Distillation

Date: Sep 16, 2026

• 🔴 [requested-research-only] (19:33) User directed work in /home/byk/Code/opencode-lore-responses-projection-fix to be research-only, with no edits, VCS operations, server use, network access, or database access.
• 🟡 (19:33) User asked to inspect the current Responses projection stack through ozrmvotw/583020fa plus its empty child, tracing raw provider events from principal and continuation parsing through validation, accumulation, deferred/transactional holding, remapping, client forwarding, and terminal reconstruction.
• 🟡 (19:33) User asked for the smallest exact implementation and a deterministic fail-first test matrix covering 3 defects: 1. unsupported indexed events or extra top-level/nested fields leaking through remapping; 2. no-index deferred principal extension frames leaking after recall; 3. high-cardinality quadratic scans in sparse candidate resolution and identity matching.
• 🟡 (19:33) User asked to distinguish what the separate private-reference candidate already fixes from what belongs in this change, without importing private-reference behavior.
• 🟡 (19:33) User requested exact current file:line/symbol evidence, existing relevant tests/helpers, expected failing behavior, a minimal safe allowlist/projection design, and an operation-counter test design.
• 🔴 (19:33) User prohibited writing code and running broad tests for this investigation.
• 🟡 (19:33) Repository setup attempt reported that skill repo-setup was unavailable; available skills were ast-grep, customize-opencode, jj-create-pr, jj-guide, jj-update-pr, and jj-workflow.
• 🔴 (19:34) Project uses Node.js >=22.5, TypeScript, and a pnpm workspace monorepo. Root package.json pins Volta Node 24.16.0, pnpm 10.28.0, and declares packageManager: pnpm@10.28.0.
• 🔴 (19:34) Tests use a temporary SQLite database through packages/core/test/setup.ts—never the production DB.
• 🔴 (19:34) Never manually bump versions or edit CHANGELOG.md; Sentry Craft handles version changes and changelog generation automatically.
• 🔴 [enforced-fail-first] (19:34) Never trust the fix without seeing it fail; regression tests must fail on the base branch and pass with the fix.
• 🔴 (19:34) Every non-trivial PR requires an adversarial correctness review before merge, following quality/REVIEW.md; review discipline includes adversarial-order state setup, fan-out registry coverage, recurring bug-class batteries, and the two-reviewer rule.
• 🟡 (19:35) Initial source search identified packages/gateway/src/stream/openai-responses.ts as the central Responses projection/accumulation implementation, including applyResponsesEvent() at line 582, indexed event handling, sparse-index validation, deferred processing, identity tracking, and terminal reconstruction.
• 🟡 (19:35) Search also identified Responses-related tests in packages/gateway/test/openai-responses.test.ts, packages/gateway/test/openai-responses-stream.test.ts, packages/gateway/test/openai-responses-recall-aware-stream.test.ts, packages/gateway/test/foreground-body-limit.test.ts, packages/gateway/test/pipeline-streaming.test.ts, and packages/gateway/test/llm-adapter.test.ts.
• 🟡 (19:36) Existing relevant test evidence includes packages/gateway/test/foreground-body-limit.test.ts:157, test "buffered Codex accepts sparse events without output_index", and packages/gateway/test/openai-responses.test.ts:843, test "drops item_reference items without breaking surrounding messages".
• 🟡 (19:38) packages/gateway/src/stream/openai-responses.ts:48-93 defines ResponsesAccState, including terminal snapshot state, rawItems: Map<number, Record<string, unknown>>, strict identity maps itemIndexById, callIndexById, and effectiveToolIndexById, active/unbound/done item sets, and accumulated items indexed by output_index.
• 🟡 (19:38) packages/gateway/src/stream/openai-responses.ts:95-127 defines SUPPORTED_RESPONSES_OUTPUT_ITEM_TYPES with: message, function_call, function_call_output, reasoning, item_reference, web_search_call, file_search_call, computer_call, computer_call_output, computer_tool_call, computer_tool_call_output, code_interpreter_call, image_generation_call, local_shell_call, local_shell_call_output, shell_call, shell_call_output, mcp_call, mcp_list_tools, mcp_approval_request, mcp_approval_response, custom_tool_call, custom_tool_call_output, apply_patch_call, apply_patch_call_output, program, program_output, tool_search_call, tool_search_output, additional_tools, and compaction.
• 🟡 (19:38) packages/gateway/src/stream/openai-responses.ts:128-130 builds RESPONSES_OUTPUT_ITEM_TYPES as a Set<string> from the supported-type list; isSupportedResponsesOutputItemType() at lines 238-242 accepts only string values present in that set.
• 🟡 (19:38) packages/gateway/src/stream/openai-responses.ts:131-251 defines phase-specific status validation through OUTPUT_ITEM_STATUSES_BY_TYPE and isValidResponsesOutputItemStatus(), with phases "added", "done", and "terminal".
• 🟡 (19:38) packages/gateway/src/stream/openai-responses.ts:254-260 begins recordExtends(actual, streamed, ignored = new Set()), which compares streamed record entries against an actual record while allowing ignored fields; this is relevant to nested-field identity/projection behavior.