Eureka: a faculty pipeline for Claude Code agents
  • Shell 79.5%
  • Python 17.4%
  • Rust 1.6%
  • Dockerfile 1.1%
  • JavaScript 0.4%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Meta Vens 9b2f51440c Lead with the failure this prevents
The README opened with the org chart when the evidence is the argument. It now
opens with the thing that goes wrong when one agent plans, implements, tests
and reports on its own work, and puts the numbers in the first screen.

Four cases from one day show what the falsifier actually catches, including a
rule everyone agreed could not be tested and a mutation history that turned out
not to exist.

One correction: the status section described the state client in the present
tense. It is not written, and the skill is not wired to it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-17 03:01:32 +02:00
references Ask what the test proves, not whether it passes 2026-09-17 02:36:04 +02:00
tools Give the harness one home 2026-09-17 02:38:42 +02:00
LICENSE Publish Eureka, the Claude Code faculty pipeline skill 2026-09-15 20:15:10 +02:00
README.md Lead with the failure this prevents 2026-09-17 03:01:32 +02:00
SKILL.md Give the harness one home 2026-09-17 02:38:42 +02:00

Eureka

An agent that plans, implements, tests and reports on its own work will hand you green tests and a clean report over a broken machine. Not because it lies, but because its tests pin its own spelling, and its report is written by the party with the most reason to believe it.

Eureka is a Claude Code skill that takes that job apart. One agent maps the change, a second executes exactly one cut of it, and a third — which did not write the code, and is told to attack it — tries to prove the second one wrong. The operator rules on anything that is genuinely a product decision. Nothing grades itself.

In the migration it was extracted from, 29 of 32 independent falsification passes found real defects that the implementing agent's own tests had passed — about 75 in all, among them deadlocks, lost writes, a wire-parity split between two runtimes, and operator rulings with no test behind them. references/postmortem-cultcache.md carries the evidence.

What that actually looks like

Four from a single day, each found by a pass that did not write the code:

  • A rule everyone agreed could not be tested. The implementer said so honestly in the entries file rather than faking a kill. A fifteen-line probe reached it in about a millisecond, and the mutant crashed an ordinary optimised build on contact. The suite and the prose agreed with each other, and both were wrong.
  • A mutation history that did not exist. Every kill a native bridge had ever reported came from harnesses living in agents' scratchpads. When a later pass went to check two of them, their definitions were gone, and the rules they had defended were defended by nothing anyone could rerun.
  • Three passes to pin one line. A rule said a wait must be the timeout the caller asked for. The first suite proved the wait was not one particular constant. The second proved it was not any constant. Then a clamp — cap the wait so shutdown gets noticed, the most ordinary way that line ever gets written — walked through the entire matrix on both platforms.
  • A document the system accepted and could never return. Admission took a record 86,751 bytes larger than the transport could carry. Measured against a real client, not inferred.

Each became a rule in references/changelog.md, with the evidence that motivated it. The skill is accumulated scar tissue, not a design.

The pipeline

  • Self routes the work, keeps the maps, and talks to the operator.
  • Imagination maps the cut: deletes first, an authority map for anything that changes ownership, and verification naming the rule each test pins.
  • Hands executes one cut, in small pushed commits, and never updates the map.
  • Soul runs against the executed commits and tries to falsify every promise, building its own probes rather than trusting the report.
  • Mind Steward keeps memory honest at phase boundaries.
  • The operator rules on real forks, and only on real forks.

It is the Claude Code counterpart of Epiphany, and shares its faculty vocabulary.

Install

git clone https://github.com/GameCult/Eureka.git ~/.claude/skills/eureka

Claude Code discovers the skill from the name and description in SKILL.md.

Layout

  • SKILL.md: the pipeline — the faculties, the loop, Self's discipline, and the git and tooling rules, each of which exists because something broke.
  • tools/eureka-mutations.ps1: the mutation harness, parameterised by repo so one copy serves every campaign.
  • references/briefs.md: brief templates for each faculty.
  • references/cut-map.md: the shape of the target document and the cut map.
  • references/postmortem-template.md and references/postmortem-cultcache.md: the template and the worked example.
  • references/epiphany-comparison-2026-09-15.md: a dated, source-grounded comparison with Epiphany.
  • references/changelog.md: every change to the skill, with its evidence.

Status

Eureka keeps its pipeline state in committed documents and memory. A campaign in Epiphany (notes/eureka-pipeline-state-target.md) is giving it typed state instead: Epiphany owns the schemas and admission, and a memory organ owns each instance's state.

That work is partly landed and partly unbuilt. The organ answers typed queries over the network today. The eureka-state client that would let Eureka agents admit and query through it is not written, and this skill is not wired to it.

Maps and postmortems are public artifacts from the first cut, with the scars left in. That is deliberate: a pipeline whose record has been tidied up is a pipeline you cannot check.