Control-plane repo for the djbclark ops suite's task/issue orchestration.
Holds the shared Beads DB (.beads/), Ralph TUI config fragments, and the
small standard-library agent-coord safety client used to report worktree
state and append claims — no application code or infrastructure config of its
own.
The client is intentionally machine-local and advisory outside the narrow
high-value gates described in
docs/agent-coordination-protocol.md.
Run it from this checkout with bin/agent-coord status --format text.
| Repo | Role | Controller epic | Routing label (metadata only) |
|---|---|---|---|
| djbclark/stayturgid | Android device automation / ops tooling | ops-djbclark-cr0 |
repo:stayturgid |
| djbclark/site-djbclark | Public site / ansible playbooks | ops-djbclark-6ub |
repo:site-djbclark |
| djbclark/site-private | Private site config + agent memory | ops-djbclark-6qp |
repo:site-private |
| frdminc/Shizuku | Fork (RikkaApps → thedjchi → frdminc) providing the Android permission broker stayturgid depends on | ops-djbclark-bk7 |
repo:shizuku |
All issues live in this repo's single shared Beads DB — not one DB
per code repo. Every per-repo Ralph controller's tracker config points
workingDir at this same checkout and scopes with trackers.options.epicId
set to that repo's controller epic (each repo's task graph lives as
children of its epic). Routing is epic-based, not label-based — an
earlier design used --label <routing label> instead, but bv --robot-next/--robot-triage were found (2026-08-01, first live run) to
silently ignore --label entirely, and beads-bv has no fallback
verification for labels the way it does for epics. See
docs/ralph-tui-setup-guide.md §0/§2 and
subsy/ralph-tui#401 for
the full story. repo:* labels are still applied to every task for
human-readable filtering/search convenience, just never relied on for
controller scoping.
This shared-DB-plus-epic scheme was chosen over true Beads federation
(bd repo add/per-repo .beads/ + distinct ID prefixes) because
everything runs on one machine — one shared directory sidesteps
Dolt-remote/JSONL sync entirely. Revisit federation only if controllers
ever need to run from different machines.
Ralph TUI (~/.bun/bin/ralph-tui) is architecturally single-git-root per
session — one cwd, one lock, one worktree manager, one merge queue.
It cannot orchestrate multiple repos as a single process. The working
pattern instead is one Ralph controller run per repo, each pointed at
that repo's own --cwd, with every controller's tracker
(trackers[].options.workingDir) pointed at this repo's .beads/ instead
of a repo-local one. That gives each repo's coding agent its own git root
while sharing one task graph across all of them.
Each orchestrated repo's .ralph-tui/config.toml sets
trackers[].options.workingDir to this repo's absolute path directly
(not a BEADS_DIR env var). That's a deliberate choice, confirmed against
beads-bv's actual source
(getWorkingDir()/detect() in src/plugins/trackers/builtin/beads-bv/index.ts):
detect()'s readiness check is access(join(workingDir, beadsDir)), and
getWorkingDir() returns the configured workingDir verbatim — so pointing
it straight at this repo's absolute path makes the check pass against the
real .beads/ here, with no local placeholder needed in the orchestrated
repos at all. subsy/ralph-tui#397
is a real, separate bug (about the BEADS_DIR env var path specifically,
when no workingDir override is configured) — not applicable to this setup
since we don't use that path, but still worth the upstream fix and worth
re-checking if this design ever changes to rely on BEADS_DIR instead.
- No application code, ansible config, or device-automation logic — that stays in the four repos above.
- No release process. This repo is not part of the ops suite's
coordinated
ops-vMAJOR.MINOR.PATCHrelease contract (seesite-private'sCLAUDE.md/docs/OPS-RELEASES.mdinsite-djbclark) — it's operational task data, versioned by ordinary commits, not tagged releases. - Shizuku's own release process (
v13.7.0-thedjchi+stayturgid-releaseNNtags, independent signing key) is untouched by any of this — only its task/issue graph is shared here.
See the full setup writeup (agent-facing, not yet migrated into this repo) for the researched rationale behind these choices.