01 · Task flow
Label-driven, in-process execution
archied polls GitHub for issues matching the configured trigger
(assignee, label, or either). Each issue becomes a SQLite-tracked
task, claimed one at a time, and routed to a workflow by issue
labels -- no LLM involved in routing. All stages run sequentially
inside the daemon process.
02 · Failure boundaries
Environmental constraints, not prompt rules
The quality gate, test-file protection (TDD fix stage), and
diff-size cap are code enforcement -- not "please don't do that" in
the system prompt. Crashed daemons re-queue stale running tasks on
startup. Parks are never silent: every parked task posts a comment
with the reason and gate output.
03 · Design decisions
Deterministic git, fresh worktrees
The model never runs git -- worktree operations (clone, branch,
commit, push, diff, cleanup) are deterministic steps, not agent
tools. Every task gets a fresh clone under
~/.local/share/archie/work/. Tokens are tracked per
task and emitted in observability events, but not budgeted beyond
the agent loop's own limits.