Colcon Dashboard

A live web dashboard for colcon build. It shows how many packages are done, building, ready, waiting, or failed. It draws the dependency graph and colors each node as the build moves through it. Each package gets its own log pane, so logs never mix.

The server uses only the Python standard library, with zero dependencies. It works on any colcon workspace, because it reads the events.log and logger_all.log files that colcon writes under log/ for every build.

Install

pipx install colcon-dashboard
colcon-dashboard --install-service

The service starts the server at login and restarts it on failure. Open http://127.0.0.1:8642/ and pick your workspace. That is the whole setup.

On Ubuntu, the apt package from the PPA does the same job and updates with the system:

sudo add-apt-repository ppa:xmfcx/colcon-dashboard
sudo apt install python3-colcon-dashboard
colcon-dashboard --install-service

Plain pip install colcon-dashboard works in any Python 3.8+ environment. To install from a checkout, replace the package name with the path to this repository. Without the service, colcon-dashboard starts the server by hand.

Quick start

Build as you always do, and open your workspace's page:

$ colcon build
$ colcon-dashboard .
http://127.0.0.1:8642/?ws=%2Fhome%2Fuser%2Fws

The dashboard needs nothing from colcon: it reads the log files colcon already writes, so any build shows up, from any terminal, with any options.

One server runs per machine, on port 8642, and serves every workspace. The page's ws query parameter picks the workspace, so each browser tab can show a different one. A file lock refuses duplicate servers, and the server keeps running after the build, so the pages stay available for the next build and for post-mortems.

Click the workspace path in the header to open the workspace picker: your recent workspaces, a path box, and a scan of your home directory for colcon workspaces.

The command line:

colcon-dashboard                              # start the server, or print its URL
colcon-dashboard ~/projects/autoware          # same, and open this workspace
colcon-dashboard --status                     # print the build status in the terminal
colcon-dashboard --list                       # the known workspaces and their URLs
colcon-dashboard --prune --keep 5             # keep 5 builds and 5 test runs
colcon-dashboard --stop                       # stop the server
colcon-dashboard --install-service            # install the systemd user service
colcon-dashboard --restart-service            # e.g. after a config.ini change

The service commands are --install-service, --uninstall-service, --start-service, --stop-service, --restart-service, and --service-status.

The server follows log/latest_build. If a new build starts in the same workspace, the page switches to it and shows a notice. colcon test runs appear in the build picker too, with their own pass/fail counts.

Configuration

The server reads ~/.config/colcon-dashboard/config.ini at startup. The file is optional, and a command line flag always wins over it. This is how service users set the port or the bind address, because the systemd unit passes no flags. After a change, run colcon-dashboard --restart-service.

[server]
host = 127.0.0.1        ; use 0.0.0.0 to reach the page from another machine
port = 8642
log_base = log
check_updates = false   ; true: ask PyPI once a day for a newer release

[builds]
auto_prune_keep = -1    ; N >= 1: keep the newest N builds and N test
                        ; runs after each build the server watches
                        ; -1 (the default): never delete anything

[ai]
claude_bin =            ; explicit path to the claude CLI

[ui]
editor_url =            ; e.g. vscode://file{path}:{line} - file:line log
                        ; lines get a link that opens your editor

What the page shows

Header

The header shows the workspace path, the build id, a LIVE / COMPLETE / FAILED / STOPPED badge, and the parallel worker count. The workspace path opens the workspace picker. The build id opens the build picker. The bell button turns on desktop notifications: the build finished, the build failed, or the first package failed. A second click adds a quiet chime. The ? button (or the ? key) opens the keyboard and mouse reference. The power button stops the server, after a confirmation. When the server stops answering, the whole top panel turns red.

The browser tab mirrors the build even in the background: the title shows [42%] workspace while the build runs, then ✓ or ✗, and the favicon changes color with it.

The workspace picker lists your recent workspaces with their build count, log size, and last build time, and shows live progress for workspaces that build now. The list sorts by the last build time, and a workspace that builds now has the newest build, so it sits on top. A star pins a favorite, and a sort menu reorders the list by favorites, build count, or log size. The picker also opens any path and scans your home directory for colcon workspaces. With check_updates = true in the config, a dismissible line appears here when PyPI has a newer release; the server makes no other network request, ever.

The build picker lists every build and colcon test run of the workspace with its log size and its outcome: a passed, failed, or aborted chip, with the done, failed, aborted, and skipped package counts. Each finished run also shows its total duration, and the delta against the previous run of the same kind: +3:12 in red for a slower run, −1:04 in green for a faster one. The outcome comes from one pass over the run's events.log, cached in a small file inside its folder. Open a run and the whole dashboard shows it, with the build query parameter in the address. The 🗑 buttons delete one run's logs, and a prune action keeps the last three of each kind (build and test). The server refuses to delete a run that is still writing its logs. The auto_prune_keep config key does the same pruning automatically whenever it sees a finished build, for every workspace the server watches.

The right side is the system strip. A colcon build can exhaust the machine, so pressure stays visible at all times:

• A per-core heatmap, one cell per core. Cell color runs from the idle gray to full blue at 100% use. Hover a cell for the exact number.
• CPU, RAM, and swap meters. A fill turns amber under pressure (CPU 85%, RAM 75%, swap 30%) and red near the limit (96%, 90%, 70%). At red the label also turns red and bold.
• Hover the CPU meter for the load average. Hover the RAM meter for the exact percentage.
• In a narrow window the strip folds behind a chart button and drops down on demand.

Progress strip

The meter fills as the build advances: green for done, animated blue stripes for building, red for failed, orange for aborted.

The tiles count each state: done out of the total, building, ready, waiting (plus a blocked count after a failure), failed, aborted, and skipped. Elapsed time, the rate in packages per minute, and an ETA sit at the end. The ETA is an estimate: the longest remaining dependency chain, or the remaining work spread over the workers, whichever is larger, with each package's duration taken from the previous build.

Dependency graph

Every package in the build, laid out left to right by dependency depth. An edge points from a dependency to the package that needs it.

Node states combine color, border, and motion:

• Done: a green wash.
• Building: a blue box that slowly breathes, each node in its own rhythm. The box fills left to right with a deeper blue as make reports [ 42%] progress, and the label stays legible on top. A faint halo glows behind the node on the backmost layer, so the active zone shows even from far out.
• Ready: a dashed blue border. All dependencies are done, and the package starts as soon as a worker becomes free.
• Next up: a slow pulse. This waiting package starts when the packages that build now finish, because no deeper dependency blocks it.
• Waiting: quiet gray. Skipped: gray with a struck label.
• Failed: solid red. Everything downstream of a failure turns red-dashed with red edges. This blast radius looks different from the packages that the stop only abandoned, which stay gray.
• Aborted: orange. The package was building when the build stopped.

Edges take color from their endpoints:

• Light green between two done packages: finished lineage.
• Solid blue into a building package, with droplets that flow along the edge at constant speed: the package consumes its finished dependencies.
• Dashed blue marching into a next-up package: what it waits for.
• Red along the failure cascade.

Interaction:

• Hover a node to see its state, phase, time, stderr count, and path, and to light its full dependency chain while the rest dims. When the previous build has a duration for the package, the tooltip compares them, and marks a clear regression.
• Click a node to open its log pane.
• The find package box highlights matching packages while the rest fades, in every layout and in the timeline. Enter and Shift+Enter jump through the matches. The / key jumps to the box.
• The state chips (failed, building, waiting, done) filter the same way by state instead of by name, and combine with the find box. One click on failed shows the whole blast radius of a broken build.
• Double-click a log tab in the dock to center the view on that package.
• Drag to pan, scroll to zoom. Fit frames the whole graph.
• ⌖ follow build keeps the camera on the packages that build now, so the action stays framed as the build moves through the graph. Pan or zoom by hand and the camera is yours again.
• The show switch picks what the graph draws: only this build's packages, or the whole workspace.
• When most package names share a prefix such as autoware_, the labels hide it. Tooltips keep the full name.

Layout modes

The layout menu picks one of three modes, and the choice persists:

• layered: the static left-to-right layout described above.
• force: a live spring simulation with the same left-to-right anchoring. Drag nodes to rearrange. The simulation cools and stops by itself.
• 3d: the build as a wavefront. Building packages share one central plane, finished discs stack to its left, and waiting discs queue to the right by dependency depth. Packages glide through the blue plane as the build advances. Left-drag orbits, shift-drag or right-drag or middle-drag pans, the wheel zooms, and the camera rotates by itself until the first grab.

Fit and follow build drive the camera in every mode. In the force and 3D modes, a spread slider in the corner scales how far the simulation spreads the nodes.

Timeline

A Gantt chart of every started package, with a time axis and a dashed now line. Bar colors match the graph states.

The order menu sorts the rows: by start time, by duration with the longest build first, by end time, or by status with failures on top. The choice persists.

The chart makes the parallelism and the long serial chains visible. It auto-scrolls to the newest bars as they start. Scroll up to release, or use the ⤓ follow new toggle. Drag anywhere to pan, like the graph.

Click a bar or a package name and its log pane opens while the graph flies to that package. From the pure timeline this switches to the side-by-side view, so the graph comes in without losing the timeline.

Side by side shows the graph and the timeline together, over a drag bar. It is the default view, and the split ratio and the view choice persist.

Log panes

The dock opens with a pinned build log tab: the whole build, as a terminal shows it, with Starting >>> and Finished <<< lines between every package's output. Click any package in the graph or the timeline to open its own tab next to it. Each tab has its own scrollback, so logs never interleave.

• A pane follows new output until you scroll up. The ⤓ follow button re-engages.
• Panes open at the tail for an instant start. ⤒ load all fetches the whole history in one click.
• Every line carries a timestamp: build-relative in the build log, job-relative in package panes. The 🕒 ts button hides them.
• The ↩ wrap button soft-wraps long lines. The choice sticks for future panes, and AI panes start wrapped because prose reads badly on one line.
• The search box filters as you type: matching lines highlight, Enter and Shift+Enter step through them, and the search stays live while the log streams.
• A selector switches a package pane between the timestamped output, stdout+stderr, stderr, stdout, and the command log.
• ANSI colors render as in a terminal: the 16 classic colors, 256-color, and truecolor, with palettes tuned per theme. Uncolored lines that match error or warning patterns still get color.
• When a package fails, its pane opens by itself and a toast points to it. A page opened on an already failed build opens the failed packages' panes too, earliest failure first.
• A ✗ failures tab appears with the first failure and aggregates every failed package: the last stderr lines of each, in failure order, with a jump link, an open-log button, and the ask-claude button. When a failed build ends, this pane comes to the front.
• The ⧉ path button in a package pane copies the package's absolute source path.
• With editor_url configured, file:line positions in the logs get a ↗ link that opens your editor at that line.
• ✕ close all at the end of the tab strip closes every tab except the build log.
• Drag the bar above the dock to resize it.

The address bar mirrors the open pane and the picked view (#pkg=...&view=...), so a copied link opens the same package pane for a colleague. The ws and build query parameters already pin the workspace and the build.

AI failure analysis

When the claude CLI is installed, the pane of a failed package shows an ✦ ask claude button. It starts a headless claude run in the workspace, with the tail of the failed log as the prompt. A new ✦ pane streams the investigation live: the files it reads, the commands it runs, and the answer.

• Nothing runs by itself. Each analysis starts with a click, and it spends your Claude usage.
• The run gets read tools approved and no edit tools, and headless claude refuses actions that need more permission.
• The transcript persists next to the package's logs, so it survives a server restart and reopens instantly.
• The input box under the transcript asks follow-up questions in the same session, through claude --resume.

Theme

The page follows the system theme. The theme button cycles system, light, and dark. Every view recolors instantly, including the 3D canvas.

How it works

colcon writes a structured event stream to log/<build>/events.log. The server tails this file and reconstructs the exact job set from it. JobQueued events carry the full dependency closure of each job, as colcon resolved it. JobStarted, JobEnded, and JobSkipped events give the state machine. TimerEvent heartbeats give liveness, so the LIVE badge does not depend on process inspection.

The graph shows direct edges, not the full closure. The server computes the transitive reduction of colcon's per-job dependency sets. Package metadata (path, build type) and the edges of packages outside the current build come from a scan of the package.xml files in the workspace.

Per-package logs come from log/<build>/<package>/. The server serves them incrementally by byte offset, so an open pane costs one small request per second.

Options

When the server binds 0.0.0.0, it prints the real LAN URL next to the loopback one. The server trusts every caller: it serves logs and deletes them on request. Expose the port only to a network you trust.

API

Workspace endpoints take a ws=<path> query parameter. Add build=<build id> to read an older build.

Limits

• One server per machine, and a lock enforces it. It serves any number of workspaces.
• A page follows its workspace's latest build, unless its address pins an older one. Finished builds stay readable until their logs get deleted.
• The server trusts the local workspace. Do not expose the port to an untrusted network.