Documentation content

Dashboard, saved queries, and filters

Capture a repository, filter and rank its open work, open an exact dashboard view, replay past decisions, and keep scoring defaults.

Capture a repository

Collect a GitHub graph with open, then query the dashboard model it saves:

issue-graph open vercel-labs/agent-browser --no-open
issue-graph query github:vercel-labs/agent-browser --json --no-open

open reads GitHub within the crawl budget. In an interactive terminal, it can offer an installed agent for clustering. Add --agent none to skip that prompt.

query reads the saved model without provider requests or authentication, so it never refreshes the data or fetches missing nodes. If no saved model exists, query returns an error. Run open first, or import a normalized dashboard model with --input model.json. --input expects a dashboard model, not the graph JSON that graph -o graph.json writes. Leave --no-save off the collection command if you want to query its model later.

Run issue-graph dashboard to browse the latest saved models with a project selector, or issue-graph runs to list them.

Choose a view

ViewUse it to
ExploreRead clusters and inspect an item with its linked context
ImpactInspect relationships and affected items in the captured graph
SwarmCompare items grouped by cluster, state, kind, or all together
RankOrder open items by discussion Heat and adjust scoring weights

Impact and the Swarm links and blast metrics count every relationship in the graph, parent and sub-issue links included, so an epic with open sub-issues lists them as follow-ups. Heat counts inbound references only, and sub-issue links are not references.

The compact Filters menu applies across views. It filters by state, item kind, clusters, Heat, and search, plus solution and review filters where the provider supports them. Context items can stay visible when they don't match a filter.

For GitHub, the project selector shows a separately fetched total of open issues and PRs when available. Other providers supply their own count and coverage. In query output, counts.captured, matched, visible, and ranked describe the saved graph and the current view. Inspect coverage before you interpret them.

Prioritize with filters

issue-graph query github:vercel-labs/agent-browser --state open --kind Issue --heat-min 50 --view rank --json --no-open
issue-graph query github:vercel-labs/agent-browser --kind PullRequest --review pending --heat-top 25 --view rank --json --no-open

Different filters intersect. Repeated --cluster values select the union of those clusters. Take cluster indices from groups[].index in a query result, because indices belong to that capture.

Heat combines comments, participants, reactions, inbound references, and age. Use it to decide what to inspect first. It does not measure severity or merge readiness. --heat-min is inclusive and can't be combined with --heat-top. --heat-top 10 and --heat-top 25 compute their threshold from the open captured items of the selected kind in the project, before the other filters apply, and ties can include more than the requested percentage. Heat filters exclude items without an open Heat score.

--solution with finds issues that have an observed closing link from an open or merged PR. --solution without means the capture didn't observe that link, so a fix may still exist.

Open the exact result

Every query returns viewUrl, captureId, historyId, the effective query.weights, capabilities, coverage, and results. Add --open to open the exact view and keep the JSON output:

issue-graph query github:vercel-labs/agent-browser --state open --kind Issue --heat-min 50 --view rank --json --open

viewUrl identifies the project, view, filters, weights, and selection, so share it unchanged, including its query string and hash. It's a local file URL, not a hosted page. opened: true means the CLI asked the OS to open the URL. It doesn't confirm that the page loaded.

The Next.js dashboard is a static export with embedded data and needs no Next.js server. When you move it, keep the HTML file, the sibling _next/ directory, and font-LICENSE.txt together. Anyone with the export can read the full embedded capture, including items your filters hide.

Explore and replay a capture

Replace CAPTURE_ID, HISTORY_ID, and ITEM_KEY with values from a query result, and take cluster indices and item keys from that same capture:

issue-graph query --capture CAPTURE_ID --cluster 0 --cluster 2 --view swarm --group cluster --metric heat --json --no-open
issue-graph query --capture CAPTURE_ID --view explore --select ITEM_KEY --json --no-open
issue-graph query --history HISTORY_ID --json --open

--capture reuses the full immutable model with the current defaults and only the filters in this command. It doesn't inherit filters from the previous query. --history restores the frozen parameters and saved view, even after the defaults or the latest model change, and it rejects scope, input, filter, and weight overrides. To change a historical view, start a new query with its captureId.

Explore can select a contextual item outside the filter matches. An Impact selection must be inside its results. --focus-cluster N focuses an Explore cluster, while --cluster N filters membership. Swarm --group and --metric require --view swarm.

Keep defaults separate from exploration

Inspect the effective weights before you change them:

issue-graph config show --provider github --scope vercel-labs/agent-browser --json
issue-graph query github:vercel-labs/agent-browser --weights comments=4,reactions=3 --view rank --json --no-open

Query --weights overrides and dashboard sliders apply only to that query or page and leave persistent defaults unchanged. Reset weights restores the weights the page opened with. To persist a preference, use config set:

issue-graph config set --weights comments=3,age=1 --json
issue-graph config set --provider github --weights reactions=4 --json
issue-graph config set --provider github --scope vercel-labs/agent-browser --weights reactions=5 --json

Precedence runs built-in, global, provider, project, then command. Partial updates keep the weights you don't specify. The five weights are comments, participants, reactions, inboundRefs, and age, and each accepts a finite value from 0 to 10. The built-in values are 3, 2, 2, 2, and 1 respectively. Before weighting, the age signal is min(12, daysOpen / 30). The total is rounded to one decimal.

Provider and project scope

A query identifies its model as provider:scope. A bare owner/repo selects GitHub. You can omit the scope only when exactly one saved model exists. Query doesn't infer its scope from the current checkout.

For another provider's normalized model, use the provider and scope fields it returns. Pass them separately to config show or config set, where --scope requires --provider. A Linear project model can use linear:WORKSPACE_ID:project:PROJECT_ID, which keeps projects and workspaces distinct.

Read capabilities before you choose filters, views, or metrics. Unsupported choices fail with an error. The shared query surface can read normalized Linear or Jira models, but it doesn't collect live Linear or Jira data.

Local storage

State lives under ~/.issue-graph/ by default. Set ISSUE_GRAPH_HOME to change that root. Weights live in config.json, immutable models in captures/, and query receipts with their static views in history/. Queries write these files locally and never change anything in a provider.

Reloading a saved page doesn't refresh its data. Captures stay until you remove them locally. Read Permissions and privacy before you store or share private results, and see Command reference for every query flag.

Manage saved models

issue-graph runs list --format json
issue-graph dashboard --no-open -o dashboard.html
issue-graph runs rm owner/repo

runs lists the latest dashboard model for each repository, and dashboard combines those models in one export. runs rm removes only the selected model from that list. It leaves graph snapshots, query captures, history, and existing exports in place. Run open again to collect a fresh model.