werkrevisor 0.19.1: billed calls as unknown repository, dry runs as local review #28
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: mi/werkrevisor#28
Reference in New Issue
Block a user
See docs/prs/2026-10-06-PR#28-pseudo-repositories.md.
🤖 Generated with Claude Code
🟡 werkrevisor YELLOW for
d4f112a, superseded by the result for754ee60🟡 werkrevisor: YELLOW / HUMAN_REVIEW
Change kind:
mixed.Reasons
Findings
[medium] Local rows missing from sums per period at
server/src/main/kotlin/de/hoennig/werkrevisor/server/Statistics.kt:203-220, hint, as a line comment[medium] Local flag stored as boolean in integer column, old rows null at
server/src/main/kotlin/de/hoennig/werkrevisor/server/UsageStore.kt:172, hint, as a line comment[low] Document mentions a token not in the diff at
docs/prs/2026-10-06-PR#28-pseudo-repositories.md:79, hint, as a line commentClassifier
8 of 8 units passed all questions (1 PR document(s)).
Usage: model `openrouter/anthropic/claude-sonnet-5.5`, LLM done, tokens counted before the call 68960, calls 1, tokens in 67863 / out 867 (reasoning included in out), duration 19.5 s, cost 0.1444 USD (reported), classifier `jev-1.13.0` calls 8, tokens in 6641, cost 0.0003 USD, finding judge `jev-1.13.0` calls 2, tokens in 2867, cost 0.0001 USD, build file judge `jev-1.13.0` calls 1, tokens in 560, cost 0.0000 USD, config: defaults + host + repository (.werkrevisor.toml at base1 answer ignored because of a confidence below [classifier] min_confidence.
f585223).Reviewed head `d4f112a5e415ccf9307f295707e226cc65bc00ca` against base `f5852232e0b7cd9959c922c9ef081d90c029daf3`, werkrevisor 0.19.0.
werkrevisor: YELLOW / HUMAN_REVIEW for
d4f112a5e415ccf9307f295707e226cc65bc00ca.The result comment of werkrevisor holds the details.
@@ -0,0 +76,4 @@- Each group of the table per repository holds `kind`: `repository`, `private`, `unknown` for the billed callsor `local` for the dry runs; the Web Component names the last three in grey italics.- The cards name the local reviews, e.g. "26 reviews (1 local)", and the recent reviews mark them with "local".- The service gets a source token for dry runs; the operator keeps it in `WERKREVISOR_STATS_SOURCE_TOKEN_LOCAL`.[low] Document mentions a token not in the diff (werkrevisor)
hint
The solution section says the service gets a source token for dry runs, kept in
WERKREVISOR_STATS_SOURCE_TOKEN_LOCAL. No code in this PR adds or uses this variable; it is configuration of the operator. The statement is unclear about whether the PR changes anything for it.@@ -183,6 +200,25 @@ class StatisticsService(.filterValues { runs -> runs.maxBy { it.timestamp }.repoPrivate == false }.keysprivate fun byRepo([medium] Local rows missing from sums per period (werkrevisor)
hint; Jev: real 0.22, severity 0.8 of 3
byRepocomputes the "local" row fromlastMonthrecords and lists repository rows with!it.local. The row "unknown repository" shows billed calls with 0 reviews. This is consistent, butbyVerdictstill includes local records andbyModeltoo, while the table per verdict counts local reviews as pull requests viagroups(...)andSums.of. SinceSums.ofexcludes local records frompullRequests, this is fine. The remaining uncertainty iscostPerPullRequestUsd, which subtractslocalCostUsdfromllmCostUsd + jevCostUsd.localCostUsdusescountedCostUsd, which equals the counted part, so the result is correct. No defect is confirmed here.@@ -162,6 +169,7 @@ class UsageStore(llmOutputTokens = rs.getLong("llm_output_tokens").takeUnless { rs.wasNull() },jevInputTokens = rs.getLong("jev_input_tokens").takeUnless { rs.wasNull() },repoPrivate = rs.getBoolean("repo_private").takeUnless { rs.wasNull() },local = rs.getBoolean("local"),[medium] Local flag stored as boolean in integer column, old rows null (werkrevisor)
hint; Jev: real 0.23, severity 1.8 of 3
localis added as anintegercolumn, so old rows hold NULL.rs.getBoolean("local")returns false for NULL, which is correct. However, therun_keydoes not includelocal: a dry run and a CI run of the same repo, PR, head and timestamp are practically unique. A real risk is low; this is only a note that a local run reporting the same key as a CI run would be dropped.Assessment of the werkrevisor hints (as of the last commit):
Fixed:
Not changed:
Statistics.kt:203-220, local rows in the period sums: the hint itself finds no defect; periods, verdicts and models count local reviews in costs and reviews,Sums.ofleaves them out of the pull requests, and the cost per pull request subtracts their counted costs.UsageStore.kt:172,localin an integer column: old rows hold NULL, whichgetBooleanreads as false, as intended. The run key holds repository, PR, head and timestamp in seconds; a dry run and a CI run with all four equal do not occur in practice.werkrevisor/cli.pygiven only as diff: the change is one line that marks a run without publishing as local; a test covers both cases.The judged build path works: the version bump in
server/build.gradle.ktsis now a hint instead of YELLOW (weakens checks 0.02, risk 0.00).🤖 Generated with Claude Code
🟡 werkrevisor: YELLOW / HUMAN_REVIEW
Change kind:
mixed.Reasons
Findings
[medium] Local dry runs reuse the CI run key at
server/src/main/kotlin/de/hoennig/werkrevisor/server/UsageStore.kt:38, hint, as a line comment[medium] Migration adds
localas nullable integer without a default atserver/src/main/kotlin/de/hoennig/werkrevisor/server/UsageStore.kt:52, hint, as a line comment[low] Truthiness precedence makes a hidden row depend on
PSEUDO_REPOSatserver/src/main/resources/static/werkrevisor-dashboard.js:227-229, hint, as a line commentClassifier
8 of 8 units passed all questions (1 PR document(s)).
Usage: model `openrouter/anthropic/claude-sonnet-5.5`, LLM done, tokens counted before the call 69006, calls 1, tokens in 67900 / out 764 (reasoning included in out), duration 16.7 s, cost 0.1434 USD (reported), classifier `jev-1.13.0` calls 8, tokens in 6663, cost 0.0003 USD, finding judge `jev-1.13.0` calls 2, tokens in 2521, cost 0.0001 USD, build file judge `jev-1.13.0` calls 1, tokens in 560, cost 0.0000 USD, config: defaults + host + repository (.werkrevisor.toml at base1 answer ignored because of a confidence below [classifier] min_confidence.
f585223).Reviewed head `754ee60c44517f32dac0b4d868c423a7a970e1c5` against base `f5852232e0b7cd9959c922c9ef081d90c029daf3`, werkrevisor 0.19.0.
werkrevisor: YELLOW / HUMAN_REVIEW for
754ee60c44517f32dac0b4d868c423a7a970e1c5.The result comment of werkrevisor holds the details.
@@ -33,1 +35,4 @@// Without the unknown LLM part, as the sums count it.val countedCostUsd get() = if (costKnown) totalCostUsd else jevCostUsdval runKey get() = "$repo#$pr@${headSha ?: ""}@$timestamp"[medium] Local dry runs reuse the CI run key (werkrevisor)
hint; Jev: real 0.36, severity 2.0 of 3
runKeyis built from repo, pr, head_sha and timestamp, and the newlocalflag is not part of it. A dry run and a CI run of the same PR commit differ only in the timestamp, so a collision is unlikely, butinsert or ignorewould drop whichever record arrives second if the timestamps are equal. The risk is small. Addinglocalto the key would remove it.@@ -45,3 +50,3 @@)private val ADDED_COLUMNS = setOf("llm_input_tokens", "llm_output_tokens", "jev_input_tokens", "repo_private")private val ADDED_COLUMNS = setOf("llm_input_tokens", "llm_output_tokens", "jev_input_tokens", "repo_private", "local")[medium] Migration adds
localas nullable integer without a default (werkrevisor)hint; Jev: real 0.62, severity 2.0 of 3
Rows written before this PR get NULL in
local.rs.getBooleanreturns false for NULL, so this is safe today. However,ADDED_COLUMNScreates every column asintegerwith no default, so the schema does not state the meaning of old rows. Missing test: the migration testa database of an earlier version gets the token columnsdoes not check thatlocalreads as false for old rows.@@ -222,3 +225,3 @@{},// A private repository keeps its row, so that the rows add up to the total, but not its name.g.hidden ? el("td", { class: "key hidden" }, "private repository") : el("td", { class: "key" }, name(g.key)),PSEUDO_REPOS[g.kind] || g.hidden[low] Truthiness precedence makes a hidden row depend on
PSEUDO_REPOS(werkrevisor)hint
PSEUDO_REPOS[g.kind] || g.hidden ? A : Bworks as intended, but thekindlookup is also applied to the verdict table, whereg.kindis alwaysrepository. The code is correct, but the sharedgroupTablenow carries repository-specific logic. This is a maintainability note only.