Agent Harness Wiki

7 min read

Nightly Src Projects Desk Raw Survey (2026-05-03)

This raw note preserves the public-safe basis for the 2026-05-03 nightly src/ projects desk. It is deliberately compact: the source tree was inspected as local evidence, not mined for publishable secrets. Names or details that would create a public-safety leak are withheld or reduced to category-level descriptions.

Survey scope and method

  • Survey root: /Users/ericfode/src
  • Survey timestamp: 2026-05-03 01:36 PDT
  • Coverage: 30 top-level directories, including hidden/scratch directories.
  • Execution shape: exactly 10 top-level Hermes survey lanes. A first all-at-once 10-task dispatch was rejected by the runtime because max_concurrent_children was 6; the real survey therefore ran as two batches, 6 lanes plus 4 lanes, preserving exactly 10 top-level survey lanes.
  • Recursive shape: the completed lanes reported orchestrator-style project splits and focused passes over purpose/docs, live-work evidence, and public-summary safety. Most lanes reported deeper delegate_task use; one child was handled directly by its project worker, and parent lanes spot-verified with local tools. No lane output is treated as magic; all synthesis below is based on filenames, docs, git metadata, and visible artifacts.
  • Evidence allowed: README/docs/plans, project manifests, git branch/status/log/diff metadata, modified/untracked filenames, mtimes, tests, run summaries, and visible artifacts.
  • Evidence excluded: secret contents, .env contents, local settings contents, raw prompt/log/trajectory material, hidden evaluator/supervisor payloads, explicit/provocative material, and directories too skeletal to support a public claim.

Ten survey lanes, safety-filtered

  1. Hidden local agent configuration; public tinygrad checkout; Lean-backed harness prototype.
  2. Sensitive reputational/social-claim notebook withheld; Dungeon Steward / cardgame1; empty creative scratch directory.
  3. deer-flow; FACEMUSIC; gas-city-but-its-just-codex.
  4. gemma4-tinygrad-opt; mixed creative-worldbuilding suite withheld pending curation; zero-commit Codex plugin workbench withheld.
  5. is-it-formal; justfooln; kettlebellsim.
  6. Skeletal local test directory withheld; local langfuse compose stack; local-hermes GGUF runner.
  7. meta-hermes wrapper with nested Gas City checkout; nnpl-external-latent-bus; nnpl-shared-bus.
  8. nnpl-typed-boundary-ir; openai-symphony; empty overengineeredlife scratch directory.
  9. silly-pi-stuff; hidden cardgame1 skill bundle under src; testing-rl.
  10. testing-rl-hermes; empty tinygrad scratch directories; tinygrad-gemma-kimi.

Public-safe lead candidates

Test-writing environments: testing-rl and testing-rl-hermes

  • testing-rl evidence: README.md, SPEC.md, WORKFLOW.md, pyproject.toml, docs/project-dashboard.md, docs/artifact-schemas.md, docs/counterfactual-verifier.md, docs/environment-contract.md, docs/software-risk-model.md, docs/lean-verification.md, testing_rl/*.py, formal/TestingRL/Plan.lean, and tests/test_*.py.
  • testing-rl git signal: branch master, one visible commit eac91d0 2026-05-02 init, dirty worktree with 23 tracked modified files and 46 untracked files; current local edits expand replay verification, artifact/training-product schemas, adapters, risk/event mining, and Lean formalization.
  • testing-rl-hermes evidence: MASTER_PLAN.md, ADVERSARIAL_RISK_REVIEW.md, docs/test-generation-rl-environment.md, docs/history-to-testgen-fixtures.md, docs/good-test-verifiable-properties.md, benchmarks/test-generation/README.md, src/testing_rl_hermes/testgen/*.py, and tests.
  • testing-rl-hermes git signal: branch main, clean worktree; recent commits add the test-generation RL environment, history-derived fixtures, and inverse-fix history mutants.
  • Public summary basis: the strongest publishable theme is an agent-testing bench moving from broad idea to explicit environment contracts, replay/evidence artifacts, and hidden-evaluator discipline. Supervisor-only benchmark details are not public copy.

Dungeon Steward / cardgame1

  • Evidence: project.godot, design/gdd/game-concept.md, design/gdd/systems-index.md, docs/plans/2026-04-05-combat-fun-roadmap.md, docs/plans/2026-04-11-deck-inspection-mvp-plan.md, docs/WORKFLOW-BALANCE-SIM.md, combat-stage and deck-inspection controllers, texture loader, smoke probes, and prototype tests.
  • Git signal: branch hermes/combat-stage-art-fallback-upstream, clean tree, ahead of upstream/main by one commit. Latest commit hardens combat-stage art presentation and asset fallbacks; recent areas include combat stage UI, deck inspection UI, texture loading, smoke probes, and image-generation prompt docs.
  • Public summary basis: browser-first Godot roguelite deckbuilder work is currently about legibility and trust: combat-stage art, deck/run presentation, asset fallback behavior, and smoke coverage.

FACEMUSIC

  • Evidence: docs/browser-instrument-architecture.md, ios/FACEMUSICNative/README.md, web/package.json, ml/README.md, ml/pyproject.toml, ml/configs/expression_forecast_v0_500ms.json, Cargo.toml, browser control/music-engine files, iOS native control files, and offline ML scaffold files.
  • Git signal: dirty main branch with tracked edits across browser and iOS face-control/conductor/music-engine paths; new untracked ML work adds expression-forecasting data contracts, configs, tinygrad training, TensorBoard visualization, scripts, and plan docs.
  • Public summary basis: face-expression-driven music remains a real interface project, now spanning browser, iOS, audio mapping, and forecasting semantics. Raw capture/session details are omitted.

gas-city-but-its-just-codex

  • Evidence: README.md, AGENTS.md, Cargo.toml, docs/usage/research-triangulation.md, docs/usage/image-first-codex-loop.md, docs/context-boards/README.md, docs/specifications/workflow-ledger-specification.md, templates, schemas, src/metaharness.rs, repo-loop gate artifacts, and smoke/validator scripts.
  • Git signal: active dirty branch codex/native-codex-ui; recent commits cover Harbor transfer reporting, native sandbox relaunch/operator UI wiring, UI showcase templates, and policy-routing semantics.
  • Public summary basis: Codex-native orchestration work is centered on workflow-ledger semantics, operator tooling, repo-loop automation, image-first context boards, and research-triangulation workflows. Local state IDs, logs, raw databases, and generated run payloads are omitted.

openai-symphony

  • Evidence: README.md, SPEC.md, LICENSE, NOTICE, elixir/README.md, elixir/WORKFLOW.md, logging/token-accounting docs, and Elixir modules for orchestrator, agent runner, workspace, Codex app-server, Linear client, and path safety.
  • Git signal: clean main tracking origin/main; visible shallow HEAD 58cf97d fix(elixir): configure Codex app-server model via config.
  • Public summary basis: an Apache-2.0 engineering-preview agent-orchestration service with an Elixir/OTP reference implementation. The safe wording should not imply production readiness or reveal operational tokens/issue details.

NNPL research cluster

  • nnpl-external-latent-bus: docs/source/tests/report artifacts support a public-safe summary of a two-space external/internal latent bus experiment, including output-side bridge probes and matched one-space comparators. No git metadata.
  • nnpl-shared-bus: docs/configs/tests/run summaries support a public-safe negative-result summary: v0 shared-bus recurrence did not show an honest recurrent advantage over preregistered controls. No git metadata.
  • nnpl-typed-boundary-ir: docs/source/tests/results support a public-safe summary of typed boundary artifacts, validation, rendering, auditability, and comparison to direct shared-bus baselines. No git metadata.
  • Public summary basis: the useful public claim is methodological, not promotional: baselines, boundaries, and negative evidence are being preserved as inspectable objects.

Research bench and side-room candidates

  • .tinygrad_research: clean upstream-style public tinygrad checkout on master, tracking origin/master; recent upstream commits around FP8 llama quantization, shape/range/special handling, one_hot, and mock interfaces.
  • another-harness: initialized git repo on main with no commits and hundreds of untracked paths; docs and layout point to Codex/Hermes harness architecture, work-item/evaluator/handoff loops, MCP/control-plane tooling, plugins, and Lean-backed modeling. High-level only.
  • is-it-formal: new Lean 4 + Python scaffold for grading claim formality across domains, with JSON examples and deterministic CLI; all visible files are untracked and no license was found.
  • justfooln: non-git research harness plus benchmark ladder for long-horizon/tool-heavy agent evaluation, with a long-autonomous-loops research brief; hidden agent workflows and recursive artifacts make it side-room material only.
  • kettlebellsim: active branch codex/reward-audit-and-swing-training; evidence points to scripted swing templates, cyclic observations, behavior-cloning warm-start, reward/retention diagnostics, and Modal probe/training workflows. Local temp/tool fragments and external secret-handling references keep it high-level only.
  • gemma4-tinygrad-opt: non-root-git Gemma/tinygrad optimization workspace with nested clean tinygrad checkout, Metal benchmarking/generation scripts, logs/prompts/evolution artifacts. Publish only as a local sandbox.
  • tinygrad-gemma-kimi: dirty opt/attention worktree for Gemma attention/JIT/memory-layout/correctness benchmarking, patch experiments, and result JSON artifacts. No benchmark numbers should be laundered into public claims.
  • local-hermes: non-git llama.cpp / GGUF runner with helper scripts and a local model artifact; publish only as a generic local serving setup.
  • langfuse: non-git local self-hosted compose stack with .env present; credential-bearing details withheld.
  • meta-hermes: top-level wrapper around a nested clean Gas City checkout; env-looking material and internal prompt/workflow surfaces mean only a narrow high-level summary is appropriate.
  • silly-pi-stuff: private-marked Pi companion UI plus an octonion-surface browser cellular automata demo; hidden prompt/config material withheld.
  • src: hidden game-development workflow/skills bundle for cardgame1; filename-derived broad characterization only.

Held back from project-specific public detail

The survey fully held back, or reduced to category-only omission, material in ten top-level directories. Reasons included sensitive identity/reputational framing, local agent settings, empty/skeletal directories, zero-commit all-untracked work that is not yet responsibly summarizable, local deployment configuration with .env signals, mixed creative material requiring curation, and hidden supervisor/evaluator or prompt-bearing workflow surfaces.

This is a public-safety filter, not a lack of inspection. The desk can say a room was surveyed without printing the contents of the drawers.

Editorial synthesis

The public-safe movement tonight clusters around four themes:

  1. test-writing and verification environments becoming explicit workspaces (testing-rl, testing-rl-hermes, is-it-formal, selected another-harness surfaces);
  2. interface projects spending effort on legibility and embodied control (cardgame1, FACEMUSIC, kettlebellsim at high level);
  3. orchestration systems moving state out of transcript fog and into ledgers, workspaces, and policy surfaces (gas-city-but-its-just-codex, openai-symphony, nested Gas City material);
  4. research benches preserving baselines, boundaries, and negative results instead of rounding them into triumphal prose (NNPL, Gemma/tinygrad benches).

That is enough for a public desk. It is also the only style of progress note that does not require clairvoyance, which remains, regrettably, absent from the toolchain.

Graph View

Backlinks