The .praxiom.yml file
Reference for per-repository configuration: where the file is read from, every key that has an effect, and what happens when the file is invalid.
A .praxiom.yml file in a repository changes how codingassist reviews that repository's pull requests. The file is optional: without it, every review uses the defaults listed below.
Where the file is read from
codingassist reads the first non-empty file among these paths, at the head commit of the pull request:
.praxiom.yml.praxiom.yaml.github/praxiom.yml.github/praxiom.yaml
The file comes from the source branch, so a pull request that edits .praxiom.yml is reviewed with its own version of the file. There are no organization-level defaults: each repository has only its own file.
Example
version: 1
review:
customContext: |
Handlers return Result<T>; flag any "throw new" used for an expected failure.
Code under src/Payments must never log card numbers.
autoFix: false
agents:
scholar:
enabled: true
topK: 5
similarityThreshold: 0.55
archaeologist:
enabled: true
languages: [csharp, typescript]
sandboxTools:
- name: ruff
cmd: ruff check /workspace --output-format sarif --exit-zero
triggerGlobs: ["**/*.py"]
parser: sarif
severity: Low
timeoutSeconds: 60Keys
Keys are camelCase. Enum values are PascalCase and case-sensitive (Low, not low), except parser.
Top level
| Key | Type | Default | Effect |
|---|---|---|---|
version | integer | 1 | Schema version. A value above 1 makes codingassist ignore the whole file |
review
| Key | Type | Default | Effect |
|---|---|---|---|
review.customContext | string | none | Free-text instructions added to the AI review prompt. Truncated to 2,000 characters. See Custom review instructions |
review.autoFix | boolean | false | When true, codingassist proposes a fix automatically for each review that keeps at least one Critical finding. When false, a fix is produced only when someone replies @codingassist fix |
agents
| Key | Type | Default | Effect |
|---|---|---|---|
agents.scholar.enabled | boolean | true | Retrieves similar code from the repository index and gives it to the AI review as context. false turns retrieval off |
agents.scholar.topK | integer | 5 | Number of similar snippets retrieved per query (minimum 1) |
agents.scholar.similarityThreshold | number | 0.55 | Minimum similarity score for a snippet to be used |
agents.archaeologist.enabled | boolean | true | Builds an impact map of the code that calls or is called by the changed code. false turns it off |
agents.archaeologist.languages | list of strings | all | Limits the impact map to these languages: csharp, python, javascript, typescript, go, php |
workItem
| Key | Type | Default | Effect |
|---|---|---|---|
workItem.provider | AzureDevOpsBoards | organization setting | Sets the work item provider for this repository. Azure DevOps Boards is the only provider that returns work items today. None keeps the organization setting |
sandboxTools
Each entry runs one command of yours in the analysis sandbox, next to the built-in analyzers. Its findings are added to the static findings the AI review reads.
| Key | Type | Default | Effect |
|---|---|---|---|
name | string | required | Lowercase letters, digits and -, at most 40 characters, not starting or ending with -, unique in the file |
cmd | string | required | Shell command. The changed files are under /workspace; write output files to /output/ |
triggerGlobs | list of strings | empty | The tool runs only when a changed file matches one pattern. ! excludes. Empty means the tool runs on every pull request |
parser | sarif, junit, json-jq, text-llm | text-llm | How the output is read. sarif and json-jq read /output/<name>.json or standard output. text-llm sends the raw output to a model to extract findings, which costs tokens |
jqExpression | string | none | Required with json-jq. Must produce an array of {file, line, message} objects, with an optional severity |
severity | Low, Medium, High, Critical | Low | Severity given to the tool's findings when the output has none |
timeoutSeconds | integer | 60 | The tool is stopped after this time and the review continues |
An invalid entry is dropped and the rest of the file still applies.
When the file is invalid
A review never fails because of .praxiom.yml. codingassist uses the defaults instead of the file when:
- the file is not valid YAML or is larger than 32 KiB
versionis above1- a known key has the wrong type, or an enum value has the wrong case
Unknown keys are ignored. The parser also accepts review.severityThreshold, review.blockedSeverities, agents.verification, agents.fixer, customCommands, ignore, workItem.projectKey and workItem.acField, but they have no effect today. A type error in one of them still discards the whole file.
What the file does not control
These behaviours are fixed today:
- The status check fails when a review keeps at least one
Criticalfinding. - AI findings with a confidence below 0.80 are dropped and never posted.
- A pull request with more than 150 changed files is not reviewed. See Why didn't codingassist comment on my pull request?