Factory docs, home
Page navigation

Accepted locally. Two tools expose the commissioned receiver without transport or model authority. Current verification and historical evidence belong to build review. Installed contract · human guide · receiver.

Intelligence: none — Two bounded tools retain original Guidance identity and current receiver authority; no model decides permission, acceptance or replacement.

Sources: provider-guidance-tools-contract.ts, provider-guidance-tools.ts. Public entry software-factory/provider-guidance-tools. The neutral contract has no runtime dependency; the facade imports canonical JSON, actual GuidanceAcceptance and shared native Promise observation. Its runtime dependencies stay within the existing receiver closure. ES2024 public declarations need no Node ambient types.

new ProviderGuidanceTools({receiverOptions}) captures the exact public receiver commission and constructs actual GuidanceAcceptance on the supplied actual Journal, ownership reader, registry, clock and qualified current authority fence. Construction and tools() are passive. The handler takes the exact original generation identity, plus authenticated actor/Product context outside arguments. No values, authority, revised command, expected revision, allowance or deadline comes from model arguments.

call(name, arguments, context) handles factory_inspect_guidance_offer and factory_apply_guidance_offer; other names are not-handled. Context/arguments/busy failures precede receiver invocation. Once invoked, malformed output or an exception is an unobserved receiver result: apply remains possibly-recorded; inspect is a read-only invocation. Complete closed receiver Outcomes retain authorised historical witnesses even on refusal. Tool result text is the same canonical JSON as structuredContent; isError never grants a retry with a new command. No replacement or release exists.

The caller owns base MCP negotiation, strict UTF-8/duplicate-member byte parsing, authentication, current connection credentials, correlation, cancellation and response-channel lifetime. Input _meta, progress tokens, task execution and generic dispatch are unsupported. Inspect is read-only; apply can write Guidance and advertises destructiveHint. Both idempotent hints mean the same original domain command, not an admission cache. They have openWorldHint false and taskSupport forbidden. No SDK, listener, model or network exists here.

Bounds: one active call and commission; arguments 1,024 bytes, context 512, commission 2,048, depth 24, structured result 4,096, complete handler 20,480 UTF-8 bytes. Oversize refuses, never truncates. Own descriptor data, including symbol data, is examined within a bounded traversal; native Promise rejections are observed but never awaited as permission. Getters and generic thenables are never invoked. Exceptions use fixed results without caller messages. The qualified receiver's native store wait is synchronous and cannot be cancelled by a request timeout.

Checks: provider-guidance-tools.test.ts, the installed consumer and exact dependency graph cover original identities, current permission, replay, corruption, possible-commit outcomes and strict native rejection handling. Disposable hosts establish local boundaries; live qualification remains separate. Build review owns results.

Native Promise fulfilment data uses the same observation depth, seen set and remaining node allowance, so nested rejected data is observed without renewing the traversal budget. Later settlement faults are contained; synchronous traversal faults reach the owning refusal classifier. Fulfilment is never awaited as permission, cannot change a returned refusal or possible-effect classification, and cannot trigger another receiver call. Strict tests distinguish direct public input/options boundaries from instrumented malformed result controls.

Shared value implementation: common contracts. Owning errors, limits, stored identities and authority remain as specified above.

Source: docs/agents/provider-guidance-tools.md