feat(ui): launch an agent with selected terminal text as context

[forketyfork]

Issue

There was no way to hand terminal output to an AI agent as context without manually copying and pasting it. Building this also surfaced problems in the adjacent, pre-existing "send diff comments to agent" feature, which shares its pending-delivery mechanism with the new flow:

• sendCommentsToAgent compared a newline-terminated PTY input string against the bare agent name, so AgentKind.fromString never matched — "send to agent" was broken for claude/codex/gemini.
• The pending-send delivery mechanism dropped a queued send entirely once its deadline passed without a confirmed agent match, instead of guaranteeing delivery like the previous implementation did.
• The selection-agent context preview silently truncated at 4 lines while the full, unbounded selection was sent to the agent, with no indication anything was cut.
• A leftover diagnostic log.warn dump fired on every modal open.
• The new pending-send state machine and the relative --log-dir resolution path had no test coverage.
• Minor UX gaps: a whitespace-only selection silently no-oped instead of giving feedback; the agent dropdown didn't restore prompt focus after closing; texture-creation errors were collapsed into an opaque error.LabelTextureFailed.

Solution

• Releasing a text selection in a terminal shows a bottom-right "Launch agent" pill.
• Clicking it opens a centered modal (selection_agent_overlay.zig) to pick an agent (Claude/Codex/Gemini), review the selected context, and write instructions.
• Launching queues a LaunchAgentWithContext action; the runtime spawns a new terminal in the source session's working directory and defers sending the composed prompt until the target agent is confirmed running via drainPendingSessionSends, a generalized pending-send state machine that also now backs the pre-existing diff-comments-to-agent flow (documented as ADR-015).
• Diff-comment agent matching now uses the bare agent name; the newline is appended only for the PTY write.
• drainPendingSessionSends now sends anyway once the deadline passes instead of dropping the queued text.
• The context preview shows a "+N more lines not shown" indicator when it truncates.
• Removed the leftover diagnostic logging.
• Added test coverage for drainPendingSessionSends (missing/dead session, deadline-based delivery) and for relative --log-dir resolution.
• Whitespace-only selection now clears visibly instead of silently no-opping; the agent dropdown restores prompt focus after closing (keyboard and mouse); texture-creation errors now propagate the real error.

Context

Documented as ADR-015 in docs/ARCHITECTURE.md.

Test plan

• Select terminal text, click "Launch agent", pick each of Claude/Codex/Gemini, and confirm the new session spawns in the correct cwd with the composed prompt delivered once the agent starts.
• With a selection longer than 4 non-blank lines, confirm the preview shows a "+N more lines not shown" indicator.
• Open the diff overlay (⌘D), write comments, and send them to a running claude/codex/gemini session — confirm delivery (previously broken for all three).
• Open the agent dropdown in the selection-agent modal, pick an item via keyboard and via mouse, and confirm the prompt field regains focus (blinking caret) in both cases.
• Select only whitespace and click "Launch agent" — confirm the pill/selection clears instead of doing nothing visible.

[Copilot]

Pull request overview

Adds a selection-to-agent workflow integrated through the UI action pipeline and runtime session management.

Changes:

• Adds a selection pill and agent-launch modal.
• Generalizes deferred prompt delivery and repairs diff-comment delivery.
• Adds documentation, tests, logging-path handling, and rendering fixes.

Validation: Not run during this review.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
src/ui/types.zig	Adds selection-agent actions.
src/ui/session_view_state.zig	Tracks selection-pill state.
src/ui/mod.zig	Exports the new overlay.
src/ui/components/session_interaction.zig	Implements pill interaction and positioning.
src/ui/components/selection_agent_overlay.zig	Implements the agent-launch modal.
src/ui/components/glyph_badge.zig	Supports custom badge colors.
src/ui/components/diff_overlay.zig	Repairs agent command delivery and clipping.
src/main.zig	Registers overlay tests.
src/logging.zig	Resolves relative log directories.
src/c.zig	Exposes SDL clip-state API.
src/app/runtime.zig	Spawns agents and manages deferred sends.
README.md	Documents selection-to-agent behavior.
docs/ARCHITECTURE.md	Adds component documentation and ADR-015.

Suppressed comments (2)

src/ui/components/selection_agent_overlay.zig:262

• This default branch consumes SDL lifecycle events while the modal is open. UiRoot.handleEvent causes the runtime to skip its event switch for every consumed event, so quit/close, resize, focus, and window-destroy events never reach their handlers. Consume the remaining input events explicitly, but return false for unrelated events as the other overlays do.

            else => return true,

src/ui/components/selection_agent_overlay.zig:214

• SDL_EVENT_TEXT_INPUT is still inserted while the agent dropdown owns focus. Pressing a printable key with the dropdown open first hits the dropdown's key-down branch, then the corresponding text-input event silently modifies the prompt even though prompt_focused is false. Ignore text input until prompt focus is restored.

            c.SDL_EVENT_TEXT_INPUT => {
                _ = self.prompt.insert(self.allocator, std.mem.span(event.text.text), host.now_ms);
                return true;
            },

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.