Upgrade Qwen3-Coder model and refactor package-body prompt for child summaries - #117
Merged
Conversation
Add a single aiDefinition (key Qwen3-Coder-30B-A3B-Instruct-UD-Q4_K_XL-C64k) pointing at the locally-present 17.7 GB UD-Q4_K_XL GGUF, and reference it from both the ai-generate (file-body) and ai-aggregate-packages (package-body) executions. - contextSize 65536: generous but CPU-feasible (native 262144; 256k prefill is unusable on CPU). - maxOutputTokens 1536: short dense bullet summaries, not long code gen. - Official Qwen3-Coder sampling: temp 0.7 / top_p 0.8 / top_k 20 / repeat 1.05. - Non-thinking model: no enable_thinking flag set. - Remove the two stale Qwen3-Coder Q4_K_M entries (C32k/C256k) that pointed at a file not present locally; update the model-selection legend accordingly. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017i6A5bMhgCSbiKibQhxwuz
Swap the Java-centric prose file-summary prompt for a dense, section-based multi-language indexer prompt (Java/C/C++/OpenCL/SQL) that emits a blockquote lead plus fixed '####' sections (Purpose, Type, Input, Output, Core logic, Public API, Dependencies, Exceptions/Errors). The File/Source %s placeholders are unchanged. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017i6A5bMhgCSbiKibQhxwuz
Swap the Java-centric prose package-summary prompt for a dense, section-based multi-language indexer prompt that synthesizes the per-file summaries into a blockquote lead plus fixed '####' sections (Purpose, Responsibilities, Key units, Data flow, Dependencies, Cross-cutting). Both %s placeholders are preserved (now labeled Package / File summaries). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017i6A5bMhgCSbiKibQhxwuz
The package-body prompt expects the already-generated per-file summaries, but buildPackageSourceText only emitted the package header plus a '#### Contents' list of child file *names*. The model therefore had nothing to synthesise. Embed each child .ai.md's actual AI body (per-file and sub-package summaries) into the %s package source, each under a '### <name>' heading so the model can attribute which summary belongs to which file. Only the child body is inlined; the single-letter child header metadata is deliberately omitted to save context. - Sub-package children reuse their aggregated package.ai.md body, labelled with a trailing '/'. - Children with a blank/unreadable body get a '(no summary available)' note instead of a fabricated summary. - When no child yields a usable body, fall back to the existing '#### Contents' name listing so the source is never empty. - No bespoke truncation: the combined source flows through the existing AiPromptPreparationSupport.preparePrompt trim/warnOnTrim budget logic. - The deterministic '#### Contents' fallback for the written package body is unchanged; package checksums (child name|c|d|x) are unaffected, so this does not trigger spurious regeneration. The package-body prompt text is unchanged — the code now matches the prompt. Add two PackageIndexerTest cases proving the %s source contains the real child bodies (per-file and bubbled-up sub-package), via a capturing provider. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017i6A5bMhgCSbiKibQhxwuz
README's TODO claimed PIT targetClasses was 'narrowed to a single class (AiCompletionParser)' — stale; the pom now gates an explicit 21-class list (~146 mutations). Update it and add the test-compile invocation + policy link. Add a 'PIT Mutation Testing' pointer to the workspace policy in CLAUDE.md. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017i6A5bMhgCSbiKibQhxwuz
…ifics Slim the CLAUDE.md PIT pointer to the uniform cross-repo form. Trim the README TODO to the repo-specific scope-expansion roadmap (toward the whole aiindex.* tree) + a policy link, dropping drift-prone counts and generic mechanics. Trim the pom PIT comment to a policy pointer plus the repo-specific rationale (whole-tree targetTests, pure-Java MockAiGenerationProvider) and drop the stale 'pitest-maven 1.25.4'. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017i6A5bMhgCSbiKibQhxwuz
bernardladenthin
had a problem deploying
to
startgate
June 25, 2026 07:19 — with 
GitHub Actions
Error
GitHub Actions
Error
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Model upgrade: Replace Qwen3-Coder-30B-A3B-Instruct-Q4_K_M (32K context) with the newer UD-Q4_K_XL variant (64K context, 17.7 GB), which is locally present and serves as the active model for both file and package summarization phases. Reduce
maxOutputTokensfrom 2048 to 1536 to keep summaries dense and on-topic.Package-body prompt redesign: Rewrite the
package-bodyprompt template to be a structured markdown indexer (matching the newfile-bodytemplate style) that synthesizes summaries from already-generated child.ai.mdbodies rather than raw source code. This reduces context usage and improves coherence.PackageIndexer enhancement: Implement child-summary embedding in
buildPackageSourceText(). The method now reads the AI body of each child.ai.mdfile (per-file and sub-package summaries) and embeds them under labelled headings in the package source text fed to the AI model. Falls back to a plain contents listing if no child bodies are available. Includes comprehensive Javadoc and new helper methods (appendChildSummaries,appendChildSummary,readChildBody,toChildDisplayName).Test coverage: Add two new integration tests (
aggregate_childFileBodies_areEmbeddedInPackageSourceText,aggregate_subPackageBody_isEmbeddedInParentSourceText) and aCapturingProvidertest fixture to verify child summaries are correctly embedded and labelled in the package source text.Documentation: Update PIT mutation-testing guidance in
pom.xmlto reference an external policy document. Add PIT section toCLAUDE.mdwith invocation instructions. UpdateREADME.mdTODO to reflect the new model and prompt design.Test plan
PackageIndexerTest.aggregate_childFileBodies_*andaggregate_subPackageBody_*)PackageIndexerTesttests pass (no regression in file/package aggregation)CLAUDE.md,README.md,pom.xmlcomments)Related issues / PRs
None.
Checklist
CONTRIBUTING.mdandCODE_OF_CONDUCT.mdhttps://claude.ai/code/session_017i6A5bMhgCSbiKibQhxwuz