9.7 KiB
作成
LocalTicketBackend によって作成されました。
State changed
Ticket を workspace-panel が queued にしました。
Decision
Routing decision: waiting_capacity
Panel Queue により routing 対象として確認したが、現在 00001KVHKWNQS が inprogress で Coder 作業中のため、この MCP foundation Ticket は実装開始せず queued のまま保持する。
Evidence:
- Ticket body / relations を確認。
- Orchestrator worktree は clean。
- queued は MCP chain とこの Ticket 群、inprogress は
00001KVHKWNQSのみ。 - matching branch/worktree は存在しない。
Next:
00001KVHKWNQSが完了して実装 capacity が空いた後、改めて reroute する。
Decision
Routing decision: implementation_ready
Reason:
- Panel Queue により、この Ticket は Orchestrator routing 対象として明示許可された。
- Ticket body は local stdio MCP server configuration / trust policy の範囲を、spawning/lifecycle を含まない config parse/validation/diagnostics Ticket として明確に定義している。
- 前回の waiting_capacity note は
00001KVHKWNQSが inprogress だったためだが、現在00001KVHKWNQSは closed で capacity blocker は解消済み。 00001KVHR3WRF自身には未解決 blocking relation はない。Incoming00001KVHR3WRY depends_on thisは後続 Ticket であり blocker ではない。- 現在 inprogress は 0 件、child implementation Pods はなし、matching branch/worktree はなし、Orchestrator worktree は clean。
- Risk domain は mcp / config / trust-boundary / secrets / process-exec だが、Ticket は no process spawning、no auto-start、secret redaction、local executable trust boundary、Plugin permissions /
pod::featureauthority separation などの invariants を明示している。bounded context check 後も implementation 前に必要な追加 human decision は見つからなかった。
Evidence checked:
- Ticket
00001KVHR3WRFbody / thread / relations / artifacts。 TicketRelationQuery(00001KVHR3WRF): no outgoing blocking dependency; incoming lifecycle Ticket depends on this。TicketOrchestrationPlanQuery(00001KVHR3WRF): previous waiting capacity note resolved by00001KVHKWNQSclosure; accepted plan recorded now。- Workspace state:
- Orchestrator worktree clean at
a5df9e37。 - queued: MCP chain remains queued。
- inprogress: 0。
- visible Pods: self + peers only; spawned children 0。
- no matching MCP implementation branch/worktree。
- Orchestrator worktree clean at
IntentPacket:
Intent:
- Add typed Profile/config support for named local stdio MCP servers and the trust-policy diagnostics around that config。
- This Ticket is intentionally config-only: parse, validate, redact, and document; do not spawn processes or implement JSON-RPC lifecycle。
Binding decisions / invariants:
- No package/workspace presence auto-start。Config alone must not spawn an MCP process。
- Local stdio MCP servers are local executables running with user OS permissions; Yoi feature authority does not sandbox OS-level side effects。
- Keep MCP config/trust policy separate from Plugin permissions and
pod::featureauthority。 - Config must be explicit and named; no implicit discovery or auto-enablement。
- Command/args/cwd/env/secret-related configuration must validate fail-closed with bounded diagnostics。
- Secret/env values must be redacted in diagnostics/log/model-facing surfaces; do not persist plaintext secrets through errors/debug output。
- This Ticket must not implement initialize/capability negotiation, tool/resource/prompt registration, streamable HTTP/OAuth/remote MCP, or subprocess spawning。
Requirements / acceptance criteria:
- A Profile/config can declare a named local stdio MCP server。
- Invalid command/env/secret config is rejected with bounded diagnostic。
- Secrets are not emitted in plaintext diagnostics/log/model context。
- Config alone does not spawn a process。
- Docs explain that configured local MCP servers are not OS-sandboxed by Yoi feature authority。
- Tests cover valid config, invalid config, secret redaction, and no auto-start。
Escalate if:
- Existing Profile/config model cannot represent MCP server declarations without a broader config/profile schema decision。
- Secret reference representation would require changing the local secret-store policy rather than referencing existing secret/env abstractions。
- Tests require actual process spawning; that belongs to a later Ticket, not this one。
Next action:
- Record
queued -> inprogressand commit Ticket records before creating the implementation worktree and spawning Coder。
State changed
Ticket body/thread, relation metadata, resolved capacity blocker, Orchestrator worktree state, visible Pods, existing branch/worktree, and bounded MCP config/trust context were checked. There is no unresolved blocking dependency, no inprogress/capacity blocker, and no missing planning decision. Accepting this queued Ticket for implementation before worktree/Pod side effects.
Implementation report
Implementation orchestration start:
- Acceptance records were committed on Orchestrator branch:
b0225e48 ticket: accept mcp stdio config trust。 - Created implementation worktree
/home/hare/Projects/yoi/.worktree/00001KVHR3WRF-mcp-stdio-config-truston branchimpl/00001KVHR3WRF-mcp-stdio-config-trustatb0225e48。 - Spawned Coder Pod
yoi-coder-00001KVHR3WRFwith write scope limited to the implementation worktree and read-only root scope only for runtime workspace identity。 - Coder task explicitly forbids root/original workspace implementation, process spawning for this Ticket, merge, Ticket close, and cleanup。
Next action:
- Wait for Coder implementation report, then inspect branch diff/validation evidence and route to Reviewer。
Implementation report
Coder implementation report received from yoi-coder-00001KVHR3WRF.
Implementation commit:
e0680cce mcp: add stdio server config
Changed areas reported:
- Added typed MCP config schema in
crates/manifest:mcp.stdio_serverlist with explicitname,command,args,cwd, andenv。- explicit env policy with
inheritandset。 - env values support
literal,secret_ref, andenv_ref。 - literal env values are redacted in
Debug。
- Added validation for duplicate names, empty/NUL command and args, cwd policy/path, env variable names, secret refs, and bounded diagnostics without secret/env value leakage。
- Integrated MCP config through Profile resolution and child manifest inheritance without adding process spawning。
- Added focused tests for valid config, invalid command/duplicate names, invalid secret refs/redaction, nonexistent command accepted as no auto-start proof, and Lua Profile MCP config resolution。
- Updated docs in
docs/design/profiles-manifests-prompts.mdwith local stdio MCP config and trust boundary。 - Updated Nix cargo hash due manifest dependency graph change。
Coder validation reported:
cargo test -p manifest mcp --lib: passed。cargo check: passed。cargo fmt --all --check: passed。git diff --check: passed。nix build .#yoi --no-link: passed。cargo test -p manifest --lib: failed on pre-existing/unrelatedplugin::tests::template_includes_current_pdk_version, reportedly dueresources/plugin-templates/rust/Cargo.tomlPDK version0.2.0vsCARGO_PKG_VERSION0.1.0。
Known deferrals:
- No MCP subprocess spawning, initialize/capability negotiation, tool/resource/prompt registration, HTTP/OAuth, or JSON-RPC lifecycle added。
- Command executability intentionally not checked at config-parse time; only structural validation is performed。
Orchestrator evidence checked before review dispatch:
- Implementation worktree is clean。
- HEAD is
e0680cce。 - Diff from acceptance
b0225e48..HEADis one implementation commit touching 8 files, about 515 insertions / 10 deletions。 git diff --check b0225e48..HEADproduced no diagnostics。
Next action:
- Dispatch Reviewer for r1 review against Ticket requirements, with focus on config schema fit, validation/redaction boundaries, no process spawning/no auto-start, Profile/config layering, docs trust boundary, and whether the reported
manifest --libfailure is unrelated or must be fixed。
Plan
Review dispatch:
- Spawned Reviewer Pod
yoi-reviewer-00001KVHR3WRF-r1against implementation branchimpl/00001KVHR3WRF-mcp-stdio-config-trust。 - Review target commit:
e0680cce mcp: add stdio server config。 - Review baseline:
b0225e48。 - Reviewer task focuses on config-only boundary, no process spawning/no auto-start, secret/env redaction, fail-closed validation, Profile/config layering, trust-boundary docs, MCP vs Plugin/feature authority separation, and the reported
cargo test -p manifest --libfailure。 - Reviewer is instructed not to edit source, commit, merge, close the Ticket, or use TicketReview directly; it will report verdict/evidence back to Orchestrator。