From ace0cd654e4538e0937fbe2ae3351da8856bcd9d Mon Sep 17 00:00:00 2001 From: Drew Stone Date: Sun, 9 Aug 2026 17:21:02 -0600 Subject: [PATCH] docs(skills): teach the shipped skills the bundle command and the prime engine --- skills/build-trace-analyst/SKILL.md | 13 +++++++++++++ skills/inspect-agent-traces/SKILL.md | 15 +++++++++++++++ 2 files changed, 28 insertions(+) diff --git a/skills/build-trace-analyst/SKILL.md b/skills/build-trace-analyst/SKILL.md index d9719cc..26d837f 100644 --- a/skills/build-trace-analyst/SKILL.md +++ b/skills/build-trace-analyst/SKILL.md @@ -48,6 +48,19 @@ Use `runAnalystBenchmark` with `registryBenchmarkRunner` from `@tangle-network/a Use AgentRx or CodeTraceBench labels when they match the finding. Compare the candidate with a no-findings baseline and relevant maintained or external analysts on identical cases and limits. +Two shipped engines are available as comparison arms. +`--analyzer prime` inlines the whole span projection into one prompt and needs no REPL. +`--analyzer halo` and `--analyzer hodoscope` drill into the trace with paged tools. +Compare an inline arm with a paged arm to separate the prompt from the navigation. + +```bash +traces analyze --last 1 --analyzer prime +``` + +Assert the served model, not the requested id. +A gateway can answer one id with another model, so a panel can report families it never ran. +Use `assertServedModel` and `assertCrossFamilyServed` from `@tangle-network/agent-eval` before you trust a per-model or cross-family number. + Use at least 20 independent cases before comparative statistics. Retain every case row and report exact-step recall, precision, F1, root-step accuracy when labeled, citation coverage and validity, clean-case false positives, repeat agreement, failures, latency, calls, all token fields, and known or missing cost. diff --git a/skills/inspect-agent-traces/SKILL.md b/skills/inspect-agent-traces/SKILL.md index 89855d9..cb08f6e 100644 --- a/skills/inspect-agent-traces/SKILL.md +++ b/skills/inspect-agent-traces/SKILL.md @@ -73,6 +73,21 @@ traces improve --harness codex --current --latest-turn --workflow \ `improve` writes findings, evidence, a report, and spans. It does not edit an agent, repository, memory store, or knowledge base. +Write one session's durable evidence directory when a later reader must cite the session: + +```bash +traces bundle --harness claude-code --session --out .traces/bundle +``` + +`bundle` copies the transcript, the derived report and spans, the `.evolve` ledger rows inside the session window, and a `manifest.json` with a SHA-256 per file. +It spends no model call. +Use it when the live stores can rotate before the reader arrives. +A missing transcript stops the assembly. +An absent optional input is recorded in `manifest.absent` with the probed path. + +The bundle holds the whole session transcript. +Do not give a bundle to a writer that must not see an earlier conclusion, because the transcript holds the text of every file the session wrote. + ## Report - State the source (`--otlp ` or the harness), selected task boundary, session and span counts, and integrity warnings.