Repository navigation
mcpx: sampling relay -- human review, validation of results, error mapping (8 conformance rows) #1955
Description
Relocated from dezren39/darwin-nix#256 on 2026-10-28 (placement audit): mcpx's single source lives in developing-today/code (pkgs/mcpx); re-home tracked at #309. Original body below; issue references inside it are historical (dezren39/nix numbering).
Gap keys sampling-review (4), sampling-validation (2), sampling-errors (1) in docs/spec/conformance-matrix.md: relayed sampling/createMessage is not offered for human review, results are not validated against the request (model/tool-use shape), and failures are not mapped to the codes the revisions specify. Sampling is deprecated in 2026-07-28; user instruction is to keep it working, add nothing new -- so scope is conformance only.
Acceptance criteria
- Relayed
sampling/createMessageis offered for human review per the sampling-review rows. - Sampling results are validated against the request (model/tool-use shape) before relay.
- Sampling failures map to the error codes the current revision specifies.
- Scope stays conformance-only: sampling is deprecated in 2026-07-28 and the standing instruction is keep-working, add-nothing-new.
Activity
developing-today commentedon Oct 28, 2026
CODE-880
dezren39 commentedon Oct 28, 2026
History carried over from dezren39/darwin-nix#256 (verbatim, oldest first):
@dezren39 · 2026-10-27T15:35:39Z
Intake review
Fleet monitor intake pass, 2026-10-27 (repo-fleet-monitor).
Related: the 2026-09-30 mcpx conformance set — parent analysis #252, siblings #253, #254, #255, #256, #257, #258, #259 (filed together from
docs/spec/conformance-matrix.mdin dezren39/nix). Treat the set as one body of work; do not re-file per-gap issues.Notes: 8 conformance rows (sampling-review ×4, sampling-validation ×2, sampling-errors ×1 +1). Effort l. The deprecation note in the body is the scope guard — this is correctness of an existing relay, not new sampling features.
Risks: Building review UI beyond the conformance rows; keep to the matrix.
Suggested approach: error mapping and result validation first (pure relay logic), human-review surface last.
@dezren39 · 2026-10-28T19:12:51Z
Triage verdict (task 2310): CONFIRMED at current main for validation + error mapping; human-review subset needs a design decision first.
Verified at developing-today/code main
35726e8e(mcpx source:pkgs/mcpx):sampling-both-validate-message-content(both sides, every spec section),sampling-client-errors-user-rejected-and-tool-shape(conformance-matrix.md:1621), and the four human-review keys.internal/mcpclient/modern.goanswer()relayssampling/createMessagestraight to the handler — params/results are not validated against the request — and every handler error returns-32603(~:296–300). Neither-1(user rejected) nor-32602is produced on this path. (Older-revision user-rejected rows are already tested; the gap is the 2025-11-25 combined row.)internal/elicit/policy.go:138: 'sampling asks a model, and the agent driving mcpx has one'. What 'human review' should mean for a headless agent relay (interrupt the agent flow? an approval audience like elicitation's?) is a design decision, not an implementation detail, so it is not in the fix packet. Drew/design: this is the decision fix(id): show world chat only to participants who may chat or read history #256 still needs.Routing: validation + error-mapping fix enqueued as task 2620 (queue
repo-health-nix-pilot, deduperepo-health:v1:fix:dezren39/nix:issue:256), conformance-only scope per the standing instruction (sampling is deprecated in 2026-07-28: keep working, add nothing new). This issue stays open for the human-review decision + the packet's matrix rows.dezren39 commentedon Oct 28, 2026
🤖 Agent note · picked up
dezren39 commentedon Oct 28, 2026
🤖 Agent note · done
queue/queued.Dedupe result (checked fresh 2026-10-28T22:53Z)
Design #1834 — "mcpx: design — sampling trust boundary (reviewable, validated, honestly error-coded)" (filed 2026-10-28T21:35Z, origin: dezren39/darwin-nix#256) scopes exactly this issue's subject, including its conformance-only scope guard, and settles the human-review design question the earlier task-2310 triage left open: review happens through the existing
mcpx elicitsurface — "no Drew direction decision is needed — nix#256's acceptance is explicit". Its three phase issues map 1:1 onto this issue's acceptance criteria:-1user-rejected,-32602tool-shape, expiry stays-32603)sampling/createMessageoffered for human review per the sampling-review rowsmcpx elicit+ matrix/register truth-up + nix#256 evidence (open, queue/queued, gated on Phase 2)Scope itself is confirmed live, not stale: the conformance matrix at current main (
6a284efd7) still cites #256 on 26 rows, and #1835/#1836/#1837 cite the current code sites (sampleViaBrokerininternal/daemon/hooks.go, the-32603flattening wrap ininternal/mcpclient/modern.go~L296-303,renderQuestionininternal/cli/elicit.go).Loose end named
Task 2620 (queue
repo-health-nix-pilot, deduperepo-health:v1:fix:dezren39/nix:issue:256) — the validation+error-mapping packet the task-2310 triage enqueued before relocation — is superseded by #1835/#1836 in the code pilot queue. It is not claimed (nix queue claimed_count 0 at check time); flagging it here for cancellation by the nix lane / fleet monitor rather than cancelling another queue's task from this lane.Closing this issue as a duplicate (GitHub state_reason: duplicate) per the triage packet's duplicate rule — scope is not shrunk; every acceptance item has a named, queued home above.