Should you use Birdview?
Use Birdview when you want to question an agent's proposed scope before authorizing a change across several modules. Its map makes the agent explain ownership, relationships, applicable rules and planned checks in a form you can inspect. That review still depends on someone checking the explanation against source.
Birdview 0.3.1 displays agent-declared snapshots without independently observing edits or enforcing filesystem restrictions. Keep Git diffs, tests and code review in the workflow. The HTML page cannot establish that the agent changed only the approved files.
What our sandbox recorded
Our run on 2026-10-02 tested commit acddbf0 in our sandbox. The full revision is acddbf0f9496f50685ed916f8be8cad7b70f32d3, and the build output identifies Birdview 0.3.1. The attached sandbox configuration records lab-node:22, 3 CPUs and 8 GB of memory.
The stored harness records installation at 15 seconds, build at 10 seconds and the test command at 28 seconds. For each command, it stores a success flag, exit code, elapsed time and timeout flag. All three recorded exit codes are zero, with no recorded timeout. Test counts are stored separately as parsed summaries.
The attached stored parser summary reports 92 passed and 0 failed. The supplied raw test footer also reports 92 passed, 0 failed and 0 skipped. Those matching figures do not establish exhaustive unique test coverage. No new execution or complete-log reconciliation was performed. The footer's runner duration and the harness's command duration describe different timing scopes.
This run did not establish successful activation in a live coding-agent session, a completed browser test run, the accuracy of a map of another repository or compliance with an approved scope. Successful repository commands do not answer those questions.
The MrKeyoor testing methodology explains how the publisher records command outcomes and parsed counts. The supplied methodology page was fetched on 2026-10-31; it does not establish which methodology version governed this older run.
Installation succeeded, but the records disagree
The stored dependency summary records 22 packages and 120 MB on disk. The raw installation output reports 14 packages added. The supplied evidence does not reconcile those figures or establish that they count the same thing.
The stored npm audit summary records zero advisories at every listed severity. Separately, the raw installation tail reports "1 moderate severity vulnerability". Both records remain relevant; the successful installation does not resolve the discrepancy. Scan coverage is unknown because the supplied record has no scannedPackages field. Recheck the dependency tree and audit report for the revision you intend to install. Advisory totals are not a security assessment.
Validation checks the artifact's consistency
At the measured revision, validation checks structure and cross-record consistency, including sequence order, valid scope, targets and ownership references. It does not prove that architectural claims are true or that referenced source files exist.
The architecture file describes modules, ownership, relationships and evidence. The constraint catalog records reviewed rules and coverage. The activity file holds the task scope, progress and check results declared by the agent. Displaying a rule does not prove compliance, and confirmation does not mean tests passed.
Under the documented workflow, confirmation happens in the conversation and material scope changes require another confirmation. Version 0.3.1 requires regenerating the HTML and refreshing the browser when the artifact changes; live transport and automatic refresh are not implemented.
Verify activation after installing the bundle
Birdview needs its complete skill bundle, including scripts, schemas, assets, references and package files. Copying only SKILL.md is insufficient. The npm package is private, so a global npm installation is not the documented distribution route.
The installation guide separates dependency installation from the doctor check and agent activation. Doctor invokes an installed validation CLI and renders an example in memory. It does not exercise every worker CLI or prove that the host loaded the skill.
Start a fresh task with a map-only request. Check that the agent reads the skill, reports any existing map, delivers a browser-readable artifact and discloses missing constraint coverage. Inspect a module's ownership and cited evidence before approving a code-changing task.
On-demand invocation is the default; auto mode is optional. Project guidance relies on the host loading AGENTS.md or, for Claude Code, CLAUDE.md. These instructions do not intercept filesystem writes.
Keep the measured version separate from current features
Our recorded build identifies version 0.3.1. The supplied current README describes version 0.4.0, including architecture review and outside-scope impact features. Those features were not verified by our run. The README at the measured commit provides the relevant historical boundaries.
Alternatives answer different needs
Code Review Graph builds a structural map with Tree-sitter and serves compact context over MCP. Choose it when the immediate need is repository context and dependency analysis. It can complement a human review artifact; its supplied documentation does not establish an equivalent conversational scope-confirmation workflow.
Superpowers is a software development methodology for coding agents built from composable skills. Its documented workflow includes design, implementation planning, testing and review. Choose it for that broader process. Before combining it with Birdview, inspect their activation and confirmation rules; the supplied evidence does not establish conflict-free operation together.
What the scores mean
These editorial ratings describe the available evidence, not measured product performance. Setup ease runs from 1 for unresolved basic installation barriers to 3 for documented installation with several manual steps and 5 for a verified path through activation. Documentation runs from 1 for missing operating guidance to 3 for basic instructions and 5 for detailed setup, workflow and limitation guidance.
Community measures documented feedback access: 1 means no route, 3 means one route and 4 means multiple routes; 5 would require evidence of responsive support. It does not measure adoption. Maturity runs from 1 for unresolved basic command failures to 3 for successful repository commands with operational behavior untested and 5 for independently verified operation across the intended workflow.
The retained scores are setup ease 3, documentation 5, community 4 and maturity 3. GitHub feedback and a QQ group support the community score's narrow meaning. The maturity score remains limited by untested agent behavior and unresolved installation and audit records.

