ARCH: перезапустить план на MCP bounded autonomy и добавить metadata pilot

2026-04-21 22:04:23 +03:00
parent bda7ca9cc1
commit 561b4ea45c
15 changed files with 1214 additions and 57 deletions
@@ -1345,6 +1345,22 @@ Module progress:

 - Big Block 5 MCP Semantic Data Agent: `100%`.

+## Reset Hand-Off - 2026-04-21
+
+The progress updates above closed the first guarded MCP discovery pilot wave.
+
+They do **not** mean the strategic autonomy target is complete.
+
+From 2026-04-21 onward the mainline continues in:
+
+- [15 - mcp_bounded_autonomy_reset_plan_2026-04-21.md](/x:/1C/NDC_1C/docs/ARCH/11%20-%20architecture_turnaround/15%20-%20mcp_bounded_autonomy_reset_plan_2026-04-21.md:1)
+
+That reset freezes the continuity/authority stabilization as sufficient and returns the project to the primary trajectory:
+
+- metadata-first self-navigation;
+- entity/schema grounding;
+- planner-selected MCP primitive chains instead of route-per-question hardcoding.
+
 ## Execution Rule

 Do not implement this plan as:
@@ -0,0 +1,201 @@
+# 15 - MCP Bounded Autonomy Reset Plan (2026-04-21)
+
+## Purpose
+
+This note resets the execution focus after the stabilization wave inside turnaround `11`.
+
+It does not cancel the continuity and authority work already done.
+
+It clarifies that the main project trajectory is not:
+
+- endless polishing of deterministic route arbitration;
+- route-per-question hardcoding;
+- a fake "agentic" mode that is actually another brittle prompt wrapper.
+
+The real trajectory is:
+
+- bounded assistant autonomy over reviewed MCP primitives;
+- proof-first discovery of 1C evidence;
+- gradual reduction of route hardcoding at the question level.
+
+## Why The Reset Is Necessary
+
+The previous wave repaired real defects:
+
+- stale carryover stopped beating the current question on critical contours;
+- guarded MCP discovery answers stopped being overwritten by stale exact/lifecycle paths;
+- broad business evaluation no longer breaks the return into the data contour.
+
+That work stays valid.
+
+But it was support work around the edge of the main target.
+
+The strategic target was always bigger:
+
+- the assistant should not need one deterministic route per business wording;
+- the assistant should be able to orient itself inside 1C through a reviewed set of MCP primitives;
+- the planner should decide which safe primitive chain to use for the current data need;
+- the evidence gate should decide what may be stated to the user.
+
+So the reset is not "we were wrong".
+
+It is:
+
+- stabilization is now frozen as sufficient;
+- the mainline returns to `MCP-first bounded autonomy`.
+
+## What Is Frozen
+
+The following is now baseline, not the mainline:
+
+- current-turn meaning authority;
+- continuity subordinated to current explicit meaning;
+- guarded discovery response replacement;
+- broad-evaluation bridge that does not destroy the next data follow-up.
+
+These seams may still receive bug fixes if they block MCP-first execution.
+
+They are not the main feature track anymore.
+
+## North Star
+
+The target is not an unrestricted free agent.
+
+The target is a bounded planner over a reviewed primitive catalog:
+
+1. recognize the business data need;
+2. pick allowed MCP primitives;
+3. execute bounded probes against 1C;
+4. aggregate evidence;
+5. answer only within the evidence gate.
+
+In short:
+
+- move determinism from `route per user wording`
+- to `catalog of safe primitives + proof workflow`.
+
+## Big Block A. Metadata-First Self-Navigation
+
+### Goal
+
+Teach the assistant to inspect the 1C schema surface before guessing a route.
+
+This is the first real step from pilot hardcoding toward self-navigation.
+
+### Scope
+
+- live execution of `inspect_1c_metadata`;
+- metadata-aware planner path that cannot collapse into `query_documents`;
+- machine-readable metadata evidence:
+  - available object sets;
+  - matching objects;
+  - available fields/sections when metadata returns them;
+  - known limitations;
+- human-safe answer draft for metadata discovery.
+
+### Why This Block Comes First
+
+Without metadata-first inspection, every later autonomy step is blind:
+
+- entity resolution is guesswork;
+- register/document choice is guesswork;
+- long-tail discovery turns back into hidden route hardcoding.
+
+### Acceptance
+
+- raw metadata wording can bootstrap discovery input;
+- planner keeps `inspect_1c_metadata` as the chosen primitive;
+- pilot executes live metadata inspection through MCP;
+- user-facing answer stays free of primitive/query/runtime garbage.
+
+## Big Block B. Entity And Schema Grounding
+
+### Goal
+
+Move from "I found some metadata" to "I can ground the user ask onto the right 1C surface".
+
+### Scope
+
+- search and resolve candidate entities through MCP instead of local tails only;
+- bind metadata findings to probable document/register families;
+- preserve ambiguity honestly when multiple surfaces compete;
+- keep machine-readable grounding evidence for downstream probes.
+
+### Acceptance
+
+- assistant can say which schema surface it selected and why;
+- ambiguity is surfaced as a bounded clarification, not silent route drift;
+- chosen surface becomes reusable context for the next primitive.
+
+## Big Block C. Planner-Selected Primitive Chains
+
+### Goal
+
+Replace one-off pilot scopes with planner-selected safe chains.
+
+### Scope
+
+- chain primitives such as:
+  - `inspect_1c_metadata`
+  - `search_business_entity`
+  - `resolve_entity_reference`
+  - `query_documents`
+  - `query_movements`
+  - `aggregate_by_axis`
+  - `probe_coverage`
+  - `explain_evidence_basis`
+- keep exact deterministic routes as fast-paths;
+- use discovery as the general path for understood long-tail questions.
+
+### Acceptance
+
+- new long-tail questions become answerable without adding a dedicated route for each wording;
+- the planner output explains the chosen primitive chain;
+- the answer gate still blocks overclaiming.
+
+## Stage 1 Started Now
+
+The first block is no longer just planned.
+
+It has started in code in this pass.
+
+Implemented in this stage:
+
+- raw metadata wording now bootstraps discovery input;
+- metadata planning stays on `inspect_1c_metadata` and no longer falls into `query_documents`;
+- the pilot executor now has a live metadata inspection slice;
+- the answer adapter can produce a user-safe metadata surface answer.
+
+This stage is intentionally narrow.
+
+It does **not** yet mean:
+
+- unrestricted Qwen3 navigation across arbitrary 1C contours;
+- automatic multi-step schema-to-entity-to-query chaining;
+- hot-runtime replacement of broad assistant behavior everywhere.
+
+It means the architecture now has the first real self-navigation primitive in production code.
+
+## Execution Rule
+
+From this point the project should prefer:
+
+- adding or strengthening reviewed MCP primitives;
+- planner-selected evidence workflows;
+- machine-readable grounding and proof contracts.
+
+And should avoid:
+
+- growing another layer of hidden route hardcoding for each new wording;
+- long stabilization detours unless they protect an MCP-first invariant;
+- fake autonomy that bypasses the evidence gate.
+
+## Bottom Line
+
+Turnaround `11` is no longer only about making the assistant feel less glitchy.
+
+The next move is larger:
+
+- make the assistant able to look into 1C through bounded MCP discovery,
+- choose its path through reviewed primitives,
+- and answer from proved evidence instead of memorized route scripts.