Agents

Each agent owns one bounded stage of the workflow and runs in its own isolated context. Only reward-tuning-agent dispatches workers of its own; every other agent is a leaf, which is what keeps dispatch depth at two.

benchmark-generator

Adds benchmark sanity scaffolding to a Python env that dependency-generator already built and verified (uv backend — host venv at <repo>/.venv/). Reads repo markdown for benchmark-level context, renders TWO scripts (random-action rollout + render-to-MP4), runs a 2-tier smoke (L1 random / L2 render), captures the suite spec into <repo>/harbor/benchmark-generator/benchmark-spec.json, and emits history.md + benchmark.md receipts. Does NOT generate train/eval scripts — that scaffolding is owned by rl-integration-generator. Does NOT modify the env config or dependencies — dependency-generator owns the environment, including the imageio[ffmpeg] extras line. The ONE deliberate env-source edit it makes is the IsaacLab dt-strip (Step 3.4): a single-line, idempotent reward-semantics normalization of the vendored RewardManager.compute().

Tools · Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion
Source · agents/benchmark-generator.md

dependency-generator

ENTRY POINT for setting up a Python simulation repo via uv on the host. Probes the repo, reads README + markdown to build an InstallationPlan, renders <repo>/harbor/dependency-generator/setup_uv.sh, executes it (creates <repo>/.venv/), runs an import smoke test, and reports back to the main thread. Does NOT recursively dispatch to sub-subagents — the main thread orchestrates the next step (benchmark-generator).

Tools · Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion
Source · agents/dependency-generator.md

dr-generator

COMING SOON — not released, and never dispatched by /harbor:task-create in this version. Authors §7 (domain randomization) of a task in a benchmark repo across THREE groups — (1) robot, (2) object, (3) observation noise. Two modes — create (replace the empty DR slot left by task-generator) and edit (overwrite an existing DR config). Discovers every available randomization term per group, wires the requested/default ones once-per-episode-per-env (mode="reset"), then smoke-checks each effective term by exact value read-back at num_envs=16. Reads task-implementation.md as a per-benchmark migration aid; relies on its own contracts (smoke template + IsaacLab DR reference). Phase A authors DR (or skips per the 3-condition gate); Phase B renders + runs the §7 smoke. Iterates up to 2× on smoke failure; ambiguity batches into a single AskUserQuestion. Writes/updates a handoff at <task_dir>/handoff-dr-generator.md (under harbor/create-task/<slug>/, next to dr-history.md).

Tools · Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion
Source · agents/dr-generator.md

reward-candidate-agent

Takes ONE candidate design from reward-tuning-agent — a bounded §1–§5 task delta plus a complete §6 reward — and carries it to a scored verdict: implement both into its own task, pass the smokes for every section it touched plus the §6 reward smoke, train + render on its assigned GPU, then score per-term curves + rendered frames into the verdict JSON. Implements only: it never designs, never re-weights to make a smoke pass, and never touches a sibling's files. Isolated by a slot clone when the pool runs more than one candidate, or working directly on the source task when it runs sequentially. Leaf agent: no Agent tool, no nested dispatch. Dispatched only by reward-tuning-agent.

Tools · Read, Write, Edit, Bash, Glob, Grep
Source · agents/reward-candidate-agent.md

reward-tuning-agent

Designs and drives the §6 tuning loop for ONE task. Each candidate it designs is a bounded §1–§5 task delta PLUS a complete reward (B1: adapt-first, magnitude budget, concrete weights/gates/composer, in-flight-aware distinctness); it dispatches one reward-candidate-agent per candidate to implement both, smoke every section touched, train + render, and score. Owns DECIDE (best-so-far, convergence, refill) and PROMOTE (winning design onto the source task, re-verified by that winner's own smoke set). Keeps pool_size candidates in flight, capped by gpus locally. Isolation follows the effective pool: sequential over a base/ snapshot when it is 1, one slot clone per candidate when it is more. Dispatched by /harbor:reward-tune (standalone) and /harbor:task-create (§6).

Tools · Read, Write, Edit, Bash, Glob, Grep, Agent, TaskStop
Source · agents/reward-tuning-agent.md

rl-integration-generator

Renders the RL training/eval/render scaffold (per-impl harbor/scripts/rl/<slug>/{train,eval,render,env_wrapper}.py + configs + rl-suite-spec + rl-integration.md receipt) into a benchmark repo whose base env (<repo>/.venv/) + benchmark-spec.json already exist. Three algorithm sources: custom_torch (a self-contained algorithm tree shipped with the plugin — no external RL-lib pip-install), stable_baseline3 (SB3-backed), local_implementation (user-provided package path or github URL — thin shims).

Tools · Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion
Source · agents/rl-integration-generator.md

rl-tuning-agent

Open-ended hyperparameter tuning loop for ONE (algorithm, task) pair on an already-set-up RL benchmark repo. Inputs: repo_path, algorithm, task, mode (local|cluster), parent tune_dir, and an optional metric weighting (default: equal weight on sample_efficiency and final_return). Iterates: default-config baseline → tricks pass → log-driven hyperparameter edits → convergence plot. Stops when the running best is not beaten for N consecutive iterations (default 3). Per-cell state lives at <tune_dir>/<wandb_project>/.

Tools · Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion
Source · agents/rl-tuning-agent.md

task-cloner

Clones an existing task into an isolated, independently-editable copy registered under a new suffixed gym id (SAME-REPO mode of the general /harbor:task-clone primitive; cross-benchmark/sim2sim migration is COMING SOON and refused in this release). Copies only the task's EDITABLE surface (env_cfg + the mdp modules the requested surface touches — default: reward), rewires the cloned cfg's imports to the copies, mirrors the source's registration mechanism for <dest>, then runs the clone smokes (build + rollout + per-term-logging). Writes a manifest listing every created file so the clone can be deleted cleanly. Callers include section A/B experiments and any flow needing an isolated task variant.

Tools · Read, Write, Edit, Bash, Glob, Grep
Source · agents/task-cloner.md

task-generator

Authors any subset of §1..§5 of a task in a benchmark repo (§1 register/scene · §2 actions · §3 reset · §4 goal+termination · §5 observation). Two modes — create (build a brand-new task with placeholder §6 reward + empty §7 DR) and edit (surgical re-author of one or more sections on a task that already builds). Reads task-implementation.md as a per-benchmark migration aid and one knowledge/references/task-generator/ file per section it works on. Step 0 scaffolds task-history.md; Phase A authors each section and records the analysis behind every design choice; Phase B renders and runs that section's smokes, looping each until it passes; Phase C gates the history with check_task_history.py and emits test-checklist.md, the list of checks that actually ran. The smoke loop is bounded by progress, not by an attempt count — it escalates via AskUserQuestion after two consecutive attempts that fail to move the measured quantity, or at 10 attempts. Ambiguity batches into a single AskUserQuestion per section.

Tools · Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion
Source · agents/task-generator.md