Pixir Harness Open Beta Developer Preview

Copy Markdown View Source

Date: 2026-06-09

Pixir Harness is available as an early developer preview. This note records the public preview scope for source installs and the narrow Hex CLI/ACP distribution path. It does not define a stable Elixir library contract.

0.1.4 Runtime Truth And Fanout Honesty Update

Pixir 0.1.4 keeps the same CLI/ACP-only Hex contract and tightens backend evidence for terminal Turn and Subagent outcomes:

  • terminal Turn failures are persisted as durable evidence instead of being collapsed into clean completion;
  • wait_agent reports structured partial outcomes when only some children finish;
  • Subagent terminal evidence, stale fanout diagnostics, and runtime-truth gauntlets make bounded fanout easier to audit;
  • ADR 0026 documents the runtime terminal-state and replay contract now exposed in HexDocs.

This release does not claim a complete scheduler, WorkflowRun runtime, production SLA, SAAAS surface, or performance superiority.

Post-release dogfood also validated developer-preview presenter confidence across the CLI, T3 Code Pixir dogfood, and Zed ACP when all presenters point at the same source-built binary and workspace. That evidence supports local dogfood and operator feedback, but it is not a public claim of strict UI parity or bundled T3/Zed support. The next hardening priority is Subagent/Workflow operability: clearer depth and timeout ergonomics, parent/child diagnostics, and honest partial/cancel/timeout outcomes.

0.1.3 ACP Registry Readiness Update

Pixir 0.1.3 keeps the same CLI/ACP-only Hex contract and adds the pieces needed for public ACP Registry validation:

  • ACP initialize now advertises terminal authentication through pixir login, while Pixir still owns OAuth and credential storage outside the ACP stdio channel.
  • ACP message projection and diagnostics preserve partial assistant evidence more honestly when Provider streams end unexpectedly.
  • Turn, CLI, Subagent, wait_agent, and diagnostics hardening now make fanout failures and timeouts visible as terminal evidence instead of clean completion.
  • Source checkouts include mix pixir.bench.fanout_gauntlet, a no-network correctness gauntlet for direct CLI fanout and parent-led Subagent fanout. It is evidence for outcome honesty, not a public performance benchmark.
  • Source checkouts include the repo-local bin/diagnose-t3-pixir-projection helper for comparing Pixir Log truth, T3 provider logs, and T3 projection storage when a response is not visible where expected. It is not part of the Hex package surface.
  • CI is split into clearer Pixir check gates and the projection diagnostic tests can run with uv or a plain Python runner.

This release is the intended Hex base for the public npx ACP Registry wrapper.

0.1.1 Runtime Diagnostics Update

Pixir 0.1.1 keeps the same CLI/ACP-only Hex contract and adds local Session diagnostic surfaces for operators:

  • pixir inspect-replay <session-id> --json inspects replay continuity from the local Log.
  • pixir diagnose session <session-id> --json emits a structured Session diagnostic verdict for replay, lifecycle, and local evidence review.
  • Replay folding now tolerates transparent lifecycle/status Events between matched tool_call and tool_result records.
  • The escript SIGINT path no longer depends on Mix.env/0 at runtime.

The repo-local Codex/Pixir operator Skills are intentionally not part of the Hex package. They remain workflow guidance for this repository, not a public Pixir package contract.

This is not a stable ecosystem release. The goal is to let early users install Pixir, run doctor, try a first terminal or ACP-driven turn, and report the first installation, auth, model, and workflow failures with useful evidence.

What Is Supported

  • Source install from https://github.com/Ranvier-Technologies/pixir-harness.
  • Hex escript installation when the pixir package is available, scoped to the same CLI/ACP runtime.
  • Terminal CLI through the installed pixir escript or a source-built ./pixir.
  • ChatGPT subscription login through pixir login, with OPENAI_API_KEY as a fallback.
  • Local append-only Session Logs under .pixir/sessions/.
  • pixir doctor --json as the first no-network diagnostic gate.
  • ACP over stdio through pixir acp.
  • Agent Skills, Subagents, Workflows, Session tree inspection, and durable compaction checkpoints.
  • Durable Provider usage accounting for token/cache evidence.
  • Initial Image Attachment support through local Session Resources and Provider projection.
  • Provider-hosted Web Search request/evidence plumbing, with opt-in smoke verification.

What Is Experimental

  • Long-running, non-blocking Subagent status and result retrieval in specific clients.
  • Workflow partial outcomes and checkpoint bundles as a polished product experience.
  • Presenter/client integrations beyond ACP stdio.
  • Networked smoke tests for model/provider behavior.
  • Skill Context Hydration, which is accepted as a design direction but not yet a public implemented surface.

What Is Out Of Scope For This Preview

  • Stable public Elixir API.
  • Broad package ecosystem or extension API promises.
  • MCP server support.
  • Packaged T3Code provider.
  • Self-update/install channel.
  • Production SLA, telemetry, or hosted service guarantees.

Newcomer Smoke

The expected first-run proof is:

git clone https://github.com/Ranvier-Technologies/pixir-harness.git
cd pixir-harness
mix deps.get
mix escript.build
./pixir doctor --json

Run this smoke against the current checkout or installed package before reporting setup results. Use pixir instead of ./pixir for package installs.

Open-Source Release Checklist

Before changing the GitHub repository visibility, maintainers should run the local gate and a history-aware secret scan against the exact ref that will become public. Prefer a clean clone of origin/main so local-only runtime refs, ignored .pixir/ state, and worktree artifacts do not hide or invent release findings:

mix check
trufflehog git file://$PWD --no-update --fail

If trufflehog is unavailable, use an equivalent tool such as:

gitleaks detect --source . --no-banner --redact --exit-code 1

Record the command and whether it scanned a clean clone or a local checkout in the release notes or PR. Do not rely only on filename scans.

Recommended public repository metadata:

  • Description: Elixir/OTP runtime for supervised coding-agent sessions with CLI, ACP, Subagents, Workflows, and replayable evidence.
  • Homepage: https://pixir.dev
  • License: MIT

Changing repository visibility is a manual release action. Do not automate it from a maintenance PR; confirm the final local gate, secret scan, and owner approval first. After the repository becomes public, enable GitHub Private Vulnerability Reporting in the repository Security settings before inviting broader security review.

Feedback To Capture

When a preview user reports a failure, ask for:

  • exact OS and Elixir version;
  • command run;
  • pixir doctor --json output with secrets removed;
  • whether they used ChatGPT subscription login or OPENAI_API_KEY;
  • the relevant .pixir/sessions/<id>.ndjson only when safe to share.