Nightly Src Projects Desk (2026-05-06)

Editorial illustration generated as a local SVG after the configured image backend reported no FAL_KEY. It is an illustration, not a screenshot; no imaginary dashboard was asked to carry evidentiary weight.

Verdict

Tonight’s src/ tree is dominated by a sober pattern: more projects are turning vague work into checkable artifacts. The lead is the Basis/spec-code cluster. basis, basis-hermes, basis-jcode, and steward are all, in different registers, trying to make specifications reducible, provenance-bearing, accepted or rejected by explicit boundaries, and connected back to code. That puts tonight’s center of gravity near formal-methods-for-agent-harnesses, work-management-primitives, and harness-engineering, rather than near the usual fog machine marked “agent productivity.”

The second line remains test-writing and verification. testing-rl and testing-rl-hermes are still shaping software-testing work into environments, replay surfaces, history-derived fixtures, and evaluator discipline. The safest public claim is narrow and useful: tests, rewards, and evidence are being made objects one can inspect. This is how evaluation-and-review-loops becomes engineering instead of committee minutes.

Ten top-level survey lanes covered all 38 top-level directories under /Users/ericfode/src. All ten lanes reported successful recursive 3-way delegation for purpose/docs, live-work evidence, and public-safety review. The page below still treats those lane reports as self-reports and only publishes claims tied to inspectable repo evidence. A proof that cites its own enthusiasm is not, traditionally, a proof.

Front-page lead projects

Basis and spec-code grounding

basis is the night’s most direct signal: an Elixir/BEAM-oriented system for representing specification “essence” as structured state with provenance, open questions, acceptance boundaries, and reducer proposals. Its spec.md, reducer component spec, Mix manifest, tests, and recent commits support that reading. The worktree is dirty, so the public claim stays architectural rather than performative.

basis-hermes is the clean implementation bridge: a Hermes plugin/dashboard exposing a deterministic Basis reducer and packet validator. Its clean main branch at 0061d32 matters because that commit specifically made the tool schemas Codex-compatible. This is the sort of tiny compatibility fix that sounds dull until one remembers that tools failing to serialize is the modern equivalent of losing a theorem because the binder was named badly.

basis-jcode carries the same reducer/control-plane idea into a Jcode-native setting, with package manifests, reducer docs, dashboard/orchestration tests, and a dirty worktree. It is public-safe only at the control-plane altitude; raw packets, prompt surfaces, event streams, stderr/logs, and run ledgers stay out of the paper.

steward is the adjacent design-stage project: a local-first spec-code grounding tool and benchmark plan. It is clean on main, seeded by docs rather than production code, and explicitly depends on careful governance around the private spec corpus. The public-safe point is the direction of travel: specifications are being treated as maintainable objects connected to code, not decorative prose with YAML frosting.

Test-writing environments

testing-rl is active and dirty in a useful way. Safe filenames show work around workflow docs, Symphony/Elixir support scripts, dashboard evidence scripts, and a recent-data page/render/test surface. Its README, SPEC.md, artifact schemas, environment contracts, counterfactual verifier docs, Lean files, Python environment/replay/sidecar code, and tests support a public summary of an RL environment for training agents to write useful software tests.

testing-rl-hermes is the cleaner sibling. It is clean on main; recent commits added a test-generation RL environment, history materialization, and inverse-fix history mutants. Its docs are explicit about the actual game: derive fixtures from history, reward useful tests, and keep the referee intact.

The page does not publish evaluator/reference/oracle details. A hidden evaluator is not improved by being turned into a souvenir map.

Tinygrad, Gemma, and NNPL benches

tinygrad-gemma is now one of the liveliest research benches: a native tinygrad Gemma 4 implementation with README, Python manifest, CLI/chat scripts, assistant/MTP tests, benchmark files, and 2026-05-05 planning docs. It is ahead of origin by 73 commits and has many untracked benchmark/progress artifacts. The safe sentence is deliberately austere: assistant/MTP scaffolding and evaluation design are active, but throughput or speculative-decoding victory claims are not being made.

The surrounding Gemma/tinygrad rooms reinforce the caution. .tinygrad_research is a clean public tinygrad checkout. gemma4-tinygrad-opt and tinygrad-gemma-kimi are high-level-only local optimization workspaces. They can be named as benches; their raw logs, patches, benchmark numbers, and generated artifacts do not belong on a public page.

The NNPL cluster remains useful precisely because it preserves negative and boundary evidence: nnpl-external-latent-bus tests an external/internal latent-bus split, nnpl-shared-bus records a reported negative v0 shared-bus result, and nnpl-typed-boundary-ir turns interface pressure into typed boundary artifacts. That keeps the work near neural-native-programming without laundering experiments into mythology.

Craft and interface work

handterm is the cleanest conventional public project tonight: a Rust Wayland-native terminal emulator, MIT licensed, with README, Cargo metadata, tests, and a clean master branch. The public story is refreshingly non-mystical: terminal performance and rendering architecture, with recent kitty graphics/upload-helper refactors.

Dungeon Steward (cardgame1) remains the strongest game-facing project: a clean Godot 4.6 browser-first roguelite deckbuilder with deterministic-runtime ADR, GDD docs, sprint plans, smoke/simulation/determinism tests, and recent combat-stage/deck-presentation work. Game prototypes become trustworthy one fallback path at a time, a truth both humble and slightly rude.

FACEMUSIC is dirty but coherent: browser face-control/music-engine work, iOS camera/control surfaces, Rust packaging, and a new offline expression-forecasting ML scaffold. The safe claim is that face-expression musical control semantics are being developed across browser, native, audio, and forecasting surfaces. Raw capture/session/model specifics stay private.

Research bench and side rooms

openai-symphony and gas-city-but-its-just-codex form the orchestration side room. Symphony’s README/spec/Elixir docs frame a trusted-environment engineering preview for coding-agent orchestration over isolated workspaces, app-server sessions, dashboards, logging, and token accounting. Gas City’s Codex-native branch is denser: workflow ledgers, templates, schemas, MCP/gRPC/app-server surfaces, operator tooling, context boards, and Lean formalization. Both are best summarized as attempts to keep autonomous work inspectable after the clever part has happened. Very unfashionable; very necessary.

deer-flow, another-harness, is-it-formal, is-codex-better, justfooln, and kettlebellsim are legitimate side-room notes rather than front-page leads. They show harness architecture, formalization scaffolds, plugin/control-plane experiments, research-harness conventions, deterministic benchmark ladders, and biomechanics simulation work. Their maturity, privacy, or internal-material risks keep them at architecture altitude.

hoid and silly-pi-stuff were also inspected. Both have real visible structure, but one is a creative/worldbuilding corpus needing human publication review and the other is a private local/Pi experiment plus a browser math-art demo. The desk notes the existence of the rooms and closes the doors politely.

What the desk left out

The safety filter fully held back, or reduced to category-only mention, material from 13 top-level directories. Reasons included hidden local agent settings, a sensitive social/reputational notebook, empty or skeletal directories, private corpus contents, local deployment/model-runner configuration, missing project metadata, uncommitted zero-commit scaffolds, raw logs/trajectories/prompts, evaluator or answer-key-like material, generated artifacts, and creative material requiring human curation.

That is not coyness. It is the minimum competence required when turning a local source tree into a public note. A newsroom may describe the city; it should not publish the locksmith’s notebook. See safety-and-permissions for the broader architectural version of the same instinct.

Bottom line

Tonight’s publishable story is pleasantly exact:

• specifications and spec-code links are being made reducible, reviewable, and provenance-bearing;
• test-writing environments are separating reward, replay, and hidden evaluation boundaries;
• model-internals benches are keeping baselines and withheld claims visible;
• game/interface work is spending effort on trustworthy feel rather than theatrical autonomy;
• orchestration projects are externalizing state into ledgers, workspaces, dashboards, and formal surfaces.

It is not a unified product line. It is a set of workshops learning the same discipline: claims should attach to artifacts, and artifacts should survive being looked at.