Read the dashboard
Find a signal. Open its source. Choose the next useful step.
Start with the decision
Open Agent work → Improvements for the next action. Open Reports to see what an agent changed, checked and learned. You do not need to turn on every metric before using PhaseDrift.
Code health describes the repository. Effort describes recorded work. Reports explain obstacles. Use them together to investigate a specific problem.
Choose the view for your question
| View | What it answers | Where to look next |
|---|---|---|
| Improvements | What should we try or review? | The change and its checks |
| Reports | What happened during a task? | The outcome, obstacle and suggestion |
| Effort | What did recorded work take? | Sessions and turns behind the chart |
| Code health | What did the scan find? | The analyzer result and affected code |
| Code explorer | Which file or function is involved? | Its source and revision |
| Delivery | Did work reach a PR and pass checks? | Imported GitHub evidence |
A first scan may give you only a baseline. A trend needs more than one point. Sampled history leaves gaps between measured commits.
Follow the complete work
A turn is a bounded agent execution segment, not a whole feature. Include retries, corrections, review, and rescue effort before judging an improvement. Native sources expose different measurements; missing tokens or runtime stay unavailable. Check host coverage →
Trials, reports, and lessons
Open Improvements to inspect the problem, applied change and result. What changed gives the short explanation; Evidence & details contains reports, checks and review history. Older direct effort comparisons are under All tracked changes where available. Pinning keeps evidence stable; recomputing creates a new version.
Read the report for the agent's account of the task. Follow its sources when evaluating a suggestion. Review and activate a scoped lesson separately so relevant agents can retrieve it at entry. Delivery is not proof of use.
Try a change worth keeping or see how improvement compounds.
Actual values or relative scale?
Actual values show recorded units. Relative scale divides each series by its maximum in the selected window. It is not a health score or a percentage improvement.
The same point can change position on a relative scale when the window's maximum changes.
| Recorded point | Window maximum | Relative position |
|---|---|---|
| 6 tool calls | 12 tool calls | 50% |
| 6 tool calls | 24 tool calls | 25% |
The recorded point is still 6 calls. To judge a change, inspect its before/after trial evidence and the work that contributed.
Optional delivery context
Read-only GitHub enrichment can add PRs, reviews, findings, and checks. Connect the correct owner/name repository using local gh authentication. Agent feedback, native effort, and delivery records answer different questions.
Inspect code health → · Investigate a missing result →
Which direction is good?
| Signal | What matters |
|---|---|
| More source code | What capability was added and whether it is easy to change |
| More tests | Whether they catch the failures that matter |
| More tool calls or tokens | Whether comparable work requires more retries or rework |
| Lower runtime | Whether the full task still meets its acceptance checks |
| More complexity or duplication | Whether related code is repeatedly hard to change |
Compare similar work before attributing a change to a fix. Input includes cached input for some hosts; output can include reasoning. Those components are not extra independent totals. See host coverage for producer differences.
Zero, missing, and reported evidence
| Result | Meaning | Next step |
|---|---|---|
| Measured zero | The check ran and found none of that issue | Inspect scope and completeness |
| Missing | The source did not provide a measurement | Read the collection or analyzer blocker |
| Agent report | The agent's account of what happened | Open its source and checks |
A first scan gives you a baseline. Later compatible scans give you a trend.