The agent that grows with you
Python aiai-agentai-agentsanthropicchatgptclaudeclaude-codecodexhermeshermes-agentllmnous-researchopenai
plugin-catalog: point skill-lineage at 0.1.3 and report the real-model result
Simin (Simon) Yuan71e7cb3abd4d
diff --git a/plugin-catalog/skill-lineage.yaml b/plugin-catalog/skill-lineage.yamlindex 09b6ed063dd..a56306ede9c 100644--- a/plugin-catalog/skill-lineage.yaml+++ b/plugin-catalog/skill-lineage.yaml@@ -1,8 +1,8 @@name: skill-lineagetitle: Skill Lineagerepo: https://github.com/simin-yuan/hermes-skill-lineage-sha: 37132da4b8fb8935f4ef07430bc854cf2bcc680c-version: "0.1.0"+sha: 8df4bcb730c7627a58ec5241c732f043a648c950+version: "0.1.3"description: >-Records where each skill version's content came from, so a skill that rewrites itself can be auditedafterwards. For every skill write — a foreground skill_manage call, a background-review edit, or a@@ -21,15 +21,17 @@ description: >-shown in `status` rather than happening silently. The plugin does not stop a write and does not judgea skill: it records and it verifies, and it does not claim a source caused an edit. Prior work onmemory lineage (MemLineage, arXiv:2605.14421) is acknowledged in the README.- Measured on 2026-10-21 against Hermes 3166d61: Python 3.14 59 passed; Python 3.11 51 passed,+ Measured on 2026-10-22 against Hermes 3166d61: Python 3.14 62 passed; Python 3.11 54 passed,1 skipped; `hermes plugins validate` reports no warning other than one known false positive (atest template that quotes a "curl | sh" instruction as data and never runs it).- A labelled synthetic evaluation — 600 generated scenarios on a held-out seed, with the acceptance- floors fixed before the run — puts precision at 0.988 for the default link set, with recall 1.00- on verbatim, paraphrase-with-operative-token and two-hop-via-curator cases, against 0.385 when- every link is recorded; the README states plainly that the paraphrases keep their operative token- by construction and that a rewrite by a real model is not yet measured. Raw output in- eval/results-seed2.json, reproducible with `python eval/run_eval.py --seed 2 --per-label 100`. A manual end-to-end run in an isolated HERMES_HOME+ A labelled synthetic evaluation — 600 generated scenarios on two held-out seeds, with the acceptance+ floors fixed before the run — puts precision at 1.000 on the second seed for the default link set,+ with recall 1.00 on verbatim, paraphrase-with-operative-token and two-hop-via-curator cases, against+ 0.385 when every link is recorded. A real model was then run through the plugin, 60 runs per plugin+ build: pre-registered false links fell from 53/60 to 6/60 and clean ones from 10/60 to 2/60+ (Fisher exact, two-sided, 4.1e-19 and 0.029), with the two batches being independent samples. Raw+ output in eval/results-seed2.json, eval/results-seed4-v013.json and eval/real/, reproducible with+ `python eval/run_eval.py --seed 2 --per-label 100`. A manual end-to-end run in an isolated HERMES_HOME(one real session creating a skill through skill_manage, thenstatus/trace/descendants/report/verify/head) passed, raw output in tests/smoke-2026-10-21.txt.Disclosure — four read-only hooks (on_session_start, pre_llm_call, pre_tool_call, post_tool_call);