Repository navigation
docs(rfc): propose gateway session interoperability - #18294
twpedersen wants to merge 2 commits into
Conversation
First Principles Review (Opus 5.5, fork) — ⏭️ skippedRevision |
UX Review (Opus 5.5, fork) — ⏭️ skippedRevision |
Design Review (Opus 5.5, fork) — ✅ PASSDesign-level review of Design-Verdict: PASS This is a docs-only draft RFC with one index row. It removes nothing and changes no runtime contract, so it does not block. Notes
[DESIGN-REVIEWED] 6924c69 Description read[DESCRIPTION-READ] 75a38827bf83a2005ca1d8bfb075b4cff27dfd7599d17a65032685b77fa41b13 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 |
GPT 6.1 Review (fork) — ✅ no blocking findingsReviewed Review detailsNo findings. |
Opus 5.5 Review (fork) — ✅ no blocking findingsReviewed Review detailsNo findings. Notes:
[OPUS-REVIEWED] 6924c69 |
93b2d8c to
7442359
Compare
7442359 to
003d43f
Compare
003d43f to
02a83ff
Compare
|
Hi @twpedersen, this RFC now conflicts with main again, only in the index table in |
28a883b to
d125ece
Compare
|
🤖 Rebased onto current The current-base audit also found that Copilot joined The full 62-command profile passed before the final unrelated upstream rebase, and bounded exact-head checks passed afterward: docs/static gates, 84 backend tests, and 12,451 frontend tests. Opus 5.5 and the available GPT-family fallback ( |
|
Hi @twpedersen, main moved again and this head (d125ece) now conflicts with it, only in the RFC index table in |
d125ece to
b63cd0c
Compare
|
🤖 Rebased onto |
bolichen97
left a comment
There was a problem hiding this comment.
Maintainer decision on this head (b63cd0c): accept the direction of this RFC as a draft. This is a comment, not an approval.
I checked the delta from 28a883b, my last review. The only change to the RFC is the new Copilot row in the harness table, plus the matching "eight" → "nine" updates in the Phase 1 tests and exit criteria. The Copilot row is fail-closed. It has a mirror carrier and whole-server narrowing, but replace/restore return the side-effect-free harness_unsupported until its user-level ~/.copilot MCP sources are disabled or refused before readiness. It becomes a supported row only once a seeded user-level server is shown not to be callable. That is the right posture.
My earlier points (owner predicate vs. authorization, bootstrap row, app-token wording, the env equality oracle, no-wait provider_busy, per-harness narrowing, and the containment-prerequisite scope) are all fixed, as noted in my previous reviews.
Before the status can move from draft to accepted:
- The separate managed-execution containment RFC must be accepted and implemented. This RFC already names it as a hard Phase 1 prerequisite.
- The six open questions must be settled: the correlation carrier,
registration_idspelling, restart posture, title events, conditional restore, and the first consumer.
b63cd0c to
d074511
Compare
iamwhatever
left a comment
There was a problem hiding this comment.
Request changes at d074511: one deferred item needs a tracking issue.
The RFC itself is sound as a draft: the maintainer accepted its direction at b63cd0c, this head is a byte-identical rebase, every earlier review item is answered, and CI is green.
Blocker
- The RFC makes a separately accepted and implemented managed-execution containment RFC a hard Phase 1 prerequisite (
), and the maintainer's review lists it as the first condition for
accepted. No issue or RFC tracks that work, and #16891 does not mention it. Please open one issue for the containment RFC and link it in the PR description (and, if you like, from the RFC'stracking-issues).
Non-blocking, easy to fix in the same push
- The README index row still says "the eight harnesses" ( ); the RFC's table now has nine rows.
— crew-pr-reviewer (r145)
|
Re: review 5764897971 (crew-pr-reviewer r145):
|
Define the owner-authenticated, protocol-neutral session contract before implementation. Separate Gateway behavior, the existing Python client, and protocol adapters into independently reviewable phases.
…s in the index Part of kirodotdev#16891
|
Thanks @twpedersen for this RFC! Fork CI hit a wall here: the fork Security Scope lane failed with |
|
Thank you @twpedersen for this thorough RFC. Gateway session interoperability is a large direction change that no maintainer has taken up, so we are closing this proposal (and its in-repo copy #31179) for now. If the need comes back, this document is a good place to restart from. |
Problem / Motivation
Goal: Record a reviewable design for owner-authorized, protocol-neutral Gateway session interoperability before implementation begins.
Trusted local clients currently cannot configure session-scoped MCP requests, answer stateless question cards, correlate turns, or consume minimal session events through one neutral contract. The existing implementation proposal reached code review before the required architectural decision was recorded.
Why it matters
A base-branch decision lets maintainers settle the trust boundary, wire vocabulary, lifecycle ownership, and phased rollout before server, client, or adapter code depends on them. It also lets later implementations reuse the existing Gateway and Python client rather than creating protocol-specific security and lifecycle paths.
Not a goal
What changed (motivation → approach → change)
This PR adds a draft RFC for the T4 architectural change tracked by #16891. It audits the current owner-authentication, question-card, MCP-projection, WebSocket, and Python-client seams; defines a protocol-neutral owner-only contract; and specifies fail-closed authorization, name-only MCP requests, non-blocking provider mutation, atomic question answers, turn correlation, bounded events, and audit-or-deny behavior.
The rollout remains default-off and separates the Gateway contract, Python-client extension, and any later protocol adapter. Phase 1 is Linux-only and requires a separately accepted and implemented managed-execution containment RFC, tracked in #29483. The document gives every selectable harness an explicit projection outcome, treats PR #14999 and PR #7415 as falsification evidence rather than dependencies, and leaves six named design questions for maintainer sign-off.
Backwards compatibility
Compatible: documentation only; no runtime contract or accepted decision changes.
Removes nothing: the PR adds one draft RFC and one index row.
Tests
b63cd0c525d39e0e40a0ef0829a5adf7fca483b7overda44457c102bcb6fc444aa7240bf3d02c9381e8b, including documentation and policy gates, 84 related backend tests, and 13,041 cross-surface frontend tests. The preceding exact-candidate full profile also passed types, lint, Electron, duplication, and memory-store checks.SettingsPage1.7 KB over budget. The identical sequence reproduces on clean currentmain; a standalone clean candidate build passes the bundle and chunk-cycle checks.b63cd0c525d39e0e40a0ef0829a5adf7fca483b7. Opus ran asclaude-opus-5.5; the GPT lane's served model was not reported.Manual verification
Confirmed the RFC body remained byte-identical through the rebase, the current-main index was preserved wholesale, and exactly one audit-history sentence and one RFC index row were restored. The branch is one clean documentation commit directly over current
main.Screenshots / video
N/A — documentation-only change with no rendered product surface.
Related Issues
Part of #16891
Prerequisite tracked in #29483
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement