any media processing pipeline. Covers 86+ processing robots for file transformation at scale.
terraform-azurerm-set-diff-analyzer
Analyze Terraform plan JSON output for AzureRM Provider to distinguish between false-positive diffs (order-only changes in Set-type attributes) and actual resource changes. Use when reviewing terraform plan output for Azure resources like Application Gateway, Load Balancer, Firewall, Front Door, NSG, and other resources with Set-type attributes that cause spurious diffs due to internal ordering changes.
technical-job-search
Use this skill when a software engineer asks for help with job search tasks: parsing or analyzing a job description, tailoring a CV/resume, writing a cover letter, evaluating a job offer, or drafting a post-interview follow-up email. Do not activate for general career advice unrelated to an active job search action.
steno-mode
Shorthand-first response compression that cuts ~40% of response tokens while preserving technical precision and exact literals. Use when the user says "steno mode", "shorthand mode", "compressed responses", "token reduction", "brief structured output", or invokes /steno. Supports four compression levels: lite, brief, court, machine. Do not trigger for requests needing polished prose such as onboarding/tutorial content, stakeholder or customer-facing copy, or teaching-focused explanations.