Inspect an agent session as a debugger notebook

Traceability

• Spec ID: harness-studio-session-debugger
• Status: Implemented

Intent

Redesign the live Run surface in @qoder-ai/harness-studio as a session-oriented debugging workbench. The surface should make an Agent Session readable as semantic work stages while keeping retained ACP/AG-UI facts inspectable. It is a debugger for observed session state, not a time-based Replay player and not a claim that the current workspace can be restored.

The supplied ACP Debugger screenshot defines the IDE-scale information density, three-column composition, light visual system, and synchronized selection language. Harness Studio keeps its own product identity, navigation, local-run boundary, and evidence wording.

Acceptance Scenarios

• AC-1: The Run view opens as a full-screen workbench with a product/session header, semantic debug toolbar, Execution Tree, Session Notebook, State Inspector, and Timeline Minimap. It does not present Play, playback speed, or any other Replay transport.
• AC-2: The default recorded sample is named 优化 Replay UI and contains one Prompt, one Plan, an Explore group with nine retained calls, a workbench.js change, a failed test, a CSS fix, a passing test, and a final Response. Read/search/inspect calls are grouped rather than rendered as nine peer cards.
• AC-3: Selecting a tree node, notebook event, or timeline marker updates one shared evidence cursor and visibly synchronizes all three regions. Previous Stop, Continue, Next Stop, Step Into, Step Over, Step Out, and Previous State update that cursor according to their labels; Step Into expands the current Explore group.
• AC-4: Stop Conditions for Changes, Failures, Permissions, Tests, and Responses can be toggled. Continue and stop navigation skip events whose stop types are disabled without mutating the retained session or workspace.
• AC-5: The notebook is organized by Prompt, Plan Revision, Execution Group, File Diff, Validation, and Final Response. Exploration details remain collapsed by default; file changes show a Before/After diff; failed and passing validations show command, status, and retained duration.
• AC-6: State Inspector exposes Changes, Files, Artifacts, Tests, Terminal, Plan, Evidence, and Raw ACP tabs. Changes answers what changed since the prior stop. Evidence distinguishes Exact, Correlated, and Inferred relationships; Raw ACP includes direction, method, ids, and trace context for the selected event.
• AC-7: The state boundary explicitly names the current cursor and available checkpoint evidence. The recorded sample offers Previous State and View History only; it does not expose Restore or Fork from Here because no restorable Workspace or Runtime checkpoint exists. Live inspection labels a UI-only pause as Soft Pause rather than claiming the Agent is stopped.
• AC-8: Existing AG-UI run submission remains reachable from the redesigned Run view. Starting a run uses the existing endpoint and terminal/tool-call reducer; the recorded sample is clearly labeled and cannot be mistaken for live backend evidence.
• AC-9: At 1440 by 900 the three primary columns and timeline remain visible without document-level horizontal scrolling. At narrower widths the inspector and tree can be collapsed or stacked, controls remain reachable, and overflow stays inside dense code/trace regions.
• AC-10: Focused tests exercise the semantic cursor model and built-page interactions. Browser verification checks synchronized selection, stop toggles, Step Into, inspector tabs, diff rendering, live-run entry, responsive containment, console errors, and a screenshot comparison against the supplied reference.

Non-goals

• Building a generic ACP Gateway, recorder, JSON-RPC proxy, or persistent session store.
• Rerunning retained tool calls, restoring Git state, resuming a native Coding Agent session, or adding Workspace/Runtime checkpoint infrastructure.
• Replacing the existing Experiment Builder, comparison Workbench, Results surface, evidence contracts, or server endpoints.
• Claiming hidden intent, causality, token usage, cost, exact duration, or file state when the current protocol evidence does not provide it.
• Adding a new host adapter or changing the supported host set.

Plan and Tasks

1. Add a pure session-debugger model for the recorded sample, semantic stops, tree projection, cursor movement, and inspector facts.
2. Recompose the Run view into the debugger shell while preserving the existing AG-UI request and reducer path behind an explicit New live run action.
3. Add icon-library-backed controls and a compact light visual system in the existing bundled application stylesheet. Avoid custom SVG/CSS icon art.
4. Add focused model tests and extend the built-app Playwright flow for synchronized cursor behavior, tabs, diff, responsive layout, and the live run path.
5. Build, inspect at desktop and narrow viewports, compare against the reference, fix actionable visual gaps, and record browser/design evidence.

Test and Review Evidence

• AC-2 through AC-7: packages/harness-studio/test/session-debugger-model.test.ts passes four focused tests for the eight-stage sample, nine-call Explore group, enabled-stop traversal, hierarchy stepping, node resolution, and cumulative file state.
• AC-1/AC-3 through AC-10: npm run harness-studio:test:browser passes four built-app flows. The Session Debugger flow selects the shared cursor, enters and exits Explore, disables Changes, continues to the failed Test, opens Raw ACP and Diff view, checks the read-only state boundary, and verifies a 1440 by 900 shell with no document overflow or console/page errors. The existing 390 px live-run flow still completes and expands a retained Tool Call.
• Package gate: npm run harness-studio:test passes 10 files and 63 tests after the production build. npx vitest run test/skills-docs/doc-link-graph.test.mjs passes six checks after regenerating docs/better-harness-doc-links.mmd; the generated graph is unchanged. git diff --check passes.
• Visual gate: the same-state desktop and focused comparison boards are recorded in the ignored .qoder/design-qa/ workspace. The ignored design-qa.md records final result: passed after fixing the initial auto-scroll hierarchy. In-app Browser inspection also exercised Step Into, Stop Conditions, Raw ACP, the live AG-UI path, compact side-panel overlays, and reported no errors.
• Risk: debugger labels can overstate execution control. Copy names the Evidence Cursor and Soft Pause boundary, and does not claim a gate or hard runtime pause.
• Risk: mock data can be mistaken for product evidence. The default session is visibly marked Recorded sample; live runs receive their own state and ids.
• Risk: dense three-column layouts can become unreadable. Columns use bounded widths, sticky chrome, local overflow, and explicit responsive collapse.
• Risk: keeping the existing live run path can mix semantic sample data with live AG-UI data. The two modes stay explicit and never merge evidence.