菜单
Use when eval-author has routed to this sub-flow for one Intake trace. Do not use for instrumenting agents, ingesting telemetry, importing a trace store, verifying ingest, or querying Intake outside Eval Author; those belong to nemo-intake.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/NVIDIA-NeMo/nemo-platform --skill eval-author-inspect-trace命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Fine-tune models on NeMo Platform with `automodel`, `unsloth`, or `rl` (all `submit`-only): HF dataset conversion, filesets, model entities, and job JSON (hyperparameters, batch, schedule, optimizer) + job polling. `automodel`/`unsloth` run SFT/LoRA as Docker GPU jobs; `rl` runs DPO (preference) or GRPO (NeMo Gym environment + reward) on a Ray cluster (Kubernetes). Covers building and converting NeMo Gym environment packages (native-v1, wheels-v1, adapter-wheels-v1) and Gym rollout-row datasets. Use for train, fine-tune, customize, SFT, LoRA, DPO, GRPO, RLHF, reinforcement learning, reward environment, NeMo Gym, verifiers, Prime Intellect, preference optimization, learning rate, epochs, or nemo customization.
Generate, validate, and measure an audit-spec coverage denominator for Eval Author. Use when the user wants a hand-editable audit.md file derived from Ethos, needs schema enforcement for declared tools, capabilities, failure cases, evidence, and references, or wants to measure which audit items one ATIF trace covers. Changes none of the user's source, and saves audit artifacts under `.eval-author/`.
Work on evaluation suites in a user's repository or understand an agent run from NeMo Intake. Owns the evidence standard that every Eval Author sub-flow follows. Use when the user asks "help me with my evals", "what's the state of the eval suite here?", "what happened in this trace?", or when you need to pick between the Eval Author sub-flows. Routes to a sub-flow and changes none of the user's source. The selected sub-flow uses the provider's supported tools and saves its findings under `.eval-author/`.
正在显示 SKILL.md
| name | eval-author-inspect-trace |
| description | Use when eval-author has routed to this sub-flow for one Intake trace. Do not use for instrumenting agents, ingesting telemetry, importing a trace store, verifying ingest, or querying Intake outside Eval Author; those belong to nemo-intake. |
| triggers | ["eval-author routed to inspect-trace","inspect the intake trace eval-author named","continue eval-author intake trace inspection"] |
| not-for | ["eval-author (use for the standard, the boundaries, and to pick a sub-flow)","eval-author-discover (use to establish whether a repository's Harbor suite is runnable)","eval-author-audit (use to validate an existing finite audit.md coverage denominator)","nemo-intake (use to instrument agents, ingest telemetry, or query Intake outside Eval Author)","nemo-experimentalist (use for an optimization experiment across many trials)"] |
| compatibility | A working nemo CLI invocation, an explicit workspace, and read access to Intake on a configured local or remote NeMo Platform instance. |
| maturity | alpha |
| license | Apache-2.0 |
| user-invocable | false |
| allowed-tools | ["Bash","Read","Write","Grep","Glob"] |
Read one Intake trace, explain its trajectory, and save an evidence-backed
report. Start this sub-flow only after eval-author selects it. For
instrumentation, ingest, or a general Intake query, use nemo-intake. Read
eval-author for the shared standard and boundaries.
Use a working nemo CLI invocation with an explicit workspace and read access to
Intake. The configured NeMo Platform instance can be local or remote. Use the
active context, a caller-supplied context, or NMP_BASE_URL and
NMP_ACCESS_TOKEN.
Resolve the CLI invocation once before the first Intake read. If the caller
provides an invocation, verify that caller-supplied invocation first. For a
caller-supplied uv run nemo invocation, add --no-sync.
Otherwise, probe these candidates in order with the --version flag:
nemo when command -v nemo succeeds..venv/bin/nemo when that executable exists in the source checkout.uv run --no-sync nemo when uv and the prepared environment provide
the CLI.Continue after a launcher error while another candidate remains. If no candidate works, quote the launcher errors and stop.
Use the resolved invocation for every command. The examples use nemo as the
logical prefix; replace only that prefix. In the report, record each expanded
command instead of the logical example.
Don't install the CLI, authenticate, change contexts, start services, or repair
the platform. Don't run uv sync, use uvx, or activate an environment. Don't
print, copy, or save credentials. The first Intake read validates access. If it
exits nonzero, quote the CLI error and stop.
Accept a bare Intake trace ID or intake://traces/TRACE_ID. Reject every other
URI scheme. If no trace is named, list recent candidates:
nemo intake traces list --output-format=json --workspace="WORKSPACE" --mode=preview --sort=-started_at --page-size=20
If the request names an agent, list recent traces that contain its spans:
nemo intake spans groups list --output-format=json --workspace= --by=trace_id --filter.agent-name= --=-started_at --page-size=20
For a time-bounded agent search, add the command's JSON --filter value with
started_at.gte and started_at.lte. Don't encode the filter. If the command
returns no candidates, report the empty result and stop. Don't widen the search
without user direction.
Read the server-computed trace rollups:
nemo intake traces get --output-format=json --workspace="WORKSPACE" --mode=detailed "TRACE_ID"
Then read every span without payloads:
nemo intake spans list --output-format=json --workspace="WORKSPACE" --filter.trace-id="TRACE_ID" --mode=summary --sort=started_at --all-pages
Use Intake's root status, duration, token, cost, span-count, error-count, model, and provider rollups. Use span timestamps and parent IDs when order matters. Don't recreate a timeline object or assume that an error span made the trace fail.
Read bounded previews for a server-side selection:
nemo intake spans list --output-format=json --workspace="WORKSPACE" --filter.trace-id="TRACE_ID" --filter.status=error --mode=preview --sort=started_at --page-size=20
Use --filter.kind or --filter.parent-span-id for equivalent selections. Read
an exact payload only when its text affects the assessment:
nemo intake spans get --output-format=json --workspace="WORKSPACE" "SPAN_ID"
Require the returned trace_id to equal the selected trace ID. Otherwise,
discard the span and report the mismatch. Don't recreate payload-length flags.
Collect unique session IDs from the summary spans. Query each relevant session:
nemo intake evaluator-results list --output-format=json --workspace="WORKSPACE" --filter.session-id="SESSION_ID" --sort=created_at --all-pages
Keep only results whose span_id belongs to the selected trace. Quote evaluator
comments when they affect the outcome.
Use exact payload text only when it changes the assessment. For every important moment, name the span ID, input or attempt, output, and relationship to nearby spans. Separate observations from interpretations.
Assign one outcome:
success: Direct evidence shows that the requested work completed.failure: Direct evidence shows that the work didn't complete or was wrong.unknown: Evidence is missing, incomplete, or contradictory.Start every finding with one category:
behavior: What the agent did without judging it as defective.issue: A result that harmed correctness, safety, cost, or latency.recovery: A later action that handled an earlier problem.uncertainty: A question that the evidence can't settle.Don't invent an issue for a successful trace. Tie every important claim to a
span ID, evaluator result ID, or source symbol.
Inspect local source only after trace evidence identifies behavior that source
can explain. Name the repository-relative path and symbol. Source can't override
recorded trace evidence. Report conflicts as uncertainty.
Use the trace ID returned by Intake. Replace characters outside letters,
numbers, ., _, and - with _. Save the report as
.eval-author/traces/intake-TRACE_ID.md, replacing any report for the same trace.
Put only source metadata and exact read commands in JSON-compatible YAML front matter. Record the workspace and a named CLI context when used. Never record a token. Then include outcome, summary, important moments, findings, and optional source evidence:
---
{"source":{"kind":"intake","trace_id":"TRACE_ID","workspace":"WORKSPACE","context":"CONTEXT_OR_NULL"},"commands":["EXACT_READ_COMMAND"]}
---
# Trace TRACE_ID
## Source
<SOURCE_METADATA>
## Outcome
<OUTCOME_AND_DECISIVE_EVIDENCE>
## Summary
<TRACE_SUMMARY>
## Important moments
- `<SPAN_ID>`: <TRANSITION_AND_SIGNIFICANCE>
## Findings
- **<FINDING_CATEGORY>**: <CLAIM_AND_EVIDENCE_IDS>
## Source evidence
- `<PATH>:<SYMBOL>`: <SOURCE_BACKED_EXPLANATION>
Omit Source evidence when you didn't inspect source. Before completion, verify every important moment and finding names evidence, the outcome uses an allowed value, and only the report changed. Leave the report uncommitted and state its path.