claude-code
4685c8db84
Generated sections (`<!-- generated by … -->` up to the next heading) are no longer checked by the writing rules, and the model and Codex are told to report nothing about them. See docs/prs/. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Reviewed-on: #122 Co-authored-by: claude-code <claude-code@javagil.de>
257 lines
11 KiB
TOML
257 lines
11 KiB
TOML
# werkrevisor host config for reviewing werkregie (mi/werkregie) on git.javagil.de.
|
|
# It is complete, so the review works without a .werkrevisor.toml in werkregie; a PR of werkregie never changes this file.
|
|
|
|
[model]
|
|
# An economy model suffices, since Jev decides and the model only advises (PR#40):
|
|
# on 4 heads with 10 known findings, MiniMax M3 found as many as Sonnet 5.5 for a tenth of the cost, but needs minutes.
|
|
id = "openrouter/minimax/minimax-m3"
|
|
# Operator-maintained prices in USD per 1M tokens.
|
|
input_price_usd_per_mtok = 0.30
|
|
output_price_usd_per_mtok = 1.20
|
|
# Longer prompts degrade the review even within the context window.
|
|
max_context_tokens = 150000
|
|
# Includes reasoning tokens: MiniMax M3 wrote up to 47050 in 211 reviews, and with 32000 it once answered nothing (PR#84).
|
|
max_output_tokens = 64000
|
|
# "low" and "none" missed a review on 1 of 4 heads each.
|
|
reasoning_effort = "medium"
|
|
# On 2026-10-08 some calls hung for over 700 s; the maker of the model first, others when it fails.
|
|
provider_order = ["minimax"]
|
|
# A call took up to 242 s; Werkator's ai-review build allows 15 minutes for 2 calls.
|
|
timeout_seconds = 300
|
|
max_retries = 1
|
|
# MiniMax reported at most 2 % more prompt tokens than o200k_base counted.
|
|
token_count_factor = 1.05
|
|
|
|
[budget]
|
|
# A review with 2 calls of 150000 input and 64000 output tokens costs at most about 0.24 USD.
|
|
max_usd_per_review = 0.25
|
|
min_key_remaining_usd = 0.50
|
|
|
|
[size]
|
|
# Changed lines and files outside the docs; above them a human looks at the PR (YELLOW).
|
|
max_changed_lines = 800
|
|
max_files = 30
|
|
|
|
[categories]
|
|
docs = ["**/*.md", "doc/**"]
|
|
tests = ["**/src/test/**", "**/src/functionalTest/**"]
|
|
main = ["**/src/main/**"]
|
|
api_spec = []
|
|
auto_approvable = [
|
|
"docs_only", "tests_only_added", "docs_and_tests", "refactoring", "feature_with_tests", "harmless_only",
|
|
"mechanical_only", "mixed",
|
|
]
|
|
|
|
[hard_yellow]
|
|
# Only secrets stay hard.
|
|
paths = ["**/*secret*", "**/*.pem", "**/*.key", ".env*", ".envrc"]
|
|
# Jev judges the build, CI and publication files, and with [path_judge] scope = "all" every other file.
|
|
judged_paths = [
|
|
".gitea/**", ".github/**", ".werkator*", "tools/**",
|
|
"dashboard/build.gradle.kts", "dashboard/settings.gradle.kts", "dashboard/gradle/**", "dashboard/packaging/**",
|
|
]
|
|
secret_scan = true
|
|
# Test data holds fake passwords; private keys and the real key formats stay flagged there too.
|
|
secret_scan_exclude = ["**/src/test/**", "**/src/functionalTest/**"]
|
|
|
|
[pr_doc]
|
|
# An open question in a changed PR document gives RED before any model call; TEMPLATE*.md and README.md never count.
|
|
globs = ["doc/prs/20*-PR#*.md"]
|
|
open_question_headings = ["Open Questions"]
|
|
|
|
[review]
|
|
extra_instructions = """
|
|
Write all findings in English.
|
|
werkregie orchestrates software development with AI agents; so far it is a dashboard in Kotlin under dashboard/,
|
|
which reads the state of coding agents, Werkator and Gitea and serves it to a browser behind an access token.
|
|
Judge whether the PR document under doc/prs/ is clear, correct and consistent with the code.
|
|
werkrevisor itself checks line length, sentences per line and sentences per paragraph of the docs;
|
|
never report these rules, not even as a style remark.
|
|
A tool writes each section of a doc from a comment `<!-- generated by … -->` up to the next heading;
|
|
report nothing about such a section.
|
|
For documentation, report only bloated or mannered prose (filler, narration of the implementation),
|
|
unclear statements, and contradictions between the PR document and the code.
|
|
Report only concrete problems with a realistic impact.
|
|
The user prompt contains, after the diff, the full new content of every changed file except the documentation;
|
|
documentation appears only in the diff.
|
|
Check each changed line against constraints, invariants and conventions stated elsewhere in these files,
|
|
e.g. in header comments, and report a violation as a key issue.
|
|
Review the diff; use the full files only as context.
|
|
Start every issue_header with the severity of the issue: [high], [medium] or [low].
|
|
[high]: wrong behavior, data loss, a security problem or a broken contract in a realistic scenario.
|
|
[medium]: a defect with limited impact, or a missing test for changed behavior.
|
|
[low]: cosmetics, wording, log noise or style.
|
|
"""
|
|
max_findings = 3
|
|
# Without Jev, one high finding or this many medium findings give ORANGE, another medium finding YELLOW;
|
|
# with Jev, finding_real_min and finding_severity_min of the host config decide.
|
|
orange_min_medium = 2
|
|
# The complete new version of each changed file follows the diff in the prompt.
|
|
# 160000 chars are about 50000 of the 150000 input tokens, counted with token_count_factor = 1.05;
|
|
# a larger prompt gets excerpts instead, then the diff only.
|
|
full_files = true
|
|
max_full_file_chars = 80000
|
|
max_full_files_chars = 160000
|
|
# PR-Agent reads AGENTS.md from the default branch, never from the PR.
|
|
repo_context_files = ["AGENTS.md"]
|
|
|
|
[writing]
|
|
# The shared workflow asks for one sentence per line and short sentences, but sets no line or paragraph limit.
|
|
# A violation is a hint in the comment and never blocks; the model still reviews the PR.
|
|
enabled = true
|
|
max_line_length = 400
|
|
max_sentences_per_paragraph = 20
|
|
# Empty means the docs category.
|
|
globs = []
|
|
|
|
[mechanical]
|
|
# A changed file whose every changed line is an import, a whitespace change or a rename does not go to the model.
|
|
enabled = true
|
|
# Indentation or line breaks carry meaning in these files, so a whitespace change there is never mechanical.
|
|
whitespace_significant = ["*.py", "*.yml", "*.yaml", "Makefile", "makefile", "GNUmakefile", "*.mk", "*.md",
|
|
"*.haml", "*.pug", "*.coffee", "*.nim", "*.fs"]
|
|
# Built-in import rules exist for java, kt, kts, ts, tsx, js, jsx, mjs, go and rs.
|
|
# A [mechanical.imports] table adds or replaces the line regexes of a file extension; an empty list switches it off.
|
|
|
|
[gate]
|
|
# Branch protection requires one approval from {bot, human_reviewers}; see the README.
|
|
yellow_gate = "approval"
|
|
human_reviewers = ["mi"]
|
|
assessment_context = "ai-review/assessment"
|
|
# Only a PR whose title changed from "WIP: " to one without gets a review; the bot then schedules the merge,
|
|
# so a GREEN PR merges at once and a YELLOW or ORANGE one with the human approval.
|
|
require_wip_transition = true
|
|
automerge = true
|
|
|
|
[engine]
|
|
# "llm": the LLM review alone; "classifier": Jev alone, no LLM; "combined": Jev, then the LLM with Jev's hints.
|
|
mode = "combined"
|
|
# Jev decides with [path_judge] scope = "all"; the model only advises, so an economy model suffices.
|
|
model_role = "advises"
|
|
|
|
[classifier]
|
|
endpoint = "https://api.typesafe.ai/v1/systemone"
|
|
model = "jev-latest"
|
|
# Operator-maintained price in USD per 1M input tokens; Jev's output is free.
|
|
input_price_usd_per_mtok = 0.042
|
|
# Jev reads about 32000 tokens of state; a larger unit is not sent and makes the review YELLOW.
|
|
max_state_chars = 60000
|
|
max_units = 80
|
|
timeout_seconds = 20
|
|
pr_doc_globs = ["doc/prs/20*-PR#*.md"]
|
|
# Choice and score answers below this confidence do not count for the verdict; calibrate it with tools/jev-eval.
|
|
min_confidence = 0.3
|
|
|
|
# Every question asks for a problem: a value above red_above gives ORANGE, above yellow_above YELLOW.
|
|
# A score's criteria run from "no problem" to "worst problem"; its value is the normalized expected level.
|
|
|
|
[[classifier.scenario_questions]]
|
|
name = "hard_to_read"
|
|
type = "score"
|
|
instructions = """
|
|
`scenario` is a Gherkin scenario from a pull-request document; `feature` and `background` give its context.
|
|
Domain experts who do not program approve these scenarios as the specification.
|
|
How hard is the scenario to understand for such a domain expert?
|
|
The text may be German or English.
|
|
"""
|
|
criteria = [
|
|
"A domain expert understands the scenario at the first reading.",
|
|
"A domain expert understands it, but must read some steps twice.",
|
|
"A domain expert needs a developer to explain parts of it.",
|
|
"A domain expert cannot tell what the scenario requires.",
|
|
]
|
|
red_above = 0.75
|
|
yellow_above = 0.5
|
|
|
|
[[classifier.scenario_questions]]
|
|
name = "more_than_one_aspect"
|
|
type = "noul"
|
|
instructions = """
|
|
Does the Gherkin scenario in `scenario` specify more than one independent business rule or aspect,
|
|
so that it should be split into two or more scenarios?
|
|
Variants of the same rule that differ only in their data count as one aspect.
|
|
A second When with its own Then that checks a different case counts as a second aspect.
|
|
"""
|
|
red_above = 0.85
|
|
yellow_above = 0.6
|
|
|
|
[[classifier.scenario_questions]]
|
|
name = "implementation_details"
|
|
type = "noul"
|
|
instructions = """
|
|
Does the Gherkin scenario in `scenario` describe implementation details instead of business behaviour,
|
|
for example class, method, table or column names, SQL, internal HTTP headers or JSON internals?
|
|
Naming the public API endpoint, the HTTP method, the response status and the business fields
|
|
counts as observable behaviour, not as an implementation detail.
|
|
"""
|
|
red_above = 0.85
|
|
yellow_above = 0.6
|
|
|
|
[[classifier.scenario_questions]]
|
|
name = "given_when_then_misused"
|
|
type = "noul"
|
|
instructions = """
|
|
Does the Gherkin scenario in `scenario` misuse Given, When or Then?
|
|
Misuse is: an action or API call in Given, an expected result or check in When,
|
|
an action in Then, or a Then without a preceding When.
|
|
And-steps belong to the keyword above them.
|
|
"""
|
|
red_above = 0.85
|
|
yellow_above = 0.6
|
|
|
|
[[classifier.scenario_questions]]
|
|
name = "vague_then"
|
|
type = "noul"
|
|
instructions = """
|
|
Does a Then step of the Gherkin scenario in `scenario` lack a checkable outcome,
|
|
for example "works", "is handled correctly" or "behaves as expected",
|
|
without saying what an observer can check?
|
|
"""
|
|
red_above = 0.85
|
|
yellow_above = 0.6
|
|
|
|
[[classifier.section_questions]]
|
|
name = "hard_to_understand"
|
|
type = "score"
|
|
instructions = """
|
|
`section` is a section of a pull-request document that a human reviewer reads before the code.
|
|
How hard is the section to understand for a reviewer who knows the project but not this change?
|
|
Markdown lists, tables, links and code identifiers are normal here.
|
|
The text may be German or English.
|
|
"""
|
|
criteria = [
|
|
"The reviewer understands the section at the first reading.",
|
|
"The reviewer understands it, but must read some sentences twice.",
|
|
"The reviewer must guess what some sentences mean.",
|
|
"The reviewer cannot tell what the section says.",
|
|
]
|
|
red_above = 0.75
|
|
yellow_above = 0.5
|
|
|
|
[[classifier.section_questions]]
|
|
name = "slop"
|
|
type = "score"
|
|
instructions = """
|
|
How much of `section` is slop: filler words, mannered or promotional prose,
|
|
narration of the implementation instead of the reasons, or repetition of what other sentences already say?
|
|
Plain lists, tables, links and code identifiers are not slop.
|
|
The text may be German or English.
|
|
"""
|
|
criteria = [
|
|
"No slop: every sentence carries information.",
|
|
"A few filler words or one redundant sentence.",
|
|
"Several sentences are filler, mannered prose, narration or repetition.",
|
|
"The section is mostly filler, mannered prose, narration or repetition.",
|
|
]
|
|
red_above = 0.75
|
|
yellow_above = 0.5
|
|
|
|
[path_judge]
|
|
# Jev judges every changed file: only a change that touches security or weakens the checks waits for a human.
|
|
judge = "jev"
|
|
scope = "all"
|
|
|
|
[excerpt]
|
|
# An economy model selects the lines of an excerpt; on 4 measured heads it never found less than the fixed window.
|
|
select = "model"
|