elizaos--eliza/scripts/training-harvest/README.md

gpt-5.5 trajectory-training harvest pipeline

Stage 2 corpus harvest for the gpt-5.5 → eliza-1 training pipeline. Runs the elizaOS test/eval corpus through gpt-5.5 (Codex subscription), captures every run as an eliza_native_v1 trajectory + pass/fail verdict, so Stage 3 can GEPA-repair the failures and Stage 4 can extract the passes into training data.

Stage 1 status: inventory + wiring + driver proven on a small slice. The full corpus is NOT run here.

Files

File	Role
build-manifest.mjs	Enumerates the whole corpus → manifest.json (families, items, run commands, trajectory landing). Discover-only; runs nothing.
manifest.json	Machine-readable corpus manifest (generated).
harvest-runner.mjs	Stage-2 driver. Consumes the Stage-1 provider incantation, iterates the manifest, runs each item, captures trajectory + verdict into harvest/.
s1-provider.example.json	Example provider-env the driver consumes ({ELIZA_CHAT_VIA_CLI:"codex", ELIZA_CLI_CODEX_MODEL:"gpt-5.5"}). Stage-1 leg S1 emits the real one.

Provider seam (how gpt-5.5-via-Codex slots in)

The scenario runner already has a first-class CLI-subscription provider (packages/core/src/testing/live-provider.ts → selectCliProvider). Setting ELIZA_CHAT_VIA_CLI=codex selects provider "cli", model gpt-5.5, plugin @elizaos/plugin-cli-inference, which reads ~/.codex/auth.json itself — no API key ever passes through eliza. The driver injects this env per spawn; it is never hard-coded. Precedence: --provider-env <s1.json> → $HARVEST_PROVIDER_ENV_FILE → inherited ELIZA_CHAT_VIA_CLI/API key → --deterministic (offline self-test).

Families & trajectory emission

• scenario — @elizaos/scenario-runner drives a real AgentRuntime+PGLite. Emits eliza_native_v1 natively via --export-native. Verdict per scenario id (report.scenarios[].status, mirrored to native row.scenarioStatus). 978 base scenario ids across 7 dirs (~10,716 with persona expansion).
• benchmark — packages/benchmarks orchestrator, 53 adapters. Emits per-benchmark result JSON today. Native trajectories require wiring: the ~25 eliza-adapter-routed benchmarks boot a real runtime; set ELIZA_SAVE_TRAJECTORIES=1 ELIZA_TRAJECTORY_DIR=<dir> then convert <dir> with scenario-runner native-export. Python-only/standard/* benchmarks don't boot the runtime and can't emit native trajectories.
• e2e — 51 *.live.e2e.test.ts / *.real.e2e.test.ts vitest lanes that drive a real runtime. Same trajectory wiring as benchmarks; verdict is coarse (vitest pass/fail per file).

Only the scenario family emits eliza_native_v1 today, so the driver executes that family in Stage 1. Benchmark/e2e are enumerated with their wiring notes for Stage 2.

Harvest layout (what Stage 2 writes)

<repo>/reports/training-harvest/gpt55/harvest/
  <family>/<dir-slug>/<item-slug>/
      report.json                 scenario aggregate report (per-scenario status)
      native.jsonl                eliza_native_v1 rows (scenarioStatus + judgeScore)
      native.jsonl.manifest.json
      verdict.json                { item, status, rows, judgeScore, exitCode }
      run/trajectories/**         RecordedTrajectory JSON (source for native export)
      stdout.log / stderr.log
  harvest-summary-<run|dryrun>-<ts>.json

Usage

# rebuild the manifest
node scripts/training-harvest/build-manifest.mjs

# driver self-test (offline, deterministic proxy)
node scripts/training-harvest/harvest-runner.mjs --deterministic --limit 3 --dry-run

# STAGE 2 real harvest (consumes S1's proven provider env)
node scripts/training-harvest/harvest-runner.mjs \
  --provider-env <s1-output.json> --family scenario

Prereq in a fresh worktree: generate the i18n keyword data once — node packages/shared/scripts/generate-keywords.mjs (gitignored build artifact; the CLI imports packages/core/src/i18n/generated/).