Repository navigation
docs(rfc): hybrid local/remote execution, three open decisions - #21613
Conversation
|
Thanks for writing this up. The RFC index has changed on main since you branched, so the branch conflicts. Could you rebase onto current main and re-add your index row? |
ca0f16d to
60b5ef9
Compare
|
Thanks. Rebased onto current main (
|
bolichen97
left a comment
There was a problem hiding this comment.
Thanks for the rebase. The branch is now one commit on 4be3332429, and GitHub reports it as mergeable.
Head reviewed: 60b5ef982a. This is a design document for a T4 issue, so it needs a design sign-off from a maintainer. This review is a comment only.
What I checked against main:
- Every "on main" claim in section 2 and in the index row that I sampled holds:
UNREACHABLE_CODESand<fieldset disabled={!connected}>inCrewChatWindow.tsx,proxy_peer_not_connectedinssh_tunnel_manager_peer.py,compute_max_subagentsreturningagent.subagent_auto_max,instances.mdsection 14.3,api_instances_send_sessionand its route,createRemoteChatMutationincreate.tsand its call inSidebarHeader.tsx, the 0.9.0 changelog entry,redact_peer_texton peer text, and noremote_subagents.pyon main. scripts/docs-lint.shandscripts/docs-lint.sh --testpass on this head.- The diff matches the description: one new RFC file and two additions to the index.
The AI review lanes have not run on this head. Five fork workflows wait for maintainer approval, so PR Readiness fails for that reason only.
Two small points are inline. Neither blocks the design discussion.
| | One Sessions list across crews, grouped per machine | On main, same preview | `useRemoteCrewChat` in `website/src/hooks/usePreviewFlag.ts`, called from `sessionSources.ts`; `read_peer_slots` in `src/kiro_crew/dashboard/handlers_instances.py` | | ||
| | Send a copy of a session to a crew | On main | `api_instances_send_session` in `handlers_instances.py`, route in `dashboard/routes/connections.py`; `docs/system-specs/modules/instances.md` §14 | | ||
| | The older turn relay and "bound local slot" model | Gone: `rfc-remote-crew-sidebar` wave 4 ([#17642](https://github.com/kirodotdev/KiroCrew/pull/17642), [#17662](https://github.com/kirodotdev/KiroCrew/pull/17662)). `remote_already_bound` and `remote_not_connected` have no hit under `src/` | `git grep` at the sha above. The 2026-09-25 triage cites both codes at an earlier sha; they no longer apply | | ||
| | Remote subagent executor: `spawn_run(executor="remote")` | **Open**, [#12821](https://github.com/kirodotdev/KiroCrew/pull/12821). Not on main: `instances.remote_subagents`, `remote_spawn` and `spawn_enforces` have no hit; `src/kiro_crew/dashboard/remote_subagents.py` does not exist. Reviewer `cruisercohen` re-reviewed head `026161ec5` on 2026-10-16 and found every point from his previous review met. A `CHANGES_REQUESTED` review of 2026-09-25 by `bolichen97` is still recorded, and the PR carries a merge conflict | `gh pr view 12821` on 2026-10-17 | |
There was a problem hiding this comment.
🟡 The #12821 row records review state that will be stale soon.
The row names who re-reviewed which head, which review is still recorded, and that the PR has a merge conflict. These facts change with each push or review on #12821. This document stays in the index as the decision record, so a reader later sees a wrong state and has no date beside it except the gh pr view cell.
... re-reviewed head `026161ec5` on 2026-10-16 ... A `CHANGES_REQUESTED` review of 2026-09-25 ... is still recorded, and the PR carries a merge conflict
Why it matters: the RFC is the durable record. Review state belongs on the PR, where it stays current.
Suggestion: keep "Open" and the "Not on main" evidence, and remove the review-state sentences. If you want to keep them, start them with "As of 2026-10-17".
There was a problem hiding this comment.
Agreed, fixed in 927d68ba09. The row keeps "Open" and the not-on-main evidence (now cited as git grep at the audited sha) and says its review state lives on the PR. The reviewer, head and conflict sentences are gone.
|
|
||
| | Phase | What | Depends on | Exit criteria | | ||
| |---|---|---|---| | ||
| | 1 | [#12821](https://github.com/kirodotdev/KiroCrew/pull/12821) as it stands: explicit remote placement, refuse when no crew is connected | Nothing in this RFC. It implements Decision 1 A, Decision 2 A and Decision 3 A as written | Merged; `spawn_run` default stays `executor="local"` and every local call behaves as before | |
There was a problem hiding this comment.
🟡 Phase 1 depends on "nothing in this RFC", but section 1 says to decide 2 first.
Section 1 (line 29) says a routing rule is only safe once it is known whose memory the routed work uses. Phase 1 is #12821, which adds remote spawns that use the crew's memory. So merging Phase 1 puts Decision 2 A into effect for spawns before a maintainer rules on Decision 2. Decision 2 A is the status quo for sessions, so this can be correct. But the two sentences read as a contradiction.
| 1 | [#12821](...) as it stands: ... | Nothing in this RFC. It implements Decision 1 A, Decision 2 A and Decision 3 A as written |
Why it matters: a maintainer who reads section 1 can hold #12821 for a Decision 2 ruling. A maintainer who reads section 6 does not. The document should give one answer.
Suggestion: say it in one place. For example: "Phase 1 extends Decision 2 A, the status quo for sessions, to remote spawns. Accepting #12821 accepts Decision 2 A for spawns." If you intend Phase 1 to wait for Decision 2, change the Depends on cell to "Decision 2 accepted". If the intent is already the first reading, I will withdraw this point.
There was a problem hiding this comment.
The first reading is the intent: Phase 1 should not wait for a Decision 2 ruling. Fixed in 927d68ba09 so both places give that answer. Section 1 now says Decision 2 gates the routing and fallback rules (1 B, 3 B), and that explicit placement (1 A) extends Decision 2 A, the status quo for crew sessions, to remote spawns, so accepting #12821 accepts Decision 2 A for spawns. The Phase 1 "Depends on" cell says the same instead of "Nothing in this RFC".
… vs Decision 2 Review on kirodotdev#21613: the kirodotdev#12821 row no longer records review state, which belongs on the PR; section 1 and the Phase 1 row now both say that Phase 1 extends Decision 2 A to remote spawns, so accepting kirodotdev#12821 accepts Decision 2 A for spawns.
UX Review (Opus 5.5, fork) — ⏭️ skippedRevision |
First Principles Review (Opus 5.5, fork) — ⏭️ skippedRevision |
Design Review (Opus 5.5, fork) — ✅ PASSDesign-level review of Design-Verdict: PASS This is a draft RFC with documentation changes only. It removes nothing and changes no default, so nothing blocks. Notes
[DESIGN-REVIEWED] 8a5134b Description read[DESCRIPTION-READ] b9e90ab75cf90e8001da00e3e1293a24da0cdc6daed849b0b6a5b8918ae931c5 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
Opus 5.5 Review (fork) — ✅ no blocking findingsReviewed Review detailsNo findings. Notes:
[OPUS-REVIEWED] 8a5134b |
GPT 6.1 Review (fork) — ✅ no blocking findingsReviewed Review detailsNo findings. |
bolichen97
left a comment
There was a problem hiding this comment.
Thanks for the update. Head reviewed: 927d68ba09.
Both points from my previous review are fixed:
- The #12821 row in section 2 no longer records review state. It keeps "Open" and the "Not on main" evidence, and says the review state lives on the PR.
- Section 1 and the Phase 1 row in section 6 now give one answer. Explicit placement does not wait for a Decision 2 ruling, and accepting #12821 accepts Decision 2 A for spawns.
I also checked the claim the Design Review asks about. The #12821 head 026161ec5c takes the strictest of the admitted mode and the requested memory_mode, so the memory_mode on spawn is tighten-only, as the RFC says.
I have no new findings. The diff is still one new RFC file and the index row. The second commit changes only the three sentences above.
This RFC is for a tier T4 issue, so it needs a design sign-off from a maintainer. This review is a comment only.
The remaining red checks are not caused by this diff. One Windows backend shard timed out in a CI routing test, and two fork workflow runs wait for maintainer approval.
927d68b to
09645ce
Compare
… vs Decision 2 Review on kirodotdev#21613: the kirodotdev#12821 row no longer records review state, which belongs on the PR; section 1 and the Phase 1 row now both say that Phase 1 extends Decision 2 A to remote spawns, so accepting kirodotdev#12821 accepts Decision 2 A for spawns.
bolichen97
left a comment
There was a problem hiding this comment.
Thanks for the rebase. Head reviewed: 09645cefb0.
The RFC file is unchanged from 927d68ba09. The only change in this PR's own diff is in the index: the row and the provenance note now say the row was re-verified at cea787cbd5. I checked the row against that commit. The symbols it names as present are present, the ones it names as absent have no hit, and the PRs it lists as open are still open. My earlier review still applies and I have no new findings.
This RFC is for a tier T4 issue, so it needs a design sign-off from a maintainer. This review is a comment only.
|
For open question 2 (instances registry as in #12821, or the A2A provider): I've put a concrete proposal on #12821 (#12821 (comment)). Merge the instances-registry path now. Fit A2A in later as a second provider under the same |
… vs Decision 2 Review on kirodotdev#21613: the kirodotdev#12821 row no longer records review state, which belongs on the PR; section 1 and the Phase 1 row now both say that Phase 1 extends Decision 2 A to remote spawns, so accepting kirodotdev#12821 accepts Decision 2 A for spawns.
fa87f06 to
05a24ff
Compare
… vs Decision 2 Review on kirodotdev#21613: the kirodotdev#12821 row no longer records review state, which belongs on the PR; section 1 and the Phase 1 row now both say that Phase 1 extends Decision 2 A to remote spawns, so accepting kirodotdev#12821 accepts Decision 2 A for spawns.
05a24ff to
0f109ac
Compare
|
This RFC proposes hybrid local and remote execution and leaves three product decisions open. It changes where work runs, which machine holds data, and how failures move between environments. Those decisions require explicit maintainer agreement before implementation. design-doc-gate: Rule D — this changes a design document, so it needs a human to agree to the design. Not approved and not rejected by this review. |
|
HOLD at cbaee78: this needs a maintainer's design decision. This draft RFC asks for three product rulings for #3278 (tier T4). The design-doc gate requires a maintainer to agree to the design, and an approval from me would not stand in for that ruling. A maintainer recording acceptance (or a change of direction) would clear it. Two Design Review notes on this head also have no disposition: the RFC front matter (L5) still says audited at 2d926d9 while the index row says verified at 69e016c, and the Phase 1 "Depends on" cell (L92) says "No ruling in this RFC" and then names a Decision 2 A consequence. A short reply or a wording fix would close both. Docs only. CI is green, and the earlier inline points were fixed in 927d68b. — crew-pr-reviewer (r123) |
bolichen97
left a comment
There was a problem hiding this comment.
Re-reviewed at 68d3a46, incremental from cbaee78.
This push adds one commit that changes four lines of the RFC and nothing else. What I checked:
- The front matter
last-auditedandaudited-at, and the body's audit sentence, now name69e016c117, which matches the index row and is on main. - The Phase 1 "Depends on" cell no longer says "No ruling in this RFC" and then names a Decision 2 A consequence. It now says accepting #12821 records Decision 2 A for spawns and implements Decisions 1 A and 3 A. That matches the paragraph at the top of section 1, so the two places now read the same way.
- I spot-checked claims at
69e016c117: the dashboard moduleremote_subagents.pydoes not exist, nothing in the Python package mentionsremote_subagents,remote_spawnorspawn_enforces,CrewChatWindow.tsxstill renders the composer inside<fieldset disabled={!connected}>, and #12821 and #3278 are still open.
No findings. The linked issue is tier T4, and accepting this RFC is a design decision for a maintainer, so I am leaving a comment and not an approval. Readiness was still running on this head when I looked.
|
HOLD at 4157f5d: this needs a maintainer's design decision. The RFC is a draft that asks for three product rulings for #3278 (tier T4), and a maintainer put the design-doc gate on it (comment). An approval from me would not stand in for that ruling. A maintainer recording acceptance (or a change of direction) in section 8 would clear it. The two wording points from my r123 review are fixed: the front matter and the index row both name — crew-pr-reviewer (r124) |
kirodotdev#3278 Scopes kirodotdev#3278 to the three rulings still open: how work leaves a local session, whose memory remote work uses, and what happens when the crew is unreachable. Each is a numbered decision with options and a recommendation. Lists what is already on main or open (verified at 69e016c) so none of it is redesigned, and orders delivery so that kirodotdev#12821 lands as-is in phase 1. Part of kirodotdev#3278
bolichen97
left a comment
There was a problem hiding this comment.
Approving at head 8a5134b.
This head is a rebase of the previously reviewed head 4c94be8 onto current main. The RFC file is byte identical, and the README change carries the same added sentence and table row; the only other line delta comes from main.
Design verdict: sign. The RFC records three conservative rulings for the hybrid local and remote execution issue: remote placement is explicit (overflow routing later and opt in), remote work uses the crew's own memory explicitly, and an unreachable crew refuses with a typed error instead of silently falling back to local execution. It changes no runtime behaviour, gives follow up work one written position to build on, and adds no long term tech debt. Cross instance memory and automatic routing stay deferred, as the RFC states.
Checked at this head: no blocking findings, two files changed, docs only, readiness passed, all checks green, AI review lanes not blocking, and GitHub reports the branch mergeable.
Problem / Motivation
Goal: Give maintainers three small, decidable rulings on hybrid local/remote execution so that #3278 can be unparked and the remaining work can proceed in phases.
#3278 is parked (
tier:T4). The 2026-09-25 triage lists three things that still need a maintainer ruling: routing heavy work out of a local session, remote-authoritative memory, and falling back to local. The 2026-10-08 comment says new concepts go throughdocs/request-for-change/. Meanwhile, pieces of the original request have landed or are in review (Chat on a crew, the crew window, Send a copy to, the remote subagent executor #12821, fetch-back #19440, folder send #19859, file hand-off RFC #19460), so the open question is now narrow.Why it matters
Without a written ruling, each of those PRs is reviewed against an unstated design, and the memory question in particular (
docs/system-specs/modules/instances.md§14.3 says memory does not travel) blocks everything that would route work to another host. A decision record lets #12821 and the follow-ups proceed against the same position.Not a goal
What changed (motivation → approach → change)
Goal → the RFC scopes #3278 to exactly the three open rulings, each a numbered decision with options and a recommendation, and lists what is already on main or open (verified at
2d926d926, definitions and callers) so nothing is redesigned. It orders delivery so phase 1 is #12821 as it stands, which means accepting the RFC does not block that PR.Changes:
docs/request-for-change/rfc-hybrid-remote-execution.md(new, statusdraft).docs/request-for-change/README.md: one index row with an honest "what is actually on main" cell, and the audit sentence extended the way earlier additions did.The RFC ships as
draft, so it records a proposal, not a decision. Per the index's own rule, no product-shape change may cite it until a maintainer records acceptance.Backwards compatibility
Removes nothing: documentation only, no default, API, config key or behaviour changes.
Tests
No code.
./scripts/docs-lint.shand./scripts/docs-lint.sh --testboth pass; everyrfc-*.mdin the directory still has an index row.Manual verification
Every "on main" claim was checked with
git grepin a worktree at2d926d926; PR and issue states were read withgh pr view/gh apion 2026-10-17. Not verifiable from the repository: whetherkiro-cli acpcan attach a cloud session, and the live merge state of the open PRs after that date.Screenshots / video
N/A — documentation only, no user-visible frontend surface.
Related Issues
Part of #3278
Related: #12821, #19440, #19859, #19460, #13031, #19092, #21546, #21547, #10035, #12942
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.