Agent Harness Wiki

6 min read

Nightly Src Projects Desk Raw Survey (2026-05-06)

This raw note preserves the public-safe basis for the 2026-05-06 nightly src/ projects desk. It is evidence-first and deliberately leaky only in the engineering sense: it says what could be inspected, what was held back, and where the public page should not pretend to know more than the files do.

Survey scope and method

  • Survey root: /Users/ericfode/src
  • Survey timestamp: 2026-05-06 01:36 PDT
  • Coverage: 38 top-level directories, including hidden/scratch directories.
  • Execution shape: exactly 10 top-level Hermes survey lanes, dispatched as 6 + 4. This preserved the required ten lanes without betting the night on a concurrency semaphore.
  • Recursive shape: all 10 lane summaries reported recursive delegation available, attempted, and succeeded: each lane split into purpose/docs, live-work evidence, and safety/public-summary eligibility, with further 3-way leaf splits where useful. The parent synthesis treats those as lane self-reports and relies only on inspectable repo evidence in the final page.
  • Evidence allowed: README/docs/plans, project manifests, branch/status/log metadata, safe modified/untracked filenames, mtimes, tests, checked-in reports, and visible artifacts.
  • Evidence excluded: secret contents, .env contents, local settings contents, raw prompt/log/trajectory material, hidden evaluator/supervisor payloads, private corpus bodies, explicit/provocative material, and directories too skeletal for a responsible public claim.
  • Illustration: the configured image backend failed with missing FAL_KEY; the page therefore uses a generated local SVG editorial illustration under queries/news-assets/. It is symbolic art, not a screenshot.

Ten survey lanes, safety-filtered

  1. Hidden local agent settings; clean public tinygrad checkout; early Lean-backed harness scaffold; one sensitive social-claim notebook withheld.
  2. basis; basis-hermes; basis-jcode; cardgame1 / Dungeon Steward.
  3. Empty creative; deer-flow; FACEMUSIC; gas-city-but-its-just-codex.
  4. gemma4-tinygrad-opt; handterm; hoid; is-codex-better.
  5. is-it-formal; justfooln; kettlebellsim; skeletal kimi-tests.
  6. Local langfuse deployment folder; local Hermes GGUF runner; meta-hermes; nnpl-external-latent-bus.
  7. nnpl-shared-bus; nnpl-typed-boundary-ir; openai-symphony; empty overengineeredlife.
  8. silly-pi-stuff; private spec-dataset-evolution-corpus; skeletal nested src; steward.
  9. testing-rl; testing-rl-hermes; empty tinygrad.
  10. tinygrad-gemma; empty tinygrad-gemma-gemini; tinygrad-gemma-kimi.

Public-safe lead candidates

Basis and spec-code grounding cluster

  • basis evidence: spec.md, AGENTS.md, mix.exs, components/spec-basis-reducer/spec.md, and tests under test/basis/. Git branch main; recent commits include 8e81759 resetting the reducer implementation to spec and 92f7eaf revising the core boundary/purpose. Worktree is dirty with modified reducer spec and untracked Elixir app/test/UI files.
  • basis-hermes evidence: README.md, plugin.yaml, pyproject.toml, dashboard/manifest.json, dashboard API files, reducer component docs, and tests. Git branch main; clean at 0061d3261329, whose commit message records Codex-compatible Basis tool schemas. Safe summary: Hermes plugin/dashboard exposing deterministic reducer and validator tools for Basis packets.
  • basis-jcode evidence: spec.md, components/spec-basis-reducer/README.md, reducer spec.md, package manifest, and orchestration/dashboard tests. Git branch main; dirty with tracked deletions in examples/UI. Safe summary: Jcode-native reducer control-plane work; raw packets, prompts, ledgers, event streams, and local run surfaces withheld.
  • steward evidence: README.md, pyproject.toml, and design docs including project charter, benchmark spec, architecture, implementation plan, data governance, modeling roadmap, and product workflows. Git branch main; clean at ba88837 with README stating design/ideation only and no production code yet. Safe summary: early local-first spec-code grounding project.
  • spec-dataset-evolution-corpus is intentionally private by README and clean at 4659608; public use is metadata/source-boundary only, not raw corpus publication.

Test-writing environments

  • testing-rl evidence: README.md, SPEC.md, pyproject.toml, docs for dashboard/artifact schemas/environment/counterfactual verifier/Hermes adapter, Lean formalization files, Python environment/replay/sidecar code, and tests. Git branch master tracking origin/master; dirty. Recent commit 139cea4 publishes the project surface; safe modified/untracked filenames point at workflow/docs/scripts plus a recent-data page/render/test surface.
  • testing-rl-hermes evidence: MASTER_PLAN.md, pyproject.toml, ADVERSARIAL_RISK_REVIEW.md, docs for test-generation environment/history fixtures/verifiable properties, testgen source, and tests. Git branch main; clean. Recent commits on 2026-05-01/02 add test-generation RL environment, history materialization, and inverse-fix history mutants.
  • Public summary basis: software-testing environments remain a front-page theme, but evaluator/reference/oracle/hidden-answer details are category-only.

Tinygrad, Gemma, and NNPL benches

  • .tinygrad_research is a clean public tinygrad checkout on master, with README, MIT license, pyproject.toml, docs, tests, and recent upstream commits around FP8 llama quantization, shapes/range/special handling, and one-hot helpers.
  • tinygrad-gemma evidence: README and pyproject.toml for a native tinygrad Gemma 4 implementation, CLI/chat scripts, tests, benchmark files, and 2026-05-05 assistant/MTP plans. Git branch main, ahead of origin by 73 commits, with no tracked changes and many untracked benchmark/progress/log-style artifacts. Safe claim: assistant/MTP scaffolding and evaluation design are active; throughput/speculative-decode claims are explicitly not made.
  • tinygrad-gemma-kimi is an undocumented dirty opt/attention local optimization repo with attention/JIT/correctness/validation filenames and patch artifacts. High-level only.
  • gemma4-tinygrad-opt is a non-root-git Gemma/tinygrad optimization workspace with nested clean tinygrad checkout, scripts, and generated/log/prompt artifacts. High-level only.
  • NNPL evidence: nnpl-external-latent-bus, nnpl-shared-bus, and nnpl-typed-boundary-ir have docs, Python manifests/source/tests, and public-safe methodological claims around external/internal latent buses, a reported negative shared-bus v0 result, and typed boundary IR. Raw results, traces, rollouts, metrics, and checkpoints are withheld.

Craft, interface, and game benches

  • handterm is the cleanest ordinary public-facing project tonight: Rust Wayland-native terminal emulator, MIT license, README, Cargo metadata, clean master at 977e709, and recent commits refactoring kitty graphics/upload helper state.
  • cardgame1 / Dungeon Steward remains a clean Godot 4.6 browser-first roguelite deckbuilder project on hermes/combat-stage-art-fallback-upstream, with deterministic-runtime ADR, GDD docs, sprint/planning docs, smoke/simulation/determinism tests, and recent combat-stage/deck-presentation work.
  • FACEMUSIC is dirty but coherent: browser face-control/music-engine files, iOS camera/control files, Rust package metadata, browser/iOS docs, and new offline expression-forecasting ML package files. Public summary is face-expression musical control semantics; capture/session/model specifics are withheld.
  • kettlebellsim is high-level only: Python simulation-first kettlebell swing biomechanics/RL toolkit with docs, tests, planning artifacts, probe/training scripts, and rollout media; remote-compute, experiment-tracking, and secret-reference details withheld.
  • hoid is high-level only: worldbuilding/narrative tooling corpus with docs, Go/TypeScript/Lean surfaces, tests, and many unpublished creative/generated artifacts. Story/canon/music/comic details are not public desk copy.

Orchestration and harness side rooms

  • openai-symphony: README/spec/Elixir docs describe a trusted-environment engineering preview for coding-agent orchestration over isolated workspaces, status dashboards, Codex app-server sessions, logging, and token accounting. Branch main tracks origin; dirty tracked work touches app-server/orchestrator/status/dashboard/presenter code and tests.
  • gas-city-but-its-just-codex: Rust/Codex orchestration-control-plane research prototype with workflow-ledger specs, templates/schemas, MCP/gRPC/app-server surfaces, operator tooling, context boards, and Lean formalization. Dirty branch; raw state/log/transcript/verifier/vendor/sandbox material withheld.
  • deer-flow: public docs/manifests support a high-level LangGraph/LangChain agent-harness summary, but local Flox/config/deployment state keeps it side-room.
  • another-harness, is-it-formal, is-codex-better, justfooln, meta-hermes, local langfuse, local Hermes GGUF runner, and silly-pi-stuff were inspected and kept to high-level or category-only treatment according to maturity and privacy risk.

Held back from project-specific public detail

The survey fully held back, or reduced to category-only mention, material from 13 top-level directories. Reasons included local hidden agent settings, a sensitive social/reputational notebook, empty directories, skeletal hidden workflow folders, private corpus contents, local deployment or model-runner configuration, missing README/docs/manifests, uncommitted zero-commit scaffolds, raw logs/trajectories/prompts, benchmark answer-key/evaluator-like materials, generated artifacts, and creative material needing human publication review.

This is a public-safety filter, not a claim that nothing is there. The desk can report that a room was surveyed without publishing the contents of the drawers.

Editorial synthesis

The public-safe movement tonight clusters around five themes:

  1. specification state and spec-code grounding are becoming first-class artifacts (basis, basis-hermes, basis-jcode, steward);
  2. software-testing environments are preserving replay/evidence boundaries rather than presenting benchmarks as magic (testing-rl, testing-rl-hermes);
  3. model-internals benches are using boundaries, baselines, and withheld performance claims (tinygrad-gemma, NNPL, Gemma/tinygrad sandboxes);
  4. craft projects continue to make feel and interface surfaces inspectable (handterm, Dungeon Steward, FACEMUSIC, kettlebellsim);
  5. orchestration projects are externalizing state into ledgers, workspaces, dashboards, and formal/control-plane surfaces (openai-symphony, gas-city-but-its-just-codex, DeerFlow).

The theme is not a product launch. It is better than that, in the limited but respectable sense: more of the work is acquiring checkable shape.

Graph View

Backlinks