A PR Bot That Stays Silent on Green Pull Requests in Signals Mode

repowise teamUpdated 7 min read
On this page

In Signals mode the repowise PR bot posts nothing on a green pull request, and one comment, edited in place, when a change carries a real signal. A bot that comments on a typo fix teaches reviewers to scroll past it, so the free default stays quiet until it has evidence worth a reviewer's time. Every push still gets a repowise review check, so the absence of a comment is a result you can read. This post covers what makes the bot speak, what it says, and where the optional AI review fits.

Why do PR bots end up ignored?

A bot that comments on every pull request trains its readers to skip it, so its one useful comment gets skipped too. The failure is structural: a bot that always speaks carries no information when it speaks.

GitHub's review flow already gives humans comments, line notes and review states (GitHub docs). A bot earns a place in that thread by adding evidence a reviewer cannot see in the diff: which callers sit outside the change, which file usually travels with this one, how this change compares to the repository's own history. Anything the diff already shows is noise.

What does "silent by default" mean?

It means Signals mode has a strict no-op path: if no signal fires, no comment is posted and nobody is pinged. The bot is still running. It analyses every pull request and keeps the result to itself until there is a reason to share it.

The comment rule is configurable per repository (comment_mode: on_signal is the default, always comments on every pull request), but the default is the product: a team installs the app and hears nothing until it matters.

How do you know a silent bot actually ran?

Read the repowise review check on the head commit. Every push gets one. It shows in progress while the review runs, then completes with the commit it reviewed, the base it compared against, and what it found: success with a title like Nothing to act on · free checks or Safe to merge · quick review when the change is clean, neutral with the number of points when there is something to look at. It never fails a build unless you turn on the merge gate.

That check is what makes silence safe to trust. A reviewer sees a green check and no comment and knows the change was read. A coding agent can wait on the check before it reports a pull request as clean, without comparing comment timestamps.

What makes the bot comment?

Signals comments when at least one of seven deterministic conditions holds. Each reads a measurement repowise already computes for the index, so there is no model judgement in the gate.

TriggerWhat fires itWhat it tells the reviewer
Net health degradesRepository health drops across the changed filesThe change made the codebase harder to maintain
File health dropsA modified file's score fallsA regression in one file is hiding inside a healthy average
Hotspot touchThe PR modifies a high-churn, high-complexity fileThe patch landed in code that already absorbs maintenance pain
Missing co-change partnerA file that usually changes with a modified file is absent from the PRHidden coupling may have been missed (routine pairings are ignored)
New dead codeThe PR creates unreachable files or unused exportsCleanup is cheapest while the branch is open
Declining fileA modified file is already on a falling health trajectoryThe PR continues a decline that is easy to stop while the branch is open
High change riskThe diff is large, scattered or unfamiliar for this repoReview this one more carefully than the last ten

Scroll the table sideways to see every column.

Improvements are shown next to problems, so cleanup gets credit, but improvements alone never trigger a comment.

How are file-level changes attributed?

The bot diffs a file's findings before and after the PR and reports the biomarkers the PR introduced in its added lines or resolved, not the file's current top findings. An author is shown what their change did, not the debt they inherited. The scoring behind it is the code health layer, with 59 detectors across three pillars (see code health).

How is change risk banded?

Change risk is scored against the repository's own commit history, not an absolute threshold, so "high" means high for this codebase. The comment carries the banded level and percentile, and the raw 0 to 10 score lives on the pull request's page. The method is described on the change risk feature page.

What does the comment look like?

The comment is short, factual and numbered by what to check. This one is a real Signals comment on a pull request in the repowise repository:

repowise PR bot comment on a pull request in the repowise repository: three things to check, covering 20 new findings across 5 files, a file getting worse, and three co-change partners missing from the PR, with the health of changed files and a link to the full report
repowise PR bot comment on a pull request in the repowise repository: three things to check, covering 20 new findings across 5 files, a file getting worse, and three co-change partners missing from the PR, with the health of changed files and a link to the full report

It leads with the decision-driving items: new findings, files getting worse, co-change partners and tests missing from the change, and a before and after health figure. Everything else, including the full file table, hotspots with dependent counts and the raw risk score, sits behind the button on the comment that opens a per-PR page with the blast radius. Pages for public repositories need no sign-in, and pages for private ones open for signed-in users with access to the repository. Every page has a plain-markdown version for agents.

The verbosity setting (compact, standard) changes how much renders, independent of whether the bot comments at all.

Why one comment, edited in place?

A pull request gets exactly one bot comment, updated on every push, so a ten-push PR produces one notification, not ten. If a later push clears every signal, the stale comment is deleted in Signals mode. With AI review on, it turns into the short clean comment instead.

Why does Signals make zero LLM calls?

The gate and the content are set arithmetic over a call graph, a git history and a health scorer, so the same diff gets the same review every time, nothing leaves your network for a model provider, and a review costs no tokens.

Signals answers a narrow question deterministically: what does this change touch, what is missing from it, and how risky is it relative to this repository. You can trace any comment to the rule that fired, which is what makes the rare comment trustworthy.

Where does AI review fit?

Some problems need a reader. A function that swallows an error, a caller that will pass the wrong shape, logic that does the opposite of what the title says: those show up in the code, not in a graph. AI review is the optional mode for them, on Pro, at about 15 cents a pull request on average from credits.

It keeps the same discipline. It reads every pull request that changes code, on every push, and speaks only when it has a point to raise. Fix pushes are read for what changed, with memory of earlier points, so nothing is raised twice. A verifier checks risky claims against the code before they post. A deep review runs once per pull request when change risk, blast radius or hotspot history call for it, or when someone comments /repowise review deep. Missing tests are listed once, in a single thread edited in place. On a clean pull request AI review posts one short comment instead of silence: that it is safe to merge, what the change does in a sentence, and what was checked (callers, tests, health, security and coverage). It is edited in place on each push, so it costs one notification per pull request. Signals mode stays silent on the same pull request.

How do I install and configure it?

Install the free GitHub App, pick your repositories, and open a pull request.

  1. Go to github.com/apps/repowise-bot and click Install. Public, private and organisation repositories work. Private repositories need Pro.
  2. The bot indexes the default branch on install, then reviews each push.
  3. Optionally add .repowise/bot.yaml at the repository root. Invalid values fall back to the layer below instead of failing a PR, and the merge-gate keys are read from the base branch so a pull request cannot loosen its own gate.
yaml
bot:
  comment_mode: on_signal   # on_signal (silent on green) | always
  verbosity: standard       # compact | standard
  apply_labels: false       # opt in to repowise:* labels
  ignore_paths:
    - "docs/**"
    - "vendor/**"

With apply_labels: true the bot applies a short set of repowise:* labels (gate-failed, risk-high, missing-tests, size-xl) and removes each when its signal clears. They are off by default because a label is public in the PR list and reads as a verdict on the author. Per-section switches (show_health, show_coupling, show_ownership and others) turn individual sections off. The review mode, AI review depth and spend caps are set in your dashboard settings, so a pull request from a fork cannot switch on a paid pass.

The same analysis runs without the GitHub App: see repowise in CI for the gates, SARIF and GitLab Code Quality output. The bot's own page is repowise.dev/bot, and a real analysis page is repowise-dev/repowise#4506.

How do I judge whether a PR bot is worth keeping?

Ask whether the team would miss its comments if it vanished tomorrow. If not, it is too loud. A silent-by-default bot passes because every comment it posts has a named reason, and its absence on a PR, next to a green review check, is itself information: nothing in the change crossed a line.

Keep the trigger set short. Every added trigger is another chance to turn signal back into noise.

FAQ

What does a silent PR bot mean?
In Signals mode it means the bot posts nothing on the common case. No comment, no ping, no label. It speaks only when a measurable condition fires, so each comment is worth reading. A status check still records that the review ran.
Does the repowise PR bot comment on every pull request?
No. In Signals mode (comment_mode: on_signal, the default) it stays quiet unless one of seven conditions fires: net health degradation, a file health drop, a hotspot touch, a missing co-change partner, new dead code, a declining file, or high change risk for that repository. With AI review on, a clean pull request gets one short comment that says it is safe to merge and what was checked, and a pull request with points gets the full review. Setting comment_mode: always comments on every pull request.
How do I know the bot reviewed my latest commit?
Look at the repowise review check on the head commit. It completes with the reviewed commit and the open points on every push, whether or not the bot commented. Tools should read its conclusion and the JSON status in its text, not the title.
Does the PR bot use an LLM?
Not in Signals mode. Signals makes zero LLM calls, so the same diff always gets the same review, no source is sent to a model provider, and a review has no token cost. AI review, an optional paid mode, adds a model pass on top of the same evidence.
How many comments does it post per pull request?
One summary comment, edited in place on every push. AI review adds a thread on each line it has a point about, and missing tests go in one thread. Longer detail lives on the per-PR page.
Is the PR bot free?
Yes on public repositories: the GitHub App is free to install, and Signals mode makes no model call and spends no credits. Private repositories need Pro ($15 a month, with $5 of credits included). AI review is billed per pull request from pay-as-you-go credits, with no per-seat review fee: about 15 cents a pull request on average, so 100 reviewed pull requests cost about $15. Each pull request has a spend cap, 60 cents by default. See repowise.dev/bot for details.
Can I run the same checks without the GitHub App?
Yes. repowise has a CI command with gates for health, doc drift, security, coverage and risk, with annotations, SARIF and GitLab Code Quality output. See repowise in CI.