werkrevisor asks the risk categories that fit werkprobe #12
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: mi/werkprobe#12
Reference in New Issue
Block a user
See the PR document.
🤖 Generated with Claude Code
🟡 werkrevisor YELLOW for
190f9cd, superseded by the result for1b349e6🟡 werkrevisor: YELLOW / HUMAN_REVIEW
Change kind:
mixed.Reasons
Findings
[low] Comment narrates measurement at
.werkrevisor.toml:8, hint, as a line comment[low] Contradiction with PR description at
.werkrevisor.toml:6-7, hint, as a line commentClassifier
0 of 0 units passed all questions (0 PR document(s)).
Usage: model `openrouter/anthropic/claude-sonnet-5.5`, LLM done, tokens counted before the call 5131, calls 1, tokens in 5000 / out 417 (reasoning included in out), duration 9.9 s, cost 0.0142 USD (reported), classifier `jev-latest` calls 0, tokens in 0, cost 0.0000 USD, finding judge `none` calls 0, tokens in 0, cost 0.0000 USD, phases: context 0.1 s, model_review 9.6 s, finding_judge 0.0 s, config: defaults + host.Reviewed head `190f9cd70a4808e1872ec44e27ad9cf8d34b3b5a` against base `1a1006ae0f1e739b7a350aa6663277f84f4a5b0e`, werkrevisor 0.30.0.
werkrevisor: YELLOW / HUMAN_REVIEW for
190f9cd70a4808e1872ec44e27ad9cf8d34b3b5a.The result comment of werkrevisor holds the details.
@@ -0,0 +3,4 @@# The host config on the review machine holds all other rules; see https://git.javagil.de/mi/werkrevisor.[path_judge]# The host config sets scope = "all", so Jev asks each changed file the risk categories of this template,[low] Contradiction with PR description (werkrevisor)
hint
The description says the 8 categories are switched off in this file and that the host config drops its
risk_templateafter the merge. The file comment says the host config setsscope = "all", which the description does not mention. The wording "as changed below" is also unclear. Until the host config is changed, the template is set in two places, and the description does not say which one wins.Fixed in
1b349e6: the comment says which categories Jev asks and that this choice replaces the host'srisk_template; the PR description now names both, and the host'sscope = "all".@@ -0,0 +5,4 @@[path_judge]# The host config sets scope = "all", so Jev asks each changed file the risk categories of this template,# as changed below.# The 9 merged PRs reached no threshold; the highest value was architecture 0.68.[low] Comment narrates measurement (werkrevisor)
hint
The comment about the 9 merged PRs and the architecture value 0.68 is a snapshot of a one-time measurement. It does not explain the configuration and becomes stale as soon as more PRs are reviewed. It belongs in the PR description only.
Fixed in
1b349e6: the measurement is only in the PR description now.Review round of
190f9cd7: both hints fixed in1b349e6. The YELLOW comes from.werkrevisor.toml, which always needs the approval of the human user.🟡 werkrevisor YELLOW for
1b349e6, superseded by the result fora2f71ce🟡 werkrevisor: YELLOW / HUMAN_REVIEW
Change kind:
mixed.Reasons
Classifier
0 of 0 units passed all questions (0 PR document(s)).
Usage: model `openrouter/anthropic/claude-sonnet-5.5`, LLM done, tokens counted before the call 5182, calls 1, tokens in 5044 / out 94 (reasoning included in out), duration 5.9 s, cost 0.0110 USD (reported), classifier `jev-latest` calls 0, tokens in 0, cost 0.0000 USD, phases: context 0.1 s, model_review 5.6 s, config: defaults + host.Reviewed head `1b349e68379f2633e092d49b6f51ff4b41062182` against base `1a1006ae0f1e739b7a350aa6663277f84f4a5b0e`, werkrevisor 0.30.0.
🟡 werkrevisor YELLOW for
a2f71ce, superseded by the result fora443120🟡 werkrevisor: YELLOW / HUMAN_REVIEW
Change kind:
mixed.Reasons
Findings
[medium] PR document contradicts AGENTS.md workflow at
docs/prs/2026-10-09-PR#12-risk-categories.md:27, hint, as a line comment[low] Comment says "has no users" but switched-off categories include
business_invariantsat.werkrevisor.toml:32-43, hint, as a line commentClassifier
4 of 4 units passed all questions (1 PR document(s)).
Usage: model `openrouter/anthropic/claude-sonnet-5.5`, LLM done, tokens counted before the call 8380, calls 1, tokens in 8211 / out 533 (reasoning included in out), duration 10.6 s, cost 0.0218 USD (reported), classifier `jev-1.13.0` calls 4, tokens in 3022, cost 0.0001 USD, finding judge `jev-1.13.0` calls 1, tokens in 1351, cost 0.0001 USD, phases: classifier 0.4 s, context 0.1 s, model_review 9.6 s, finding_judge 0.3 s, config: defaults + host.1 answer ignored because of a confidence below [classifier] min_confidence.
Reviewed head `a2f71ce59470df748bdc55b84330202da2acd044` against base `1a1006ae0f1e739b7a350aa6663277f84f4a5b0e`, werkrevisor 0.30.0.
werkrevisor: YELLOW / HUMAN_REVIEW for
a2f71ce59470df748bdc55b84330202da2acd044.The result comment of werkrevisor holds the details.
@@ -0,0 +29,4 @@name = "api_compatibility"action = "off"# Its databases live in throwaway containers: no data of an installation can be lost or need a correction.[low] Comment says "has no users" but switched-off categories include
business_invariants(werkrevisor)hint
The comment above the first five categories says "no users, logins or request input", while
business_invariantsis switched off under the "throwaway containers" comment. Invariants of the application logic are independent of where the data lives, so a Spring app with domain logic can still violate them. The justification in the file does not support switching this category off, and the PR document gives no further reasoning.Fixed in
a443120:business_invariantshas its own reason, in the file and in the PR document: werkprobe has no business rules, its application only reports which server it reached and that the init script ran.@@ -0,0 +24,4 @@`external_integration`, and the hints `performance` and `architecture`.- `AGENTS.md` describes the project, the build, the writing style and the pull-request workflow, as in werkcrappit;`CLAUDE.md` imports it.- From this PR on, each PR has a document under `docs/prs/`.[medium] PR document contradicts AGENTS.md workflow (werkrevisor)
hint; Jev: real 0.79, severity 1.2 of 3
The Solution section says "From this PR on, each PR has a document under
docs/prs/", but the document is namedPR#12and its Problem section says werkprobe has "no PR documents". That agrees with the diff. However,AGENTS.mdintroduces a workflow that requires## Open Questionsand## Follow-up PRssections in the PR document, and the document has neither. This is harmless today, but the first PR document already deviates from the rules it introduces.AGENTS.mddoes not require the two sections; they hold open questions and follow-ups when there are any.a443120says so inAGENTS.md, so the wording cannot be read as a requirement.Review round of
a2f71ce5: both hints fixed ina443120. The YELLOW comes from.werkrevisor.toml, which always needs the approval of the human user.🟡 werkrevisor YELLOW for
a443120, superseded by the result for50066b5🟡 werkrevisor: YELLOW / HUMAN_REVIEW
Change kind:
mixed.Reasons
Findings
[low] Contradiction with AGENTS.md at
docs/prs/2026-10-09-PR#12-risk-categories.md:16-18, hint, as a line comment[low] Unclear Measurement Section at
docs/prs/2026-10-09-PR#12-risk-categories.md:30-35, hint, as a line commentClassifier
4 of 4 units passed all questions (1 PR document(s)).
Usage: model `openrouter/anthropic/claude-sonnet-5.5`, LLM done, tokens counted before the call 8005, calls 1, tokens in 7821 / out 531 (reasoning included in out), duration 9.3 s, cost 0.0210 USD (reported), classifier `jev-1.13.0` calls 4, tokens in 3030, cost 0.0001 USD, finding judge `none` calls 0, tokens in 0, cost 0.0000 USD, phases: classifier 0.2 s, context 0.1 s, model_review 8.6 s, finding_judge 0.0 s, config: defaults + host.1 answer ignored because of a confidence below [classifier] min_confidence.
Reviewed head `a443120af47b3ea2793c0de7ecbbab01f0a6fbaa` against base `1a1006ae0f1e739b7a350aa6663277f84f4a5b0e`, werkrevisor 0.30.0.
werkrevisor: YELLOW / HUMAN_REVIEW for
a443120af47b3ea2793c0de7ecbbab01f0a6fbaa.The result comment of werkrevisor holds the details.
@@ -0,0 +13,4 @@## The Solution- `.werkrevisor.toml` chooses `jvm-spring-backend`.[low] Contradiction with AGENTS.md (werkrevisor)
hint
The PR document says
.werkrevisor.tomlreplaces therisk_templateof the host config, "which drops it after the merge". AGENTS.md states that.werkrevisor.tomlis read at the head ofmain, never from a PR. So the choice has no effect before the merge. The document does not say clearly whether this PR's own review uses the old or the new categories. The Measurement section says nothing about which categories this PR itself was judged with.Fixed in
50066b5: the PR document says that this PR's own review still asks all 16 categories.@@ -0,0 +27,4 @@`CLAUDE.md` imports it.- From this PR on, each PR has a document under `docs/prs/`.## Measurement[low] Unclear Measurement Section (werkrevisor)
hint
The Measurement section ends with "Nine PRs are too few to tune a threshold, so this PR changes none." The document never says that thresholds are tunable here, or what the default thresholds are. The value
data_correction0.54 is shown for a category this PR switches off, and the section does not explain why it is listed. A reader cannot tell from this section why the categories chosen here are right.Fixed in
50066b5: the section names the thresholds, says that the switched-off categories hide no finding of the past PRs, and that.werkrevisor.tomlcould change a threshold.Review round of
a443120a: both hints fixed in50066b5. The YELLOW comes from.werkrevisor.toml, which always needs the approval of the human user.🟡 werkrevisor: YELLOW / HUMAN_REVIEW
Change kind:
mixed.Reasons
Findings
[low] Contradiction with AGENTS.md at
docs/prs/2026-10-09-PR#12-risk-categories.md:16-18, hint, as a line comment[low] Inconsistent PR document rule at
AGENTS.md:74-78, hint, as a line commentClassifier
4 of 4 units passed all questions (1 PR document(s)).
Usage: model `openrouter/anthropic/claude-sonnet-5.5`, LLM done, tokens counted before the call 7957, calls 1, tokens in 7769 / out 513 (reasoning included in out), duration 10.6 s, cost 0.0207 USD (reported), classifier `jev-1.13.0` calls 4, tokens in 3133, cost 0.0001 USD, finding judge `none` calls 0, tokens in 0, cost 0.0000 USD, phases: classifier 0.3 s, context 0.2 s, model_review 9.8 s, finding_judge 0.0 s, config: defaults + host.Reviewed head `50066b5c272df0ca0d30150e5814b035315cf7c2` against base `1a1006ae0f1e739b7a350aa6663277f84f4a5b0e`, werkrevisor 0.30.0.
werkrevisor: YELLOW / HUMAN_REVIEW for
50066b5c272df0ca0d30150e5814b035315cf7c2.The result comment of werkrevisor holds the details.
@@ -0,0 +71,4 @@- YELLOW and ORANGE wait for the approval of the human user.- RED needs a new commit that fixes the cause.### PR Document[low] Inconsistent PR document rule (werkrevisor)
hint
The "PR Document" section says each open question without
[x]gives RED, but it never names the heading## Open Questionsformat or the section## Follow-up PRsas required structure for the PR document. The new PR document contains neither section, so the rule is unclear on whether they must be present. The text states both sections are optional, but the first sentence then talks about items "under## Open Questions" as if it is defined elsewhere.No change: the section names both headings and says that both are optional;
## Open Questionsis the heading werkrevisor's host config looks for.@@ -0,0 +13,4 @@## The Solution- `.werkrevisor.toml` chooses `jvm-spring-backend`.[low] Contradiction with AGENTS.md (werkrevisor)
hint
The PR document says "AGENTS.md describes ... as in werkcrappit" and that "each PR has a document", but its text claims werkprobe has "no agent instructions" in the problem section while the host config is described as choosing the template. More concretely, the document states the host config "drops"
risk_templateafter the merge, which is an unverified claim about external config. Reviewers cannot confirm it from this repository.No change: the problem section describes the state before this PR, when werkprobe had no agent instructions. Dropping
risk_templatefrom the host config is the operator's step after the merge; it cannot be seen in this repository, and until then the repository's choice wins.Review round of
50066b5c: both hints answered without a change. The YELLOW comes from.werkrevisor.toml, which always needs the approval of the human user.