Repository navigation
Unwatched rigs: knowing whether a send was actually submitted (asking before any PR) #1249
alxledesma
started this conversation in
Ideas
Replies: 2 comments 2 replies
|
Thanks, @alxledesma, for asking about scope before opening a PR, for offering the work in small pieces, and for the plain evidence and limits, including that the real-agent runs weren't unattended. We've read the whole post and will come back here; we don't have an answer on scope yet. |
2 replies
|
One more data point for your scope decision, no reply needed. Reading 0.6.10, several of its fixes address problems my fork handles in a different way:
Each upstream release means I re-solve the same problems on my side, which is why I'd rather contribute pieces than keep a diverging fork. I'm happy to write a short side-by-side for any of these if it would help. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Short version. I want to leave a rig of Codex and Claude Code seats working for hours unwatched. On a private fork of v0.6.8 I've built a state machine for that. I directed the work, and Claude models wrote and reviewed it. I'm not asking you to take it whole. My first question: would you take one focused PR that reports a send as
verifiedonly when the agent's input box shows the text was submitted, as a step toward settling the rows in #165? It can sit behind a setting if you prefer. Upstream behaviour below is as of 0.6.10 (3001cc43). I've read 0.6.9 and 0.6.10 for overlap but haven't rebased.What stops me today
verifiedmeans the first 40 characters of the text appear once more after Enter than in a capture taken before the paste (the post-Enter check insession-transport.ts). Text left unsubmitted in the input box can therefore pass (Titlerig sendcan returnok:truewhile the message is left as a draft in the target pane #14), and stall checks can miss unsubmitted wakes (Stall checks miss unsubmitted wakes and flag waiting or working seats #331).delivered-ack-pendingeven after it's claimed and closed, anddeliveryOutcomeis never written (lastNudgeResult stuck at delivered-ack-pending after claim and closure; deliveryOutcome never written #165; fix(queue): resolve an unconfirmed nudge result to claimed when the row is claimed #1094 improved the display).First PR: input-box evidence for delivery
What a user gets: in a wake result or
rig send --verify,verifiedmeans the exact text was in the agent's input box before Enter, and after Enter the box was clear (or showed Claude Code's queued-message label), on the same pane process. Anything else reads as unconfirmed, including a layout the parser doesn't recognise. A TUI update should therefore produce unconfirmed results, not falseverifiedones.On your #48 trade-off: it changes only what is reported. It reads the box once, in the existing pause between paste and Enter, and that reading never decides whether Enter is sent. Nothing is held, refused or repeated. You'd see two differences: a draft left in the box no longer counts as
verified, and a submission the old check missed can now be confirmed. Other runtimes keep a render check, but it compares the whole text instead of the first 40 characters.What it includes: input-box parsers for Claude Code and Codex, built from real captured screens, plus unit and integration tests. arteries.md notes that the delivery stub has no input behaviour, so I'd add that to the stub, including a draft that never gets submitted. Writing
deliveryOutcome(#165) could come in the same PR or later. That field's values only cover the gateway today, so I'd follow your lead.Limit: it's tested at unit and integration level but hasn't run in a live rig yet.
Second PR, if wanted: maintenance hold
rig daemon start --maintenance-hold <reason>starts the daemon for an upgrade, restore or investigation. Under the hold it makes no automatic sends or recoveries and refuses the mutating API. The hold survives restarts untilrig daemon maintenance release. Without the flag, the only change is that the daemon checks for a hold at start, and treats itself as held if it can't read one. It adds one migration and touches the same start path as #1061, so I'd test the two together.Later pieces, only if you want them
My fork also has crash and restore safety (a send with an unprovable outcome is recorded as unknown and never replayed), opt-in coordinated rigs (one seat holds a time-limited lease, and a handover cuts off the old holder's sends), owner pause, and a bounded fix-and-review loop that ends in an owner decision. These are bigger, may fit better with the bounded agent runs design recorded in #27 or saved operating modes than in core, and can stay on my fork.
One compatibility note: #1074 works with today's behaviour. My fork waits for several independent checks, including a second sample, before it treats a seat as gone. That only matters if recovery ever acts on
detachedautomatically.Evidence and limits
Questions
I won't open a PR until you say yes on scope. You're welcome to build any of this your own way instead. I'll stay around to fix anything I contribute.
All reactions