Repository navigation
Replies: 4 comments 1 reply
|
Scope note: PR 1 here (runner -> daemon identity/model/context sidecars) is the same proposal as #2846, where the maintainers said they would confirm fit before anyone builds. Treat #2846 as the home for that half; this thread is for the daemon -> runner channel (cancel, wake). I closed #2863, which I opened too early; the reference branch stays on the fork (smota:feat/community-runner-observability, rebased on main). |
|
Thanks, @smota, for a clear proposal for the other direction: a command channel from the daemon to a community runner for "cancel" and "wake", with the four gaps it would close, who gets what, and use examples. We'll confirm here whether it fits before anyone builds it, and the scope answer is worth having before you put time into code. As you said, the runner-to-daemon half (identity, model and context sidecars) rides #2846, so this thread is for the "cancel" and "wake" channel. Two small corrections so we start from the same place: the existing command is |
|
@mvschwarz — a follow-up clarification to separate the generic command contract, current implementation choices and remaining decisions. This complements the runner-to-daemon observations in #2846 and follows the original integration discussion in #2854. Scope: a command contract for opted-in community runners, not a Nodecode-specific integration. The daemon handles capability checks, routing, launch identity, correlation and deadlines; each runner translates commands into native operations. The dsh reference runner and Nodecode connector are separate consumers. Existing Claude/Codex one-shot wake paths are unchanged.
The last point emerged from implementation: Nodecode currently starts a turn when idle and returns a busy/recent-output snapshot otherwise. That is consumer behavior, not an agreed generic contract, and it differs from a cold one-shot wake. In-flight wake waits also do not survive daemon restart. References: daemon branch at Recorded proof includes dsh reference-runner protocol tests with a fixture native backend, plus earlier private-daemon Nodecode checks for idle/busy wake and mid-turn cancellation. Those live checks were not repeated for this clarification. The YAML scenario-grammar gap is documented in the commit; we are not claiming a scenario pair or durable in-flight recovery. Alignment requested: does this generic cancel/wake boundary fit, and which transport/reply/busy semantics should it use? Scope remains pending; the implementation stays on the fork until agreement, adaptation, rebase and fresh verification. Memory, handover and learning workflows remain outside it. |
|
Yes to a structured channel per run for input and wake, alongside turn-bound cancellation. This is new implementation work, not a claim that the existing permission long-poll already provides that channel. The channel needs ordered delivery, acknowledgements and duplicate handling, including reconnect and uncertain outcomes. It should use the per-seat typed/native route choice. Typing stays available; native delivery starts opt-in and becomes the default for a program only after it has been proved there. Cancellation should name the turn and distinguish requested, cancelled, not-this-turn and not-supported. Withdraw a pending permission request first. OpenRig should not synthesize a typed /abort or use a per-seat command file for this; the person's own Ctrl-C and /abort remain available. Runner-reported identity remains labelled as such, as in #2846. We'd like to prove the input/wake and cancellation behavior with two independently authored runners, OpenRig's Pi runner and the dsh example, including the busy and reconnect cases. That's the proof target, not a claim that those paths already pass. |
Uh oh!
There was an error while loading. Please reload this page.
Motivation
Today a community runner only reports. The daemon sets env at launch, the runner posts hooks, and that is the whole conversation. OpenRig's own docs name what this costs a seat on a community harness:
public-what-you-can-do.md:651)rig transcript --wakecan ask Claude/Codex seats, not community seats (:187)community-harness-runtime-adapter.ts:45)I hit all four building the nodecode harness (#2854, smota/openrig-nodecode). The agy runner and the dsh example have the same ceiling. None of it is harness-specific.
Value
rig seat cancel Xworks on every seat, not only built-ins.rig transcript --wakeanswers from any seat. Seat identity shows verified.unconfirmed/awaiting-decisionstates.Use examples
Non-breaking by construction
commands: true(format 1, default false). Presets without it behave exactly as today;rig seat cancelon such a seat refuses with "harness declares no command channel", same style as current unsupported-runtime errors.OPENRIG_COMMAND_PATH. Runners that ignore it lose nothing.session_identity,model,context_usage) are additive on/api/activity/hooks; unknown families were already ignored.Impact
domain/community-harness-presets.tscommandsflagadapters/community-harness-runtime-adapter.tscancel/wakeroutes/activity.tsclirig seat cancel,transcript --wakeroute to community seatspackages/test-systemdsh example runnercommunity-harness-presets.md,harness-runner-contract.mdNothing touches queue, projection, migrations or restore arteries.
PR breakdown
smota/openrig@feat/community-runner-observability. Ready to open on your nod (will rebase on current main first).cancelandwake, preset flag, two CLI verbs, dsh consumer, scenarios. Independent of 1. One PR because cancel and wake are one channel with two message types.Proposal
I implement both PRs, each with scenario pass + seeded-fail pairs and docs, nodecode as a second real consumer alongside dsh. I'd like your call on two things before building PR 2: file-based channel vs unix socket, and whether
wakeshould reuse/api/activity/hooksor get its own route.All reactions