feat: add planner dashboard TUI (/planner-dashboard)#21
Merged
Conversation
Full-screen, theme-aware, adaptive dashboard for the planner: - Segmented stage progress ribbon with per-stage theme colors and in-segment labels, filling left-to-right as stages complete - Marquee ticker (route, active task, branch) scrollable with ←→ - Scrollable task list with status icons and selection follow - Detail panel with current step, branch, and per-stage timings derived from the timer checkpoints - Live refresh from disk while open; adapts to terminal resize; compact single-column layout on narrow terminals - Pure model + renderers in dashboard-model.ts (unit tested with an identity palette); interactive shell in dashboard.ts - Footer/widget hint pointing at /planner-dashboard - Add @earendil-works/pi-tui dependency for TUI primitives Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Reserve a blank cell on each side of each stage label and render the label with a readable treatment (bold/text/muted) instead of the stage fill color, so "INIT", "DISCOVERY", etc. stay legible against the ▰/░ progress fill. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Rebuild /planner-dashboard as a full-screen workspace (Option 3): - Top band: stage ribbon + status + marquee ticker - Middle: live model conversation rendered from session entries (projectSessionEntries + renderTranscript), planner tool calls and results collapsed to one line by default, expandable - Bottom: our own input line that sends to the model via pi.sendUserMessage, plus focus tabs (input / chat / tasks) - Tab cycles panes; tasks pane expands the task list + stage timings; chat pane scrolls and toggles expand-all; live refresh while open - Pure transcript module (chat-view.ts) unit tested with identity palette Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Auto-open the workspace once per session for planner-worktree sessions via a session_start hook (covers create/resume/improve without touching the fragile session-switch handoff). Reopen with /planner-dashboard; close to fall back to the plain chat. - Remove /planner-preview (the workspace replaces it). - README: document /planner-dashboard, drop /planner-preview. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fixes the footer overlap and mouse-wheel-into-scrollback issues, and makes the TUI configurable: - Render the workspace as a fixed top overlay (overlay: true) instead of an inline scrollback component, so the mouse wheel no longer drags it off-screen and the native footer stays visible below it. - Size the overlay to terminal rows minus a configurable footer reserve. - Add a `workspace` settings group: enabled, autoOpen, footerReserveRows (schema, defaults, manager merge/normalize, /planner-helper docs). - openPlannerWorkspace loads settings; auto-open respects autoOpen and the master enabled switch; footer reserve is tunable if the footer overlaps or leaves a gap. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Stream assistant output live: subscribe to message_update/message_end and append the in-flight assistant message to the transcript, so tokens appear as they are generated instead of only after the turn commits. - Replace the glyph-fill stage ribbon with a clean bracket style (INIT › INTAKE › [DISCOVERY] › …): active stage bracketed + bold in its theme colour, completed stages coloured, pending dimmed. Adapts to width. - Redraw only when something visible changed (content, clock second, coarse marquee step, focus/input) instead of every tick — removes the choppy repaint and steadies the ticker. - Document the workspace keys, settings, and Pi keybindings.json in the /planner-helper report and SETTINGS.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
In the discovery-first /planner-improve flow, discovery runs before the goal is drafted. But planner_goal_decide(approve) always routed to discovery/scan_project_structure (the create-flow assumption), so improve plans re-ran the whole discovery stage after approval (re-scan, "questions already submitted", second compact). Now approval routes creationMethod=improve plans straight to planning/read_context, and the state machine allows intake/await_goal_approval → planning/read_context for improve. Create plans are unchanged (goal first, then discovery). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…auses - Sanitize transcript text (expand tabs, strip ANSI/control bytes) so the expand (x) view no longer breaks the frame on tool output that contains tabs or escape codes. - Dashboard clock + stage timings now show live elapsed time (activeMs plus time since last disk sync) instead of the rarely-synced persisted value, so they tick in real time. Timer sync interval is threaded from settings. - Restore a smooth one-cell-per-tick marquee when the ticker overflows; keep the calm render-on-change behaviour when it fits. - Timer pauses only while waiting on the user (goal approval, discovery questions, result acceptance, user decisions). It now counts intake drafting and compaction for honest active timing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…row ribbon - Streaming now redraws per message_update event (token by token) instead of on the poll tick, so the chat fills in smoothly like Pi's own chat. - Clock + stage timings recompute live in-memory each draw (applyLiveTiming) and disk reload drops to ~3s, so the terminal shows full-time elapsed without per-second disk reads (persisted timer still syncs every 10 min). - Inherit Pi keybindings: app.thinking.toggle (Ctrl+T) hides/shows thinking rows; app.tools.expand (Ctrl+O) expands/collapses tool output. - Stage ribbon falls back to a compact [ACTIVE] n/total indicator when the terminal is too narrow for the full ribbon, so the active stage is always visible. - Detail-panel note now word-wraps and ellipsises instead of hard-clipping. - Docs: workspace keys + inherited bindings in /planner-helper and SETTINGS.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Header clock now uses the live-recomputed model instead of the value from the last disk reload, so it ticks every second (was stepping in ~3s reload increments). - Chat pane: End jumps back to the live tail (newest output), Home to the top — so after scrolling up you can return to the streaming view. - Input accepts pasted text (bracketed paste handled, ANSI/control bytes stripped, newlines folded to spaces). Image paste is not supported in the workspace (Pi's image paste targets its built-in editor); documented. - Docs: paste + End/Home keys in /planner-helper and SETTINGS.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- The transcript now anchors to an absolute top line while scrolled up, so newly streamed/committed output appends below without yanking the view to the bottom. End re-follows the live tail; Home jumps to the top. - Project only a trailing window of session entries (HISTORY_WINDOW=400) and cache the projection, rebuilding only when the windowed slice changes instead of every 180ms tick. Scrolling to the top loads the next older chunk, so long sessions never project the whole conversation at once and the per-tick CPU/alloc cost stays bounded. - renderTranscript switches from scroll-from-bottom to atBottom + absolute topLine and returns the resolved top line for the caller to track. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
At a disabled compact boundary (task compaction is off by default), compact_task wants finish_step, not request_compact. The guidance never named the target step, so the model called finish_step with the current step as the target and hit invalid_next_step, then thrashed fail/retry/start before recovering. - Lifecycle: the disabled-compact and generic finish_step messages now state the exact target via getAllowedNextPlannerPositions, and the disabled-compact message explicitly says not to call request_compact. - Orchestrator: blocked-tool reasons now include the lifecycle guidance (with the target), so the model sees the right next step on the block. - status: compact_task rule explains both paths (enabled -> request/complete compact; disabled -> finish_step to execution/select_next_task). - Test: state machine advances compact_task -> select_next_task and rejects finishing into the same step. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- The auto-scrolling ticker forced a full repaint every ~180ms, which is what drove the dashboard to ~40% CPU. Replace it with a static, clipped context line (active task / branch / blocking note) — no per-tick repaint. Idle now redraws ~once/second (clock only); when nothing changes the tick is a cheap no-op. - Throttle streaming-driven redraws to ~12fps so a fast token stream cannot drive an unbounded repaint rate. - Make the workspace keys configurable via settings workspace.keys (focusNext/up/down/pageUp/pageDown/jumpBottom/jumpTop/expand/submit/exit), matched through matchesKey; Ctrl+C always exits. Pi's keybindings.json only accepts Pi action ids, so our keys live in planner settings. - Docs updated. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When final checks revealed a needed code edit (e.g. a biome import-order
error in a new test file), the model was stuck: run_final_tests cannot edit
project files, and fail_step/retry_step only re-run the same step. There was
no path back to an editing step, so the model thrashed.
- State machine: execution/run_final_tests now branches to
{capture_skill (forward), implement_task (back to fix)}.
- status: run_final_tests rule explains to finish_step into implement_task
for fixes (not fail_step), and that this step cannot edit project files.
- Test: run_final_tests allows the implement_task branch.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.