Agent Harness Wiki

7 min read

Nightly Src Projects Desk Raw Survey (2026-05-07)

This raw note preserves the public-safe basis for the 2026-05-07 nightly src/ projects desk. It records what was inspected, where the safety filter intervened, and where the desk should not pretend that a filename is the same thing as a theorem. Tempting, but no.

Survey scope and method

  • Survey root: /Users/ericfode/src
  • Survey timestamp: 2026-05-07 PDT
  • Full top-level directory count: 38.
  • 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 parent synthesis treats those as lane self-reports and grounds public claims in inspectable evidence.
  • Coverage correction: the first lane inventory used the 36 visible top-level directories. A post-dispatch full root audit found two hidden top-level directories. To preserve the exact-ten-lane constraint rather than invent an eleventh lane, the controller performed a read-only hidden-directory supplement: one local hidden settings directory was held back, and one clean upstream tinygrad checkout was summarized from README/git metadata.
  • 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 plus hidden-dir supplement

  1. another-harness; one sensitive social-claim notebook withheld; basis; basis-hermes.
  2. basis-jcode; cardgame1 / Dungeon Steward; empty creative; deer-flow.
  3. FACEMUSIC; gas-city-but-its-just-codex; gemma4-tinygrad-opt.
  4. handterm; hoid; is-codex-better; is-it-formal.
  5. justfooln; kettlebellsim; skeletal kimi-tests; local langfuse folder.
  6. local-hermes; meta-hermes; nnpl-external-latent-bus; nnpl-shared-bus.
  7. nnpl-typed-boundary-ir; openai-symphony; empty overengineeredlife.
  8. silly-pi-stuff; private spec-dataset-evolution-corpus; skeletal nested src.
  9. steward; testing-rl; testing-rl-hermes.
  10. empty tinygrad; tinygrad-gemma; empty tinygrad-gemma-gemini; tinygrad-gemma-kimi.
  11. Hidden-dir supplement, not an extra lane: hidden local settings withheld; hidden clean upstream tinygrad research checkout summarized only as tinygrad context.

Public-safe lead candidates

Spec-code and Basis cluster

  • basis evidence: spec.md, mix.exs, components/spec-basis-reducer/spec.md, Elixir runtime/source/server files, tests, and recent commits. Parent git check: branch main, head 031340a on 2026-05-06, dirty with 6 tracked changes and 1 untracked file. Safe summary: an Elixir/BEAM prototype for turning overcomplete specifications into structured, provenance-bearing Basis state and proposals.
  • basis-hermes evidence: README.md, plugin.yaml, pyproject.toml, dashboard/manifest.json, reducer component docs, and tests. Parent git check: clean main at 0061d32 on 2026-05-05, whose message records Codex-compatible Basis tool schemas. Safe summary: Hermes plugin/dashboard exposing deterministic Basis reducer and packet validator surfaces.
  • basis-jcode evidence: reducer README/spec, package manifest, CLI/dashboard server files, tests, and recent commits. Parent git check: main at 4b1e621 on 2026-05-05, ahead/dirty by lane report with tracked deletions. Safe summary: Jcode-native reducer control-plane work; raw packets, ledgers, events, and run materials withheld.
  • steward evidence: README.md, pyproject.toml, project charter, benchmark spec, architecture, implementation plan, data governance, modeling roadmap, workflows, and decision log. Parent git check: clean main at ba88837 on 2026-05-05. Safe summary: design-stage local-first spec-code grounding tool.
  • is-it-formal is a small no-commit Lean/Python scaffold for classifying formalization strength; public-safe as a concept, but not yet release-like.

Test-writing and evaluator environments

  • testing-rl evidence: README.md, SPEC.md, pyproject.toml, workflow docs, dashboard/artifact schemas, environment/counterfactual verifier/Hermes adapter docs, Lean files, Python environment/replay/sidecar code, and tests. Parent git check: master at 139cea4 on 2026-05-04, dirty with 5 tracked changes and 3 untracked files. Safe summary: a software-testing RL environment where replay, evidence, and evaluator boundaries are first-class.
  • testing-rl-hermes evidence: MASTER_PLAN.md, pyproject.toml, docs for deterministic test-generation environments and history-derived fixtures, benchmark fixture suite, source, and tests. Parent git check: clean main at 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. A hidden answer key is not improved by being laminated.

Tinygrad, Gemma, and NNPL benches

  • tinygrad-gemma evidence: README.md, pyproject.toml, source package, tests, scripts, docs, and CI. Parent git check: main at 8128f4b on 2026-05-07, no tracked dirt and 57 untracked local/generated artifacts. Safe summary: native tinygrad Gemma 4 implementation with generation, tokenizer/KV-cache, multimodal, CLI/chat, training, and int8-related surfaces; raw benchmark logs/checkpoints/prompts and performance claims withheld.
  • .tinygrad_research evidence: clean hidden upstream tinygrad checkout on master at 87378331e on 2026-04-21, with README, pyproject.toml, and license. Safe summary: tinygrad context only, not a new local project story.
  • tinygrad-gemma-kimi evidence: dirty opt/attention repo at 8d23d35 on 2026-04-26 with attention/JIT/correctness/validation filenames. High-level optimization-workspace mention only.
  • gemma4-tinygrad-opt evidence: non-root-git Gemma/tinygrad optimization workspace with scripts and nested tinygrad checkout. 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, and public-safe methodological framing around external/internal latent buses, shared-bus negative-result posture, and typed boundary IR. Raw results, traces, rollouts, metrics, and generated artifacts withheld.

Craft, interface, game, and simulation work

  • handterm evidence: README.md, MIT license, Cargo metadata, clean master at 977e709 on 2026-04-19. Safe summary: Rust/Wayland terminal emulator focused on low latency, rendering architecture, and shared-process scaling.
  • cardgame1 / Dungeon Steward evidence: README.md, project.godot, GDD docs, deterministic-runtime ADRs, balance workflow docs, smoke/simulation/determinism tests. Parent git check: clean branch hermes/combat-stage-art-fallback-upstream at 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, iOS native README, ML README/manifest, Vite/React/Tone/MediaPipe package surface. Parent git check: main at f6cf6cf on 2026-04-19, dirty with 15 tracked changes and 2 untracked entries. Safe summary: face-expression musical control semantics across browser, iOS, audio, and forecasting surfaces; capture/session/model specifics withheld.
  • kettlebellsim evidence: Python manifest, planning-system README, cyclic-control docs, tests/scripts, and branch codex/reward-audit-and-swing-training at b075200 on 2026-04-09 with only untracked temp/probe files. Safe summary: simulation-first kettlebell biomechanics and training-incentive research toolkit.
  • hoid was inspected but reduced to high-level creative/worldbuilding/tooling mention. Unpublished creative drafts, generated review data, story/canon/music/comic details, and local state stay out of public copy.

Orchestration and harness side rooms

  • openai-symphony evidence: README/spec/Elixir docs for coding-agent orchestration over isolated workspaces, status dashboards, Codex app-server sessions, logging, and token accounting. Parent git check: main at 58cf97d on 2026-04-27, dirty with 9 tracked changes. Safe summary: engineering-preview orchestration bench; uncommitted details/logs omitted.
  • gas-city-but-its-just-codex evidence: Rust workspace, workflow-ledger specs, templates/schemas, MCP/gRPC/app-server surfaces, operator tooling, Swift/macOS UI, and Lean formalization. Parent git check: codex/native-codex-ui at 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, silly-pi-stuff, and the private spec corpus 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 the hidden local settings directory, one sensitive social-claim notebook, empty/skeletal directories, local deployment/model-runner folders, private corpus contents, internal workflow/assistant configuration, scratch/meta workspaces, generated artifacts, prompt/log/trajectory materials, evaluator-like payloads, benchmark raw outputs, model/checkpoint artifacts, and creative material needing human curation.

That is not coyness. It is the basic hygiene of turning a local source tree into a public note: describe the workshop, not the locksmith’s notebook.

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