feat(evaluator-sdk): agent-skill injection + SkillUsedMetric for Fabric agent-eval#625
Draft
SandyChapman wants to merge 1 commit into
Draft
feat(evaluator-sdk): agent-skill injection + SkillUsedMetric for Fabric agent-eval#625SandyChapman wants to merge 1 commit into
SandyChapman wants to merge 1 commit into
Conversation
…ic agent-eval Add agentskills.io skill injection to the Fabric agent-eval runtime for A/B skill evaluation, a SkillUsedMetric, and a runnable example. Draft: the live example currently needs the pre-`Fabric`-rename NeMo-Fabric SDK; the FabricClient->Fabric migration is tracked separately. - AgentSkill (directory-backed agentskills.io bundle) + install_skill: stage a skill per task and wire it to the harness -- native `skills.paths` overlay for Hermes, `.agents/skills/<name>/` placement for Codex. FabricAgentRuntime gains a `skill=` param and a `with_skill()` clone for confound-free A/B; skill provenance (SkillProvenance) is stamped into trial metadata. - SkillUsedMetric: skill_present / skill_used booleans to flag a failure to use an injected skill (best-effort trajectory detection). - fix: include `provider` in the Fabric workspace/trajectory environment overlays so a config with no base environment serializes (was FabricConfigError "missing field `provider`"). - run_id folded into the Fabric evidence path so A/B runs don't collide. - examples/skill_eval: runnable baseline-vs-treated A/B over a skill required to pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: Sandy Chapman <schapman@nvidia.com>
Contributor
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Adds agent-skill evaluation to the Fabric agent-eval runtime: inject an
agentskills.io skill into an eval and A/B the same
taskset with vs. without it.
AgentSkill+install_skill(runtimes/fabric/skills.py) — adirectory-backed agentskills bundle, staged per task and wired to the harness:
native
skills.pathsoverlay for Hermes;.agents/skills/<name>/placementfor Codex (which discovers it there). Namespaced under
<name>/so it can'tcollide with task-seeded workspace files.
FabricAgentRuntimegains askill=param and awith_skill()clonefor confound-free A/B (baseline vs. treated differ in exactly the skill).
Skill provenance (
SkillProvenance: name, content hash, mode, adapter, location)is stamped into trial metadata.
SkillUsedMetric—skill_present/skill_usedbooleans, to flag afailure to use an injected skill (best-effort trajectory detection; strongest
for Codex filesystem discovery).
run_idis folded into the Fabric evidence path (<root>/<run_id>/<task>)so A/B runs don't collide.
providerin the workspace/trajectoryenvironmentoverlaysso a config with no base
environmentserializes (wasFabricConfigError: missing field 'provider').examples/skill_eval/— a runnable baseline-vs-treated A/B where the skill(the "Supercool Coding Guidelines":
supercool_prefix + Frencheniemeparam) is required to pass
follows_guidelines.Status: draft (WIP)
Saving current state. Not ready to merge:
mainrenamedFabricClient→Fabric(config-first run model, norun(profiles=...)), sothe existing Fabric runtime (
runtime.py, and feat(evaluator-sdk): sandboxed containerized Fabric runner (AALGO-321) #604'scontainer_runtime.py)is incompatible with current
mainand the live example can't run against it.The
FabricClient→Fabricmigration is a separate, coordinated effort and istracked on its own branch.
tools/lint/lint-all.sh+ a live end-to-end run are pending before thisleaves draft.
Verification
test_fabric_skills.py,test_skill_used_metric.py,test_fabric_runtime.py— 42 passed;
ruff+ focusedtyclean.providerfix was verified against a liveFabricClient().plan(...)repro(serializes clean) when a compatible Fabric core was installed.
🤖 Generated with Claude Code