Repository navigation
feat(pi): mount the crew member panel on pi member threads - #22813
Conversation
pi joins ACP_BACKENDS_MEMBER_PANEL on its own evidence (harness-parity H6): the array is client-composed (mirror-less branch) and reaches pi through the gate extension's sealed handoff, and the routing is the enforced VERIFIED_GATE_EXTENSION. A live capture off pi-acp 0.0.34 over pi 1.2.0 (test/fixtures/acp_frames/pi/member-panel-mount-live.jsonl) shows the resolver's own panel element registered, its tools listed, a panel_publish call gated by session/request_permission, and the panel server's own answer. pi stays out of ACP_BACKENDS_MEMBER_CAPABILITIES, and the set's comment now says why: no mirror records a consumed spec, a saved spec's own servers reach pi only as pooled stubs, and "tools": [] is not enforced there. Tests pin the exclusion and each reason.
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: pi joins one capability set on its own live capture, and no code path or frontend surface changes. Notes
[DESIGN-REVIEWED] 7b8aa00 Description read[DESCRIPTION-READ] 6439c33e6a41b8741b69dce9b1d56935956ee58a72b122dcc8c70f7bcf2678c5 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 |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing to check. What this change shipsInventory (5 items) — 5 justifiedIntent: let a crew member on the pi harness publish its own panel in its DM thread, as it already can on the other member-dispatch harnesses. This is an ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 7b8aa00 Description read[DESCRIPTION-READ] 71a8728fbdbe0dd791c41286ca1e314a9247403b3af15b76e9a0c2f4567f6355 That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Problem / Motivation
Goal: A crew member's DM thread on pi gets its own panel, like it does on claude, opencode, goose or deepseek.
pi already held a member session key and the session-control tools (
ACP_BACKENDS_MEMBER_DISPATCH), butACP_BACKENDS_MEMBER_PANELleft it out because no panel mount had been driven on it (harness-parity H6 asks for each capability on its own evidence). So a pi member could not publish its webview, and the set's own comment inbackend_cards.OFF_CARD_SETS("same as ACP_BACKENDS_MEMBER_DISPATCH's") was untrue for pi.Why it matters
pi is the only member-dispatch harness without the panel. A member running there loses the one surface an operator reads to see what it is holding and what needs them.
Not a goal
ACP_BACKENDS_MEMBER_CAPABILITIES). pi stays out; this PR writes down why (see below). It needs the spec's servers projected into the handoff and the zero-tool ban enforced first, as its own change.What changed (motivation → approach → change)
ACP_BACKENDS_MEMBER_PANELinagent_sdk/backends_members.py. Both preconditions the set is argued on hold for pi: the array is client-composed (the mirror-less branch builds it and the gate extension's sealed handoff carries it, identity env intact), and the routing (VERIFIED_GATE_EXTENSION) is enforced. No code path changes: the panel append already runs on the mirror-less branch, gated by this set.test/fixtures/acp_frames/pi/member-panel-mount-live.jsonl(+ replay snapshot), off pi-acp 0.0.34 over pi 1.2.0: the client's own resolver composedkirocrew-dashboard+kirocrew-panelfor a member key on pi, pi registered both and listed all ten panel tools, apanel_publishcall raisedsession/request_permissionthrough the gate, and the panel server's own answer reached the turn.ACP_BACKENDS_MEMBER_CAPABILITIES's comment now says why pi is not a member: no mirror means_session_agent_specis never recorded, so confirming would refuse every saved-spec member; a saved spec's own servers reach pi only as pooled stubs; and"tools": []is not enforced on pi. So a saved-spec member on pi is refused before spawn (capability_harness_unsupported), never run narrower or wider than saved.Backwards compatibility
Compatible: one backend joins a set; nothing that works today stops working.
Removes nothing: the diff only adds a set member, comments, tests and a fixture.
Tests
test/test_member_mount_pi.py: pi is in the panel set; the live capture shows the panel tool called, gated once by the gate envelope (nonce matches the header), and answered bykirocrew-panel; the captured array is what the resolver composes; a member array now carries dispatch + panel, both with identity in the sealed handoff;agent.crew_panel: false, adisabledpanel server, and a narrowed panel tool each withhold the panel only; an ordinary session still carries nothing.member_backend_gaprefuses pi (deepseek passes the same input); a spec with its own MCP server does not place it on pi; pi is not in the zero-tool-ban set; and pi records no consumed spec, soconfirm_consumed_projectionwould raisecapability_runtime_unverifiedon a matching saved intent.test/test_member_mount_deepseek.py: the pi branch test now expects dispatch + panel.prove.py: PROVEN (reverting the set change fails the tests).Manual verification
HARNESS_BASE_REF=origin/main python3 scripts/check_harness_parity.pyand--test: pass.scripts/local-gate.py --base origin/main: related backend + frontend sets pass; flake ratchet passes. semgrep is not installed on this host; CI's SAST lane runs it.pi/README.md: this host refuses user namespaces, so it has no sandbox backend and Crew refuses to spawn pi unsandboxed. No gateway-created pi session (and so no pi member turn in the dashboard) can run here. The capture was driven by a scratch stdio driver outside the product path, and the panel server's answer isidentity_unattested: its token reached a disposable local gateway, which refused to vouch for a session it did not create. The publish route itself reads no backend.Screenshots / video
N/A: the diff touches no frontend file. A screenshot of a pi member thread showing the panel would need a gateway-run pi session, which this host cannot start (see above).
Related Issues
No issue linked, by the owner's standing preference not to open tracking issues.
Checklist