You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Phase 5 docs (T017 + T018, 2026-10-31): SECURITY.md is now the shipped security posture — the
invariant verbatim (the same sentence every skill
edition carries), the posture in plain language (CI on
GitHub-hosted runners only; self-hosted runners reachable
solely behind the gatekeeper's workflow_dispatch; the
gatekeeper's verified-write and exactly-once dispatch
rules; the backchannel as dev/recovery only, carrying no
job traffic), and adopter checklists for token scoping and
runner isolation — alongside the existing reporting policy. README.md opens with the two-minute story requires (the problem, the scars that shaped the design,
what Syncyard deliberately does not do), the security
section links SECURITY.md above the quickstart, and
per-host quickstarts cover Windows, WSL2 (via a Windows
host), Linux, macOS, and Docker — including the fleet rule
that cross-job payloads are never staged in a WSL venue's /tmp (it does not persist between jobs; stage in the WSL
home directory or the job's working tree instead).
— shutdown protection and recovery (owner
direction, 2026-10-31): a host-side shutdown guard that
delays a non-forced shutdown while active work drains
(bounded checkpoint window, going-down record written
first with the affected run ids, outbox notification),
boot recovery that returns runners without an interactive
logon (boot shape is an owner gate — service-mode would
amend the runner rule), a generalized supervised-work
vehicle from the production anchor pattern, gatekeeper
reconciliation of every interrupted run against GitHub's
terminal state, and a recovery report that states loss
plainly. The spec's protection level is delay + drain +
checkpoint + notify, never prevent — forced shutdowns and
power loss are named outside the envelope. Specified only
in this change; implementation follows in phases.
Phase 4 — the agent skills pack: skills/ ships eight editions generated from one shared
source (skills/_source/contract.md via skills/build_skills.py) in the Amendment 1 ranking —
Meta Muse (primary), Claude Code, GitHub Copilot, OpenAI
Codex/ChatGPT, Cursor, Google Gemini/Antigravity,
generic-agent, and a CLI-only fallback. Each edition carries
the security invariant verbatim; the T016 currency
battery (skills/test_skills.py, in CI) pins every edition
to the repo VERSION, the invariant, and the shared source
(drift fails the build).
— public release program (owner direction,
2026-10-30): configurability as a single config file (lane
count/names, watermarks, budgets, endpoints, transport
ceiling), a tested host matrix (Windows, WSL2, macOS, Linux,
Docker), permanent backchannel hardening (self-test, stale
alarm, chaos battery), an agent skills pack with Meta Muse
as the primary vehicle, per-requirement test batteries, a
shipped security posture, and release 0.3.0 when green. A
GitHub gh-adoption pitch brief is prepared under and is not sent without the owner's explicit word. Specified
only in this change; implementation follows in phases.
Amendment 1 (2026-10-30): ships Syncyard as
a GitHub CLI extension (AXIOVEX/gh-syncyard, thin Go
dispatcher, official precompile action); agent-skill
editions ranked by market share (Meta Muse primary, Claude
Code, Copilot, Codex/ChatGPT, Cursor, Gemini/Antigravity,
generic). Phase 1 slice 1: config/syncyard.example.yaml + gatekeeper/config.py loader/validator with the test_config.py golden battery (25 checks, all passing);
additive — the gatekeeper does not read config yet.
Phase 7 SHIPPED 2026-10-31: AXIOVEX/gh-syncyard v0.1.0 — the GitHub CLI extension.
Thin Go dispatcher over the installed gatekeeper
(lookup: $SYNCYARD_GK / $SYNCYARD_HOME / ~/syncyard /
~/.syncyard / beside the binary / ./gatekeeper);
precompiled binaries via the official precompile action;
install proof: gh extension install AXIOVEX/gh-syncyard
then gh syncyard status against the live fleet.
Phase 1 slice 2 (T003/T004, 2026-10-31): the
gatekeeper now reads its effective configuration at
startup — lane set, runner labels, per-lane concurrency,
watermark tables, budgets, and dispatch knobs all bind
from syncyard.yaml deep-merged over the built-in
defaults (the pre-config constants verbatim); invalid
config fails closed with a named error. gatekeeper/test_golden.py proves byte-identical
dispatch decisions under defaults vs the fleet config,
and runs in CI.
Phase 1 COMPLETE (T005, 2026-10-31): the fleet
gatekeeper now runs on the configuration system — wired
code deployed behind a queue snapshot (534 records,
open-set 5) with exact reconciliation, and the fleet syncyard.yaml in place beside the gatekeeper.
Phase 2 begins (T006, 2026-10-31): hosts declare
the transport workflows they serve (execute/grab) and
their identity markers (a per-type contract); the
gatekeeper routes by host and refuses, at submit/adopt,
any job addressed to a venue that does not serve its
workflow — no silent fallback. Config validation covers
the new shape; the golden battery pins the routing
matrix across the fleet hosts plus a config-only macOS
host.
T007 (2026-10-31): the shipped transport
workflows are generalized — a lane input addresses the
runner, hardcoded label lists are gone, and each file
states its host contract; execution bodies unchanged
from the proven production shape. New contract battery
in CI pins dispatch-only triggers and lane binding.
Phase 2 COMPLETE (T008/T009, 2026-10-31):
native Linux and macOS transports (one shared script
body, drift-checked), a Docker transport (caller image,
declared mounts, results land on the host for grab
return), and gatekeeper --image/--mounts submit
options with host-config image fallback — all pinned by
golden argv checks and a hosted smoke battery
(ubuntu + macOS + a docker-required job).
Phase 3 COMPLETE (T010–T013, 2026-10-31): the
backchannel agent self-tests at start-up and reports
the result in every heartbeat; gk status shows the
heartbeat age with a stale alarm past five minutes;
a 16-check chaos battery (duplicates, corruption, STOP,
kill/restart) runs the real agent in CI; and the
recovery runbook ships the four cases with
pre-written command bodies.
— failsafe backchannel (owner override order,
2026-10-29: Syncyard must have a failsafe backchannel for when
Syncyard itself is down or blocked, with a dedicated channel
just for Syncyard development in the Desktop Access chat).
A dedicated channel syncyard-dev in the private repo tbitcs/syncyard-backchannel (channels/syncyard-dev/cmd.json, out.json, heartbeat.json), reserved exclusively for
Syncyard development/recovery by the Desktop Access chat — no
program job traffic, ever, and no other chat may post to it.
The desktop agent (backchannel/agent.sh) runs as a WSL
systemd user service independent of the runners, the tray, and
the gatekeeper, syncs over git (no REST quota), executes only
new, hash-verified command ids, honours a STOP kill switch,
and never writes gatekeeper queue records. The VM-side client
(backchannel/bc.py, stdlib only) sends through the GitHub
contents API (works during VM write waves), refuses to
overwrite an unacknowledged pending command unless --force,
and verifies every send by read-back; offline tests live in backchannel/test_bc.py. Named recovery use cases: restart
the packaged tray in the owner's session, read tray.log,
restart a runner lane, read-only network diagnostics.
Periodic Syncyard status-check duty, fleet-wide: every agent that uses Syncyard checks status before
submitting desktop work and at least once per hour during an
active desktop campaign (gk status, otherwise the
backchannel heartbeat/status); a stale heartbeat (>5 min) or a
down tray/runner is reported to the Desktop Access chat, never
worked around. SKILL.md carries the duty and a Failsafe
Backchannel section.
Job protection on any Syncyard change:
begin with a queue snapshot (record count, aggregate sha256
of queue files, open-job list) and end with reconciliation —
the open-job set after must equal the set before, or every
difference is named and explained. No job may be lost
silently.
Fixed
Gatekeeper record updates are atomic swaps:
a job record's file is only ever replaced by renaming a fully
verified temp file over it — the record path is never opened
for writing. The pre-direct-write fallback truncated
the standing record before verification could fail, which is
how records were lost in the 2026-10-28 write waves despite
verified-retry. A failed save now leaves the prior record
standing untouched (or nothing, for a new record — no husk
can exist), sweep-internal saves go through a safe wrapper
that ledgers save_failed and continues the pass, and a job
whose pre-dispatch save fails is skipped rather than
dispatched unrecorded. Proven by the fault-injection battery
in gatekeeper/test_fr013.py.
Gatekeeper refinement (same night): the failed-write
cleanup removes only content that is not a valid record (0-byte
or partial husks); a failed UPDATE leaves the last valid record
standing and fails loudly. The first cleanup compared bytes and
deleted a live record on a failed dispatch update - found and
fixed under the pause-first fix discipline.
Gatekeeper durable writes: every queue, ledger
and control-file write is now write-verify-retry — unbuffered
write, read-back byte verification, bounded retries with backoff,
loud failure when unverifiable. The loader skips unreadable
records with a warning, and a submission reports queued only
after the loader verifies the record written. Defect escalation
2026-10-27: the host VM intermittently refuses writes and can
silently drop a buffered write, leaving a 0-byte file.
Added
Tray offline state is unmistakable: with no runners up (or one
faulted) the tray mark renders in grayscale under its red slash;
the color states mean at least one runner is live.
Tray icon and menu: the notification
icon is the official mark loaded from a copy embedded in the
program itself — clean when all runners are up, amber slash when
some are down, red slash when none are up or one has faulted. The
menu gains Start all and Stop all above the per-lane controls.
Runner directories move under the user profile
(%USERPROFILE%\Development\Syncyard\runners\<lane>) — nothing
Syncyard-owned sits in the drive root. Tray and watchdog defaults
and the setup docs follow; the tray installer is rebuilt against
the new locations.
T007 — a real Windows installer for the tray: SyncyardTray-Setup-0.1.0.exe (NSIS; per-user, silent /S mode,
registered uninstaller), published as a GitHub Release asset with
a SHA-256, plus assets/logo.ico and the tray icon now carrying
the project mark with a state badge. README and SETUP.md document
the download route and manual from-source installation.
— windowless runner supervision: the Syncyard Tray app
(runner/tray/SyncyardTray.cs + install-tray.ps1) supervises the
runners from the owner's tray — hidden starts, per-lane state,
capped restarts with an OWNER NEEDED alert, stray-console hiding,
logon autostart. setup-runner.ps1 now starts runners hidden too,
and the tray subsumes the planned watchdog scheduled task.
Gatekeeper push notifications: jobs register a chat with --notify-chat (on submit or adopt) and receive started (with
expected length), completed (marker verdict), and W2-escalation
pushes via the new gk outbox command — no polling and no new
infrastructure.
closed: the polish pull request merged 2026-10-27 under
the owner's one-time protection pause, with protection verified
identical afterward.
New project logo: a marshalling-yard mark (three lanes converging through
one gate) as assets/logo.svg, with rendered assets/logo.png and assets/social-preview.png alongside it. Replaces the previous emblem.
CONTRIBUTING.md (development setup, the self-hosted security invariant
for contributions, review and changelog discipline) and CODE_OF_CONDUCT.md (Contributor Covenant 2.1).
README badges for CI status and the latest release.
gk.py status now leads with queued/running counts and an ISSUES
section (watermark breaches, failed or pending log harvests, stale
dispatches, completions without a marker verdict), and takes a --repo flag that adds project-repo health: latest CI runs, open
Dependabot alerts, and open Dependabot pull requests.
A daily health check reviews the status board, the ledger, CI, and
Dependabot, and reports issues to the operating chat.
The project governance process governance: a project constitution
(.specify/memory/constitution.md) and numbered specs under specs/ (001 gatekeeper dispatch, 002 runner lanes, 003 status /
logging / health, 004 WSL2 management, 005 fleet cutover,
006 repository and governance), each with an aee-claims.json
recording every requirement as a claim with evidence and a
falsification test.
The SKILL.md currency rule, stated in SKILL.md and CONTRIBUTING.md:
SKILL.md is updated in the same change as any behavior change.
Gatekeeper pause/resume: gk pause --reason <text> holds dispatch
of program jobs (the watch pass continues; status shows the pause
banner; sweep summaries print PAUSED) and gk resume clears it.
Jobs submitted with --maintenance dispatch during a pause,
bypassing the active cap and lane serialization but never the
duplicate check — for operator work on the fleet itself.
Gatekeeper stop/start and purge, completing the lifecycle CLI: gk stop is the hard halt (sweeps return before spending any API
budget — no watch, no dispatch) and gk start clears both a stop
and a pause. gk purge removes terminal (completed/cancelled) job
records, dry-run unless --yes, with an optional --older-than-days filter and --ledger to archive the ledger
to a fresh one.
Fixed
Log harvest now falls back to the REST logs archive when gh run view --log cannot read a run (observed in the field: some
runs' logs are unreadable through the CLI but fine through the API).
The harvest records which source supplied the log.
Job-state saves fall back to a shell write when the host denies the
native temp-file write path, so sweeps keep persisting state through
intermittent write restrictions.
Changed (logging)
Sweeps now also write their one-line summary to a human-readable
rotating log (gatekeeper.log, 1 MB, one backup) alongside the
append-only JSON ledger.
Changed
License: the project is now licensed under the Apache License,
Version 2.0 (owner decision, 2026-10-27), superseding the
all-rights-reserved terms the 0.1.x releases shipped with. The
historical release entries below are unchanged.
Project renamed from Sallyport to Syncyard. The repository is now AXIOVEX/syncyard; GitHub redirects the previous repository URL.
Documentation, the agent skill, workflow and client repository
constants, setup defaults, templates, and release metadata now use
the Syncyard name.
README now describes the project in terms of synchronized runner
lanes: a marshalling yard for CPU and GPU jobs dispatched through
GitHub Actions.
The version stays at 0.1.1 for now, by owner decision: no version
bump until the project is fully landed, passes full CI, and has
served as the fleet's daily driver for a few days.