DRM

Apocha

The agent does the work. The controller decides what ships.

An open-source controller for coding agents. It grew out of the harness that built this site, and every change it accepts carries a receipt. View the code on GitHub ↗

HumanWrites the task
ControllerDecides what ships
AgentWorks through scoped tools
Reviewer agentJudges the verified change

Jump to any step

  1. A person writes the task: the paths the agent may change and the checks that decide.
  2. The controller makes a private clone of the repository at the base commit.
  3. The agent edits only through harness tools: no shell, no built-in tools.
  4. The task's checks run in an OS sandbox, from a catalog the agent can't edit.
  5. A separate reviewer agent judges the verified change, bound to its exact digest.
  6. Findings go back for a bounded number of repairs, or to a person.
  7. A passing change is committed with a receipt; land checks every receipt.
107runs recorded, every outcome counted
92 of 95tasks delivered
82 of 92passing runs needed no repair
111changes the earlier harness delivered to this site

Apocha developing itself, Oct 5–16, 2026: one machine, one operator, not a benchmark. The earlier harness, Sep 26–Oct 3: 62% passed review with no repair, 91% within one.

01The contract

A task is a contract,not a prompt.

Before any agent runs, a person bounds the work: what it is for, where it may write, what it must not touch, which checks decide, and how many repairs it gets.

task manifestassembled from a published receipt

Illustrative layout. Field values come from the published receipt; paths are summarised, not shown.

01 · fieldTask

A task names the paths the agent may change and the checks it must pass. Check IDs select commands from a closed catalog in project configuration read from the base commit; tasks cannot supply commands.

02What it guarantees

An agent can't changewhat judges it.

Each guarantee is enforced by the controller, not left to the agent's judgment.

ENFORCED BYCHECK CATALOG

Checks it can't rewrite

Checks come from a closed catalog in the project config, read from the base commit. A task names check IDs but can't supply commands, and agents can't edit the config or protected paths.

REFUSED

ENFORCED BYTEST POLICY

Tests change only with evidence

To change an existing test, the controller runs the old test on the new code. If it still passes, the edit must be a declared refactor or keep every assertion; otherwise it's rejected.

EVIDENCE

ENFORCED BYHARNESS TOOLS

No shell, no built-in tools

Agents read, search, patch and run named checks only through the harness. Each runtime adapter disables its own tools, and a live probe checks whether a runtime can escape.

SCOPED

ENFORCED BYOS SANDBOX

Checks run sandboxed

Agent-written code runs in a fresh copy of the tree, without .git or secret-looking files, under macOS Seatbelt or Linux bubblewrap, with no network by default.

ISOLATED

Infrastructure errors never count as passes.

03Receipts

Every accepted changecarries a receipt.

The receipt is a Git note on the commit: the task, the config, the checks that ran, the review decision and the exact change they were bound to. Receipts can be signed. apocha land moves a branch forward only through commits whose receipts match.

The first change Apocha delivered to this site: the System page moved to Apocha's own address.

Receipt

Delivery receiptordinary

Apocha gets its own address

PASSED

Move Apocha to /apocha, turn /system into a handoff for old links, and add a 404 page.

judge-inputspass
allowed-scopepass
protected-pathspass
test-editspass
portfolio-contentpass
portfolio-typecheckpass
portfolio-lintpass
portfolio-testspass
portfolio-buildpass
test-changespass
changed17 files · +286 −55
repairs0
reviewerPASS
promotionlanded
Delivered by Apocha on 2026-10-22: checks ran in the macOS Seatbelt sandbox and a reviewer from a different provider approved the change. Exact revisions and digests are withheld.
Before: the page titled Portfolio Code-Delivery System, with 'The system decides what ships.'After: the Apocha page at its own address, with 'The controller decides what ships.' and a link to the code on GitHubBeforeAfter

Real screenshots and excerpts from this delivery.

04Who decides

People decide.The controller enforces.

The human

  • Sets the scope and the checks
  • Approves claims, costs and releases
  • Resolves conflicts and repeated failures
  • Never turns a failing check into a pass

The controller

  • Isolates every candidate
  • Runs only the checks the task named
  • Binds each review to the exact change
  • Owns every state change and cleanup

05Evolution

From one site's harnessto an open-source tool.

  1. Aug–Oct 2026
    The harness behind this site

    Built around the open-source Pi agent to ship changes to this portfolio. It delivered 111 agent-written changes; 62% passed review with no repair.

  2. Oct 4
    Extracted

    Pulled out of the portfolio as a standalone tool that any repository can configure.

  3. Oct 5
    Apocha v0.1

    Published as open source: Codex, Claude Code or Pi as runtimes, workflows as data, receipts as Git notes.

  4. Oct 7
    v0.2

    Signed receipts, a verify-only GitHub Action, a deny-by-default macOS sandbox and an optional reviewer from a different provider.

  5. Since
    Tests need evidence

    Existing tests change only when the controller's evidence supports it, because editing tests is the easiest way for an agent to cheat.

06Next

Where it goesfrom here.

  1. 01

    Publish the evaluation suite: replays of real changes with hidden tests, and cases it should refuse.

  2. 02

    Sandbox the agent runtime process itself, not only the checks.

  3. 03

    Open it to more repositories than its own and this site.

It doesn't sandbox the agent runtime process itself, and it never pushes, opens pull requests or deploys.

Separate systemHow this site's content is published
See the workMulti-Agent Collaboration for Mathematical Reasoning