Memory composer and source context

Traceability

• Spec ID: memory-composer-context
• Status: Implemented
• Request: remove duplicated More settings, merge the two footer rows into the Memory input and show the file context used by the analysis.

Intent

Give Memory one bounded input surface and make the supplied source visible in its conversation without claiming an Agent performed a filesystem read.

Acceptance Scenarios

• AC-1: Model/mode/reasoning controls appear once. More contains only additional options and disappears when there are none; settings failures remain retryable.
• AC-2: State, Agent label, keyboard hint and Close session share a row inside the input. Stop, queueing, draft preservation and independent Compare inputs work.
• AC-3: After the first request starts, the transcript shows the frozen source title/path and source line information with a return-to-source action. It is labelled supplied context, not a successful Agent tool call. Changing the selected Memory entry never relabels the source of a running conversation.
• AC-4: Explicit tool file locations are visible when tools are collapsed and remain linked to their own status. No filesystem reads or inferred paths.
• AC-5: Wide, compact, narrow, light/dark, keyboard, overflow and error checks pass in the browser, followed by a Studio build and local Desktop refresh if idle.

Non-goals

Changing Memory access, snapshot limits, ACP permissions, source content, workspace indexing, publishing, packaging or committing.

Plan and Tasks

1. Remove core settings from More and retain extra-option search.
2. Let the shared composer own its caption and caller-specific session controls.
3. Add supplied-source disclosure to the transcript; expose declared tool paths.
4. Extend settings and Memory browser tests, and preserve the Compare work.

Test and Review Evidence

• AC-1: a 22-model fixture with only core options renders one model control and no More popup. The extra-option fixture retains Fast mode, retryable errors and keyboard dismissal without duplicating Model.
• AC-2: Memory tests verify the status, Agent name and Close session inside the input, plus Stop, subsequent turns, retained drafts and no external footer row. Compact Memory exposes keyboard instructions on the input rather than a second footer row. Shared conversation/queue/IME/attachment tests continue to pass.
• AC-3: Memory fixtures inspect the actual sent prompt and the frozen transcript source. Switching the reader leaves the source tied to General Tips; Open source selects that original tab. The source record identifies supplied context.
• AC-4: one path projection now feeds collapsed tool titles, Compare and mentions. The prior CSS hid every tool summary; only ordinary argument summaries remain hidden. A browser assertion verifies the file path has a visible width.
• AC-5: light/dark screenshots reviewed at wide/compact/narrow sizes, with visible focus, bounded overflow and no page/console errors. The 10 Memory scenarios and the affected conversation/stream/Compare suites passed. Studio build, 30 focused unit tests and eight document-link tests passed; independent Desktop smoke passed with native NSXPC and a sandboxed renderer.

Review Readiness Check

The request and the opened source are confirmed scope evidence. The existing server memoryAnalysisPrompt supplies the frozen line-numbered snapshot; this change exposes that source without introducing a file read or claiming an Agent read it. User controls, errors and capability checks remain. Tests cover the More duplication and original-context navigation. AI: Codex implementation. No commit, package or installation was made; the existing desktop Debugger conversation was preserved while a separate temporary Desktop smoke ran.