Each tick lists skipped items with terminal and reason (concurrency n/m with
the holding runs, serialize group holder, busy_process not started by agentd,
relay down, no route for assignee, start failed) in heartbeat.json "skipped".
A reason is logged (agentd.log / cron.log) only when it is new for that item;
the last set is kept in the state file as "skips".
- config/agentd.yaml is now the agentd-01 live config (state /var/lib/plane-agentd, logs /var/log/plane-agentd);
config/agentd.oregon.yaml keeps the oregon dispatcher settings for rollback (cron/agentd-tick.sh uses it).
- cron/agentd-tick-lxc.sh: the LXC cron entry.
- tools/win10/: win10-ps.sh / win10-put.sh (vault client dir via AGENT_ACCESS_HOME); relative helpers_dir
resolves against the repo, default tools/win10.
- config/protocols/gwent-board-protocol.md: copy of gwent/design/board-protocol.md used in prompts.
- agentd.log_dir (default state_dir).
- smoke --watch: create the item and wait for the scheduled dispatcher instead of ticking locally.
Before the launch call the planned run (run id, terminal, handle, bookkeeping) is saved as a launching
marker. If the launch call fails or agentd dies before recording the run, the next tick probes the run
directory (exit.txt / launch.pid / pid.txt): evidence -> adopt the run; nothing -> relaunch with the same
run id in <run_id>-r2 (-r3 ...); probe unreachable -> keep the marker, which holds the terminal slot.
Transports refuse to clean a run directory holding a pid file. A terminal whose relay fails is skipped
for the rest of the tick (heartbeat.down) while other terminals carry on.
The dispatcher moves to an LXC without agent CLIs (design 9), so every terminal is
reached through Warpgate. The relay call (agent-access.py run --target <t>) lives in
one helper; the target comes from the terminal config.
os linux/darwin (PosixWarpgateTransport): each primitive is one bash -c call. Files
go up as base64 (chunked above 60k), start_run stages prompt.md + run.sh and launches
in a single round trip; the runner is detached with setsid nohup (pid == pgid == sid,
$! kept in launch.pid); kill signals the process group and session (TERM, then KILL),
skipped once exit.txt exists. macOS detaches with perl POSIX::setsid.
os windows keeps the win10-ps.sh / win10-put.sh helper scripts (target win10) and
schtasks, now through the shared call() helper.
oregon is now warpgate-ssh (target oregon, run_root ~/.gwent-board/agentd-runs);
the local transport stays for co-located setups and tests.
executors.<x>.tier_defaults maps a model to the tier used when no label sets one;
a tier:<name> label sets it explicitly and tier:none drops the tier argument.
- runbook: how executor binaries are resolved (bins or PATH) and what that
assumes on oregon (no codex installed) and win10; cancellation and
max_progress_rounds in the item lifecycle; secrets scrubbed from children.
- design: blocked-by ordering is a TODO (not read by the scheduler); failure
classification, success-when-tagged, one-time immediate rebuild, retries
outside the thread-round budget, ID-based new-comment filter.
- executor contract: how core reads ExecResult, kill used for cancel, Windows
quoting assumptions, codex exec resume without -C/-s.
- A run whose item leaves In Progress is killed, gets a [worker] cancel comment,
frees its terminal slot and serialize group and keeps its thread id; if the
kill fails the slot stays taken and the cancel is retried next tick.
- Transient retries are the same logical round: they do not count toward
max_rounds_per_thread and do not trigger a max-rounds rebuild.
- New project setting max_progress_rounds (default 12): after that many
consecutive [progress] rounds the item goes to Waiting (answerer) with a note.
- retry/rebuild markers are consumed only after the run launched, so a failed
start no longer leaves an In Progress item that is never picked again; a
rebuild is no longer reported as a retry.
- State is saved after every finished round and every launched run.
- Failure classification: transient patterns only look at the tail, and an
exit-0 failure is logic (never retried) unless the tail shows a transient error.
- First-line tag: tolerate BOM and a leading code fence / rule line.
- Start comment notes when a handoff summary was attached.
- Executors and relay helpers get an environment without any configured
Plane token / webhook secret (design 8: executors hold no Plane credentials).
- Windows runner: Set-Location -LiteralPath, and a failed cd writes exit=97
instead of running the agent in the wrong directory.
- schtasks launch: escape apostrophes and pass the script path as \"...\" so
paths with spaces survive PowerShell 5.1 native quoting.
- State and heartbeat files are written temp + fsync + rename.
- Tests for the above plus codex run/resume command shapes and unique run dirs.
One tick polls in-flight runs, posts results, classifies the first line and
moves state/assignee; failures get a structured note with transient retry and
backoff, timeouts are killed and handed over, resume failures and long threads
rebuild with a handoff summary. Then the lowest Todo items are dispatched within
terminal concurrency and serialize groups, and a heartbeat is written.
CLI: tick, new, show, comment, state, assign, smoke, check-config.
Plane: paginated issue/comment/state/label/member access, one client per
terminal identity. Transports: local (setsid-detached runner) and warpgate-ssh
(win10-ps.sh / win10-put.sh relay, schtasks launch, taskkill on timeout).
Executors share a runner shape (prompt.md, exit.txt marker) with codex,
claude-code and shell adapters for last message and thread id.
config/schema.json mirrors design §5 (terminals, executors, projects with routing,
retry, timeouts). agentd/config.py loads YAML, validates with jsonschema when
available or a small built-in draft-07 subset, adds cross-reference checks, and
maps token_key to the environment variable injected by agent-access exec.
jj.wang