Explorer for GitHub Copilot
Proof of concept, v0.1.70. Explore an idea through a branching question map inside the GitHub Copilot desktop app. Start with /explore, answer the questions that matter, research options when you are unsure, and keep going as deep as you need. Questions expand into workspaces inside the installed map. History previews saved revisions and named checkpoints; Undo previews the latest committed change before you restore it.
Usage dashboard compares current maps, saved answer states and structure, newly observed canvas visits/committed question growth, and retained owned-helper outcomes and usage. UTC date filters apply to observations and receipts, not current stock. Scans are bounded and explicitly refreshed; missing usage and incomplete coverage stay visible. It preserves your editor and drafts. See metric definitions, inventory, privacy and retention.
Commands opens a keyboard-accessible palette for existing map, history, export and review actions, plus current-journey or saved-library node search. Use Arrow Up/Down, Enter and Escape; unavailable actions explain why. Search covers saved questions, answers, categories and journey topics. Private answer draft text requires opt-in on each opening; research, reminder edits, files and history are excluded. Current results and library pages mount at most 200 rows. Saved-library pages retain the local search limits (16 files, 32 MiB, discovery capped at 2,000 directory entries); coverage and continuation are shown. Refresh after files change. Existing draft, preview, request and source checks still apply when opening a result.
The optional Ctrl+Shift+P shortcut is off by default and lasts only for this canvas. It works outside text fields only after Explorer knows a different configured voice shortcut, and ignores IME, repeats and other dialogs. Copilot or the OS may reserve it; the Commands button always remains available. Opening the palette cancels active dictation while retaining inserted draft text. Palette navigation does not save answers or start model work; helper entries open their existing review screens.
Save a local-calendar review date and/or manual revisit condition on a question. Filter reminders in the question list or Library, including archived journeys, and use Revisit to return to a source. Acknowledge, reschedule/reactivate or clear with History while keeping answer certainty and ancestor review separate. Conditions are evaluated by you; no monitoring, notifications or calendar writes run. See reminder semantics and export coverage.
Reviewed implementation handoff
Choose Implementation handoff, explicitly select saved decisions or open questions, and choose Brief, Specification, Plan, Prompt or Tasks. Optionally include recorded research sources and structured claims; these remain unverified AI interpretation with their original provenance and limitations. Preview the exact input before generating, then edit and explicitly accept the proposal before copying or exporting its Markdown. Your map and answer drafts stay intact.
Global guidance and all user constraints remain included. Unselected nodes, implicit ancestors, project-file excerpts/references/snapshots, drafts, private references and earlier briefs/handoffs/checks are excluded. Research summaries, recommendations, proposed answers and follow-up prose are also excluded. The handoff never executes code, creates external tasks or sends to another chat. See handoff contracts and limits.
Recovery
Open Recovery to retry a saved read and incomplete history finalization, inspect a validated saved version, or open it as a separate recovery copy. Inspection does not change the active journey. Copies preserve the original file, history and original-journey drafts, including malformed bytes and draft tombstones. A damaged bound journey can reach Recovery before any map loads. The preview shows up to 100 nodes; the copy contains the whole validated version.
A failed save before atomic replacement keeps the earlier committed map. If replacement succeeded but history finalization failed, the newer saved content remains authoritative. Retry reads that disk state; it never replays model inference. After a safe read, retry an answer save through the ordinary editor. Research requires a fresh explicit Research action with the usual account, model and context checks.
Recovery shows only recorded helper lifecycle evidence. An exited recorded owner proves interrupted coordination, not that an orphan helper stopped. A live or reused PID alone is unknown; only a currently tracked operation in this process is shown as active. Dismiss hides a known interrupted/failed notice for this canvas and retains all receipts and journals. Legacy journal pointers are never followed. Export observations are limited to this canvas: generation failure, response interruption or response finish. None confirms that the browser saved a file; Retry export is explicit.
All six cooperative writer locks use validated owner records. Only a known local PID proven absent may be quarantined. Legacy, malformed, foreign-host, live/reused-PID, inaccessible-owner or abandoned-recovery-claim locks stay untouched. A timeout is not proof of death. Each lock retains at most one .abandoned record; another abandoned writer stays blocked pending storage review. No broad cleanup is performed. See recovery contracts for bounded discovery, evidence retention and remaining manual checks.
Install from GitHub
In the GitHub Copilot desktop app, open Customize, use the marketplace settings control, and add:
https://github.com/AndrewGodlewsky/Copilot-Explore
Choose the main branch if prompted. Find Explorer in the copilot-explorer marketplace and install it. The exact plugin identifier is explorer@copilot-explorer.
Or install through Copilot CLI:
copilot plugin marketplace add AndrewGodlewsky/Copilot-Explore
copilot plugin install explorer@copilot-explorer
Start a new session and use /explore <your idea>. The repository includes the complete runnable plugin; no build or npm install is needed for the GitHub installation.
If this marketplace is already registered from a local checkout, switch it to GitHub with the following commands, stopping if any command fails:
copilot plugin uninstall explorer@copilot-explorer
copilot plugin marketplace remove copilot-explorer
copilot plugin marketplace add AndrewGodlewsky/Copilot-Explore
copilot plugin install explorer@copilot-explorer
Exploration files are stored separately and remain intact. For later updates, use Update in Copilot's installed plugins or copilot plugin update explorer@copilot-explorer, then start a new session.
Install a development checkout
Requires GitHub Copilot desktop with canvas support and Node.js 22 or later for development commands. The plugin itself uses the SDK/runtime supplied by Copilot; there are no npm runtime dependencies or separate AI keys. The development tests use jsdom.
For reproducible synthetic performance measurements, run npm run benchmark; see the baseline guide for fixtures, output, and native-check limits.
From this repository:
npm run install:copilot
The installer checks the package, registers this directory as the copilot-explorer marketplace, installs explorer@copilot-explorer, and lists plugins. It uses copilot from PATH. If that command is unavailable, set $env:COPILOT_CLI_PATH to the absolute path of your inspected, installed Copilot executable, then rerun. It does not download a CLI.
Equivalent commands:
copilot plugin marketplace add (Get-Location).Path
copilot plugin install explorer@copilot-explorer
copilot plugin list --json
Keep the checkout in place: a local marketplace loads it live. Start a new Copilot desktop session after installation, then send:
/explore I want to create a neighborhood tool-sharing library
/explore without a topic opens the entry form. If skill discovery has not refreshed, start another session or ask: Open the installed Explorer canvas. Do not use /create-canvas, which would create another extension.
Local marketplace installs can show a generic native tab icon due to a host issue observed in Copilot 1.4.2. The canvas is still usable. Prefer the GitHub marketplace installation above for normal use; see integration notes.
Use Explorer
While creating a journey, open Add exploration guidance to choose a mode or enter an objective and success criteria. Understand maps concepts and unknowns; Decide compares choices and tradeoffs; Design explores outcomes and feasibility; Diagnose investigates symptoms and causes; Challenge tests assumptions and failure conditions. Understand is the default. Open Journey details above an existing map and use Edit guidance to change these settings. Mode guides future questions and isolated research; earlier answers, research and explanations stay intact. Objective and success criteria remain separate inputs. Cancel pending Copilot work before editing guidance, and dismiss an old question preview before accepting new suggestions.
File holds Library, import, recovery, exports and Finish. Tools holds decision helpers and the Usage dashboard. The Settings gear opens Voice, Map and motion, Keyboard, and Local data and diagnostics from any view. Journey details holds guidance, mode, progress and AI context controls. Within a question, the answer, compact Record control, Save & continue and More options stay up front. More options opens a full-width drawer below the map with Explore, Research, Answer details and Context tabs. Expand the drawer for more room. It remembers drafts and your place; Escape closes it without closing the question. See the card workspace design and experience checks.
-
Use
/exploreto open your library, or/explore <idea>to start a journey. Search saved topics, pin favorites, or show archived journeys. Open continues a journey in the current chat; New exploration starts another without replacing the saved map. Copilot creates questions in several directions around a new idea. -
Select a question in the map or searchable Outline. The node expands in its branch for writing; research and follow-up controls are in its screen-level drawer. Combine text, saved answer state, research available, needs review, answer drafts, importance and reminder filters. All selected filters must match the same node. Legacy nonempty answers are Tentative; missing importance is Unranked. Search matches question, saved answer and category, ignoring case and repeated whitespace. Outline shows an exact matching count and pages of at most 200 nodes; Reset clears its filters. Selecting a result reveals it through folds or focus and preserves your editor drafts. In Library, choose Saved questions and answers to search node content plus journey topics; use Archived journeys for archived-only scope. Private answer draft presence is separate from Include private answer draft text, which starts unchecked. Only that opt-in allows draft text matches or snippets. Active Outline includes local and conflicted answer/state drafts for existing questions; Library includes persisted answer/state drafts matched to the exact saved file identity. Removed-question drafts, reminder edits, missing-direction drafts, reminder conditions and research prose are outside search. Draft search never adds text to AI input, exports or diagnostics. Library node search reads a local page on request, with no remote index or background polling scan. Each page examines at most 16 files and reads at most 32 MiB of document data, returning at most 200 snippets. Next search page continues inside long journeys as needed; counts describe that page, not the whole library. Discovery considers at most the first 2,000 directory entries in filesystem order and prominently warns when more entries were not searched. Unreadable files and draft-store failures show incomplete coverage. Refresh restarts after file changes; changed continuation sources and stale result opens are rejected. Ordinary topic browsing uses its existing separate library path. Open More options > Context > Context & navigation for Focus branch in a question's workspace, including a leaf, to show its root and ancestor path with that subtree. Show whole map in the map heading returns to the wider view and keeps the selected question reachable. Collapse branch hides a node's descendants; Expand branch restores them. Folds within a focused subtree still work, and unrelated folds return when you leave focus. Folding an ancestor above the focused branch exits focus. Close question closes the answer editor instead. Outline searches the whole map; page through results or filter to find a question. Selecting one outside focus exits focus and reveals its ancestor path. Focus and folding change only this canvas view; saved content, drafts, research, history and exports remain complete.
-
Open More options > Research, then use Research options for alternatives, pros/cons, a recommendation, and uncertainty. A separate Copilot researcher works inside this node. Open each claim's evidence disclosure for supporting paraphrases, source links, recorded read time and limitations. These are selected model assertions, not an exhaustive verification of the comparison; reasoning and unknown support are labeled. Use recommendation, Use this option, or Add selected details appends suggestions to your editable draft. Undo addition restores the previous draft until you edit it. Nothing is saved as your answer until you choose Save & continue or explicitly save and explore deeper.
-
Write your answer. New text defaults to Tentative; change Answer state in More options > Answer details when needed. Save & continue records text and state together, then opens an existing eligible next question without asking a model. If none remains, it keeps you on your saved answer. Explore deeper, in the drawer’s Explore tab, explicitly saves and asks for follow-up questions on that branch. You can add a direction manually there.
-
Decision brief opens an explicit Generate decision brief action. A separate tool-free helper summarizes committed guidance, decisions, open questions and direction seeds from the root within the same 128 KiB context budget. Coverage shows omitted records; omitted branches have not been examined. Review the six sections, uncertainty and source links, edit statements if needed, then Accept reviewed brief or discard. User edits keep original links, which are not automatically reverified against new wording. Acceptance changes only the brief artifact, never answers or answer states. A changed captured source makes proposals stale and accepted briefs historical; refresh and regenerate explicitly. Use Export reviewed brief for a separate Markdown file. The complete native and Markdown record also retains the accepted brief, and history can restore an earlier one. Missing historical sources remain labeled unavailable. No drafts, private references, saved research, question rationale or prior brief enter synthesis; reasons for choices must come from user text or remain unknown.
-
Check answers explicitly examines a bounded journey summary of committed answers, open questions, direction seeds and guidance. It may suggest a possible tension, an unresolved choice or an untested assumption, with links to the saved sources and a way to resolve it. Source states show whether an answer was decided or tentative. Dismiss a finding locally or go to its source; no answer or answer state changes. The check shows omitted records, so no findings means only that no possible issues were found in the reviewed context. A change to any sent source makes its results stale and requires an explicit rerun. Results and dismissals stay in this canvas only; reload, journey replacement or a new check may clear them.
Challenge decisions is a separate deliberate action, independent of Challenge exploration mode. Select saved answer records in the context review, inspect exact input strings and omitted records, then choose Start reviewed challenge. At least one included nonempty saved answer is required; tentative, assumed and needs-review states retain their meaning. The owned helper has a 180-second operation allowance and zero tools, returning at most 12 assumptions, counterarguments, overlooked alternatives or evidence gaps. Findings are model hypotheses, with references to your inputs; no prior research is sent and no new evidence is consulted. Use original-answer navigation to inspect existing Claims and evidence separately. Closing the context preview preserves drafts and caret. Results share the transient answer-check slot and dismissal/source-freshness rules; a rejected start preserves the earlier request and its Cancel control. A new ordinary check still uses the unchanged consistency-check rules. Usage receipts are retained separately as Explorer Challenge Review, without a fabricated price estimate.
-
Export maps offers complete Markdown and
.explore.json, portable maps, and an offline read-only viewer. For the viewer, select exact nodes and optional saved answers/research, review the full included-content preview, then download one searchable HTML file. Finish exploration saves the record and opens these export choices. Use Library to reopen it from any canvas/session, or Open file to import an exported journey.
Answer states distinguish intent from evidence. Decided, Tentative and Assumption require text. Need evidence, Not relevant and Revisit later can have an optional note or no text. Clear the text yourself before saving Unanswered. State selection, dictation and research additions remain drafts until Save; changing state alone is recoverable. If another canvas changed the saved journey, use Refresh saved journey, then explicitly keep the draft or discard it in favor of the saved answer.
Clean My Answer uses only your current nonblank answer text (up to 20,000 JavaScript characters) in a separate, tool-free Copilot helper. It requests conservative punctuation, clarity and repetition edits while preserving meaning, uncertainty, names, numbers and voice. Review the original and proposal: a model can still change meaning. Apply updates the local draft; Reject keeps your text. Your answer state stays unchanged, and neither action submits an answer or generates questions. Normal Copilot usage applies.
Help with this question offers Rephrase, Give an example, and Explain terms beside an expanded question. The original stays visible. Examples are illustrative, not your answer. The separate tool-free model helper receives only the selected saved question, exploration topic and guidance, and complete nearest saved ancestor records that fit its 16,000-character context limit. It does not receive your unsaved draft, selected answer, other branches, research or past help. Help is transient: at most one result for each action stays beside the active question through ordinary refreshes, then disappears when its saved source changes, you change nodes or journeys, or reload. Cancel a running help request before starting Clean My Answer, and cancel running cleanup before starting help. Neither action edits your answer or state. Use Explore options when you need sourced or current information. Normal Copilot usage applies.
You can keep editing during cleanup. Text changes, including dictation, cancel older results; changing only the answer state preserves the proposal. Starting cleanup cancels active dictation. Undo cleanup restores the captured original only until another text edit or navigation. Same-question polling preserves review and undo, but switching questions/journeys, leaving the editor or reloading ends them. The ordinary local draft remains recoverable; cleanup adds no historical draft store.
The root is the idea and is excluded from question counts. Answered means nonempty saved text; resolved means Decided or Not relevant without a needs-review flag; deferred means Revisit later; unresolved means total questions minus resolved. These counts overlap. Changed ancestor text or state marks descendants for review. Older nonempty answers are Tentative, and empty answers are Unanswered; no old answer is promoted to Decided. See answer-state semantics.
The library shows pinned journeys first, then most recently opened journeys. Files saved before the library existed use their last edited date until you open them. Archive hides a journey from the default list; it never deletes the file. Importing an identical saved journey opens the existing one. A changed file with the same identity offers Open saved or Import as new copy, preserving both versions. Answer drafts save locally after about 500 ms of inactivity and flush before journey switches. Save or clear manual-question drafts and cancel pending Copilot work before switching. Back to current map returns to unfinished drafts.
Typed answers, accepted dictation, research additions, and undo are recovered independently for each journey and question. Saving draft, Draft saved locally, and Draft could not be saved describe local draft persistence; Answer saved or State saved means an explicit save, with the meaning you chose. If the question, committed answer or saved state changed, Explorer keeps the saved answer and offers Use recovered draft, Keep editor draft, or Discard draft. Drafts for removed questions remain available to copy or discard. Another canvas's newer draft is never overwritten silently.
A failed write keeps the editor text and blocks journey switching until retry succeeds or you explicitly Discard unsaved editor text; that option leaves any previously stored draft intact. Save file exports committed answers and states only; finishing requires resolving local drafts. Manual-question text is still held only in memory. Closing or hiding the page attempts a final flush, but abrupt termination can lose keystrokes within the debounce window or an unfinished write. Wait for Draft saved locally before closing. The draft store is limited to 8 MiB and 10,000 journey/question records, including retained clear markers. Malformed files are preserved and need manual restoration; there is no cloud backup. See draft storage and recovery.
Pan by dragging the background, a compact card, or empty space in an expanded question heading. Question text can be selected and copied. Middle-button dragging also works over the editor. Scroll over the map to zoom around the pointer, or use Ctrl/Command+wheel (including trackpad pinch) anywhere on the map. Normal scrolling inside an expanded node scrolls its contents. The +/? controls zoom up to 300%; click the percentage to return to 100%. Close question keeps your place and shows a saved-answer preview; Fit collapses the workspace and fits the currently visible nodes. Show whole map leaves branch focus without closing the editor. Center returns to the selected node at reading size. Keyboard users can Tab to nodes (including those off screen), navigate the outline, or press Escape to close the outline or collapse a question. With the map focused, arrow keys pan, +/? zoom, 0 resets zoom, and Home fits the currently visible map. Window resizing preserves your zoom and position. Ctrl/Command + Enter saves the answer without generating questions. Revising an answered node preserves its descendants and marks them for review.
Choose Outline view for a text hierarchy and the same selected-node editor: answer, set answer state, review importance, research, and fold branches without returning to the map. Map view restores the expanding-node canvas. This choice lasts for the open canvas; view changes do not add history. The editor stays in place, preserving drafts, disclosures and caret. A pointer click on the view toggle keeps editor focus and active dictation; moving keyboard focus outside the editor keeps the existing dictation cancellation boundary. Research and private proposals remain attached to their original source.
Outline shows up to 200 rows per page in depth-first order. Folds hide descendants in the hierarchy; Search and filters searches the full journey, including folded branches, and labels results with their original level and parent. Native Tab/Enter controls, Previous/Next matches and capped indentation keep deep maps navigable. Filters and page changes do not hide the selected editor. Show selected in outline clears filters, expands its ancestors and reveals its page; explicit source links do the same. Pure view toggles retain filters and map camera. The minimap remembers its visibility choice while suspended in Outline. Narrow panes stack the hierarchy above the editor with separate scrolling; native readability, IME and microphone behavior still require manual checks.
Generated questions first appear in a not saved preview inside the requesting node. Select a subset, optionally adjust wording, then use Accept selected to add them in one undoable change, or Dismiss all. Dismiss before switching journeys, restoring history or generating again. Empty maps retain Generate questions for an explicit retry. A saved source change makes the preview stale; dismiss, refresh and generate again. Draft answers are preserved and do not become generation context. Accepted questions keep their original wording/category and adjustment provenance. Why suggested? in the preview and Why this question? in an expanded question show the optional model explanation of its uncertainty and consequence for your goal. It describes the original suggested wording and context, even after you edit wording or change the objective or an earlier answer. Older questions have no invented explanation; this is neither a confidence score nor an importance rank. The added-question link and static outline highlight make new nodes discoverable while respecting focus, folds and reduced motion.
New generated questions may also carry Low, Medium or High importance and a concise Why? explanation tied to the goal, constraints and decision consequences. Importance describes potential impact, not confidence, answer certainty or completion. Map cards and Outline show text labels; use Importance in this list to combine a rank with text and reminder filters while keeping the full map available. Old/manual questions stay Unranked; ideas and direction seeds are not ranked. Assessments remain historical model reasoning for the original wording and context after edits or guidance/ancestor changes, without automatic reassessment. Native/Markdown, portable maps and selected offline viewer nodes retain recorded ranks, reasons and original wording; the viewer preview shows the exact included content.
Suggest next question offers one optional existing question using a local heuristic. Save an objective first. Current question wording must share a supported literal word with the saved objective, success criteria or user constraints. More distinct constraint matches come first, then objective/success-criteria matches, then historical High/Medium/Low/Unranked importance, then stable question ID. The explanation shows the actual matched words and available historical reasons and original wording. Old importance and rationale are not reassessed for changed guidance. This is not semantic analysis or a claim about the optimal next action. Unicode words are case-normalized; words shorter than three characters and common English stop words are excluded. There is no stemming or synonym matching, so short terms and other languages can miss useful questions.
The heuristic scans the complete saved map (at most 5,000 nodes) and complete objective, criteria and constraints (at most 16,000 characters). It uses saved unresolved question states; directions, ideas, decided/not-relevant questions, deferred questions and all needs-review sources are excluded. Reminder urgency is separate. Drafts, research, project files and private references are not consulted. Open question deliberately reveals its ancestors, exits unrelated focus and preserves independent answer, reminder and direction drafts. Dismiss does not open another suggestion: that unchanged question remains excluded even after another Suggest click. Dismissals stay in this canvas, one source record per current node; unrelated edits, reminders and polling do not reset them. Changes to that question's saved wording, answer/state, review flag or historical assessment/provenance, or to saved guidance, permit a fresh suggestion. Removed sources are pruned; switching journeys or reloading clears dismissals. Visible results refresh with saved snapshots, and removed/excluded sources lose their action. Nothing calls a model, saves an answer or generates a question. Missing goals, missing literal matches, dismissals, deferred/review sources and genuinely resolved maps have distinct explanations. Review the map or use Finish exploration deliberately when useful; more nodes are not proof of progress.
Previews are temporary canvas memory, excluded from exports, history and native map context. Closing or restarting the canvas discards them; polling preserves selections and wording edits. A renderer reload can recover the original batch while the canvas remains alive, but loses browser-only wording/selection edits. See question preview contracts.
Open an ordinary question and choose Pin this question as a reference to keep its saved question and answer near another editor. The optional References disclosure holds up to three sources, shows saved answer state and review flags, and offers Go to source or Remove reference. Missing sources after a history restore stay visibly unavailable until removed or restored. Pins are private per-journey view preferences, survive reopening, and do not alter answers, history, exports or AI context. If another canvas edits the pins, Refresh & retry references deliberately reloads the latest revision before applying your choice. See pinned reference storage and API.
Initial questions and Explore deeper are queued in the attached Copilot chat. Research uses a fresh SDK session with the same verified account and selected enabled Copilot model; normal Copilot usage applies. It receives your saved question path, answers and states with a dedicated research skill, and checks web search/fetch availability before inference. The main chat receives no research transcript. Status shows preparation, search, reading and comparison; Cancel stops the research helper. Question cancellation still only discards the result, since Explorer cannot stop the main chat turn.
Save & continue remains available while research runs. New independent answers survive its completion, including another canvas's committed edits. Changing an answer or state that was included in the research snapshot saves your newer input and discards stale research. Explore deeper, guidance edits and journey/history changes still require waiting or cancelling. A replacement research run waits for its previous helper to stop; unconfirmed shutdown blocks further research until release is confirmed, then you can explicitly retry. Conflicting research saved by another canvas is preserved.
Open Compare options inside saved research to compare its actual options in a horizontally scrollable table. Choose from Description, Pros, Cons and exact saved detail labels (at most 35 criteria). Duplicate detail values stay separate; missing cells say Unknown. Low, Normal and High are your personal priorities: they order rows without computing scores, choosing a winner or changing the saved recommendation. Populated cells are saved model judgments with unknown cell-level source support; the separate claim/source disclosures do not prove those values. No extra model request runs. Choices and selected details survive node navigation in this panel, then reset on journey switch, reload or observed research/source replacement. Hiding a criterion keeps its selected details. Add selected details to draft, Use recommendation and Use this option preserve your existing words and answer state; edit or undo the addition before explicitly saving.
Research offers Quick (60 seconds / 8 tool attempts), Standard (180 seconds / 24 attempts, default), and Deeper (300 seconds / 40 attempts). Time starts at owned execution and includes setup; denied attempts count. Shutdown and accounting may continue afterward. Choose balanced evidence or primary-source preference and optional recent-source guidance in the review; preferences do not guarantee source availability, freshness or accuracy. It reads public web pages, compares practical options and validates its returned data and source URLs. New research links bounded claims to sources that were successfully fetched during that request. Read timestamps come from tool activity for the requested URL, not the model; redirect destination and publication date remain unknown. Supporting paraphrases are model interpretations, not automatically verified truth. Legacy results have unknown claim support and retrieval time. Imported read metadata remains labeled imported, including in history and complete exports. See claim evidence limits and provenance. Use Ask about this research to clarify a recommendation or explore a hypothetical change through an explicitly reviewed isolated follow-up. Original research and saved answers remain unchanged. Up to four complete exchanges (64 KiB) stay in the active canvas; prior evidence keeps its original provenance. See follow-up bounds and retention. If pages cannot be consulted, the result is clearly labeled model reasoning. Missing account, model or search access produces an error without falling back to the main chat. Save relevant draft answers and states before research so they are included. The research snapshot retains complete records within 131,072 UTF-8 bytes for actual instructions plus prompt, including identities, omission counts and research preferences. It reports omitted records and refuses a target/guidance combination that cannot fit. Only the records actually sent become source dependencies.
Map detail changes with zoom. At a distance, cards emphasize short labels, status, importance and branch counts; moving closer reveals fuller questions and saved-answer excerpts. Extremely distant Fit uses a schematic overview: use Outline or zoom in to read individual nodes. Full question and answer text remains available to assistive technology. The expanded editor stays at reading size while you zoom, preserving draft text, caret, reminders and dictation. Zoom stays anchored at the pointer; use Center, select a source, or focus an editor control to bring it back into view. Fit/Home deliberately collapse the editor and retain its drafts. Detail thresholds have small hysteresis bands to avoid flicker and repeated layout at their boundaries. Zoom and detail are view state and do not change saved documents, history or exports.
Show minimap opens a compact overview below the canvas. It shows every node in the current visible layout, including offscreen cards, folded branch markers, branch focus, the selected node and a clipped viewport box. Click or drag to move the viewport, or focus the overview and use arrow keys. Home, Enter or Space centers that layout without changing zoom or selection; Escape hides the overview and returns focus to its toggle. Hide minimap reclaims the row. This visibility choice lasts for the canvas, including journey switches, and resets on reload. The row never overlays the editor; short panes can scroll to the controls. Overview navigation is instant and does not change documents, history or exports. Moving focus to its controls cancels active dictation and keeps already inserted draft text. Use Outline for individual questions and Center to return to the readable editor.
Settings > Map and motion switches ordinary wheel/trackpad input between pointer-anchored zoom (the default) and two-axis pan. Ctrl/Command+scroll always zooms; Shift+vertical scroll pans horizontally in Pan mode when the device supplies no horizontal delta. Pixel, line and page deltas are normalized. Sensitivity from 0.25 to 3 scales wheel input only; buttons and keyboard steps remain fixed. Scrolling inside an expanded question stays native unless Ctrl/Command is held. Motion follows the operating system by default; Reduce stops the recording spinner and meter transitions and uses instant scrolling, while Full allows them even when the OS requests reduced motion. Map transforms are already instant. Settings are stored in bounded navigation-settings.json in Explorer's data directory, separately from journeys, drafts, history, voice settings and Copilot settings. A corrupt or unreadable settings file is preserved; the map uses safe defaults until the file is repaired and settings are refreshed. Reset to defaults requires Save.
Local storage in Settings > Local data and diagnostics or the library/welcome screen shows bounded category sizes, duplicate hardlink lengths, declared retained history and incomplete inventory. Choose a keep-history preview target (5, 10, 25 or 50), then Preview cleanup and separately Remove listed history versions. Only captured old unnamed snapshots are eligible; journeys, drafts/tombstones, named checkpoints, undo/newest recovery, usage, locks, intents and unknown files stay protected. Protected history can exceed the target. Saving the preference never deletes anything. Removal is irreversible; exact-candidate retries report failures and unknown outcomes. The 15-minute preview/observed-outcome ledger lives only in this canvas; restart never resumes cleanup. See storage bounds, ownership and interruption semantics.
Voice dictation
Select Record beside Your answer, wait for Listening, then speak hands-free. Text appears in your local draft as the model decodes it. Stop ends capture and finishes buffered audio; Cancel stops remaining work and keeps text already inserted. Cancelling during loading captures no audio. You can also focus the answer and use your Copilot voice shortcut (for example Ctrl+Space): hold and release in push-to-talk mode, or press twice in toggle mode. A deliberate shortcut press can finish a button-started recording; releasing a shortcut key cannot stop one. Explorer reads your shortcut and recording mode from Copilot without changing either setting. Review the text before saving an answer.
The bar beneath the answer distinguishes three stages:
| Stage | What you see | What is happening |
|---|---|---|
| Loading model | Loading indicator and a reminder to wait | The helper loads the selected local model. The microphone has not started yet. |
| Listening | Moving audio bars, elapsed recording time and saved chunk count | Your microphone is recording while the local model decodes speech. Silence stays flat. |
| Transcribing | Finishing indicator and any stop reason | The microphone has stopped. Buffered audio and pending draft writes are finishing. |
One model load and continuous audio request serve each recording. Ordered text chunks arrive about once a second or after 512 characters, without splitting a Unicode surrogate pair. These are transport chunks, not complete words or finalized utterances: the installed model exposes decoded tokens but no speech-boundary markers. Repeated words and the model's spacing are preserved. The final aggregate is checked against those tokens and never inserted again.
Saving draft and Draft saved locally distinguish visible text from recoverable text. Explorer acknowledges a chunk only after the existing local draft store accepts it. A failed write keeps the visible text, stops further dictation and exposes Retry saving draft or the existing conflict recovery controls. Chunks never commit an answer or add answer history. Cancel, Escape, edits, caret moves, leaving the answer and voice controls, or switching questions stop remaining work and retain inserted text. The indicators respect reduced-motion preferences.
Recordings end at 10 minutes, or after 3 minutes without decoded text (including silence). Capture also stops explicitly if the model falls five seconds behind, a draft chunk is not acknowledged within five seconds, or text buffering reaches its limit. A four-second audio-delivery stall reports potentially incomplete audio; the recorder cannot detect every hardware drop. Finishing captured audio has a 45-second limit. The complete answer remains limited to 20,000 characters; an exceeding chunk is rejected whole. Review retained text and record the remainder after a failure. See the technical bounds.
Microphone and model settings
Open the Settings gear, then Voice. No question needs to be open:
- Microphone: choose the system default or a named input device. Explorer checks the name again when recording starts, so changing device order does not select a different microphone. A disconnected or duplicate device name produces an explanation instead of silently switching inputs.
- Downloaded voice model: keep Follow Copilot or choose a supported downloaded model. The list reads actual local model folders, not the online catalog. Unsupported speech models are visible but disabled. Download additional models through Copilot, then use Refresh devices and models.
- Save voice settings applies your choices to subsequent recordings and future Explorer sessions. These preferences belong to Explorer; they do not change Copilot's microphone or model selection. Opening Settings cancels active dictation before changing preferences. Unsaved choices survive switching questions and closing Settings.
This Windows x64 preview supports the inspected Nemotron English, Nemotron Spanish, and Nemotron 3.5 Multilingual CPU formats with installed Foundry SDK 2.2.1, PvRecorder 1.2.9, and matching desktop libraries. English is the model verified on this machine. Spanish and multilingual compatibility is based on source inspection; their weights were not downloaded or exercised here. Whisper and Parakeet use a different audio path and cannot currently be selected. See the voice integration findings.
From shortcut to answer
Explorer handles recording inside its canvas and reuses Copilot's installed libraries and downloaded model weights. It does not route dictation through the main chat composer. The owner confirmed the v0.1.9 adapter works in the installed app; the earlier v0.1.8 claim of automatic native canvas dictation was incorrect.
sequenceDiagram
autonumber
actor You
participant UI as Explorer answer box
participant Server as Explorer localhost server
participant Files as Settings and model files
participant Helper as Explorer voice helper
participant Mic as Selected microphone
participant Draft as Local draft store
You->>UI: Focus the answer box
UI->>Server: Check voice availability
Server->>Files: Read Copilot shortcut and Explorer preferences
Files-->>Server: Shortcut, recording mode, selected model
Server-->>UI: Ready and configured shortcut
You->>UI: Press and hold Ctrl+Space
UI->>UI: Check focus and shortcut, then remember draft and cursor
UI->>Server: Start with a unique recording ID
Note over UI,Server: Token-protected HTTP on this computer only
Server->>Helper: Launch isolated helper with installed Copilot runtime
Helper->>Files: Load installed voice libraries and selected model weights
Note over Files,Helper: No model download or copying of weights
Helper->>Mic: Resolve device and start PvRecorder
Helper-->>Server: Listening
loop While you speak
Mic->>Helper: Audio frames, kept in memory
Helper->>Helper: Stream 100 ms PCM frames through one loaded model
Helper-->>Server: Measured audio level
UI->>Server: Poll recording status
Helper-->>Server: One ordered decoded-token chunk
Server-->>UI: Pending chunk, levels and elapsed time
UI->>UI: Check context, draft and caret; update owned insertion
UI->>Server: Persist local draft with revision
Server->>Draft: Atomic draft write
Draft-->>UI: Durable draft acknowledgment through server
UI->>Server: Acknowledge chunk sequence
Server->>Helper: Release pending chunk
UI-->>You: Text and local draft save status
end
You->>UI: Release Space or Ctrl
UI->>Server: Finish this recording
Server->>Helper: Stop capture and drain
Helper->>Mic: Stop and release microphone
Helper->>Helper: Verify final aggregate equals decoded tokens
Helper-->>Server: Completed, or visible failure with prior chunks retained
Server->>Helper: End process and clean temporary catalog
UI->>Server: Poll for completed result
Server-->>UI: Completion for this recording ID
UI-->>You: Review the local draft before saving an answer
PvRecorder captures mono, 16-bit PCM audio at 16 kHz. Foundry Local runs the speech model while capture continues. Only loudness measurements, status and text chunks travel between Explorer's helper, local server and answer UI; raw audio stays in the helper's memory. Queue lifetime PCM is capped at 19.2 MB, with a separate 160 KB recorder buffer; these are audio limits, not a total process-memory guarantee. Lost connections and timeouts stop capture, and recording IDs plus chunk sequences prevent stale or repeated insertion.
Files and privacy
Committed objective, success criteria and optional exploration mode are stored on the version-1 journey document and included in JSON and Markdown exports. Older files open in Understand mode without adding a field to the saved file. The objective is limited to 2,000 characters and success criteria to 4,000; invalid explicit guidance and modes are rejected on import. Mode changes are undoable guidance revisions and clear a finished state while preserving old content. Guidance influences future questions and research, but finishing remains your choice.
Journeys autosave as versioned JSON documents under ~/.copilot/plugin-data/explorer/ (COPILOT_HOME is respected). EXPLORER_DATA_DIR can override this for testing. New filenames hash the stable journey ID; canvas/session identity is only an attachment. Legacy .explore.json files are discovered in place. Opening a journey from another canvas uses the same file and revision protection. Finish also saves a Markdown sibling there. Export names use an exploration identifier; every node's ID and parent relationship are preserved in the resumable file. Pin/archive/recent metadata and canvas bindings live separately in library.json; opening or pinning never changes the journey's edited timestamp. See the library storage contract.
Question/answer/research data is stored locally as plain text. Generating questions sends only the current map context to the attached chat; the library listing is not added to that context. Research sends a bounded saved-context snapshot to its isolated Copilot session and may include relevant terms in web searches. Selected project files are read only after explicit approval in Challenge decisions and sent only to that tool-free review. Other operations omit them. The plugin loads no external UI assets. Downloaded files contain the exploration text. Importing a journey keeps the previously open journey in the library. Local answer drafts are stored separately in plain text and never included in exports, question generation, or research input. Only explicitly saved text and states enter the map; only Decided records a user decision.
Research usage receipts are retained under ~/.copilot/plugin-data/explorer/usage/session-state/, respecting the same directory overrides. They contain model, token/credit metrics and accounting IDs, without prompts, answers, account names or credentials. The temporary research conversation is removed after the helper stops and usage is retained. Cleanup/accounting failures preserve recovery files rather than losing receipts; usage/recovery-<session>.json identifies the temporary directory when a pointer can be saved. See integration notes.
Dictation reads Copilot's voice settings record and local voice assets. Explorer's microphone and model choices are saved in voice-settings.json in its data directory, separately from exploration exports and Copilot settings. Audio stays in memory in a separate helper and is processed locally; no recording file is saved. Temporary Explorer catalog metadata is removed when the helper exits. Foundry's native library can emit operational telemetry. Dictated text follows the same save/research behavior as typed answers.
There is no fixed depth limit. This POC supports up to 5,000 nodes, 20,000 characters per answer, 9 MiB of saved map data, and 10 MiB imports. The saved-data limit keeps exports reloadable. Very large maps or prompts can hit host/context limits sooner. A crashed process may leave a .lock file; fully exit Copilot before removing only the matching lock, then reopen the map. Corrupt files are preserved and reported, never silently reset.
Unchanged maps use conditional polling, avoiding repeated map downloads and parsing. A connection gap shows Reconnecting while the editor and local draft remain available. Polling retries automatically with a bounded delay and requests a full snapshot after reconnect. Failed document actions are never replayed automatically; refresh the saved map before retrying an action whose response timed out, because it may already have saved.
Develop and verify
npm ci
npm test
npm run check
npm run preview
Preview prints a local URL to open in Brave. It uses the real UI and persistence, but has no Copilot connection; AI actions report that limitation. Manual questions, answers, imports and exports work. No browser opens automatically.
The expanding node is integrated into the installed canvas from v0.1.7. The earlier layout studies remain available with npm run prototype:layouts for design reference; those pages use in-memory example data. Use npm run preview to inspect the real shipped interface locally.
The interaction research examines six product precedents and relevant studies. It records why the owner preferred the expanding node. Its earlier concept sketches remain available for reference.
DOM interaction tests run the shipped HTML/JavaScript against the real local server and persistence controller, with geometry stubs because jsdom does not render a visual browser. They cover in-map forms, drafts, saves, research results and follow-up navigation. Layout tests separately check variable-size boxes and branch separation.
For the real SDK/runtime smoke tests, set COPILOT_SDK_DIR to the installed SDK directory containing index.js, and COPILOT_CLI_PATH to its paired runtime executable, then run npm run e2e and npm run e2e:research. Disposable homes and loopback model/MCP fixtures check plugin discovery, HTTP serving, queued chat delivery, canvas callbacks, isolated research tools/context, cancellation, accounting and export/import without paid inference or browser automation.
npm run package creates a clean, dependency-free installable directory under dist/. Register that directory as a local marketplace if you prefer a snapshot. There is no build step; the checked JavaScript/HTML/CSS is the shipped code.
For each release, run npm version <next-version> --no-git-tag-version --ignore-scripts to update the root package and both lockfile version fields. Update plugin.json, .github/plugin/marketplace.json, and com.github.copilot/extensions/explorer/package.json to the same version. Change only the current-version and installation-check sections marked below and near the top of this README; leave historical version references in place. Prepend a release-notes.json entry with the new version and a user-facing summary of at most 400 characters, preserving older entries. Run npm test, npm run check, and npm run package before committing. The check rejects missing or mismatched current release fields, and packaging runs it again.
Manual acceptance in Copilot:
- Update to v0.1.70, start a new session, invoke
/explore, and confirm the bottom label reads Explorer v0.1.70. In a narrow pane, choose each mode in creation and Edit guidance; check its description, active label, keyboard focus, retained draft, and old question explanations after a change.
- Open Settings > Map and motion in a narrow pane. Try wheel zoom and pan with a mouse and trackpad, Ctrl/Command+scroll over an answer, Shift+vertical scroll in Pan mode, and a system reduced-motion change. Check the mode hint, focus return, and recording feedback in light and dark themes.
- Select a question and confirm it expands within the map, with no separate question sidebar. Collapse it and check its saved-answer preview.
- Pin a saved question, open another question, and expand References in a narrow pane. Check long answer disclosure, source jump with an unsent draft, keyboard focus after Remove reference, and unavailable-source text after restoring an earlier history version.
- Open Challenge decisions, select/exclude saved answers, review the exact tool-free scope, cancel/retry, and inspect hypothesis labels and source navigation. Confirm empty selection cannot start, native keyboard/narrow-pane layout is usable, and drafts survive. Local fixtures do not prove live model reasoning quality.
- Open Check answers in a narrow pane. Run it on a saved journey, inspect possible findings and omitted coverage, dismiss one, reveal a source outside the current focus or behind a fold, and confirm an unsaved draft remains intact. Change a sent answer in another canvas, refresh and confirm the old result is stale before an explicit rerun. Live model judgments still need human review.
- Collapse a branch with an active descendant editor, confirm focus moves to the visible ancestor, then select the hidden question from Outline. Check that its draft text and state return, the other branches stay as they were, and the compact hidden/unresolved counts remain readable in a narrow pane.
- Focus a deep branch and a leaf, use Fit visible map, then Show whole map. Check root and ancestors remain visible during focus, unrelated branches return with their prior folds, and the answer editor and caret remain usable in a narrow pane. Select a hidden question from Outline and check focus exits.
- Focus the answer, select Record, wait for Listening, and speak for more than two minutes. Confirm ordered text appears at the earlier caret, becomes Draft saved locally, and never commits an answer. Stop should finish buffered speech; Cancel should keep inserted text. Repeat with the configured shortcut, silence, deliberate repeated words, caret changes, model lag and draft recovery. Check Escape, Tab through controls, moving to References or another question, and narrow-pane button focus. Automated synthetic audio does not establish these microphone/UI checks.
- Open Settings > Voice, choose a microphone and downloaded model, save, and reopen Explorer to check persistence. Disconnect a selected device and confirm Explorer explains the problem. Check that the recording bars respond to your speech, stay flat in silence, and stop during transcription.
- Answer a question and confirm relevant deeper questions appear.
- Research another question; confirm progress appears in the node and the main chat stays quiet. Inspect relevance, source links and uncertainty. Add a recommendation and selected details to an existing draft, undo an addition, then save explicitly. Cancel another research run and confirm your answer is preserved.
- Export both files, open the JSON in a new canvas, and confirm the full map returns.
- Open Library in a different chat, search/pin/archive/unarchive a journey, then reopen it. Confirm questions arrive in the new attached chat. Try importing a changed export with the same identity and verify both duplicate choices. Type an answer, wait for Draft saved locally, reopen the journey and restart Explorer to check recovery. Check changed-answer and cross-canvas recovery choices, explicit discard, and manual-question switch guards.
- Check the narrow layout, keyboard focus, pan/zoom, and light/dark appearance. Show and hide the minimap at 280px width and in a short pane; verify its viewport box, selected/fold/focus markers, pointer cancellation, arrow keys and visible schematic recovery hint. Confirm the editor and map controls remain reachable and late cancelled dictation does not insert text.
Automated runtime fixtures cannot establish desktop rendering, actual model research quality, or successful slash-command discovery in your visible session. Those checks remain manual.
See AGENTS.md for development rules and the assessed Copilot practices imported from Usage Insights and Enhance Prompt.
History retains up to 50 snapshots and 128 MiB per journey, with up to 10 protected named checkpoints. Older unnamed snapshots are pruned first. Restoring preserves local drafts for explicit recovery and creates a new revision when content changes. See history storage and recovery for the commit protocol, limits and recovery instructions.
Suggest a missing direction from a branch's Missing a direction? disclosure. Add & explore saves your original text and asks Copilot to develop questions underneath it. Seeds are labeled as user directions, stay separate from answers and question counts, and remain available for retry after generation fails or is cancelled. Accept proposed questions in the preview before they enter the map. Schema and behavior.
Pinned constraints and AI request scope
Edit guidance to add up to ten user-supplied constraints, each with up to 1,000 JavaScript characters. Text stays exact; these statements are not verified facts or answers. They are separate from private visual reference pins and remain in native files, history and complete Markdown. Empty legacy metadata stays absent. Guidance changes require cancelling an active generation/research request; staged questions and help are invalidated when their captured sources change.
Research options now opens a local context review before an explicit Start reviewed research action. Inspect required guidance, pins and the selected question, exclude optional ancestors or other answers, and inspect the exact prompt and instructions. Depth and source preferences refresh this exact preview without starting inference. Saved results retain these settings; older results show unknown settings. Choices are request-local and expire after ten minutes; changed saved sources require refresh and a new explicit start. No project files or source attachments are included. Drafts remain untouched. See research context review for limits and lifecycle.
AI request context offers ancestor-path-only or additional-decision scope for the next request. Checking coverage or changing scope does not start inference. Topic, objective, success criteria, mode, all constraints and the active node are mandatory. Nearest ancestors are considered first, then other saved stateful questions by tree distance and stable ID. Optional records are omitted whole with counts and reasons; no answer or constraint is truncated. Mandatory overflow requires shortening the selected answer or guidance/pins.
The 128 KiB (131,072 byte) ceiling measures UTF-8 bytes in the actual Explorer-authored SDK prompt and instruction strings, including serialized JSON escaping and the packaged research instructions. It is not a token estimate, provider limit or measure of existing attached-chat context. Native get_exploration still exposes the full current committed map to the attached chat. Payload exclusions do not erase information already in chat. Drafts, private visual references, library metadata, history and audio are excluded from these prepared payloads. Generation also uses bounded child question/direction cues and separately labeled historical model explanations to avoid repeats and preserve reasoning continuity. Path-only scope excludes child cues. These are never user decisions; research does not acquire those generation-only fields.
Question help retains its smaller 16,000 serialized-character input bound, adding only explicit pinned guidance and matching freshness checks; it can fail if mandatory guidance cannot fit. It still excludes the selected answer and other branches. Clean My Answer remains answer-only. Visible native layout, microphone behavior and live model quality require manual checks.
Large maps mount compact cards near the viewport while keeping the selected editor and the current keyboard-focused control connected. Pan and zoom reuse stable layout geometry; resizing, folding and changed editor dimensions refresh it. Use Outline to reach any logical node and Fit to recover the visible map. Existing document/node limits still apply; fitting an entire map can mount every visible card.
Diagnostics and support
Open Settings > Local data and diagnostics > Diagnostics for read-only host, speech and owned-helper prerequisites. Retry checks refreshes observations without recording, inference, sign-in prompts or downloads. Ready describes observed prerequisites; model loadability, live capture and future provider/tool access remain unknown. Installed speech component compatibility stays cached until Explorer restarts.
Export support JSON explicitly downloads only the categories shown in the dialog. It excludes user content, identities, paths, device names and raw errors. Helper shutdown uncertainty remains a blocker even after a recovery notice is dismissed. See the support schema and detection limits. Native layout, microphone operation and live provider quality still need owner verification.
Selected project files
In Challenge decisions, enter up to five relative text/source paths and choose Read / refresh selected files. Review exact Unicode excerpts, source hashes and included/total bytes before starting. Changed or missing sources require explicit refresh. Files only ground this tool-free challenge; questions, web research/follow-ups, other helpers and exports omit them. Selection/results are transient, but native SDK helper journals may retain input. Limits, freshness and retention.
Portable map exports: use Export maps for Mermaid, JSON Canvas, SVG or a downloadable print document with a manual Brave Save as PDF path. Choose the whole saved journey or the selected branch and all descendants, independently of folds and focus. See portable export scope, limits and fidelity. Complete native and Markdown archives remain available.