Commit History
-
v1.0.2 W4 SRS: forward-only disclosures for ripgrep, dubbo, go-zero, element, nushell (19 hunks, 16 files)
Hydrapse +1: Claude Fable 5- ripgrep milestone_seed_a6e0be3_1_sub-01: --passthru exempt from the match limit, worded as retention of the public pre-existing behaviour (SearcherBuilder::passthru, Searcher::passthru, --passthru). - dubbo M003.2: FR1/FR2 name the Mutiny publisher/subscriber shape the M003.3 tests bind (incl. isCancelled() on the publisher); "following the existing Reactor implementation pattern" removed. - go-zero M023: the capability is required, the form is the variadic StartAsync(c ...Config), the only one that compiles against all five live vintages. - element milestone_seed_be3778b_1 FR3: the rendering the four descendants' snapshots freeze (public VisualListItem / destructive). - nushell: the source-level compatibility paragraph added to the 12 graded SRSs that lacked it; the parse_block signature clause to M02_parser, core_development.1, core_development.3. No published number changes (future trials only). Evidence and the verification of every hunk: SWE-Milestone reeval/v…3500478 -
v1.0.2 W3 SRS: scikit-learn M12.1 FR28 names every image placeholder; companion clauses for M01, M12.4, M12.5
Hydrapse +1: Claude Fable 5The image ships forward-compatibility placeholders (validate_data, _check_n_features, the estimator-check helpers, _NotAnArray, assert_docstring_consistency, skip_if_no_numpydoc, _yield_masked_array_for_each_param, sklearn/utils/_test_common/) so that future-vintage test modules collect. The SRS now lists them, states that a raise NotImplementedError body suffices and that graded tests never call them. Forward-only: future trials see the disclosure; published numbers are corrected by the W3 waivers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>501081a -
v1.0.2 W3: scikit-learn collection-death waivers (issue #26), 2,115 entries in 6 lists (M01, M12.1-M12.5)
Hydrapse +1: Claude Fable 5Whole-file waivers for test files whose imports are satisfied only by the image's [ENV-PATCH] placeholders: the e2e evaluation tree is the agent's snapshot over the base image, the placeholders are gone, the whole module dies at collection (the reference END state fails collection too; >= 90 % of published e2e cells constantly dead for the file while the mstone control with the placeholders passes). 249 invalid_fail_to_pass + 1,866 invalid_pass_to_pass, reason per id naming the dying symbol; 11 files / 18 groups (scope study: SWE-Milestone .swe-milestone/issue26_filter/ SCOPE_REPORT.md). The 20 M12.5 sklearn/tests/test_common.py F2P ids of the study are not landed: SRS FR10 specifies ClassifierTags in full and the group is 20 of 47 F2P obligations (maintainer decision C, configuration b). M12.5_filter_list.json is new (the milestone had no list; the 7 published M12.5 derivatives were orphans). Applies to the v1.0.x corpus; withdrawn as a batch once the evaluation tree keeps the environ…f8f1732 -
v1.0.2 W1: cross-universe conflict waivers (issue #25), 860 invalid_pass_to_pass entries in 67 lists; drop 4 misplaced dubbo M024 entries
Hydrapse +1: Claude Fable 5Blanket per-universe waivers for P2P obligations that contradict the golden/effective behaviour mandated by another milestone (no single tree satisfies both universes): 848 census-backed C1 entries (strict cohort of reeval/xuniv_20260820 CENSUS.md 4.11, 195 ids, execution or mechanical proof) + 12 element C3 leakage entries derived from the shipped filter lists and equal test-body hashes. 17 lists created (13 ripgrep, 2 nushell, 1 element, 1 navidrome), 50 appended. Entry shape {test_id, reason, class, owner_milestone, evidence, census}; reversible (withdrawn when the evaluation contract changes, docs/maintainers/v1.0.2-release-plan.md). dubbo M024: the 4 invalid_fail_to_pass entries were P2P ids (ineffective copies of the P2P entries of the same file) and fail the new filter-list validator; deleted. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>c0f4f48
-
ripgrep: pin SummaryKind::Quiet's name and early-stop behaviour
maintenance_fixes_1_sub-01 FR1 redefines what a "match" means for SummaryKind::Quiet but says nothing about quit_early(), and the golden implementation resolves that by renaming the variant to QuietWithoutMatch. Neither the rename nor the early-stop grouping is bound by this milestone's own effective tests, so both readings pass here — while sibling milestones, whose END is base plus only their own commits, keep a frozen test that constructs SummaryKind::Quiet by name and asserts it stops after the first match. Under continuous evaluation the tree is cumulative, so the reading decides every later cell. Measured on two glm-5.3 runs of the same model: the run that moved Quiet out of the early-stop set failed that frozen test in 12 of 13 cells (reliable 67.76%), the run that left it in failed none (73.14%, 6 resolved vs 1). Removing that single test from the first run's results accounts for +7.37 reliable points — more than the entire gap between them. Renaming is the one reading that c…72bac10 -
element-web v1.0.1 ENV-PATCH overlays (envfix_v101_20260818)
878fcd6 -
nushell: pin the frozen public-API signatures the GT tests call
Test code is authoritative and is restored at evaluation time, so any signature it calls positionally is a contract. Three milestones left such a signature under-specified; agents that changed it hit a compile error that killed the whole crate's test binary, scoring the cell 0 with zero tests run and cascading to every later cell in the same tree. Evidence from the published corpus (278 dead nushell cells classified from their archived build logs): execute_xpath_query 61 cells / 11 trials (G04_1ddae02 FR3) parse_block 19 cells (core_development.2 FR3) fragment / table 20 cells (core_development.3 FR4) G04_1ddae02 FR3: pin execute_xpath_query's four-parameter signature; the new --output-* flags are read from `call` inside the function. Replaces a soft note ("prefer not to change signatures", present since 2026-03-19) that 61 cells violated anyway — naming the function and giving the full signature is what works, as core_develop…0ea8c72
-
nushell: SRS interface contracts for G05_0b8531e + DAG CSV semantics note
- srs/milestone_G05_0b8531e: Interface Contracts appendix. IC1 makes the SplitRead relocation additive in nu-protocol (the split_read module in byte_stream.rs remains present with SplitRead name-resolvable inside; a nu-utils re-export suffices; no duplicated implementation). IC2 pins the strings module files (strings/shared.rs, strings/unique.rs) and requires sformat!/uformat! to be invocable unqualified there. Adjudicated G1/G2/G3-established (blinded probes: 3/3 REMOVE without IC1, 3/3 KEEP with it); repairs the anchor loss that zeroed five base-vintage cells for layout-faithful agents and the macro-scope compile failure that zeroed the origin cell's N2P set. - additional_dependencies.csv: comment row documenting that column semantics match dependencies.csv (target depends on source; the loaded direction is authoritative — several rationale texts read inverted). The harness loader now skips '#' rows. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Clau…dee33e0 -
element-web v1.0.1 ENV-PATCH overlays (envfix_v101_20260818, ruling B-prime)
19 eval-side overlay Dockerfiles per docs/image-patching.md, ruling B-prime: - every image lacking it: hrp 5.2.2 nested 13-payload closure (pure addition) - sub-02/sub-03: full offline END-vintage yarn install (599112e precedent) + linkify-element presence re-add on the rebuilt tree + WASM link check + shipped-state restore - e662c19/7ff1fd2 ONLY: compound-design-tokens 4.0.0 -> 4.0.1 (own SRS contract alignment, decoupled non-blocking canaries) - linkify-element untouched everywhere else (presence semantics; §6.0 amendment supersedes exact-pin); no icon-subpath asserts (retracted theory: see campaign evidence/INCIDENT-20260818-compound-misclassification.md) Detection: env_deps_staging/<range>/PHASE4_REPORT.md + CORRECTION (harness). Campaign: SWE-Milestone-log/reeval/envfix_v101_20260818 (no promotion).2723e6e
-
dubbo: enable Java-only residual prune
9e09325
-
dubbo: filter carried Triple-REST cluster (13 P2P) from post-M006 milestones
Per DeepCommit-Env issue #38 (variance-reduced comparison): the 13 exact carried Triple-REST regression P2P tests (RestGroupVersionTest group-version x8, RestProtocolTest override-mapping x4, ambiguous-mapping x1) are added to invalid_pass_to_pass of M011/M016.1/M017/M018/M019/M025 only. M006 and all earlier milestones keep them charged, so an unfixed REST regression still costs the run at/before M006 while its downstream cascade is not re-counted 6 times. Fixed-early runs (whose REST failures are all pre-M006) are unaffected. codex-verified exact 13 (naive substring matches 4 false IDs - excluded). Leaderboard effect (7 dubbo trials rise; #1 Opus-4.8 / #2 GLM-5.2 unchanged): minimax +14.7, qwen +10.0, kimi-k2.6 +8.1, sonnet-4.6 +6.1, gpt-5.2 +5.1, opus-4.5 +3.6, opus-4.6 +1.9. Backups in reeval/dubbo_rest_filter_20260719/. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BXhU3BbzS4HgWvgvXeB1t1e85fa1d
-
dubbo: raise JUnit default timeout 60s -> 180s (audit: legit tests run up to 100s)
codex round-2 audit found 22 legitimately passing methods above 60s in the recorded runs (ConnectionTest.connectSyncTest up to 63s with a source-level 100s allowance; BitListTest.testConcurrent at 74-79s) - the 60s bound would have converted them into false P2P failures (e.g. kimi M025 0.1111 -> ~0.1053). 180s clears the 100s source allowance with headroom while still deterministically bounding genuine hangs (which are infinite). No completed measurement in the campaign was produced under a bound that clipped a legitimate test. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BXhU3BbzS4HgWvgvXeB1t1632d357 -
dubbo: JUnit method-level default timeout in every milestone test recipe
The await-bounding ENV-PATCH covers hazard #1 (MutinyClientCallsTest's unbounded negative-path await) but not the rest of the hang family: opus-4.7-xhigh's M003.3 snapshot deterministically hangs MutinyServerCallsTest.testOneToMany_success at an unbounded future.get() (agent drops the completion signal), and synchronous agent-internal blocking can never be bounded from test source at all. Surefire's fork timeout is an unreliable boundary by the instrument's own admission - fork-shutdown timing decides between losing one class report and wedging the whole reactor until the evaluator's 3600s wall. Add -Djunit.jupiter.execution.timeout.default=60s to all 26 dubbo milestone test recipes: any test method blocking >60s becomes one deterministic TimeoutException failure and the reactor continues. Golden-neutral (goldens finish in ms-s). Validated: opus-4.7 M003.3 now completes 7095 tests (score 0.6495 vs zeroed cell before); equivalence re-run of kimi's already-measured M003.3 reproduces F2P…2713b9f -
dubbo: bound the Mutiny negative-path await wherever the test exists
The MutinyClientCallsTest stabilization (await().indefinitely() -> atMost(10s)) was gated on MILESTONE_ID == M016.1 in both the compatibility script and the test-compat Dockerfile, while the same test ships in other checkpoints' trees (M003.3 END, M025 both states). An implementation that drops the mocked failure then hangs the whole reactor at the unpatched checkpoints until the evaluator's outer 3600s timeout, fail-closing the entire cell (kimi-k3 M003.3/M025, 2026-07-18; three-way adjudication in SWE-Milestone-data/reeval/dubbo_kimi_deepaudit_20260718/ADJUDICATION_CONSENSUS.md). Gate on the runtime fact instead: patch whenever the test file exists in the checked-out state. Golden semantics unchanged (correct implementations observe the failure immediately); hanging implementations become one deterministic failing test instead of a zeroed cell — identical to the treatment M016.1 already had. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/co…a78dddd
-
element maintenance_ui_ux: filter notification-panel empty-state [Pinecone]/[Dendrite] as environment-invalid P2P
0 passes across 31 trials / 116 executions in the evaluation container generation (deterministic 120s timeout on the labs-flag-gated Notifications button, identical on near-baseline agent code; stable 3/3 in the capture environment). Audit 2026-07-17, element panel flaky verification; pre-change filter_list backed up under reeval/element_rightpanel_20260717/. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>6229b2a
-
Cleanup: drop tracked *.bak build droppings; ignore /reeval/ and *.bak
26 build-time backup files/dirs (nushell dockerfile .bak dirs, dubbo/scikit Dockerfile and test_config .bak) had been committed by accident; none are read by any pipeline. reeval/ is the temporary re-evaluation campaign area and never dataset content. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Kv8v4m6ePchrw4dxfYarSke69c095 -
Release 1.0 data finalization
- SRS revisions: dubbo (24), go-zero (8), ripgrep (1) — adversarial-audit era contract cleanups. - Environment-repair dockerfiles now matching the rebuilt v1.0 images: dubbo M001.2/M004/M006/M017 (+ evaluation_post_snapshot.sh Maven closure, test-compat harness), element 599112e/feature_enhancements/maintenance_ui_ux (Dockerfile.v1.0 + apply_patches), nushell G01_48bca0a/core_development.4. - go-zero evaluation hook: dockerfiles/evaluation_post_snapshot.sh (go.mod backfill v3: require diff/align + go-directive alignment, legacy-gated) wired via config yaml evaluation_post_snapshot_script. - Residue-prune metadata (navidrome/go-zero), filter lists, classification and test_config updates; README. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Kv8v4m6ePchrw4dxfYarSkcac0e1a
-
navidrome+go-zero: pin prune_extensions=[.go] (phase-1 language scope)
Residue prune phase 1 deletes only Go source; navidrome's ui/src TypeScript front-end is never pruned (evaluator review F3). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NhK2BVApsdR7gAKhE3VrNWa383f21 -
navidrome+go-zero: enable residue_prune, add prune_keep_list facts
Phase 1b of the residue-prune rollout (spec: SWE-Milestone docs/residue-prune-spec.md). Facts live here per the data/code split; the mechanism ships with the evaluator. Keep-list: navidrome core/mock_library_service.go (confirmed scaffolding, ~20 graded P2P), go-zero core/stores/dbtest/sql.go (watch item, SRS-mandated contract). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NhK2BVApsdR7gAKhE3VrNW0420d17
-
Validate and normalize benchmark data
b5cc3b6
-
Rename EvoClaw -> SWE-Milestone across dataset card and assets
- README: pretty_name, badges (website swe-milestone.com, GitHub DeepCommit-ai/SWE-Milestone), prose, dir-tree example, clone URL, setup-guide link - assets: new SWE-Milestone banner; rename evoclaw_example.png / evoclaw_icon.png to swe_milestone_* - Citation kept as-is: the arXiv paper (2603.13428) is still titled "EvoClaw: ..." and the reference must match the published title Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NhK2BVApsdR7gAKhE3VrNW9c86e1f -
navidrome v0.57.0_v0.58.0: SRS interface contracts, flaky reclassification, v0.91-rc1 ENV-PATCH Dockerfiles
Benchmark refinement round (see DeepCommit docs/backlog/benchmark-refine-plan.md): - SRS (m001, m003 sub-01/02/03, m007): add "Interface Contracts" appendices pinning the names/signatures that hidden acceptance tests bind to, so any implementation can match them (whitebox-test pathology, fix class 4b) - classification (sub-01, sub-03): reclassify 5 flaky ArtistRepository tests out of the none_to_pass scoring set (verified over 3 self-grade rounds); normalize escaped Unicode in test ids - dockerfiles (m002, m004): v0.91-rc1 incremental ENV-PATCH images fixing GT-END trees that failed the go build gate (evaluator silently graded START); patches touch no graded test, oracles unchanged - .gitignore: exclude self_grade* working dirs Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NhK2BVApsdR7gAKhE3VrNW6a883eb
-
Update README.md to correct dataset clone URL for EvoClaw data
24b4c7f
-
dubbo M016.1: fix FR10 wording (InputStream → JarFile)
The original FR10 told agents to "wrap InputStream resources in try-with-resources blocks in JarScanner", but JarScanner.java contains zero InputStreams. The only Closeable target in scanURL() is the JarFile returned from JarURLConnection.getJarFile(). Weak models (GLM-5.1) freelanced and wrapped the JarURLConnection cast itself, which is not AutoCloseable, breaking dubbo-native's mvn compile and cascading ~5779 P2P tests via Maven's fail-fast reactor. This is a 2-word substitution (InputStream → JarFile in 2 places): no defensive hand-holding, no verbatim code snippets, no extra warnings — just a factually correct description of the resource leak. Empirical effect (verified in claude-code_glm-5.1_001 trial): - GLM-5.1 now writes the correct ground-truth fix on JarScanner - M016.1 P2P_missing: 5779 → 0 - M016.1 N2P: 0/6 → 6/6 - M016.1 score_reliable: ~0 → 0.50 - dubbo trial Score: ~3% → 24.8% Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>71e13b2
-
Add repo-specific config files for evaluation (test_framework, build_command)
Migrated from AgentBench/config/ — these are needed by evaluator.py for test ID normalization (go_test), compilation checks, and struct field fixes. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>b1f0e96
-
Improve go-zero SRS solvability: add missing API names and signatures
Add exact internal symbol names to SRS where eval tests reference unexported identifiers that agents cannot derive from SRS alone: - M003: Specify int64 type parameter for RollingWindow consumers - M013: Add hashRing field type, conns field, and error message prefix - M018: Specify sensitive.go as the file for the Sensitive interface - M019: Add doMarshalJson function name (FR5), limitDetailedBodyBytes constant (FR6), and AddAllowHeaders signature (FR4) - M025: Specify ScheduleImmediately method name Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>99975a2 -
Add base Dockerfiles, build configs, and .gitignore for all repos
Add dockerfiles/base/ with Dockerfile, test_config.json, and env_setup_prompt.md for ripgrep, element-web, navidrome, scikit-learn, and go-zero. Add build_command.txt where applicable. Update nushell base Dockerfile permissions. Add .gitignore to exclude e2e_trial/. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>55ae933
-
Add milestone Dockerfiles for all 7 repos
Include Dockerfiles and related build artifacts (env_setup_prompt, tag_patches, test_config) for reproducing milestone Docker images. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>58a2f3b
-
Add DockerHub badge, remove self-referencing HuggingFace badge
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>7bc93a9 -
Clarify classification.json structure: stable_classification and grading fields
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>37bb141 -
Remove SRS pipeline intermediates, keep only SRS.md per milestone
Removed 1799 files: generation prompts, LLM logs, verification reports, analysis docs, and patches. The harness only reads SRS.md. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>b486587 -
Use banner instead of icon, add HuggingFace badge
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>f36b8c4 -
Add dataset card with statistics, example DAG, and badges
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>6158e11
-
Sync SRS, test_results, and metadata from AgentBench
Update all 7 repositories with latest verified SRS documents, test classification/filter_list files, and workspace metadata. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>0083415
-
Flatten directory structure: remove DAG version and SRS version layers
caa49e1 -
Initial upload: workspace data for 7 repositories
c4764ec -
initial commit
Verified signature66a98a5