Skip to content

Latest commit

 

History

History
70 lines (57 loc) · 4.7 KB

TODO.md

File metadata and controls

70 lines (57 loc) · 4.7 KB

TODO — srcmorph reactor (Maven plugin now srcmorph-maven-plugin, formerly llamacpp-ai-index-maven-plugin)

Open work items for this repo. Cross-cutting tracking lives in ../workspace/crossrepostatus.md; items here are repo-specific or this repo's slice of a cross-cutting initiative.

Completed work is not recorded here. It lives in git history and in crossrepostatus.md; a finished item is deleted from this file rather than annotated, so everything below is genuinely still open. The plan's table keeps its row numbers (other notes name them): a finished row is deleted, the numbers of the others stay.

Plan (2026-11-05): every open item

Each row is one item of the sections below (the "Details" column names its section). Who: Claude = can be done without the owner, Owner = waits for a decision or an action of the owner (the proposal is the default once the owner agrees), External = waits for something outside this repository.

# Item Proposal Who Details
1 CodeQL #270, awaitNanos ignored false positive: dismiss it on GitHub Owner PR #269
3 SonarCloud: the analysis of PR #269 reads, that of main does not: main was analysed as a short branch (Branch name: main, type: short in the run of 2026-11-04), the main branch master last three months ago the owner renamed the main branch to main (2026-11-07); left: one analysis of main as the long branch (the push of the merge of #269, or a re-run of the last SonarQube run on main), then its issues are read Owner PR #269
38 Use jllama's ChatStreamAssembler (not in llama 6.0.0) switch with its release External Other open items
40 Merge PR #269 merge Owner PR #269

Open

PR #269 (branch feat/profiles-v3-kinds-and-safety)

  • CI green but the Claude review action and FOSSA; the owner merges, then this section goes.
  • CodeQL #270 (row 1, RequestRateLimiter: the return value of awaitNanos ignored) is a false positive: the loop in acquire() re-checks the window after every wake-up; the owner dismisses it.
  • SonarCloud (row 3): once main is analysed as the long branch, its issues are read. Of PR #269's code smells left: the cognitive complexity of nine methods and an eight-parameter method (refactors to report, not to apply), "test methods should not be public" (the owner's guides decide the other way) and a few false positives (comments read as code).

Known limits, for the next model round (no open work)

Measured and written up; listed so a later model round checks them again:

  • gpt-oss-20b's map-reduce loses the middle of a large file at random (C2, remeasure-2026-11-06/); a file of that size wants a stronger model.
  • gpt-oss-20b leaves a batch file's goto labels out with either prompt (37 %); GPT-6 Luna summarises a Log4j 2 file's dotted keys as groups with the ops prompt (properties 93 % against 100 % through the fallback) (remeasure-2026-11-07/).
  • The self-test's huge-pom keeps the generic prompt: no real POM of a megabyte was found to measure file-body-xml on (big-pom takes it since C3).
  • The secrets rule's properties pattern is a heuristic: it misses a plural or a qualifier after the key (passwords=, password.prod=), credentials in a URL, an account-key and a value on a continuation line, and skips a message bundle that labels a password field.
  • plot.py's heatmap counts "Seed counts" from numbersStated also for a failed row and divides by four; score.summarize does neither. No row of the card is affected.

Other open items

  • Use jllama's ChatStreamAssembler instead of srcmorph's own stream assembly. The stream setting of the openai-compatible provider needs a Server-Sent Events body assembled into a chat completion; java-llama.cpp has that as net.ladenthin.llama.json.ChatStreamAssembler (bernardladenthin/java-llama.cpp#595), but in no release yet, so OpenAiCompatibleProviderSupport.assembleStream and its constants are a copy, marked @ReplaceWithLibrary. When net.ladenthin:llama is raised to the release that carries it, ReplaceWithLibraryTest fails: let completionJson call ChatStreamAssembler.parseStream (toCompletionJson, or hand its ChatResponse on directly) and turn its InvalidStreamException into the IOException the provider throws; delete the marked code, its tests in OpenAiCompatibleProviderSupportTest, this item, and, being the last marker, ReplaceWithLibrary with its test. The stub-server tests in OpenAiCompatibleGenerationProviderTest stay and prove the switch.