[wchwawa]

TLDR

The live Workbench gate exercised the 18 tools through one nokv mcp process and retained mcp-transcript.jsonl. Every tool call is now its own nokv workbench <tool> process graded against the native CLI contract, the tool definitions come from nokv schema, and the pre-#423 ledger admits the resulting cli-transcript.

Problem

scripts/workbench/live_workbench.py is the required nokv-workspace live gate for the 18-tool workflow and the C06 root-authority probe, yet it reached the product through the deprecated sidecar: one long-lived MCP session for the workflow, two more for the peer root and its reconnect, and a failed MCP initialize as the wrong-AgentId probe. Its evidence role was mcp-transcript, which the digest-pinned ledger required from the live-workbench producer, so the workflow could not be qualified over the primary surface without a ledger-policy change.

Merge after #587, which moved the object-namespace gate and introduced the shared contract grader; this PR targets main and carries #587's commits until #587 merges, after which its diff is its own two commits. It is the second of three gate migrations and links no issue: it is a test-harness chore with no product behavior change.

Consequences

The sidecar stayed load-bearing in CI, Gate 0 evidence described a deprecated transport, and the acceptance document had to declare the native 18-operation transcript unsatisfiable as specified.

Example

Before, the gate wrote a JSON-RPC tools/call line to a nokv mcp child and read structuredContent back; a replayed create-only put was an isError result whose content said AlreadyExists.

After, the gate runs nokv ... workbench workbench_put_file '{...}'. The replayed create-only put exits 4 with one stderr line {"status":"error","code":"AlreadyExists",...,"class":"precondition_failed","next_action":"reread_then_retry"}, and a read under a wrong --agent-id exits 2 with AgentBindingMismatch, invalid_input, fix_input before any Workbench request is sent.

Solution

Each ToolStep runs through nokv workbench <tool> and is graded with the shared grader: a success is one JSON result without internal keys; an expected failure names its Agent code plus the contract class and next action, and must arrive as one stderr line with the exit code of that class. The plan's two AlreadyExists steps and the probe's NotFound read carry those expectations. The 18 tool definitions and contract.json come from nokv schema, checked against the frozen Rust-owned schema by the existing validator. Owner readiness is observed as operation status of an unknown id answering not_found rather than a connection failure, the probe the CLI contract gate already uses.

The C06 probe keeps its reads and writes across the two roots. The reconnect read is a fresh process that re-admits the peer root from the control store; the wrong-AgentId probe is a graded AgentBindingMismatch refusal whose message must not disclose either AgentId. cli-transcript.jsonl replaces mcp-transcript.jsonl. The typed producer path accepts one transcript role per producer, cli-transcript for the CLI and mcp-transcript for the historical LingTai gap; the ledger's live-workbench producer now requires cli-transcript, and the ledger policy digest is re-pinned.

Gain

Two of the three sidecar-bound gates now run on the primary surface. Gate 0's 18-operation transcript is a native CLI transcript, and the acceptance document no longer carries the "unsatisfiable as specified" caveat.

Tradeoffs

Three expectations are restated at the CLI level: the peer session's tools/list check was an MCP handshake artifact and becomes one nokv schema validation; "rejected before MCP initialize" becomes "refused at root admission as AgentBindingMismatch"; a reconnect is a fresh process rather than a second stdio session. The ledger policy digest changes, which is the deliberate policy change the acceptance document called out. The restore-composition gate still launches the sidecar, so nokv mcp cannot be deleted yet.

Validation

• python3 scripts/workbench/live_workbench_test.py (25 tests, OK); typed_live_qualification_test.py (5, OK); qualification_receipt_test.py, qualification_invocation_manifest_test.py, pre423_contract_ledger_test.py, qualification_aggregate_test.py, source_bound_producer_test.py (OK); python3 scripts/workbench/pre423_contract_ledger.py PASS on the re-pinned digest; the whole workbench-contract job sequence (py_compile list, every *_test.py, bash -n start_rustfs.sh, the unittest set, qualification_invocation_check PASS) run locally.
• Live run on this Mac with etcd 3.7.1 and the pinned RustFS image, using the rust.yml step's arguments: workbench_workflow.status == "PASS", acceptance_gates["0"].status == "NOT QUALIFIED", 50 CLI transcript entries, contract.json with 18 tools from nokv schema. The failures arrived exactly as expected: two AlreadyExists exit 4 precondition_failed/reread_then_retry, one NotFound exit 3 not_found/do_not_retry, one AgentBindingMismatch exit 2 invalid_input/fix_input.
• Not run locally: the GitHub nokv-workspace job and the Phase 1 receipt step, whose live-workbench invocation reports NOT QUALIFIED before executing and now publishes a cli-transcript.jsonl gap stub; that path is covered by the receipt and typed unit tests.