Skip to content

PHASE_4.md

Latest commit

 

History

History
125 lines (98 loc) · 5.61 KB

PHASE_4.md

File metadata and controls

125 lines (98 loc) · 5.61 KB
Raw

Phase 4 — plan

Work deferred from Phases 2–3, plus items discovered while mining pingdotgg/t3code's ClaudeAdapter.ts. Read alongside NOTES.md (which captures spike-era limits) and the design handoff under design_handoff_conclave/.

Phase 3 follow-up (shipped)

These were carry-overs from Phase 3 that have since been implemented. Listed here so the original scoping is still discoverable.

Transcript persistence — shipped

  • messages table is written to on every append (SessionManager.PersistMessage).
  • Schema: (id, session_id, role, content, tools_json, created_at, seq, claude_uuid) (migration v3, extended in v9).
  • Loaded lazily on session activate via LoadTranscriptIfNeeded.

Diff stats — shipped

  • WorktreeService.ComputeDiff parses --numstat + --name-status.
  • SessionManager.RefreshDiff is invoked after each ResultEvent.
  • Cached in DB columns (diff_files/add/del) so the sidebar still has numbers when the worktree is unreachable.

Cancel button — shipped

  • SessionVm.CancellationSource set by ClaudeService for the duration of a turn.
  • ShellVm.CancelActiveTurn cancels the linked CTS, which kills the child process.

Per-assistant-message UUIDs — shipped

  • messages.claude_uuid column (migration v9), populated from the AssistantEvent's uuid field. Used by fork-at-message to map our row back to claude's JSONL chain.

Token / cost display — shipped

  • Session.TotalCostUsd accumulated from each ResultEvent.TotalCostUsd. Surfaced on SessionVm.TotalCostFormatted in the right panel.
  • Per-event usage breakdown is not surfaced (input vs cache_read vs cache_create); could still be added if the difference becomes interesting.

Partial-message streaming — shipped

  • --include-partial-messages is passed unconditionally; ClaudeService.HandleStreamDelta appends text_deltas to the live TranscriptMessageVm.
  • Tool-use input deltas are still ignored; we wait for the final assistant event for those.

Tool-result meta: per-tool parsers — partial

  • READ: line count, GLOB/GREP: match count, EDIT/WRITE: collapses success preamble to "ok". BASH: still falls back to first-line trim — exit-code + duration parsing is still TODO.

Phase 4 — shipped

Plan view — shipped

  • SessionVm.Plan is an ObservableCollection<PlanItemVm>. TodoWrite invocations are parsed by SessionManager.ParsePlanJson and pushed via SessionVm.ReplacePlan.
  • Persisted via sessions.plan_json (migration v4).
  • Header / progress bar / row states render per PlanView.axaml.

Logs view — shipped

  • SessionVm.Logs is a 500-entry ring buffer of LogLineVm ((ts, level, message)).
  • Lifecycle, informational stream events, and errors are routed through ClaudeService.Log. Not persisted (intentional).

Permission handling — option 1 shipped

  • Per-session permission_mode (migration v5) passed to claude via --permission-mode. UI selector: Prompt / Auto-accept edits / Full access.
  • Still open: structured AskUserQuestion and ExitPlanMode flows. They render as ordinary tool calls today and the model will wait indefinitely for a response we can't give. Either (a) implement an MCP-based canUseTool callback, or (b) embed Node + Agent SDK. Both deferred. Sessions show a warning log line when these tools fire so the user knows why claude is hanging.

PR card — shipped

  • GhService.TryGetPullRequest runs gh pr view; refreshed on session load and after each turn. Hidden when gh isn't installed/authenticated.
  • Session.PrNumber/PrState/PrMergedAt cached (migrations v2 + v7) so the card paints immediately on load.

Empty state — shipped

  • EmptyState.axaml renders when !HasActiveSession. Right panel auto-collapses via MainWindow.ApplyResponsiveLayout.

Crosscutting nice-to-haves

Shipped

  • Filter pills filter (ShellVm.ApplyFilter).
  • Project rename (SessionManager.RenameProject + sidebar context menu).
  • Auto-expand project on filter/search match.
  • Claude version gate (ClaudeCapabilities.Detect + AtLeast).
  • Auto-cleanup of merged sessions (AutoCleanupService, settings v8).

Still open

  • ⌘⏎ submit in the new-session modal — visual hint exists, key handler doesn't.
  • Auto-expand project on new session — currently the sidebar shows the new session under its project but the chevron stays collapsed if the user had it collapsed.
  • Capability probe — t3code's trick of starting a query, reading init, and aborting. Useful for an About dialog and to disable unavailable models in the selector.

Architectural open question (deferred indefinitely)

Subprocess spawn vs Agent SDK. t3code gets a lot for free from the Node SDK: multi-turn streaming via a single long-lived query() fed by an async iterable, permission callbacks, per-message lifecycle events. Our per-turn CLI-spawn approach works but costs a process-start per turn and gives us less granular control.

Three paths:

  1. Stay with CLI (current). Replicate features manually. Simplest packaging; no Node runtime dependency. What we're doing now.
  2. MCP-based augmentation: CLI base + local MCP server for permission prompts and structured I/O. Medium lift. Would unblock the AskUserQuestion / ExitPlanMode flows above.
  3. Embed Node + use Agent SDK: heaviest; introduces a second runtime in the bundle. Cleanest feature parity.

Revisit this choice when either (a) we hit a CLI limitation we can't work around (e.g. need true per-tool permission callbacks and MCP doesn't cut it), or (b) platform bundling via a Node sidecar proves cheap on all three target OSes.