Repository navigation
Luna v0.3.0-beta.2: Agent Kernel Upgrade
Pre-release
Pre-release
Luna v0.3.0-beta.2: the agent kernel
This release reworks how Luna delegates, plans and checks its own work. The aim is a smaller, stricter core: fewer delegation tools, limits that hold under load, and less prompt spent on tool text. It updates from v0.2.4-beta.9, the last tagged release, and replaces v0.3.0-beta.1, whose workflow sandbox had a call/return race (#50).
pi install git:github.com/SATUNIX/luna@v0.3.0-beta.2
Inside pi, run /update.
Highlights
- Plans instead of chains. Multi-step work runs as a small inline
workflow_runplan:- A deterministic executor decides which step runs next.
- Review gates loop within bounds, and plans can be resumed.
- A step can expand the plan with more steps. Each expansion is checked and charged like the original plan.
- Plans take no authority: they can only narrow tools, and cannot set a model, extensions, a runtime or skills.
- Per-request plan budgets. Each effort tier has a
workflowAgentsallowance per user message: E1 0, E2 1, E3 8, E4 16, E5 64. It comes on top of thesubagentlimit.- A run's worst case is charged before launch and never refunded.
- A plan that does not fit is refused with
plan-allowancebefore any run directory is created.
- One query and control surface.
workflow_statuslists script, YAML and background subagent runs, finished ones included.workflow_controlstops a run, or every run withall: true./workflowslists runs;/workflows panelopens the panel.
- A PASS is bound to the checkout. A recorded
/verifyor review PASS goes stale when the workspace revision changes, so completion stays blocked until the work is verified again. - Patch writes respect protected paths.
git apply,git amandpatchcount every file their patch names as a write target, in both protected-paths and the tool firewall. Patch forms the guard cannot read fail closed. - A smaller footprint.
- The delegation tool text drops from 6,266 to 4,730 characters.
- A metered baseline now guards each profile's tool text.
litegets compaction settings for 32k-context models.
Fixed
- Calls a workflow script sends as it returns no longer start: they are cancelled with the run, however the pipe splits the runner's output (#50).
- A verify or review PASS can no longer outlive the checkout it checked.
- The website's install command names a tag that exists.
- The web console survives a pi child that exits before reading its input.
/remote,/console, autonomy runs and the workflow sandbox work under a compiled pi.
Migration
| Removed | Replacement |
|---|---|
subagent chain mode |
workflow_run {steps} |
subagent_status |
workflow_status |
subagent_stop, /subagent-stop |
workflow_control with action: "stop" |
/subagents |
/workflows |
/reflect, /recover |
a bare /boost (/boost status reports it) |
confirmProjectAgents |
agentScope with the project trust prompt |
delegator role |
removed; use plans |
No aliases are registered, so a custom skill that names a removed tool gets an unknown-tool error.
- Subagents nest one level by default. A custom role that delegates needs
PI_KIT_SUBAGENT_MAX_DEPTH=2. - A model-started YAML run is charged to the tier's
workflowAgentsallowance instead of thesubagentlimit. - YAML files are checked when they are read.
- A
gateon a parallel member, or awhenon a parallel group, is refused. - A YAML step can no longer use
subagent.
- A
- A run with omissions reports
partial. - Resume reuses nothing outside git, or after any workspace edit.
Requirements
Node.js 22.19 or newer, git, and pi 0.85.1 or newer. The npm package is not published; installs come from the git tag.
The full list is in the changelog and the migration guide. This is a beta: report problems, and send vulnerabilities privately as SECURITY.md describes.
What's Changed
- chore: consolidate on SATUNIX/luna as the only repository by @SATUNIX in #24
- Release v0.2.4-beta.11 by @satunix-pisystem-ci-pipeline[bot] in #23
- fix(autonomy): make self-improve fit for long unattended runs by @SATUNIX in #25
- chore: ship the settings drift check publicly and show the OpenSSF badge by @SATUNIX in #26
- fix(deps): pin minimatch 11.0.1 for GHSA-g3hv-75wg-m249 by @SATUNIX in #28
- fix(autonomy): make the acceptance gate pass in its own container by @SATUNIX in #29
- fix(autonomy): make resume work for a new supervisor process by @SATUNIX in #30
- ci: shard check:all four ways, parallel --jobs, CI on integration PRs by @SATUNIX in #36
- feat(autonomy): parallel self-improve streams: upstream sync, steward board, record-aware review by @SATUNIX in #37
- deps: Bump com.android.application from 9.4.1 to 9.5.0 in /apps/android in the android group by @dependabot[bot] in #35
- integration/luna: autonomous cycles (luna-improve-01) by @SATUNIX in #38
- fix(autonomy): let a cycle add a smoke test without opening every control path by @SATUNIX in #39
- fix(autonomy): never pass NOTIFY_SOCKET to the container engine by @SATUNIX in #41
- Update cat animations with new frames and intervals by @SATUNIX in #44
- deps: Bump typescript-eslint from 8.72.0 to 8.72.1 in the dev-toolchain group across 1 directory by @dependabot[bot] in #32
- test(custom-footer): fix the cat smoke after the new animation table by @SATUNIX in #45
- docs: rename the project to Luna and add the animated overview by @SATUNIX in #46
- docs: animated SVG README, a Proof-style website and its GitHub Pages deploy by @SATUNIX in #47
- The agent kernel: one execution contract for delegation, workflows and autonomous runs by @SATUNIX in #48
- Release v0.3.0-beta.1 by @SATUNIX in #49