fix: preserve distinct MCP structured results

[JuliaF1988]

Summary

This PR is narrowly scoped to preserving distinct MCP result data and refining exact duplicate suppression. It follows the circuit-breaker policy already merged in #109245.

Current head ef82a0b6181d83ad5912a5aeb45f6631c38c046f is rebased onto upstream/main at 61e730cc0b7594eeb8e92fd8a56e4259ba87cfe6.

Current scope

• Preserve distinct MCP content and structuredContent instead of discarding authoritative machine-readable data whenever usable text exists.
• Collapse only exact JSON dual-emissions, so identical data is emitted once.
• Compare JSON recursively and type-stably: 1 != true, 0 != false, with the same rule inside nested objects and arrays. False negatives retain duplicate context; false positives would lose data.
• Preserve an isError=true result as the established top-level human-readable error plus complete structuredContent and permitted/non-reserved _meta.

Breaker policy

This PR follows merged #109245 and does not change the general isError=true breaker policy:

• isError=true remains an application-error circuit-breaker strike.
• Repeated application/tool errors may open the breaker.
• A breaker opened by an application-only streak uses upstream rejected/reachable wording.
• Transport failures retain upstream unreachable semantics; one transport strike in a mixed streak selects that wording.
• Successful non-error calls reset breaker state according to upstream behavior.

Result projection and breaker accounting are separate concerns: preserving structured error data does not make the result a breaker success.

Post-RPC output-schema validation

MCP SDK output-schema validation is handled separately from ordinary isError=true results. MCP SDK 2.0.0 still validates after send_request() completes and still emits the two specifically matched RuntimeError signatures for missing or invalid structured output. Only those narrow post-response cases reset the transport breaker; arbitrary RuntimeError and genuine transport/session/timeout failures remain strikes.

Final rebase reconciliation

The final maintenance rebase onto 61e730cc0b7594eeb8e92fd8a56e4259ba87cfe6 was conflict-free. Intervening upstream changes compose without duplicating or superseding this PR:

• MCP write-capable tools are not auto-retried after mid-flight session expiry (no duplicate side effects) (salvage #88821) #88821 adds at-most-once behavior for write-capable MCP calls after mid-flight session expiry. Its read-only classification, reconnect handling, outcome-uncertain result, and breaker behavior remain intact around this PR result projection.
• Failed turns, CLI errors, gateway and cron notices explain themselves and name the fix (message audit, core) #111600 contributes adjacent user-facing wording only; its current upstream text is preserved.
• Earlier interlocks refactor(agent-core): one Retry-After parser, reset-grammar table, token estimator, head/tail truncator and NO_PROXY matcher; LLM traffic honours CIDR/wildcard NO_PROXY #109539 and Multiplexed profiles keep their own sessions, MCP identities and served-profile status (post-merge reviews on #108294/#108352; salvage #109428 #109434 #109483 #110167 #109930) #110271 remain composed as documented.

No original #105116 hunk was dropped as upstream-equivalent. The final diff remains limited to five intended files and three commits.

Validation

Validated at exact head ef82a0b6181d83ad5912a5aeb45f6631c38c046f:

• Focused breaker, projection, real-SDK, and MCP write-capable tools are not auto-retried after mid-flight session expiry (no duplicate side effects) (salvage #88821) #88821 session-expiry tests: 54 passed.
• Complete current MCP-named test scope, including all tests/tools/test_mcp*.py, refresh/setup/reconciliation paths, and all new mainline MCP regressions: 1,254 passed, 2 skipped, 0 failed across 115 files.
• Real in-memory MCP SDK path, rerun independently: 1 passed.
• Ruff on every changed Python file: PASS.
• git diff --check: PASS.
• Repository contributor attribution audit against the current upstream base: PASS.

Hosted CI, Docker, and Nix workflows for this exact head are action_required, pending upstream maintainer approval. They were not manually triggered or bypassed.

Provenance and interlocks

• MCP tool results stop paying twice for dual-emit servers (kimi-code#3234 port) #96807 is the credited carrier of the existing content / structuredContent arbitration and fix(mcp): send content or structuredContent to the model, never both MoonshotAI/kimi-code#3234 duplicate-context work. This PR only narrows suppression to exact JSON equality so distinct structured data survives.
• fix(mcp): breaker opened by tool errors tells the model the calls were rejected, not that the server is unreachable (#11113) #109245 owns the current application-error strike policy and the rejected/reachable versus unreachable breaker wording; this PR preserves that implementation.
• fix(mcp): degrade outputSchema validation failures instead of dropping results #101344 is complementary ownership for completed-RPC output-schema validation behavior.
• MCP isError errors with nested JSON bodies are double-encoded, hiding actionable messages from the model #47867 / fix(mcp): extract readable message from nested JSON isError bodies #47871 own adjacent isError human-readable error extraction; this PR preserves the machine-readable structured payload without replacing that work.
• fix(mcp): recognize nested application failures #74262 is complementary downstream classification for application failures in non-error envelopes.

This branch should now be left untouched pending upstream maintainer action. Further rebases are warranted only on maintainer request, actionable hosted-CI failure, a confirmed merge conflict at merge time, or direct upstream supersession/conflict.

[andrexibiza]

Reviewed exact head dd0efb83639193617647a1de39f160b903feb50e against live main 478d772f2cc6421fda1d115137b8da92cd4f7d1a. The branch is currently 1 ahead / 71 behind live main. I traced the result renderer, breaker accounting/recovery path, the two changed MCP suites, exact-head Actions state, and the related MCP result/ownership lineage.

The central breaker correction is sound in principle: a completed tools/call that returns isError=true proves transport reachability and should not poison a server-level outage breaker. The explicit transport-exception negative control is good, and the new native-ID / nested structured-result / hard-cap cases are useful coverage. There are two data-contract blockers before this is safe to land; both are inline.

Verification state: the PR reports 672/672 local MCP tests plus Ruff/diff checks, but the exact head currently has CI, Docker, and Nix all completed as action_required, so this one-commit train is 0/1 hosted-green. Also, the new regressions exercise mocked sessions. The repo's network-I/O contract calls for a real-path behavioral proof; this does not require the external Defender deployment. A local/in-memory MCP SDK peer through the real handler path is enough to prove: (1) an isError=true round trip leaves the breaker closed, (2) its machine-readable error payload survives projection, (3) distinct text + structuredContent survive while exact dual-emits dedupe once, and (4) a genuine transport exception still strikes the breaker. Please rebase and rerun that proof on live main as well; tools/mcp_tool.py has moved since this branch's merge base.

Interlocks / provenance:

• Merged #96807 is the credited carrier of the current content/structuredContent arbitration and the Kimi #3234 dedup port. This PR is a deliberate refinement of that policy for lossy summaries, not a new origin for the dedup shape; retain #96807's credit and its no-double-context/drop-notice goals.
• #101344 is complementary work on the same defect-class principle: an SDK output-schema failure after a completed RPC is not evidence that the server transport is down.
• #47867 / #47871 own the adjacent isError nested-error readability problem. This PR should compose with that work by retaining machine-readable error data, not replace its attribution or independently fork message extraction.
• #74262 is complementary downstream classification for application failures inside isError=false envelopes; transport health and application outcome remain separate axes.
• #104534 is complementary profile-qualified MCP connection ownership and also touches mcp_tool_handlers.py in lookup/check paths. Its hunks are not duplicate ownership of this renderer change, but whichever lands second needs a fresh rebase and the combined MCP suite.

This is good, focused work on a real MCP boundary. Once the two projection gaps below are closed and the exact-head train is actually green, the breaker/result split will be materially stronger. 🚀

[JuliaF1988]

Rebased on current main (c8aa5608c) at head 621f7b21ff. Both prior landing blockers remain fixed; the current head passes 715 MCP tests including the real in-memory SDK path, Ruff, diff-check, and an exact-head Docker build. Ready for re-review.

[teknium1]

Retriggering CI (zero-job dispatch failure on the rebased head) — reopening immediately.

[teknium1]

Upstream corroboration: MoonshotAI/kimi-code#3654 (merged Sep 8) lands the same refinement of their #3234 absolute-suppression rule this PR makes to our #96807 port — suppress structuredContent only when a content text block parses to the deeply-equal JSON value, with conservative numeric comparison so noncanonical number spellings keep both copies. Their independent convergence on exact-equality-only dedupe (after shipping and reverting semantic-compare and size-heuristic variants) is good evidence this is the right arbitration policy. CI was stuck on a zero-job dispatch for the rebased head; retriggered and runs are now executing.

[teknium1]

Heads-up: the breaker-policy part of this cluster got a ruling in #109245 — tool isError payloads stay breaker strikes (#10447/#109180); only the open-breaker message changes. The other scope in this PR is unaffected by that ruling and still under review.

[JuliaF1988]

Reconciled the PR metadata and exact head f60c2e16b3 with merged #109245: isError=true remains a breaker strike, including application-only streaks that open with upstream's rejected/reachable wording. The remaining PR scope is structured-result projection plus exact, recursive type-stable JSON deduplication. Current MCP validation is 901/901 green across 77 files; hosted CI, Docker, and Nix still await upstream maintainer approval.