Session performance analysis

Traceability

• Spec ID: session-performance-analysis
• Status: Implemented; local validation
• Request: add a Sessions performance subpage after examining Qoder slow sessions; implement analysis in Rust wherever practical.

Intent

Explain where a retained session spent time. Preserve the existing Sessions overview and add Performance beneath Sessions. Lead with long execution and waiting intervals, distinguishing model requests, tool dispatch, shell execution, hooks, approval, nested agents, and missing evidence.

Acceptance scenarios

• AC-1: Sessions exposes Overview and Performance as secondary navigation, with project-scoped deep links and the existing shared date filter. Selection, filtering, and the route survive back navigation.

• AC-2: The Rust evidence host owns Qoder execution-log discovery, event pairing, duration calculations, interval unions, and evidence-backed finding codes. Node owns transport/project binding; React owns presentation. Desktop reuses its existing evidence host. A host without the capability shows unavailable.

• AC-3: A Bash lifecycle of 349.182 seconds with a 108 ms shell interval shows both values, with dispatch wait and hooks distinct. It is never labelled as 349 seconds of command execution.

• AC-4: A 35,308,824 ms explicit approval duration is retained even if permission event timestamps differ by 5 ms. Reported duration and timestamp basis remain visible; approval does not become command execution.

• AC-5: Concurrent hooks, tools, and nested agents are not added twice into total occupied time. Category totals are labelled as overlapping; the timeline retains containing spans and evidence identities. Unpaired/ambiguous events leave explicit incomplete coverage and never fabricated pairings.

• AC-6: Model start/completion uses stable request ids where possible, with a unique turn/loop/request-index fallback for retries. Retry signals remain visible. No first-token/chunk timing means TTFT and streaming stay unrecorded.

• AC-7: Session wall span, completed-turn time, approval time, and unattributed intervals remain distinct. Late SessionEnd does not make a completed task appear to have executed overnight. Date filtering uses observed activity.

• AC-8: The page provides a sortable session list, per-turn timeline, category filters, bounded event rows, an evidence detail pane, and explicit loading, empty, unavailable, incomplete, and error/retry states in English and Chinese.

• AC-9: Verify wide 1440x900, compact 1024x768, and narrow 390x844 layouts, keyboard selection/focus, hover/selected states, back navigation, 200% reflow, reduced motion, dark/light themes, overflow, and browser console/page errors.

• AC-10: File reads and response sizes are bounded, symlinks/path traversal do not expand workspace scope, browser requests cannot select host-home paths, and project changes invalidate pending responses. Native path handling uses Path/PathBuf and shared Windows/UNC slug rules. Return only timing metadata and bounded redacted labels; prompts/commands/results are not raw evidence.

• AC-11: Include observed subagent count, individual durations, cumulative work, interval-union time, and peak concurrency. Link explicit fork lifecycles or Agent tool invocations to parent spans where the evidence permits; a nested turn is attached only with an unambiguous identity/containment match. Do not count the wrapper and its child twice or confuse accumulated agent time with elapsed parent waiting. Unlinked subagent evidence remains explicit.

• AC-12: Follow the requested macOS Storage composition: one segmented time bar and compact category rows, with timing notes and execution records collapsed. Every segment and row expands proportional category details and calls inline; the overview and sibling categories remain present. The bar represents combined category activity time with category-internal interval unions, and separately states observed elapsed time. Categories can run concurrently. Subagent rows show actual union time even when all work overlaps another category; deeper call proportions use labelled cumulative invocation time. Rust computes both bases. Validate against the actual better-harness Qoder corpus.

Non-goals

New Coding Agent adapters, changing existing session/commit analysis, upstream Qoder instrumentation, network/service internals diagnosis, automatic tuning, publishing, installing, or changing release metadata. Other agents have explicit timing coverage boundaries; this first detailed timing adapter is Qoder.

Plan and tasks

1. Add a bounded sessions.performance capability to the existing Rust evidence host, using independent performance modules rather than extending the legacy session summary schema or duplicating analysis in JavaScript.
2. Cover timing semantics with synthetic logs representing the inspected Qoder cases, plus malformed/missing/concurrent data and native path fixtures.
3. Add a typed Studio timing contract and project-bound read-only API, reuse the desktop host, and wire the development launcher to a built native service.
4. Add a secondary Sessions route and a docked performance workbench, reusing DESIGN.md tokens and the existing Phosphor icon set. Keep category labels and evidence accessible without hover; paginate long lists instead of mounting the full event corpus. Compact hides the trailing evidence pane until a row is selected; narrow navigates list -> session -> evidence with visible Back.
5. Build and exercise Rust, transport/API, routing and browser tests. Re-run the representative local Qoder corpus through the actual Rust binary, then review readiness against this spec and the changed-file boundary.

Test and review evidence

Local validation: Rust performance semantics (11 tests), Studio build/typecheck, focused API/transport/routing/i18n/project/scale tests (38 tests), and native-backed Playwright Storage overview/drilldown checks at 1440x900, 1024x768 and 390x844. Chinese/dark mode, keyboard entry and focus restoration, reduced motion and 200% zoom also exercised. Screenshots are in the ignored Studio test-results directory. The real Qoder corpus contains 141 sessions; prior model/tool/hook timings match the native output. Root preview health and Canvas module returned HTTP 200; doc-link graph tests passed (8 tests). No installed Desktop or non-macOS runtime claim is made. The implementation was authored with Codex in this task.

Review readiness: user request is the visible scope evidence; no Story id was provided. This spec covers the native analyzer, server binding and Storage UI. The commit scope excludes release/version metadata and unrelated customization work. Missing native support, timing coverage and concurrent work have explicit representations.

Risk: request ids can change on retry; loop identities can collide across nested agents; permission events may be emitted at completion; hooks may lack unique command ids; lifecycle completion can lag actual tool execution. These are coverage/basis facts, not grounds for guessing a more specific cause. Missing native binary or older native capability must remain visible. Local macOS execution and simulated Windows path tests are not Windows/Linux runtime proof.

The user authorized committing this feature after reviewing the live preview. The commit includes only this feature's implementation, spec, and tests.