CodeStory packet capability: bounded first slice
The user authorized a new research and implementation effort on 4 September 2026. The goal is unchanged: one repository question produces a compact, source-backed evidence packet that materially reduces host exploration without lowering correctness. Prior compiler, ETR-1 and STR-1 stops and receipts remain immutable.
This first PR-sized slice tests source-only reference-packet usability. It does not integrate production behavior or select a model. Reuse the frozen fragments, publication metadata, optimizer and burned development questions. Reauthenticate source. Assemble actual <=16-row/16-KiB packets. Keep oracle annotations and selection receipts out of the reader input. Run a no-tool fixed reader, validate its execution and citation shape, and give source-supported answers to the existing independent verifier.
Use the first two case IDs in each of the four frozen groups, original wording only, for the initial eight-case diagnostic. This sampling rule is fixed before any reader output. A packet-reference result is not retrieval, selection, product or release qualification. No changes to old questions, annotations, source, public budgets, or old decisions.
Immediate budget: focused implementation/tests and a synthetic no-tool canary first. At most sixteen diagnostic model requests in the initial slice, serial or bounded pairs, 120 seconds per request; stop the batch on an execution violation. No corpus request before a clean pushed source head. No production rewrite, new training, model download, broad proof, or release gate in this slice. Reuse one implementer and the existing adversarial verifier. Subsequent work must identify the causal gap and specified mechanism before more source or model runs; no equivalent retry or example-level tuning.
Owning code: benchmark scripts and focused Node tests, stacked from STR-1 only to reuse experiment helpers. Existing production source remains unchanged. Reader capability and local candidate inference are distinct: the diagnostic reader may use the configured Codex model; this does not establish local engine latency. The reference packet includes source rows, not oracle labels, annotated explanations, or generated answer claims.
Next permitted mutation: create the child issue and isolated branch, add the source-only reader boundary tests, then implement the smallest preparation/reader path. The research/implementation goal remains active beyond this bounded slice; unsuccessful mechanisms are recorded and replaced only on new causal evidence within the newly authorized effort.
CodeStory packet capability: bounded first slice
The user authorized a new research and implementation effort on 4 September 2026. The goal is unchanged: one repository question produces a compact, source-backed evidence packet that materially reduces host exploration without lowering correctness. Prior compiler, ETR-1 and STR-1 stops and receipts remain immutable.
This first PR-sized slice tests source-only reference-packet usability. It does not integrate production behavior or select a model. Reuse the frozen fragments, publication metadata, optimizer and burned development questions. Reauthenticate source. Assemble actual <=16-row/16-KiB packets. Keep oracle annotations and selection receipts out of the reader input. Run a no-tool fixed reader, validate its execution and citation shape, and give source-supported answers to the existing independent verifier.
Use the first two case IDs in each of the four frozen groups, original wording only, for the initial eight-case diagnostic. This sampling rule is fixed before any reader output. A packet-reference result is not retrieval, selection, product or release qualification. No changes to old questions, annotations, source, public budgets, or old decisions.
Immediate budget: focused implementation/tests and a synthetic no-tool canary first. At most sixteen diagnostic model requests in the initial slice, serial or bounded pairs, 120 seconds per request; stop the batch on an execution violation. No corpus request before a clean pushed source head. No production rewrite, new training, model download, broad proof, or release gate in this slice. Reuse one implementer and the existing adversarial verifier. Subsequent work must identify the causal gap and specified mechanism before more source or model runs; no equivalent retry or example-level tuning.
Owning code: benchmark scripts and focused Node tests, stacked from STR-1 only to reuse experiment helpers. Existing production source remains unchanged. Reader capability and local candidate inference are distinct: the diagnostic reader may use the configured Codex model; this does not establish local engine latency. The reference packet includes source rows, not oracle labels, annotated explanations, or generated answer claims.
Next permitted mutation: create the child issue and isolated branch, add the source-only reader boundary tests, then implement the smallest preparation/reader path. The research/implementation goal remains active beyond this bounded slice; unsuccessful mechanisms are recorded and replaced only on new causal evidence within the newly authorized effort.