Run isolated eval and grading calls using CC 2.1.81 --bare mode. Constructs claude -p --bare invocations for skill evaluation, trigger testing, and LLM grading without plugin/hook interference. Use when running eval pipelines, grading skill outputs, benchmarking prompt quality, or testing trigger accuracy in isolation.
Instrucciones de origen · Vista previa de solo lectura
name
bare-eval
compatibility
Claude Code 2.1.220+
description
Run isolated eval and grading calls using CC 2.1.81 --bare mode. Constructs claude -p --bare invocations for skill evaluation, trigger testing, and LLM grading without plugin/hook interference. Use when running eval pipelines, grading skill outputs, benchmarking prompt quality, or testing trigger accuracy in isolation.
Run claude -p --bare for fast, clean eval/grading without plugin overhead.
CC 2.1.81 required. The --bare flag skips hooks, LSP, plugin sync, and skill directory walks.
When to Use
Grading skill outputs against assertions
Trigger classification (which skill matches a prompt)
Description optimization iterations
Any scripted -p call that doesn't need plugins
When NOT to Use
Testing skill routing (needs --plugin-dir)
Testing agent orchestration (needs full plugin context)
Interactive sessions
Prerequisites
# --bare requires ANTHROPIC_API_KEY (OAuth/keychain disabled)export ANTHROPIC_API_KEY="sk-ant-..."# Verify CC version
claude --version # Must be >= 2.1.81
Quick Reference
Call Type
Command Pattern
Grading
claude -p "$prompt" --bare --max-turns 1 --output-format text
Trigger
claude -p "$prompt" --bare --json-schema "$schema" --output-format json
Streaming grade
claude -p "$prompt" --bare --max-turns 1 --output-format stream-json
Optimize
echo "$prompt" | claude -p --bare --max-turns 1 --output-format text
Force-skill
claude -p "$prompt" --bare --print --append-system-prompt "$content"
@-file in prompt
claude -p "grade @fixtures/case-1.md against rubric" --bare (CC 2.1.113 Remote Control autocomplete)
Long harness runs (CC 2.1.199+): set CLAUDE_CODE_RETRY_WATCHDOG=1 for unattended eval batches — it raises the default retry count for non-capacity transient errors to 300 and lifts the cap of 15 on CLAUDE_CODE_MAX_RETRIES, so an overnight grading run survives transient API blips instead of dying mid-batch.
--output-format stream-json
Newline-delimited JSON events (one per token/tool-call) — lets a runner score partial output or abort early on a failing probe without waiting for the full response.
claude -p "$prompt" --bare --max-turns 1 --output-format stream-json \
| while IFS= read -r line; do# line is a single JSON event; inspect $.type == "content_block_delta"
jq -r 'select(.type == "content_block_delta") | .delta.text' <<< "$line"done
Use stream-json over json when:
grading long outputs and you want incremental scoring,
piping into another CLI step-by-step (e.g. ork:eval-runner),
you need per-token timing data alongside the content.
OrchestKit's eval scripts (npm run eval:skill) auto-detect bare mode:
# eval-common.sh detects ANTHROPIC_API_KEY → sets BARE_MODE=true# Scripts add --bare to all non-plugin calls automatically
Bare calls: Trigger classification, force-skill, baseline, all grading.
Never bare:run_with_skill (needs plugin context for routing tests).
CC 2.1.119: --print honors agent tools: / disallowedTools: (M122)
Before CC 2.1.119, --print mode ran with the full default tool set regardless of the agent's frontmatter tools: and disallowedTools:. Bare-eval grading was effectively ungated — graders could call any tool they wanted, even if the agent definition restricted them.
As of 2.1.119, --print enforces the agent's declared tool surface. Implications for eval design:
Consequence
Action
Eval graders that relied on unrestricted tool access may now fail
Audit grader prompts for tools they actually need; whitelist explicitly via the agent's tools: frontmatter
Eval results match interactive runs
Reproducibility improves — grading what the model can actually do, not what it could do in an unsandboxed --print
--agent <name> also honors permissionMode in --print
Permission-gated tools (Bash, Edit) require either permissionMode: acceptEdits or explicit allowlists in the agent definition
Migration test:
# Run an eval against an agent with a deliberately tight tools: list.# Graders that previously called Read/Bash freely will now fail unless those# tools are declared on the agent.
claude -p "$prompt" --bare --print --agent grader-test
If the grader fails with a "tool not permitted" error, add the required tool to the agent's tools: frontmatter and re-run.
CC 2.1.121: CLAUDE_CODE_FORK_SUBAGENT=1 for grader determinism (#1545)
Before CC 2.1.121, the env var only worked in interactive sessions. As of 2.1.121, non-interactive paths (claude -p, SDK) honor it too — each grader invocation gets a fresh forked subagent context.
The cross-eval state-leak problem this fixes:
Without forking, sequential claude -p --bare graders inherit harness state:
Inherited
Symptom
memory MCP query cache
grader sees stale hit from previous run; same fixture grades differently
.claude/chain/*.json on disk
grader for "implement" thinks "explore" already ran (file is from previous test)
ToolSearch deferred-tool cache
first grader's MCP loads bleed into next grader's tool registry
model picker pref
grader N inherits --model=opus from grader N-1
This produced ~5–10% retry rate and non-reproducible scores — the eval baseline drifted between runs, engineers chased phantom regressions.
Fix:tests/evals/scripts/lib/eval-common.sh exports CLAUDE_CODE_FORK_SUBAGENT=1, so every script that sources it (run-trigger-eval, run-quality-eval, run-agent-eval, optimize-description, etc.) gets forked graders automatically. The CI workflow .github/workflows/skill-eval.yml also sets it at the workflow level (its predecessor orchestkit-eval.yml was retired 2026-08-01 — its grading phases rated description prose, not behavior). Older CC silently ignores the env var (no-op).
Determinism contract: running the same grader on the same fixture twice in a row produces the same score. Verified by tests/evals/scripts/test-grader-determinism.sh.
workflows/skill-fitness.js is a runnable dynamic-workflow template — the workflow-backed complement to the static conformance grader (scripts/eval/conformance-check.mjs). It fans out one isolated-context agent per skill to score fitness (freshness / router-clarity / structure) and synthesizes a ranked scorecard, catching qualitative drift a static grep can't (description/body count mismatches, duplicate headings, install-specific absolute paths, version drift). Run it with the Workflow tool:
Treat it as a template, not a verbatim script — adapt the SKILLS list and rubric per use. Cost is real (~50k tokens/skill; scoring all ~112 is ~6M tokens), so pass an explicit batch via args. Static-first: run conformance-check.mjs (zero tokens) to pre-filter, then this harness for the judgment grep can't make.
Holdout Bake-Off Grading
The holdout-promotion gate grades a champion and a challenger SKILL.md over the same sealed holdout via bare-mode forked graders — the canonical consumer of the determinism contract above: identical grader + identical ork-rubric/1.0 + identical sealed set, with CLAUDE_CODE_FORK_SUBAGENT=1 so the only variable is the version under test. Both --bare constraints apply (requires ANTHROPIC_API_KEY, bills tokens directly → on-demand / CI only). Run it with bash tests/evals/scripts/run-skill-eval.sh --holdout-promote <skill>.
