Learningsoon
How Angada records your feedback, carries finding outcomes forward, and what deeper learning is coming.
Learning is how Angada adapts to your team. Today it records your feedback on every finding and carries finding outcomes forward across pushes. Learning that turns that history into a review policy, team style, and a managed set of learnings is coming.
What works today
Finding outcomes carry forward across pushes
When you resolve or dismiss a finding, that outcome sticks. After you push new commits, a finding that matches an earlier one stays resolved or dismissed instead of coming back. Angada edits the same comment in place rather than posting a new one, so a thread you already handled does not reappear. Rebases and force-pushes do not resurface it.
For how reviews track across pull request updates, see Re-runs and continuity.
Your feedback is recorded
Every action you take on a finding is recorded as a signal. Angada keeps these as raw signals. It does not read a 👍 as "do more of this" or a 👎 as "do less". Explicit do-more / do-less meaning for reactions is coming soon.
| Signal | What you do | What Angada records |
|---|---|---|
| Reaction | React to a finding comment | The reaction as a raw signal, with no do-more / do-less meaning |
| Reply | Reply in the finding thread | The reply as a signal on that finding |
| Resolve | Resolve the thread | The finding was addressed |
| Unresolve | Reopen a resolved thread | The outcome was reverted |
| Dismissal | Dismiss the finding | The finding was rejected |
Guide the reviewer with Review instructions
Review instructions are the explicit guidance channel today. A workspace admin sets them and the reviewer reads them on every pull request. To set them:
- Open Settings → Review.
- Enter your guidance in Review instructions.
- Wait for the saved status before leaving the page.
Write instructions that explain the reason, not just the rule. The reviewer applies a stated reason more consistently than a bare prohibition.
| Instruction | |
|---|---|
| ❌ | We don't do this. |
| ✅ | We avoid wildcard imports because they hide which symbols a file depends on. |
Review instructions accept up to 4,000 characters, are applied whole on every review, and may themselves generate findings when code contradicts them. For the full setup and how these differ from the comment header, see Custom context.
Track progress
Angada reports how your team acts on findings so you can see whether reviews are landing.
| Metric | What it measures |
|---|---|
| Addressed rate | Share of findings your team resolved by changing the code |
| Acknowledged rate | Share of findings your team engaged with, by resolving or dismissing |
Quieter known-noisy rules
When a rule produces findings your team keeps rejecting, Angada can mute that rule or lower its weight so it stops crowding out signal. This tuning lowers ranking; it does not suppress a finding outright.
Coming soon
These are in development. None of the surfaces below exist today. Treat them as the direction Learning is heading, and request access to follow along.
Angada proposes your review policy
Coming soon.
Instead of handing you an empty config file, Angada will read your repo signals and feedback history and propose a .angada.yml policy as a diff you accept or edit, with a stated reason for each key. This is the headline of where Learning is going.
Team preference and style learning
Coming soon.
Angada will learn a pooled, calibration-only style from your senior reviewers and apply it to future reviews. Today Angada only learns what to de-prioritize, not your team's positive style.
Adaptive suppression and a learning lifecycle
Coming soon.
Rejected-finding patterns will improve, merge, and retire over time, with an "exceptions / do not apply when" guard so a learning stops firing where it should not. Today a dismissal only lowers a look-alike's ranking at consolidation; nothing is suppressed, and nothing is retired.
See and manage what Angada learned
Coming soon.
A settings surface will let you review each learned item, see how often it has de-prioritized a look-alike, and delete any you disagree with. The first version will be a list with delete; it does not exist yet.
Privacy and isolation
What Angada learns in one workspace is never applied in another. Your feedback, outcomes, and any learned preferences stay scoped to the workspace that produced them.