[twarp 07] respec: port deleted Agent Mode UI instead of rebuilding

[timomak]

Why this re-spec exists

The first implementation attempt, PR #67, rebuilt the Claude Code panel entirely from GPUI primitives (Flex::column(), Container::new(...).with_background_color(...), appearance.ui_builder().span(...)/.link(...)) — zero lines from the deleted Agent Mode rendering layer, directly against the 7a port-and-adapt mandate. The result looks nothing like Warp's Agent Mode:

• plain-text "tool cards" with no per-tool affordances;
• a similar unified diff rendered as untinted plain spans (no +/- tinting, no hunk headers — none of feature 05's treatment, despite PRODUCT §33);
• a static Flex::column with no UniformList and no bottom-stick auto-scroll;
• plain-text assistant output instead of feature 03's markdown (§18);
• a WorkspaceAction::ClaudeCodePanel(...) dispatch workaround because direct typed-action dispatch "dropped silently" — symptom-treatment, not the fix.

This PR regresses feature 07 from impl-in-review back to spec-in-review and rewrites the plan as a per-file porting guide an implementer cannot misread. It changes only roadmap/ docs — no code.

Investigation (the main work)

I read every deleted UI file from fea2f7ea (the feature-02 spec commit, predates all deletions) and the kept crates/claude_code driver. Findings:

The original TECH.md table conflated two distinct deleted surfaces that were never composed together:

• app/src/ai_assistant/ — the simple Warp AI Q&A panel (prompt→answer transcript, markdown + code blocks; coupled to Requests/AIClient/ServerApi/InputSuggestions). No tool cards, thinking, todos, or diffs.
• app/src/ai/blocklist/ — Agent Mode (the tool-card / diff / thinking / todo surface that warp.dev/agents/claude-code shows), deeply coupled to AIBlock/AIBlockModel/AIAgentOutputMessageType/BlocklistAIActionModel.

Per-file decision matrix (full version + git show recovery commands in TECH.md):

Source @ fea2f7ea	Verdict	Why
inline_action/{inline_action_header,requested_action,requested_script,inline_action_icons}.rs	Port (clean leaves)	Zero crate::ai:: service coupling — take strings + Rc<dyn Fn> callbacks. The reusable card chrome.
ai_assistant/transcript.rs render_message + utils.rs markdown split	Port + reparent	Markdown prose + code blocks; swap Requests → Transcript::items().
markdown_parser::parse_markdown + FormattedTextElement	Reuse on master	AI-agnostic; feature 03's live stack (§18).
code_review_view.rs diff renderer / crate::code::inline_diff::InlineDiffView	Reuse (feature 05)	Themed read-only +/- diff (§30–§33).
block/view_impl/todos.rs	Port leaf + bridge	Reuse header+icons; swap AIConversation/AIAgentTodo → our TodoItem/TodoStatus.
block/view_impl/{output,common}.rs	Extract helpers	output::render is a 700-line match over AIAgentOutputMessageType; lift only FormattedTextElement + collapsible/format_elapsed_seconds.
inline_action/code_diff_view.rs (3190 ln)	Do NOT port	Interactive accept/reject editor; needs AIBlockModel+BlocklistAIActionModel+telemetry+history-revert; delegates real diffing to inline_diff. Wrong target — reuse feature 05.
inline_action/requested_command.rs (1646 ln)	Rewrite leaf	Stateful View bound to BlocklistAIActionModel/TerminalModel/CodeEditorView; rebuild from the clean chrome.
block/view_impl/header.rs	Do NOT port	Agent-block chrome tied to TerminalAction/DirectoryContext; our header is net-new.

Dispatch fix (the #67 bug): GlobalSearchView — a working left-panel tab — is itself a TypedActionView that dispatches its own actions directly. So #67's "actions drop unless routed through WorkspaceAction" was a focus-wiring defect, not inherent. The ported panel is a self-dispatching TypedActionView + an on_left_mouse_down focus-grab (the ai_assistant/panel.rs pattern); the WorkspaceAction::ClaudeCodePanel forwarder is deleted.

What survives PR #67 vs what is discarded

• Keep: the headless crates/claude_code driver crate (lib.rs Transcript/TranscriptEvent/TranscriptItem; driver.rs subprocess + defensive JSONL parser + SIGINT + stdin writer; sessions.rs) — 19 passing tests — and the registration scaffolding (ToolPanelView::ClaudeCode, toolbelt button, render arm, ⌘⌥K via custom_tag_to_keystroke, compute_left_panel_views push). Some pure data helpers (tool_input_summary, diff_for_tool) survive as bridge code.
• Discard: the entire primitive panel body + the WorkspaceAction::ClaudeCodePanel forwarder.

Doc changes in this PR

• TECH.md — postmortem + implementer stop-sign; corrected "what 02 deleted" table; per-file decision matrix; git show recovery commands; stub/co-port + TranscriptItem→leaf bridge layers; re-derived port-shaped 7b–7h (7b = "panel shell + ported transcript with a stub event source", visually testable end-to-end with no driver); "visually matches Agent Mode" acceptance gate; risks updated (top risk is now "re-rebuilding from primitives").
• PRODUCT.md — behavior contract unchanged except making visual consistency with Agent Mode an explicit acceptance gate (Figma note + new smoke step 28). The 62 invariants are otherwise intact.
• STATUS.md — regression narrative, [twarp 07b] Claude Code panel: resurrect view + event model #67 postmortem, cleared 7b–7h ticks (work shipped against the wrong approach), re-derived sub-phases, dispatch/two-surface decisions recorded.
• ROADMAP.md — row 07 spec column adds this re-spec PR; impl column cleared.

Relationship to PR #67

This re-spec supersedes #67's approach. #67 should be closed (leaving the kept crates/claude_code driver to be brought into the next impl branch). Per repo convention I'm not closing #67 — that's the owner's call.

Notes

• Spec-only; no code compiled/changed. Fork CI ending "cancelled" is expected (no matrix runners), not a failure.
• Implementation is a separate branch + PR in a future chat — this PR is the plan to review first.

🤖 Generated with Claude Code