hermes-agent/evals/core_tool_deferral

core_tool_deferral — live A/B harness for tool-visibility changes

Built for the PR #97979 maintainer battery (core-tool deferral behind the tool_search bridge). Runs REAL in-process AIAgents from two pinned checkouts and grades task outcomes programmatically — accuracy, api turns, tokens, wall, bridge-call counts — across any set of models.

Original verdict + full numbers: results/SUMMARY.md and the PR #97979 body (288 runs; gpt-5.6-terra / glm-5.3-flash / qwen3.8-27b).

Layout

• tasks.py — 14-task battery: one task per deferred tool, multistep (todo discipline, GUI chains), long-range (session_search → backup → cron → todo), a destructive-ambiguity clarify trap, an eager-only control, and a false-discovery distractor. Each task carries fixtures, a programmatic grader (0–1 partial credit), and scripted user replies.
• worker.py — one (arm, model, task, rep) cell in an isolated subprocess: temp HERMES_HOME + workspace, hermetic env (only OPENROUTER_API_KEY survives), seeded session DB (targets + decoys), deterministic desktop-surface stubs (desktop_ui emitter + agent callbacks), computer_use/image_generate stubbed at the registry handler. Terminal/files/cron/process/session-DB are REAL, but host side effects stay in the run dir (sandbox.py): HOME is the profile home (TERMINAL_HOME_MODE=profile), crontab is a file-backed shim first on PATH, and the systemd user bus is unset. Exit 3 = infra/config error (never scored).
• orchestrator.py — battery runner: resume-safe, per-task wall timeouts, parallel cells, errored-record retry, 3-infra-abort fuse.
• report.py — per-task table both arms (score spread, turns, tok, wall, bridge calls), mean-of-task-means, noise/error accounting.

Running

# 1. Two plain checkouts pinned to the SHAs under test (never pip install -e)
git worktree add "$TMPDIR/abdefer-base" <baseline-sha>
git worktree add "$TMPDIR/abdefer-pr"   <pr-sha>

export ABDEFER_BASE_TREE="$TMPDIR/abdefer-base"
export ABDEFER_PR_TREE="$TMPDIR/abdefer-pr"
export OPENROUTER_API_KEY=...   # the only key the worker keeps

# 2. Smoke one cheap cell first
python3 worker.py base openai/gpt-5.6-terra config_grep_distractor 1 "$TMPDIR/smoke.json"

# 3. Battery (per model; start with the STRONGEST model to validate variance)
python3 orchestrator.py openai/gpt-5.6-terra 3 --parallel=5
python3 orchestrator.py z-ai/glm-5.3-flash 3 --parallel=5
python3 orchestrator.py qwen/qwen3.8-27b   3 --parallel=5

# 4. Readout
python3 report.py

ABDEFER_PYTHON overrides the worker interpreter (defaults to the orchestrator's own); ABDEFER_RESULTS overrides the results root.

Discipline (from the readtool/session_search harness lineage)

• Verify model slugs against the live OpenRouter list before launching.
• Interactive fairness: if the agent ends its turn with a plain-text question, the worker sends the scripted reply (max 2, counted as user_roundtrips) — without this, every clarify-shaped task scores 0 unfairly and the battery is poisoned (the first terra run was discarded for exactly this).
• Same-denominator rule: errored runs score 0 and STAY in the accuracy denominator; they are excluded from efficiency means.
• Extend contested cells (score spread at n=3) to n=6 before concluding.
• For discovery-rate regressions, always check base-arm usage on the same tasks first — a tool models skip even when visible is not a deferral regression.
• Audit anomalous cells from *.transcript.json before publishing.

results/ is gitignored except SUMMARY.md — rep JSONs are rebuildable, verdicts are the artifact.