Normalize Qoder dialogue before Inspector renders it

Traceability

• Spec ID: inspector-qoder-dialogue-normalization
• Status: Implemented
• Extends the Qoder evidence path used by Harness Inspector.

Intent

Qoder transcript rows store assistant content as structured arrays containing thinking, text, and tool_use items. The provider currently serializes the whole assistant message object when it cannot find a top-level string. Session Detail consequently renders raw role/content/input JSON as an intermediate response, even though the same tool request is already normalized into the Tool calls lane.

Qoder must project structured message items into provider-neutral dialogue before Inspector receives them. Readers should see readable retained assistant text and structured tool calls, never transport JSON duplicated as prose.

Acceptance Scenarios

• AC-1: A Qoder assistant thinking item becomes readable intermediate text, without the surrounding role, content, or item-type JSON envelope.
• AC-2: Qoder text and output_text items retain their text in the normalized assistant event and Session Detail response flow.
• AC-3: A tool_use-only assistant row contributes its structured tool request event but no assistant prose, raw tool input, or serialized JSON note.
• AC-4: Mixed assistant content preserves textual item order while tool items remain exclusively in the structured tool lifecycle.
• AC-5: Malformed or unknown structured items do not make the provider throw and do not fall back to serializing the transport object.
• AC-6: Existing Qoder user prompts, tool request/result pairing, file attribution, redaction, and cross-provider dialogue behavior remain intact.

Non-goals

• Parsing Qoder transport JSON in Inspector UI code.
• Rendering raw tool inputs, tool results, credentials, or absolute paths as assistant prose.
• Changing Qoder's native session files or resuming a session.
• Reclassifying provider-neutral tool names or evidence confidence.

Plan and Tasks

1. Replace the Qoder provider's object-level JSON.stringify fallback with a structured message-text projector.
2. Recognize retained thinking, text, and output_text strings; ignore tool_use, tool_result, and unknown structured objects in dialogue text.
3. Add provider tests for thinking, text, tool-only, mixed, and unknown items.
4. Run Qoder session normalization, commit/session dialogue, Inspector, and package-boundary regression checks.

Test and Review Evidence

• AC-1..AC-5: focused tests in test/sessions/session-analysis.test.mjs inspect normalized events rather than matching provider source text.
• AC-3/AC-6: the existing Qoder transcript tool lifecycle test must still pair tool_use and tool_result with observed latency.
• Dialogue regression: test/sessions/commit-session-link.test.mjs and test/reporting/harness-inspector.test.mjs.
• Real-file smoke: render the workspace session shown in the reported defect and verify Session Detail contains no serialized assistant envelope.
• Real-file result: a 40-session Qoder report retained readable intermediate text for the reported session with zero raw assistant envelopes; the browser Session Detail showed the readable rows, no role/tool_use JSON, and no console warning or error.
• Full regression: npm run check passed 94 files / 1324 tests, followed by package verification of 521 npm entries and 543 runtime-zip entries.

Risks

• Content loss: an unrecognized future Qoder textual item could be omitted. The projector accepts conservative string/text fields while refusing object serialization, and fixtures pin known variants.
• Privacy: thinking text is retained only as the provider's existing bounded intermediate dialogue projection; tool input and result objects remain out of assistant prose and continue through existing redaction/tool seams.
• Duplicate evidence: the parent assistant event and embedded lifecycle events share a source row. Only the lifecycle events may carry tool facts.