Agent Harness Wiki

7 min read

Nightly Src Projects Desk Raw Survey (2026-05-08)

This raw note preserves the public-safe basis for the 2026-05-08 nightly src/ projects desk. It is a survey basis, not a romance novel for repositories. The method was deliberately prosaic: inspect files that exist, cite git state where present, and decline to turn private drawers into public exhibits.

Survey scope and method

  • Survey root: /Users/ericfode/src
  • Survey timestamp: 2026-05-08 PDT.
  • Full top-level directory count: 38, including hidden directories.
  • Execution shape: exactly 10 top-level Hermes survey lanes, dispatched as 3 + 3 + 3 + 1.
  • Lane recursion: all 10 lane summaries reported that delegate_task was available and used. Each lane reported a 3-way split for purpose/docs, live-work evidence, and safety/public-summary eligibility, plus a further one-level 3-way recursion by its sub-inspectors. The controller treats those as lane self-reports and grounds public copy only in inspectable evidence returned by the lanes.
  • Evidence allowed: README/docs/plans, 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, benchmark raw outputs, checkpoints/model artifacts, and directories too skeletal for a responsible public claim.
  • Illustration: the configured image backend failed with FAL_KEY environment variable not set; the page therefore uses a generated local SVG editorial illustration under queries/news-assets/2026-05-08-project-desk-hero.svg. The SVG parsed as valid XML. It is symbolic editorial art, not a screenshot.

Ten survey lanes

  1. Hidden local settings; hidden upstream tinygrad research checkout; another-harness; 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; local-hermes; scratch 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; internal nested src skill scaffold; steward.
  9. testing-rl; testing-rl-hermes; empty tinygrad.
  10. tinygrad-gemma; empty tinygrad-gemma-gemini; dirty tinygrad-gemma-kimi optimization workbench.

Public-safe lead candidates

Spec-code and Basis cluster

  • basis evidence: spec.md, Mix manifest, reducer component specs, runtime/source/server files, docs, tests, and clean git state. Lane git check: clean main...origin/main, head a5544e0 on 2026-05-07, latest message Split reducer UI into separate app entrypoints. Safe summary: an Elixir/BEAM project for turning overcomplete specifications into structured, provenance-bearing Basis state while keeping Markdown as a review projection.
  • basis-hermes evidence: README.md, plugin.yaml, pyproject.toml, reducer component docs, tests, CLI/tool/dashboard surfaces. Lane git check: clean main, head 0061d32 on 2026-05-05, latest message fix: make basis tool schemas codex-compatible. Safe summary: Hermes plugin/dashboard exposing deterministic Basis reducer and packet validator surfaces.
  • basis-jcode evidence: reducer README/spec, package manifest, CLI/dashboard files, tests, and recent commits. Lane git check: main...origin/main [ahead 10], head 4b1e621 on 2026-05-05, dirty with tracked deletions. Safe summary: Jcode-native reducer control-plane work; raw run ledgers, packets, prompts, and state materials withheld.
  • steward evidence: README.md, pyproject.toml, project charter, data governance, architecture, implementation plan, modeling roadmap, workflows, and decision log. Lane git check: clean main, head ba88837 on 2026-05-05, and docs explicitly say the project is design/ideation with no production code yet. Safe summary: design-stage local-first spec-code grounding tool.
  • spec-dataset-evolution-corpus evidence: README, license/compliance notes, aggregate reports, and fail-closed export posture. Safe summary: metadata and aggregate counts only; raw corpus content remains private.
  • is-it-formal evidence: README, Lean/Lake scaffold, JSON-to-Lean loader, CLI grader, examples, and no secret markers reported. Safe summary: small Lean/Python scaffold for classifying formalization strength; useful, but no-commit and early.

Test-writing and evaluator environments

  • testing-rl evidence: README, SPEC, environment contract, non-cheating test-writer docs, project dashboard, Lean files, Python environment/replay/sidecar code, and tests. Lane git check: master...origin/master, head 139cea4 on 2026-05-04, dirty with tracked workflow/Symphony files plus untracked recent-data page work. Safe summary: a software-testing RL environment where replay, evidence, hidden-reference boundaries, and counterfactual evaluation are central.
  • testing-rl-hermes evidence: MASTER_PLAN.md, package metadata, deterministic test-generation environment docs, history-derived fixture docs, benchmark fixture suite, source, tests, and reports. Lane git check: clean main, head 6cbca51 on 2026-05-02. Safe summary: executable/prototype companion for deterministic test-generation and hidden-reference style grading.
  • Hidden evaluator/reference/oracle details remain category-only. Publishing an answer key would be a curious way of proving one understands evaluation.

Tinygrad, Gemma, and NNPL benches

  • tinygrad-gemma evidence: README, package/docs/test/script surface, benchmark framing docs, target-JIT plan, and recent commit history. Lane git check: main...origin/main [ahead 93], head 11470a3, no tracked modifications, 57 untracked local/generated benchmark artifacts. Safe summary: native tinygrad Gemma 4 implementation with chat/API, tokenizer/KV-cache, multimodal, training, and target-JIT/Metal benchmark work. Raw benchmark logs, checkpoints, prompts, model artifacts, and speed claims withheld.
  • .tinygrad_research evidence: clean hidden upstream tinygrad checkout on master...origin/master, head 87378331e on 2026-04-21. Safe summary: tinygrad upstream context only, not a local project launch.
  • tinygrad-gemma-kimi evidence: dirty opt/attention repo, head 8d23d35 on 2026-04-26, with attention/JIT/correctness/validation filenames and no README/LICENSE. High-level optimization-workbench mention only.
  • gemma4-tinygrad-opt evidence: non-git Gemma/tinygrad optimization workspace with scripts, large logs, prompt/test files, nested tinygrad checkout, credential/proprietary-notice risk markers by lane report. High-level only; raw logs/prompts/results withheld.
  • NNPL evidence: nnpl-external-latent-bus, nnpl-shared-bus, and nnpl-typed-boundary-ir have docs, source/tests, manifests or architecture notes, result/report artifacts, and explicit methodological framing around external/internal latent buses, shared-bus negative-result posture, and typed boundary IR. Raw traces/results/checkpoints and generated artifacts withheld.

Craft, interface, game, and simulation work

  • handterm evidence: README, MIT license, Cargo metadata, optimization/remain-work docs, tests, clean master, head 977e709, with recent commits extracting Kitty upload/graphics state. Safe summary: Rust/Wayland terminal emulator focused on low latency, renderer architecture, and shared multi-window host design.
  • cardgame1 / Dungeon Steward evidence: project.godot, GDD docs, design/production materials, deterministic-runtime concepts, tests, and clean branch hermes/combat-stage-art-fallback-upstream...upstream/main [ahead 1], head a9a8ef6 on 2026-04-15. Safe summary: browser-first Godot roguelite deckbuilder with deterministic combat and combat-stage art fallback polish.
  • FACEMUSIC evidence: browser architecture docs, expression-forecasting plan, web package metadata, iOS/audio/control surfaces, and dirty active work. Safe summary: privacy-sensitive face-expression musical control prototype; raw captures, sessions, model outputs, and personal data surfaces withheld.
  • kettlebellsim evidence: Python manifest, AGENTS summary, Isaac Lab MVP spec, RL retention plan, swing-training branch, tests/scripts, and rollout artifact filenames. Safe summary: simulation-first kettlebell biomechanics and training-incentive research toolkit; local temp/probe/rollout details withheld.
  • hoid was inspected but held back to high-level creative/worldbuilding/tooling category. Unpublished creative drafts, generated review data, story/canon/comic/music details, and local state stay private.

Orchestration and harness side rooms

  • openai-symphony evidence: README, Elixir implementation docs, Mix manifest, license, app-server/session/dashboard/logging/token-accounting docs, and dirty tracked Elixir/dashboard/test changes. Lane git check: main at 58cf97d on 2026-04-27; shallow repo; dirty with 9 tracked changes. Safe summary: engineering-preview coding-agent orchestration bench.
  • gas-city-but-its-just-codex evidence: README, Rust workspace manifest, workflow-ledger and project-formalization docs, schemas/templates, MCP/gRPC/app-server surfaces, operator tooling, state/docs/scripts, and Lean formalization. Lane git check: branch codex/native-codex-ui, head 198aefc on 2026-04-21, dirty with 7 tracked changes and many untracked files. Safe summary: Codex-native durable workflow/control-plane research prototype; runtime state, logs, transcripts, and generated artifacts withheld.
  • another-harness, is-codex-better, justfooln, deer-flow, meta-hermes, local-hermes, local langfuse, and silly-pi-stuff were inspected and kept to high-level/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 hidden local settings, one sensitive social-claim notebook, empty/skeletal directories, local deployment/model-runner folders, private corpus bodies, internal workflow/assistant configuration, scratch/meta workspaces, generated artifacts, prompt/log/trajectory materials, evaluator-like payloads, benchmark raw outputs, model/checkpoint artifacts, privacy-sensitive capture data, and creative material needing human curation.

This is not timidity. It is just the ordinary discipline of making a public page from a private source tree: the fact that something is interesting does not make it publishable.

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, private corpus metadata);
  2. software-testing environments are preserving replay/evidence/evaluator boundaries rather than flattening benchmarks into slogans (testing-rl, testing-rl-hermes);
  3. model-internals benches are moving behind explicit performance and artifact gates (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, app-server sessions, and formal/control-plane surfaces (openai-symphony, gas-city-but-its-just-codex, DeerFlow).

The theme is not launch energy. It is the quieter thing: more claims are acquiring shape, provenance, and places where they can fail.

Graph View

Backlinks