Current authority — unified 0.17.6, 2026-09-19
The full product program remains open. Current maintainer-approved scope and comparator amendment supersedes the scheduling, completion and approval claims in the historical body below. Candidate packet diagnostics do not satisfy adaptive acceptance. Product convergence and valid independent acceptance must precede final qualification; promotion to main still requires explicit maintainer approval. Comparator recovery and control repairs are supporting prerequisites.
Dev is 21cfc9ab797e6082ff4c45cdf6835f91094d5572; main remains 90d49cedbd91fde0d70b8bc612e81d54a56a571b. The former release freeze is canonically invalidated. Luna is permitted for future CodeStory tests, preserving a consistently frozen model across comparative arms and original historical model identities.
The records below are preserved verbatim as history, including original failures and superseded completion/fence claims. Their historical labels do not authorize new execution or close this program.
The unified 0.17.6 program includes the accepted maintenance foundation and the original general-evidence research, design, implementation, and acceptance obligations. Historical 0.17.7 filenames describe this same release. Breadth owner #2173 remains open; #2116 stays outside the program.
Current integration is ecb98f61afb01edc47da7517e217d7a7792bfba5 (tree 9efe9eccdea08f62018e2a88740702d3be38dc13); main remains 90d49cedbd91fde0d70b8bc612e81d54a56a571b. Traversal fixes #2188/#2191, Python eligibility #2193, Java dependency closure #2195 and bounded parser-cache encoding #2197 are merged. #2190/#2192/#2194/#2196 are closed. Candidate22 is built and isolated installed; independent Cursor/Grok byte audit and static control acceptance are recorded. The original CPython Q01 in fresh batch008 failed at the unchanged disk ceiling: 3,265,662,976 charged bytes against 3 GiB after 121.023 seconds, 21 preparing responses, zero complete packets and zero answering models. Q02–Q20 were not admitted. The budget is closed. Original normal drain verification failed; separate later native-image/controller absence released the current slot, with clean unchanged CPython source. #2187 remains open. A separate bounded observation confirmed all 3,491 parser-cache rows were compressed and exposed 570 MB of retained WAL beside the 1.122 GB main database. Its budget is closed and it does not qualify original Q01. #2198 now owns the isolated disposable-WAL retention fix, preserving WAL, transaction batching and the final durable publication fence. Sol implements; Terra challenges pinned readers, profile isolation and safe reclamation. Candidate21 and all prior failures remain preserved.
The accepted unpublished maintenance comparator remains candidate17, foundation SHA-256 b23351870c2382df2108a23331e44bf25fb6ecfdfe85bec68868689dafa6a3c7. The September 12 amendment accepts increased input when output decreases: Run03 saved 3,076 output tokens (3.43%) and 8.97% wall time, with all arms 10/12. Preserve the original evaluator input-gate FAIL and all earlier failed attempts. Candidate18/19/20 do not replace this comparator.
Remaining product delivery:
- Synthesize research and diagnostics now into a capability gap map. Separate correctness defects, missing relationships/artifacts, composition failures, host omissions, and external/runtime evidence requirements. Source adjudication of 1,360 questions, 5,334 anchors and 68 roots is development preparation, not product accuracy; four software and ten study answers remain incomplete, and all 160 study questions retain external-evidence requirements.
- Demonstrate or improve locate, inspect, follow, join, compare, bounded exploration, adaptive composition, and evidence-limit reporting through existing primitives. Each slice names the user-visible improvement, supporting cases, owning component, applicability/inputs/provenance/ambiguity/failure/stopping contract, smallest credible verification, and uncertainty. Existing demonstrated behavior can satisfy an obligation without new code.
- Require the same missing capability in three unrelated repositories across two domains before a shared abstraction. Local correctness fixes need no invented abstraction. Keep repository selectors, question-family handlers, hidden prompt policy, and evaluator knowledge out of production; preserve runtime/indexer/store/retrieval/pure-compiler ownership.
- Demonstrate full adaptive investigations: explanation/teaching, actionable review, debugging, optimization, architecture research, auth/permissions, and UI; across languages, frameworks, configuration, deployment, cross-service/multi-repository systems and non-symbol UI artifacts. Update conflicting canonical skill rules where evidence warrants. Packet completion is separate from answer sufficiency and complete host-task quality.
- Continue bounded visible-bank diagnostics and the preselected stratified model development panel while feeding findings into product decisions. Reuse the existing harness and raw evidence; return to the smallest relevant public interaction when testing machinery becomes the main work.
Approved team: root/Astra high owns product decisions, Git, integration, controls and execution; Sol high owns one isolated source slice; Terra high independently challenges it; Luna medium maintains capability coverage and evidence. Escalation to xhigh is reserved for demonstrated causal/identity/architecture ambiguity. Existing independent Cursor/Grok obligations remain and are scheduled without a fifth concurrent agent. Join each wave before opening another.
Native/GPU work has one root-assigned slot; Cargo commands are serial. CPython retains 20-second RPC, 30-second case, 180-second preparation, 900-second root, 120-second drain and 3-GiB added-disk limits. Closed attempt budgets stay closed. FastAPI Q06's two failures remain closed; a third requires new causal evidence and a reviewed changed approach.
After product convergence, freeze independent unseen design before new repository/question authoring: twelve targets, seventy-two tasks, three arms, 216 sessions. Preserve workflow/domain requirements, no-loss and critical-claim gates, material-gain thresholds, and separate resource estimation/budget freeze. Engineering model assignments do not change the identical frozen participant configuration across comparison arms.
Final source stabilization, bump-before-calibration, constant-only lineage, qualification, installed/native proof and canonical release closeout follow convergence. Maintainer approval is required before main promotion. A traversal fix, completed question bank, merge, or package alone cannot close this program.
Current source update: #2193 merged the Python producer eligibility repair at 16578987a4131978b037da1ef88306b75758cafd, with unchanged reviewed tree 398165ea912d475e7147bbb6b157e33462bb9975. The original error was reproduced before the fix; 174 proof-resolution tests, five non-vacuous hostile checks, independent Cursor/Grok review and applicable draft CI pass. Source child #2192 is complete; #2187 remains open for installed CPython verification. Candidate21 build and isolated installation completed within their caps. Its original CPython Q01 was interrupted at the unchanged 3-GiB disk ceiling after 87.758 seconds; zero packets completed, Q02 was not admitted, and 59 selected questions remain unrun. Normal drain failed; current native absence and controller exits were verified separately. The budget is closed. Retained checkpoint inspection identified about 832 MB of uncompressed parser-cache data as a potential storage improvement. Compression is being implemented under #2196, not a proven repair for the candidate21 peak; #2187 remains open.
Current execution owner is root for candidate22 preparation and #2187 verification. PR #2197 merged after Terra and Cursor/Grok exact-head acceptance and all applicable CI. A runtime legacy-test fixture failure on its first b5 revision remains preserved; the revised fixture retains real nonempty legacy artifacts and the original rollback checks. Source review does not establish physical disk savings or a completed packet. Reverify installation/runtime/source identity before a new original CPython attempt; all previous attempts and budgets remain closed.
The reconciled capability map explicitly records implemented behavior, measured limitations and unmeasured workflow obligations. Development-panel preselection and a separate model/resource budget still need a current concrete manifest; bounded search found only an expired historical six-session pilot, which is not reused. Unseen acceptance retains four or more cross-service/infrastructure targets, AWS/IAM and Kubernetes/Terraform, UI, eight or more domain groups and substantial review coverage. Its material-gain gate remains eight extra successful tasks of 72, or the preselected 20% efficiency measure at equal-or-better correctness across at least four domain groups.
Java source closeout: #2195 merged the exact reviewed tree after 177 proof-resolution tests, 16 Java-focused tests including independent overlays, two existing complexity tests, Terra acceptance, Cursor/Grok source acceptance and applicable draft CI. The store guard is unchanged. This closes #2194 source work and does not close Bazel runtime verification, #2187, #2173 or #2135. Development-panel preselection and resource/model budget must be concrete before execution; the panel belongs during product iteration. Independent unseen design and final qualification follow convergence.
Development-panel preparation now has a root preselection of 13 existing visible tasks across Caddy, FastAPI, Vite, Bazel and Terraform/EKS for a proposed 39-session comparison. Terra low is proposed consistently across three arms. This is not run admission: concrete review inputs, runtime/UI permissions, grading, order, resource estimates and budgets remain prerequisites. FastAPI Q06 is excluded under its two-failure rule; labels alone do not establish optimization or actionable-review support.
Current authority — unified 0.17.6, 2026-09-19
The full product program remains open. Current maintainer-approved scope and comparator amendment supersedes the scheduling, completion and approval claims in the historical body below. Candidate packet diagnostics do not satisfy adaptive acceptance. Product convergence and valid independent acceptance must precede final qualification; promotion to main still requires explicit maintainer approval. Comparator recovery and control repairs are supporting prerequisites.
Dev is
21cfc9ab797e6082ff4c45cdf6835f91094d5572; main remains90d49cedbd91fde0d70b8bc612e81d54a56a571b. The former release freeze is canonically invalidated. Luna is permitted for future CodeStory tests, preserving a consistently frozen model across comparative arms and original historical model identities.The records below are preserved verbatim as history, including original failures and superseded completion/fence claims. Their historical labels do not authorize new execution or close this program.
The unified 0.17.6 program includes the accepted maintenance foundation and the original general-evidence research, design, implementation, and acceptance obligations. Historical 0.17.7 filenames describe this same release. Breadth owner #2173 remains open; #2116 stays outside the program.
Current integration is
ecb98f61afb01edc47da7517e217d7a7792bfba5(tree9efe9eccdea08f62018e2a88740702d3be38dc13); main remains90d49cedbd91fde0d70b8bc612e81d54a56a571b. Traversal fixes #2188/#2191, Python eligibility #2193, Java dependency closure #2195 and bounded parser-cache encoding #2197 are merged. #2190/#2192/#2194/#2196 are closed. Candidate22 is built and isolated installed; independent Cursor/Grok byte audit and static control acceptance are recorded. The original CPython Q01 in fresh batch008 failed at the unchanged disk ceiling: 3,265,662,976 charged bytes against 3 GiB after 121.023 seconds, 21 preparing responses, zero complete packets and zero answering models. Q02–Q20 were not admitted. The budget is closed. Original normal drain verification failed; separate later native-image/controller absence released the current slot, with clean unchanged CPython source. #2187 remains open. A separate bounded observation confirmed all 3,491 parser-cache rows were compressed and exposed 570 MB of retained WAL beside the 1.122 GB main database. Its budget is closed and it does not qualify original Q01. #2198 now owns the isolated disposable-WAL retention fix, preserving WAL, transaction batching and the final durable publication fence. Sol implements; Terra challenges pinned readers, profile isolation and safe reclamation. Candidate21 and all prior failures remain preserved.The accepted unpublished maintenance comparator remains candidate17, foundation SHA-256
b23351870c2382df2108a23331e44bf25fb6ecfdfe85bec68868689dafa6a3c7. The September 12 amendment accepts increased input when output decreases: Run03 saved 3,076 output tokens (3.43%) and 8.97% wall time, with all arms 10/12. Preserve the original evaluator input-gate FAIL and all earlier failed attempts. Candidate18/19/20 do not replace this comparator.Remaining product delivery:
Approved team: root/Astra high owns product decisions, Git, integration, controls and execution; Sol high owns one isolated source slice; Terra high independently challenges it; Luna medium maintains capability coverage and evidence. Escalation to xhigh is reserved for demonstrated causal/identity/architecture ambiguity. Existing independent Cursor/Grok obligations remain and are scheduled without a fifth concurrent agent. Join each wave before opening another.
Native/GPU work has one root-assigned slot; Cargo commands are serial. CPython retains 20-second RPC, 30-second case, 180-second preparation, 900-second root, 120-second drain and 3-GiB added-disk limits. Closed attempt budgets stay closed. FastAPI Q06's two failures remain closed; a third requires new causal evidence and a reviewed changed approach.
After product convergence, freeze independent unseen design before new repository/question authoring: twelve targets, seventy-two tasks, three arms, 216 sessions. Preserve workflow/domain requirements, no-loss and critical-claim gates, material-gain thresholds, and separate resource estimation/budget freeze. Engineering model assignments do not change the identical frozen participant configuration across comparison arms.
Final source stabilization, bump-before-calibration, constant-only lineage, qualification, installed/native proof and canonical release closeout follow convergence. Maintainer approval is required before main promotion. A traversal fix, completed question bank, merge, or package alone cannot close this program.
Current source update: #2193 merged the Python producer eligibility repair at
16578987a4131978b037da1ef88306b75758cafd, with unchanged reviewed tree398165ea912d475e7147bbb6b157e33462bb9975. The original error was reproduced before the fix; 174 proof-resolution tests, five non-vacuous hostile checks, independent Cursor/Grok review and applicable draft CI pass. Source child #2192 is complete; #2187 remains open for installed CPython verification. Candidate21 build and isolated installation completed within their caps. Its original CPython Q01 was interrupted at the unchanged 3-GiB disk ceiling after 87.758 seconds; zero packets completed, Q02 was not admitted, and 59 selected questions remain unrun. Normal drain failed; current native absence and controller exits were verified separately. The budget is closed. Retained checkpoint inspection identified about 832 MB of uncompressed parser-cache data as a potential storage improvement. Compression is being implemented under #2196, not a proven repair for the candidate21 peak; #2187 remains open.Current execution owner is root for candidate22 preparation and #2187 verification. PR #2197 merged after Terra and Cursor/Grok exact-head acceptance and all applicable CI. A runtime legacy-test fixture failure on its first b5 revision remains preserved; the revised fixture retains real nonempty legacy artifacts and the original rollback checks. Source review does not establish physical disk savings or a completed packet. Reverify installation/runtime/source identity before a new original CPython attempt; all previous attempts and budgets remain closed.
The reconciled capability map explicitly records implemented behavior, measured limitations and unmeasured workflow obligations. Development-panel preselection and a separate model/resource budget still need a current concrete manifest; bounded search found only an expired historical six-session pilot, which is not reused. Unseen acceptance retains four or more cross-service/infrastructure targets, AWS/IAM and Kubernetes/Terraform, UI, eight or more domain groups and substantial review coverage. Its material-gain gate remains eight extra successful tasks of 72, or the preselected 20% efficiency measure at equal-or-better correctness across at least four domain groups.
Java source closeout: #2195 merged the exact reviewed tree after 177 proof-resolution tests, 16 Java-focused tests including independent overlays, two existing complexity tests, Terra acceptance, Cursor/Grok source acceptance and applicable draft CI. The store guard is unchanged. This closes #2194 source work and does not close Bazel runtime verification, #2187, #2173 or #2135. Development-panel preselection and resource/model budget must be concrete before execution; the panel belongs during product iteration. Independent unseen design and final qualification follow convergence.
Development-panel preparation now has a root preselection of 13 existing visible tasks across Caddy, FastAPI, Vite, Bazel and Terraform/EKS for a proposed 39-session comparison. Terra low is proposed consistently across three arms. This is not run admission: concrete review inputs, runtime/UI permissions, grading, order, resource estimates and budgets remain prerequisites. FastAPI Q06 is excluded under its two-failure rule; labels alone do not establish optimization or actionable-review support.