@furongjun1999/dsh-memory 0.4.11 → 0.5.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +552 -465
- package/codebuddy/CODEBUDDY.md +11 -3
- package/codebuddy/README.md +92 -90
- package/codebuddy/mcp.json +27 -27
- package/docs/GBrain/345/217/257/345/200/237/351/211/264/347/202/271_/347/201/265/346/236/242/350/220/275/347/202/271/344/272/244/346/216/245_20260919.md +169 -169
- package/docs/Pi/345/217/257/345/255/246/344/271/240/344/274/230/347/202/271_/347/201/265/346/236/242/345/244/247/350/204/221/346/224/271/350/277/233/344/272/244/346/216/245_20260915.md +144 -144
- package/docs/README.md +143 -111
- package/docs/discipline/harnesses.yaml +244 -226
- package/docs/discipline/templates/full.md.tmpl +61 -61
- package/docs/discipline/templates/rules.mdc.tmpl +68 -0
- package/docs/discipline/templates/skill.md.tmpl +23 -23
- package/docs/eval/AGI/344/270/203/347/273/264/350/257/204/345/210/206/346/212/245/345/221/212_md_cg_v1.0.md +299 -299
- package/docs/eval/AGI/344/270/203/347/273/264/350/257/204/345/210/206/346/212/245/345/221/212_md_cg_v2.0.md +239 -239
- package/docs/eval/bench_lingshu_self/bench_self.py +140 -0
- package/docs/eval/bench_lingshu_self/self_bench_result.json +404 -0
- package/docs/eval/bench_lingshu_self//347/201/265/346/236/242/350/207/252/345/272/223/347/253/257/345/210/260/347/253/257/346/243/200/347/264/242/345/256/236/346/265/213_v1.0.md +38 -0
- package/docs/eval//344/270/215/345/217/257/351/235/240/346/200/247/350/220/275/345/234/260_P0_v1.0.md +185 -0
- package/docs/eval//345/256/236/351/252/214/346/226/271/346/241/210_/345/255/246/344/271/240/351/227/255/347/216/257AB/344/270/216/346/250/252/350/257/204_v1.0.md +174 -174
- package/docs/eval//346/225/205/351/232/234/346/263/250/345/205/245/345/256/236/346/265/213_v1.0.md +422 -0
- package/docs/eval//346/250/252/350/257/204_/345/205/255/345/256/266100/351/242/230/344/270/255/350/213/261/345/217/214/346/237/245_v1.0.md +223 -223
- package/docs/eval//347/253/257/345/210/260/347/253/257LoCoMoQA/345/220/214/345/217/243/345/276/204/345/257/271/347/205/247_v1.0.md +100 -0
- package/docs/eval//347/253/257/345/210/260/347/253/257/345/271/262/346/211/260/346/261/240/350/257/204/346/265/213_/347/241/256/345/256/232/346/200/247/350/243/201/345/206/263vsLLM_judge_v1.1.md +197 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v1.md +156 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v10.md +210 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v11.md +227 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v12.md +203 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v13.md +233 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v14.md +191 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v15.md +213 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v16.md +214 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v17.md +199 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v2.md +156 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v3.md +152 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v4.md +128 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v5.md +114 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v6.md +192 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v7.md +187 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v8.md +207 -0
- package/docs/eval//347/274/272/351/231/267/346/214/226/346/216/230_/350/207/252/344/270/273/350/277/255/344/273/243_v9.md +203 -0
- package/docs/{mdcg → hive}//344/273/244/347/211/214/344/270/216/350/247/222/350/211/262/346/235/203/350/201/214/345/210/206/347/246/273_v0.1.md +165 -160
- package/docs/hive//344/273/244/347/211/214/350/257/255/344/271/211/344/277/256/346/255/243_/344/273/262/350/243/201/344/275/215_v0.1.md +76 -0
- package/docs/{mdcg → hive}//345/255/220/344/273/243/347/220/206/351/205/215/347/275/256/346/240/207/345/207/206_v0.5.md +236 -236
- package/docs/hive//345/256/211/345/205/250/345/256/241/350/256/241/345/256/236/351/224/232_v0.1.md +202 -0
- package/docs/hive//346/243/200/347/264/242/346/224/266/346/225/233/345/256/236/346/265/213/344/270/216S1b/350/256/276/350/256/241_v0.1.md +43 -43
- package/docs/hive//346/243/200/347/264/242/347/256/227/346/263/225/345/217/243/345/276/204/345/257/271/347/205/247_v0.1.md +85 -0
- package/docs/hive//346/243/200/347/264/242/350/267/257/345/276/204/344/270/216/350/256/244/347/237/245/347/273/223/346/236/204/345/245/221/347/272/246_v0.1.md +102 -102
- package/docs/hive//350/234/202/345/267/242M6_ingest/345/256/236/346/226/275/350/256/241/345/210/222_v0.1.md +47 -0
- package/docs/hive//350/234/202/345/267/242/345/217/214/345/256/236/344/276/213/344/272/222/351/252/214_/350/256/276/350/256/241/345/256/232/347/250/277.md +519 -503
- package/docs/hive//350/234/202/345/267/242/345/267/245/344/275/234/350/256/260/345/277/206_/351/241/271/347/233/256/350/256/241/345/210/222.md +85 -85
- package/docs/hive//350/234/202/345/267/242/350/256/276/350/256/241_/347/220/206/350/256/272/345/257/271/351/275/220_v0.1.md +191 -0
- package/docs/hive//350/234/202/345/267/242/350/277/255/344/273/243_/345/256/217/350/247/202/344/270/216/347/276/244/344/275/223/350/260/203/345/272/246_v0.1.md +90 -0
- package/docs/images/lingshu-moonlight-covenant-poster-preview.jpg +0 -0
- package/docs/images/lingshu-moonlight-covenant-poster.png +0 -0
- package/docs/mdcg/D_meta_/345/267/245/347/250/213/345/214/226/346/226/271/346/241/210_v0.2.md +216 -216
- package/docs/mdcg/README/350/257/246/347/273/206/347/211/210_v0.4.10.md +646 -646
- package/docs/mdcg/lingshu_tutorial.html +14449 -14449
- package/docs/mdcg/release_v0.4.11.md +49 -0
- package/docs/mdcg/release_v0.4.5.md +55 -55
- package/docs/mdcg/tool_table_v0.3.0.md +117 -117
- package/docs/mdcg//345/205/250/345/272/223/344/273/243/347/240/201/350/257/204/345/256/241/344/270/216/346/235/241/344/273/266/345/214/226/346/263/250/351/207/212_/350/256/241/345/210/222_v0.1.md +600 -600
- package/docs/mdcg//345/212/237/350/203/275/350/260/203/347/224/250/346/230/240/345/260/204/350/241/250_v0.1.md +40 -40
- package/docs/mdcg//345/215/225/345/205/203/350/207/252/346/210/221/351/224/232/347/202/271_/347/263/273/347/273/237/346/217/220/347/244/272/350/257/215/346/240/207/345/207/206_v0.3.md +275 -275
- package/docs/mdcg//345/217/221/345/270/203/351/227/250/347/246/201/351/223/276_v0.1.md +54 -0
- package/docs/mdcg//346/272/220/347/240/201/347/272/247/346/236/266/346/236/204/345/256/241/350/256/241_GPT/346/211/271/350/257/204/345/257/271/347/205/247_v1.0.md +130 -130
- package/docs/mdcg//347/201/265/346/236/24282/345/267/245/345/205/267_/345/212/237/350/203/275/346/225/264/347/220/206/344/270/216/350/277/201/347/247/273/346/230/240/345/260/204_v0.1.md +278 -278
- package/docs/mdcg//347/201/265/346/236/242/350/256/260/345/277/206/345/212/250/350/257/215/345/215/217/350/256/256_v1.0-draft.md +172 -172
- package/docs/mdcg//347/274/272/345/217/243/345/215/225_P0/346/224/266/345/217/243_v0.1.md +270 -270
- package/docs/mdcg//350/256/244/347/237/245/345/233/276_G4-G8/347/274/272/345/217/243/350/243/201/345/256/232/345/215/225_v0.1.md +491 -491
- package/docs/mdcg//350/256/244/347/237/245/345/233/276_/346/235/241/344/273/266/347/251/272/351/227/264/345/220/210/346/210/220/344/270/216/347/224/237/346/225/210/346/235/241/344/273/266/345/217/243/345/276/204_v0.1.md +340 -340
- package/docs/mdcg//350/256/244/347/237/245/345/233/276_/347/264/242/345/274/225/344/270/216/345/267/245/347/250/213/350/247/204/350/214/203/345/214/226_/350/256/241/345/210/222_v0.1.md +588 -588
- package/docs/plans/GridWorld/346/234/200/345/260/217/351/227/255/347/216/257/350/247/204/346/240/274_v0.1.md +176 -176
- package/docs/plans//345/221/275/345/220/215/346/262/273/347/220/206_/351/241/271/347/233/256/350/256/241/345/210/222.md +81 -81
- package/docs/swarm//350/234/202/347/276/244/344/272/222/350/201/224_v0.1.md +704 -704
- package/docs/swarm//350/234/202/347/276/244/345/220/214/351/224/231/346/243/200/346/265/213/345/256/236/351/252/214/345/215/217/350/256/256_v0.1.md +242 -242
- package/docs/theory//344/270/215/345/217/257/351/235/240/345/256/232/347/220/206/344/270/216/345/244/261/346/225/210/344/274/230/345/205/210/346/241/206/346/236/266_v0.3.md +365 -0
- package/docs/theory//345/215/225/347/272/277/347/250/213/344/270/216/346/263/250/346/204/217/345/212/233/351/233/206/344/270/255_/346/227/240/344/272/211/350/256/256/347/220/206/350/256/272/346/226/207/346/241/243_v1.0.md +129 -129
- package/docs/theory//345/271/266/345/217/221/345/277/205/347/204/266/346/200/247/347/220/206/350/256/272_v0.2.md +134 -134
- package/docs/theory//346/231/272/350/203/275/350/256/2723.4.md +5260 -5260
- package/docs/theory//346/246/202/345/277/265/345/210/206/345/261/202/345/257/271/351/275/220/350/241/250_v0.1.md +145 -145
- package/docs/theory//347/220/206/350/256/272_/346/234/272/345/210/266_/344/273/243/347/240/201_/345/256/236/351/252/214_/347/274/272/345/217/243/347/237/251/351/230/265_v0.1.md +144 -144
- package/docs/theory//347/220/206/350/256/272/344/273/223/346/213/206/345/210/206/344/270/216/347/231/275/347/256/261/347/237/245/350/257/206/345/272/223/345/206/205/350/277/201_v0.1.md +198 -198
- package/docs/theory//350/256/244/347/237/245/344/273/243/347/220/206/344/270/216/346/224/266/346/225/233/347/273/223/346/236/204_/346/235/241/344/273/266/350/256/272/351/207/215/346/236/204_v0.1.md +221 -221
- package/docs/world_model//344/270/226/347/225/214/346/250/241/345/236/213_/347/245/236/347/273/217/347/275/221/347/273/234/345/272/225/345/261/202/346/236/266/346/236/204_v1.0.md +237 -237
- package/docs//344/270/215/345/217/257/351/235/240/346/200/247/347/220/206/350/256/272_v0.1.md +343 -0
- package/docs//345/217/221/345/270/203/344/273/266/345/233/236/346/272/257/350/257/264/346/230/216_v0.1.md +82 -82
- package/docs//345/267/245/344/275/234/347/272/252/345/276/213_/350/256/244/347/237/245/345/233/276/346/235/241/347/233/256_v1.1.json +28 -2
- package/dsh/README.md +82 -82
- package/dsh/cordis.yml.example +139 -139
- package/dsh/update-lingshu.bat +11 -11
- package/lib/bridge.d.ts +9 -0
- package/lib/bridge.js +35 -0
- package/lib/hooks.js +36 -2
- package/lib/index.js +7 -1
- package/lib/lib/roleplay_web.js +116 -29
- package/lib/lib/token_store.d.ts +7 -1
- package/lib/lib/token_store.js +12 -3
- package/md_cg/__init__.py +7 -7
- package/md_cg/audit.py +379 -368
- package/md_cg/autonomy.py +287 -287
- package/md_cg/backfill.py +1328 -1327
- package/md_cg/backfill_bigdomain.py +34 -34
- package/md_cg/backfill_bucket_zh.py +35 -0
- package/md_cg/bench6_arms.py +410 -410
- package/md_cg/bench6_common.py +230 -230
- package/md_cg/bench6_competitors.py +212 -212
- package/md_cg/bench_axis_domain.py +257 -257
- package/md_cg/bench_blind_comp.py +308 -308
- package/md_cg/bench_e2e_judge.py +532 -0
- package/md_cg/bench_e2e_locomo_qa.py +368 -0
- package/md_cg/bench_e2e_qa.py +256 -0
- package/md_cg/bench_en_atoms_public.py +230 -230
- package/md_cg/bench_governance.py +348 -348
- package/md_cg/bench_lme_zh.py +410 -410
- package/md_cg/bench_locomo.py +121 -121
- package/md_cg/bench_locomo_zh.py +450 -450
- package/md_cg/bench_locomo_zh_public.py +147 -147
- package/md_cg/bench_longmem.py +112 -112
- package/md_cg/bench_membench.py +632 -632
- package/md_cg/bench_p0.py +149 -149
- package/md_cg/bench_progressive.py +287 -287
- package/md_cg/bench_role_views.py +238 -238
- package/md_cg/bench_task_ab.py +243 -243
- package/md_cg/bench_task_ab_llm.py +408 -408
- package/md_cg/bench_unified_en.py +204 -204
- package/md_cg/bench_zh_mad.py +601 -601
- package/md_cg/blindspot_tickets.py +123 -123
- package/md_cg/branches.py +301 -285
- package/md_cg/build_postings.py +73 -73
- package/md_cg/ccgc.py +1006 -948
- package/md_cg/census.py +132 -132
- package/md_cg/chain.py +315 -300
- package/md_cg/codeindex.py +531 -531
- package/md_cg/coldverify.py +292 -292
- package/md_cg/comment_gate.py +337 -337
- package/md_cg/cond_compose.py +190 -190
- package/md_cg/cond_facts.py +154 -154
- package/md_cg/cond_template.json +106 -106
- package/md_cg/condition_anchor.py +142 -142
- package/md_cg/conformance.py +726 -726
- package/md_cg/consistency.py +717 -717
- package/md_cg/consolidate.py +1537 -1439
- package/md_cg/corpus.py +110 -110
- package/md_cg/crosscheck.py +1098 -1097
- package/md_cg/crypto.py +3 -1
- package/md_cg/d_meta.py +310 -310
- package/md_cg/datapath.py +78 -18
- package/md_cg/docindex.py +473 -473
- package/md_cg/eval_common.py +575 -575
- package/md_cg/evidence.py +4 -2
- package/md_cg/evolution.py +477 -477
- package/md_cg/export.py +222 -220
- package/md_cg/forgetting.py +581 -581
- package/md_cg/fsutil.py +377 -329
- package/md_cg/hotcache.py +48 -7
- package/md_cg/hyperedge.py +251 -251
- package/md_cg/identity.py +390 -390
- package/md_cg/insight.py +500 -500
- package/md_cg/interop.py +338 -0
- package/md_cg/judgment_manifest.py +177 -0
- package/md_cg/lexicon/build_cedict_en_zh.py +329 -329
- package/md_cg/lexicon/build_standard_en.py +171 -171
- package/md_cg/lexicon/expand_en_zh.py +211 -211
- package/md_cg/lifecycle.py +272 -272
- package/md_cg/linkref.py +280 -280
- package/md_cg/links.py +140 -107
- package/md_cg/mcp_server.py +403 -55
- package/md_cg/md_whitebox.py +345 -345
- package/md_cg/mdcg.py +783 -233
- package/md_cg/mdcos.py +500 -70
- package/md_cg/metacognition.py +591 -591
- package/md_cg/migrate.py +119 -119
- package/md_cg/migrate_aeis.py +221 -221
- package/md_cg/migrate_roleplay.py +293 -293
- package/md_cg/migrate_wisdom_graph.py +360 -360
- package/md_cg/mreview/__init__.py +25 -25
- package/md_cg/mreview/__main__.py +110 -110
- package/md_cg/mreview/bundle.py +178 -178
- package/md_cg/mreview/candidates.py +262 -262
- package/md_cg/mreview/govern.py +694 -693
- package/md_cg/mreview/locate.py +939 -939
- package/md_cg/mreview/pipeline.py +728 -728
- package/md_cg/mreview/rules/duplication.json +21 -21
- package/md_cg/mreview/rules/field_coverage.json +54 -54
- package/md_cg/mreview/rules/source_license.json +21 -21
- package/md_cg/mreview/rules/template_flow.json +21 -21
- package/md_cg/mreview/ruleset.py +252 -252
- package/md_cg/nodefile.py +575 -575
- package/md_cg/pooling.py +484 -472
- package/md_cg/postings.py +300 -298
- package/md_cg/predict.py +1100 -1100
- package/md_cg/progressive.py +123 -123
- package/md_cg/protect.py +272 -272
- package/md_cg/protocol/md_cg_gate.proto +33 -33
- package/md_cg/protocol.py +372 -372
- package/md_cg/provenance.py +582 -582
- package/md_cg/reach.py +453 -453
- package/md_cg/readcache.py +143 -0
- package/md_cg/reconcile.py +228 -0
- package/md_cg/refindex.py +833 -833
- package/md_cg/refine.py +604 -604
- package/md_cg/review_cli.py +170 -0
- package/md_cg/roleviews.py +89 -89
- package/md_cg/routing.py +393 -365
- package/md_cg/run_tests.py +211 -0
- package/md_cg/scrub.py +13 -3
- package/md_cg/security.py +128 -18
- package/md_cg/self_state.py +1029 -1029
- package/md_cg/selfreport.py +152 -151
- package/md_cg/semantic/__init__.py +10 -10
- package/md_cg/semantic/canonical.py +122 -122
- package/md_cg/semantic/en_normalizer.py +364 -364
- package/md_cg/semantic/en_zh_map.json +28694 -0
- package/md_cg/semantic/export_en_zh_map.py +64 -0
- package/md_cg/semantic/unify.py +45 -0
- package/md_cg/semantic/zh_en_atoms.py +139 -139
- package/md_cg/signer.py +7 -4
- package/md_cg/sources.py +816 -582
- package/md_cg/statushdr.py +179 -179
- package/md_cg/stg.py +54 -37
- package/md_cg/subgraph.py +729 -729
- package/md_cg/sustain.py +35 -5
- package/md_cg/tasks.py +470 -470
- package/md_cg/test_access_hints.py +147 -0
- package/md_cg/test_action_derive.py +203 -203
- package/md_cg/test_audit_rotate.py +270 -270
- package/md_cg/test_autonomy.py +143 -143
- package/md_cg/test_bench_governance.py +102 -102
- package/md_cg/test_blindspot_tickets.py +166 -166
- package/md_cg/test_branch_discard_tombstone.py +136 -0
- package/md_cg/test_branches.py +13 -3
- package/md_cg/test_ccg_perturb.py +184 -184
- package/md_cg/test_ccgc.py +433 -433
- package/md_cg/test_census_prune.py +81 -81
- package/md_cg/test_chain_read_isolate.py +168 -0
- package/md_cg/test_cond_compose_anchors.py +76 -76
- package/md_cg/test_cond_match.py +165 -165
- package/md_cg/test_condition_anchor.py +81 -81
- package/md_cg/test_d_meta.py +412 -412
- package/md_cg/test_datapath_device_name.py +203 -0
- package/md_cg/test_datapath_root.py +199 -199
- package/md_cg/test_emit_negtail_cache.py +156 -0
- package/md_cg/test_en_pipeline.py +22 -2
- package/md_cg/test_gain_gate.py +212 -212
- package/md_cg/test_govern_directread.py +421 -0
- package/md_cg/test_health_scale.py +173 -173
- package/md_cg/test_hive_ingest.py +285 -0
- package/md_cg/test_hot_cold.py +215 -215
- package/md_cg/test_hyperedge.py +245 -245
- package/md_cg/test_i26_empty_first_write.py +116 -0
- package/md_cg/test_i27_e041_identity.py +128 -0
- package/md_cg/test_i28_hotcache_prodpath.py +122 -0
- package/md_cg/test_i32_hotcache_env_key.py +218 -0
- package/md_cg/test_identity_attribution.py +96 -15
- package/md_cg/test_index_durability.py +17 -3
- package/md_cg/test_interop.py +95 -0
- package/md_cg/test_interop_judgment.py +228 -0
- package/md_cg/test_issue39_utf8_stdio.py +273 -0
- package/md_cg/test_lifecycle.py +309 -309
- package/md_cg/test_linkref.py +306 -306
- package/md_cg/test_links_concurrent_write.py +188 -0
- package/md_cg/test_lock.py +43 -43
- package/md_cg/test_md_access_parity.py +255 -255
- package/md_cg/test_md_writepath.py +345 -345
- package/md_cg/test_mdstore_search_parity.py +160 -0
- package/md_cg/test_merge_upsert.py +168 -0
- package/md_cg/test_mr_m2.py +587 -587
- package/md_cg/test_mr_m3.py +710 -710
- package/md_cg/test_mr_m4.py +485 -485
- package/md_cg/test_n123_derive_expiry_chain.py +205 -0
- package/md_cg/test_n130_verify_falsified_protect.py +185 -0
- package/md_cg/test_n131_merge_gate.py +205 -0
- package/md_cg/test_p0.py +250 -250
- package/md_cg/test_p1.py +316 -316
- package/md_cg/test_p10_identity.py +173 -173
- package/md_cg/test_p11_consistency.py +233 -233
- package/md_cg/test_p12_metacognition.py +212 -212
- package/md_cg/test_p13_encryption.py +241 -241
- package/md_cg/test_p14_sustain.py +249 -249
- package/md_cg/test_p15_scrub.py +280 -280
- package/md_cg/test_p16_self_state.py +301 -301
- package/md_cg/test_p17_predict.py +354 -354
- package/md_cg/test_p18_whitebox.py +171 -171
- package/md_cg/test_p19_migrate_roleplay.py +149 -149
- package/md_cg/test_p1x_ref_root.py +160 -0
- package/md_cg/test_p20_evolution.py +315 -315
- package/md_cg/test_p21_tokens.py +293 -270
- package/md_cg/test_p22_theory.py +175 -175
- package/md_cg/test_p23_links.py +311 -311
- package/md_cg/test_p24_evidence.py +227 -227
- package/md_cg/test_p25_weights.py +156 -156
- package/md_cg/test_p26_refindex.py +416 -416
- package/md_cg/test_p27_docindex.py +16 -7
- package/md_cg/test_p28_refcheck.py +305 -305
- package/md_cg/test_p29_session_ingest_export.py +354 -333
- package/md_cg/test_p2_mcp.py +3 -0
- package/md_cg/test_p3.py +11 -2
- package/md_cg/test_p30_maintain.py +330 -330
- package/md_cg/test_p31_insight.py +534 -534
- package/md_cg/test_p32_backfill.py +7 -1
- package/md_cg/test_p33_ccg_wiring.py +293 -293
- package/md_cg/test_p34_crosscheck.py +331 -331
- package/md_cg/test_p35_conditioned_claim.py +252 -252
- package/md_cg/test_p36_kp_align.py +230 -230
- package/md_cg/test_p37_condition_space.py +248 -248
- package/md_cg/test_p38_concurrent_flush.py +102 -0
- package/md_cg/test_p38_contextualize.py +273 -273
- package/md_cg/test_p39_verify_flow.py +153 -0
- package/md_cg/test_p39_vision_evidence.py +369 -369
- package/md_cg/test_p40_refine_worklist.py +241 -241
- package/md_cg/test_p41_evolve_patrol.py +224 -224
- package/md_cg/test_p42_provenance.py +269 -269
- package/md_cg/test_p43_pooling.py +412 -398
- package/md_cg/test_p44_md_whitebox.py +231 -231
- package/md_cg/test_p45_session_identity.py +219 -219
- package/md_cg/test_p46_unit_scope.py +272 -272
- package/md_cg/test_p47_session_view.py +316 -0
- package/md_cg/test_p4_fuzzy.py +223 -223
- package/md_cg/test_p5_semantic.py +226 -226
- package/md_cg/test_p6_consolidate.py +440 -387
- package/md_cg/test_p7_goals_recent.py +202 -202
- package/md_cg/test_p8_subgraph_chain.py +200 -200
- package/md_cg/test_p9_forget_protect.py +231 -231
- package/md_cg/test_predict_beta.py +135 -135
- package/md_cg/test_preflight_failclosed.py +100 -100
- package/md_cg/test_progressive.py +146 -146
- package/md_cg/test_propose_tail_index.py +157 -0
- package/md_cg/test_protocol.py +243 -243
- package/md_cg/test_reach.py +378 -378
- package/md_cg/test_reach_keys.py +201 -201
- package/md_cg/test_read_clip.py +141 -141
- package/md_cg/test_read_scope_b27.py +277 -0
- package/md_cg/test_readcache_default_on.py +168 -0
- package/md_cg/test_readcache_precise_inval.py +270 -0
- package/md_cg/test_readcache_prodpath.py +203 -0
- package/md_cg/test_reconcile_v0.py +294 -0
- package/md_cg/test_retr_gates_prodpath.py +140 -0
- package/md_cg/test_retr_s1.py +344 -340
- package/md_cg/test_retr_s1b.py +276 -209
- package/md_cg/test_retr_s3.py +194 -194
- package/md_cg/test_retr_s4.py +163 -163
- package/md_cg/test_retr_s5.py +200 -200
- package/md_cg/test_retr_s6.py +157 -157
- package/md_cg/test_retr_s7.py +392 -384
- package/md_cg/test_retr_s8_time.py +369 -316
- package/md_cg/test_retr_s9_edges.py +286 -286
- package/md_cg/test_retr_s9_entity_ctx.py +9 -3
- package/md_cg/test_retr_score_once.py +208 -0
- package/md_cg/test_review_conformance.py +367 -367
- package/md_cg/test_review_onepass.py +170 -0
- package/md_cg/test_role_views.py +354 -354
- package/md_cg/test_rrf_graph_seed_cache.py +154 -0
- package/md_cg/test_security_audit.py +155 -0
- package/md_cg/test_security_audit_b26.py +161 -0
- package/md_cg/test_security_audit_v21.py +250 -0
- package/md_cg/test_sem_noise.py +242 -242
- package/md_cg/test_semantic_canonical.py +16 -2
- package/md_cg/test_session_isolation.py +168 -0
- package/md_cg/test_snapshot_autoclose.py +187 -0
- package/md_cg/test_subproc_encoding.py +192 -192
- package/md_cg/test_sustain_mutual.py +153 -153
- package/md_cg/test_tail_watermark_race.py +208 -0
- package/md_cg/test_tasks.py +409 -409
- package/md_cg/test_tenant_env_override_warn.py +139 -0
- package/md_cg/test_tenant_registry_corrupt_warn.py +151 -0
- package/md_cg/test_tool_face.py +189 -189
- package/md_cg/test_transfer.py +180 -180
- package/md_cg/test_trust.py +361 -361
- package/md_cg/test_twophase.py +286 -286
- package/md_cg/test_v14_fixes.py +38 -20
- package/md_cg/test_validity_filter.py +280 -280
- package/md_cg/test_verify_answer.py +138 -138
- package/md_cg/test_verify_dirty_reconcile.py +157 -0
- package/md_cg/test_wisdom_md_store.py +292 -292
- package/md_cg/test_writelimit.py +197 -197
- package/md_cg/test_writepipe.py +214 -214
- package/md_cg/theory.py +6 -3
- package/md_cg/tokens.py +85 -14
- package/md_cg/tool_face.py +260 -260
- package/md_cg/trust.py +986 -950
- package/md_cg/twophase.py +231 -231
- package/md_cg/units.py +668 -667
- package/md_cg/vision_evidence.py +667 -666
- package/md_cg/weights.py +624 -624
- package/md_cg/whitebox.py +527 -527
- package/md_cg/whitebox_kb/__init__.py +37 -37
- package/md_cg/whitebox_kb/aeis_core/__init__.py +42 -42
- package/md_cg/whitebox_kb/aeis_core/semantic.py +280 -280
- package/md_cg/whitebox_kb/aeis_core/textutil.py +13 -13
- package/md_cg/whitebox_kb/engine.py +310 -310
- package/md_cg/whitebox_kb/seed_knowledge//346/231/272/350/203/275/350/256/2723.4.md +5260 -5260
- package/md_cg/whitebox_kb/wisdom/browser_units.py +2631 -2631
- package/md_cg/whitebox_kb/wisdom/causal_discover.py +432 -432
- package/md_cg/whitebox_kb/wisdom/chat_engine.py +1506 -1506
- package/md_cg/whitebox_kb/wisdom/compiler_code_units.py +3033 -3033
- package/md_cg/whitebox_kb/wisdom/condition_algebra.py +112 -112
- package/md_cg/whitebox_kb/wisdom/condition_frame.py +315 -315
- package/md_cg/whitebox_kb/wisdom/condition_kb.py +108 -108
- package/md_cg/whitebox_kb/wisdom/conflict_map.json +4445 -4445
- package/md_cg/whitebox_kb/wisdom/core/lexer.py +512 -512
- package/md_cg/whitebox_kb/wisdom/core/name_checker.py +1023 -1023
- package/md_cg/whitebox_kb/wisdom/cspmn.py +258 -258
- package/md_cg/whitebox_kb/wisdom/csre.py +264 -264
- package/md_cg/whitebox_kb/wisdom/danmaku_audit.py +252 -252
- package/md_cg/whitebox_kb/wisdom/distilled_condition_units.json +6417 -6417
- package/md_cg/whitebox_kb/wisdom/docs/WB-EVAL-20260902b.json +1950 -1950
- package/md_cg/whitebox_kb/wisdom/docs/WB-EVAL-20260902c.json +1296 -1296
- package/md_cg/whitebox_kb/wisdom/docs/whitebox_capability_graph_demo.json +59 -59
- package/md_cg/whitebox_kb/wisdom/graph_db_units.py +3089 -3089
- package/md_cg/whitebox_kb/wisdom/knowledge_points.py +318 -318
- package/md_cg/whitebox_kb/wisdom/md_access.py +470 -470
- package/md_cg/whitebox_kb/wisdom/md_store.py +276 -251
- package/md_cg/whitebox_kb/wisdom/migrate_wisdom.py +476 -476
- package/md_cg/whitebox_kb/wisdom/multilang_ir.py +128 -124
- package/md_cg/whitebox_kb/wisdom/navigate.py +248 -248
- package/md_cg/whitebox_kb/wisdom/neural_retrieve.py +224 -224
- package/md_cg/whitebox_kb/wisdom/os_units.py +2735 -2735
- package/md_cg/whitebox_kb/wisdom/pattern_separation.py +376 -376
- package/md_cg/whitebox_kb/wisdom/prereq_map.json +364 -364
- package/md_cg/whitebox_kb/wisdom/python_code_units.py +2766 -2766
- package/md_cg/whitebox_kb/wisdom/role_solidified.json +7 -7
- package/md_cg/whitebox_kb/wisdom/route_memory.py +244 -244
- package/md_cg/whitebox_kb/wisdom/scene_reconstruction.py +148 -148
- package/md_cg/whitebox_kb/wisdom/snr_report.json +40 -40
- package/md_cg/whitebox_kb/wisdom/test_code_compose_domains.py +7541 -7541
- package/md_cg/whitebox_kb/wisdom/test_compiler_self_bootstrap.py +354 -354
- package/md_cg/whitebox_kb/wisdom/test_ecosystem_assembly.py +158 -158
- package/md_cg/whitebox_kb/wisdom/test_ecosystem_demos.py +193 -193
- package/md_cg/whitebox_kb/wisdom/test_graph_db.py +292 -292
- package/md_cg/whitebox_kb/wisdom/test_python_self_bootstrap.py +160 -160
- package/md_cg/whitebox_kb/wisdom/trigger_words_index.json +4366 -4366
- package/md_cg/whitebox_kb/wisdom/verifier.py +1297 -1297
- package/md_cg/writelimit.py +356 -356
- package/md_cg/writepipe.py +20 -8
- package/package.json +101 -96
- package/skills/plugin.json +54 -54
- package/skills/skills/designer-perspective/SKILL.md +158 -158
- package/skills/skills/designer-perspective/references/01-observation-position.md +66 -66
- package/skills/skills/designer-perspective/references/02-structure-recognition.md +62 -62
- package/skills/skills/designer-perspective/references/03-direction-judgment.md +55 -55
- package/skills/skills/designer-perspective/references/04-qualification-verdict.md +72 -72
- package/skills/skills/designer-perspective/references/05-condition-attribution.md +74 -74
- package/skills/skills/designer-perspective/scripts/designer.py +545 -545
- package/skills/skills/designer-perspective/tests/cases.jsonl +17 -17
- package/skills/skills/designer-perspective/tests/selftest.py +61 -61
- package/skills/skills/lingshu-browser/SKILL.md +60 -60
- package/skills/skills/lingshu-compiler/SKILL.md +56 -56
- package/skills/skills/lingshu-compiler/units/analyze-type-infer/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/check-name-real/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-assign/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-expr-tree/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-full-pipeline/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-func-def/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-if-then/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-logic-expr/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-recursive/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-scope/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-type-check/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compile-while/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0010c4bf/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0355bffb/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0361708a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-054a0414/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-05a1691a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-05eeed1e/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0622a1f6/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-08b54217/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0a62b70c/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0ab24d00/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0e093688/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0e82b966/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0ee9b9b9/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-0f3787e6/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-1028685f/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-11897630/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-16661b9b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-1f722303/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-25be1262/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2a76ba07/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2dbea54a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2df52f16/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2ec2c9d2/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-2f8c8f39/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-38378ac7/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-39457d2e/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-39fb5926/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-47f4fbfa/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-4a8cd1f1/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-4b230b6b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-4cbbda95/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-4d5a68ff/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-56dc9bdd/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-57e76ebe/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-61ff016b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-63eae588/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-64c4224b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-66377955/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-674ab4f6/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-6af76fd6/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-6d979db2/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-6de326c0/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-6f1c0eea/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-724d6c9b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-7690177f/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-7b229a7d/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-7f471b3a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-815cad08/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-83c61634/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-8c798c81/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-8dd747ac/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-90324d0a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-94b8d72d/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-94f12231/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-95937c16/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-98a5625b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-98b4c42f/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-98e3894a/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-9bdfe4b8/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-9d9b4e83/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-9f7be5ad/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-a6daf076/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-a7995a0c/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-a8399248/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-aabbd099/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-afe169d8/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-b0678bda/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-b09bd196/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-b1396e23/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-b9b31ce0/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-c10264a7/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-cb1e8e4b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-ccafd438/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-cda9c262/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-ce648068/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-cf5776a4/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-d974e5d3/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-e3979fd3/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-eb1cf2b5/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-ecb30d5b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-f8c8b24b/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-f99fedbe/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-fa8ff5f7/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/compiler-fe1b058d/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/lex-chinese-program/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/lex-dao-de-jing/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/lex-nine-chapters/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-arithmetic/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-array-ops/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-closure-call/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-closure-create/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-compare/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-cond-jump/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-cond-space/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-exception/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-func-call/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-loop-run/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-profiling/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-refcount/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-run-loop/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-short-circuit/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-stack-guard/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-stack-ops/SKILL.md +45 -45
- package/skills/skills/lingshu-compiler/units/vm-trust-accum/SKILL.md +45 -45
- package/skills/skills/lingshu-graph/SKILL.md +63 -63
- package/skills/skills/lingshu-net/SKILL.md +48 -48
- package/skills/skills/lingshu-os/SKILL.md +64 -64
- package/skills/skills/lingshu-pylang/SKILL.md +71 -71
- package/src/bridge.ts +33 -0
- package/src/hooks.ts +38 -2
- package/src/index.ts +526 -518
- package/src/lib/datapath.ts +326 -326
- package/src/lib/mdcg_client.ts +413 -413
- package/src/lib/mutual.ts +428 -428
- package/src/lib/prompt_safety.ts +62 -62
- package/src/lib/python_path.ts +71 -71
- package/src/lib/roleplay_web.ts +116 -29
- package/src/lib/token_store.ts +13 -3
- package/src/tools.ts +212 -212
- package/zcode/AGENTS.md +11 -3
- package/zcode/README.md +41 -41
- /package/docs/{mdcg → hive}//344/270/273/344/273/243/347/220/206/345/255/220/344/273/243/347/220/206/350/256/260/345/277/206/346/236/266/346/236/204/350/256/276/350/256/241.md" +0 -0
|
@@ -1,729 +1,729 @@
|
|
|
1
|
-
# -*- coding: utf-8 -*-
|
|
2
|
-
"""记忆评审流水线 · 级 4 蜂巢并发对接(spec 构造/投递/收卷)+ 级 5 意见落库治理。
|
|
3
|
-
|
|
4
|
-
真源:docs/记忆评审系统_立项设计与施工交接_20260915.md §5(spec/result 契约)
|
|
5
|
-
§7(反面清单:评审不直接改知识层 / 意见必过 verify 令牌 / 不确定 DEFER)
|
|
6
|
-
|
|
7
|
-
五级分工:级 1-3(M1)=候选→捆包→规则装配,产出 (pkg, asm) 对;
|
|
8
|
-
级 4(本模块)=每对 → hive spec → submit → poll 收卷(真并发、文件协议留痕);
|
|
9
|
-
级 5(本模块)=意见格式校验 → verify 令牌过 writepipe 六道闸落 contextual。
|
|
10
|
-
|
|
11
|
-
纪律是结构性的,不靠自觉:
|
|
12
|
-
· 评审不改知识层:落库层固定 contextual,verify 令牌 layers_allow 不含 knowledge
|
|
13
|
-
——库层 require_layer_write 拦截,任何拦截器都绕不过(§7 反面清单第 1 条)。
|
|
14
|
-
· 自验禁止:产出意见者(reflect)与落库裁决者(verify)不得同一 actor,判据
|
|
15
|
-
复用 crosscheck.detect_self_verify(不另造判据,保持真源单一)。
|
|
16
|
-
· 不确定即 DEFER:解析失败/格式越界一律不落库(Precision over noise)。
|
|
17
|
-
· 零写入面:worker 只读 spec/context、只写自己的 result.json;认知图写入只发生在
|
|
18
|
-
收卷后的 apply_opinions(本模块,走 writepipe 通道)。
|
|
19
|
-
|
|
20
|
-
运行:python -m md_cg.mreview.pipeline --help
|
|
21
|
-
"""
|
|
22
|
-
from __future__ import annotations
|
|
23
|
-
|
|
24
|
-
import json
|
|
25
|
-
import os
|
|
26
|
-
import re
|
|
27
|
-
import subprocess
|
|
28
|
-
import tempfile
|
|
29
|
-
import time
|
|
30
|
-
import uuid
|
|
31
|
-
|
|
32
|
-
from .. import crosscheck as CC
|
|
33
|
-
from .. import tokens as TK
|
|
34
|
-
from .. import writepipe as WP
|
|
35
|
-
from ..security import AccessDenied, Principal
|
|
36
|
-
from . import candidates as CD
|
|
37
|
-
from . import bundle as BD
|
|
38
|
-
from . import ruleset as RS
|
|
39
|
-
|
|
40
|
-
__all__ = [
|
|
41
|
-
"VERDICTS", "OPINION_LAYER", "DEFAULT_MODEL", "TERMINAL_STATES",
|
|
42
|
-
"hive_exe", "build_packages", "build_prompt", "build_spec", "dump_context",
|
|
43
|
-
"submit", "submit_many", "poll", "wait_jobs", "read_result", "collect",
|
|
44
|
-
"parse_opinions", "validate_opinion", "validate_opinions",
|
|
45
|
-
"verifier_principal", "opinion_doc", "apply_opinions", "run_package",
|
|
46
|
-
"run_batch", "main",
|
|
47
|
-
]
|
|
48
|
-
|
|
49
|
-
# ---- 常量 ----------------------------------------------------------------
|
|
50
|
-
|
|
51
|
-
VERDICTS = ("ACCEPT", "DEFER", "REJECT", "BLINDSPOT") # 意见四态(资格裁决同源)
|
|
52
|
-
OPINION_LAYER = "contextual" # verify 令牌 layers_allow=rejected/contextual
|
|
53
|
-
OPINION_BASIS = "measurement" # 意见的验证基底:确定性规则命中 + 复核
|
|
54
|
-
TERMINAL_STATES = ("done", "error", "timeout", "killed")
|
|
55
|
-
DEFAULT_MODEL = "glm-4-flash"
|
|
56
|
-
_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
|
|
57
|
-
|
|
58
|
-
SYSTEM_PROMPT = """你是记忆评审流水线中的【验证单元】。
|
|
59
|
-
|
|
60
|
-
职责边界(违反即无效):
|
|
61
|
-
1. 只对给定节点清单逐条裁决,不得引入清单外的节点。
|
|
62
|
-
2. 只输出意见,不改写任何记忆——你没有写入通道,也不需要。
|
|
63
|
-
3. 不确定就 DEFER(硬纪律):宁可漏判,不可错判。
|
|
64
|
-
4. 只有拿得出证据时才 REJECT;说不出证据就 DEFER。
|
|
65
|
-
5. ACCEPT 表示「已确认适用」,不是「没发现问题」——正条件无法确认时用 DEFER。
|
|
66
|
-
|
|
67
|
-
四态语义:
|
|
68
|
-
- ACCEPT 已确认适用
|
|
69
|
-
- DEFER 条件不足/无法确认(默认落点)
|
|
70
|
-
- REJECT 有据否定(冲突/来源许可不足/重复且劣于既有)
|
|
71
|
-
- BLINDSPOT 当前观测位置看不见(不是"不知道",是"这个位置判断不了")
|
|
72
|
-
|
|
73
|
-
输出契约(严格 JSON,无额外文字、无 markdown 围栏):
|
|
74
|
-
{"verdicts": [{"node_id": "...", "verdict": "ACCEPT|DEFER|REJECT|BLINDSPOT",
|
|
75
|
-
"reason": "一句话理由", "evidence": "证据(REJECT 必填)"}],
|
|
76
|
-
"summary": "整包一句话结论"}
|
|
77
|
-
清单中每个节点都要有一条裁决,一条不多一条不少。"""
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
# ---- 级 4-A · spec 构造 ---------------------------------------------------
|
|
81
|
-
|
|
82
|
-
# 生效条件:exe 为真值时返回 str(exe),否则 HIVE_EXE 经 (os.environ.get("HIVE_EXE") or "").strip() 后非空则返回该值,否则按 os.name 返回 os.path.join(_ROOT, "hive", "target", "release", "hive.exe" 或 "hive");
|
|
83
|
-
def hive_exe(exe=None) -> str:
|
|
84
|
-
"""蜂巢可执行文件:显式参数 > HIVE_EXE > 仓内 release 构建。"""
|
|
85
|
-
if exe:
|
|
86
|
-
return str(exe)
|
|
87
|
-
env = (os.environ.get("HIVE_EXE") or "").strip()
|
|
88
|
-
if env:
|
|
89
|
-
return env
|
|
90
|
-
name = "hive.exe" if os.name == "nt" else "hive"
|
|
91
|
-
return os.path.join(_ROOT, "hive", "target", "release", name)
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
# 生效条件:name 为假值(None/""/0/False 等)时以 "pkg" 为输入,否则 str(name),结果经 re.sub(r"[^0-9A-Za-z_.\-]", "_", ...) 替换并截取前 120 字符;
|
|
95
|
-
def _safe(name) -> str:
|
|
96
|
-
return re.sub(r"[^0-9A-Za-z_.\-]", "_", str(name or "pkg"))[:120]
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
# 生效条件:path 与 obj 给定后,若 os.path.dirname(path) 非空则 os.makedirs 该目录,随后以 UTF-8、ensure_ascii=False、indent=2 将 obj 写入 path 并返回 path;
|
|
100
|
-
def _dump_json(path: str, obj) -> str:
|
|
101
|
-
d = os.path.dirname(path)
|
|
102
|
-
if d:
|
|
103
|
-
os.makedirs(d, exist_ok=True)
|
|
104
|
-
with open(path, "w", encoding="utf-8") as f:
|
|
105
|
-
json.dump(obj, f, ensure_ascii=False, indent=2)
|
|
106
|
-
return path
|
|
107
|
-
|
|
108
|
-
|
|
109
|
-
_ENTRY_KEYS = ("ref", "node_id", "proposal_id", "origin", "layer", "tags", "role",
|
|
110
|
-
"importance", "evidence_count", "verification_basis",
|
|
111
|
-
"lifecycle_state", "content_hash", "issue_kinds", "evidence",
|
|
112
|
-
"excerpt")
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
# 生效条件:pkg/asm/workdir 给定,base 取 _safe(pkg.get("bundle_id"))(bundle_id 假值→"pkg"),entries 取 pkg.get("entries") or [] 的前 int(entry_limit) 项且仅保留 _ENTRY_KEYS,写出 workdir/base.bundle.json 与 workdir/base.rules.json(asm 原样),返回这两个 join(workdir,...) 路径(workdir 相对时不保证绝对);
|
|
116
|
-
def dump_context(pkg: dict, asm: dict, workdir: str, *, entry_limit=64) -> list:
|
|
117
|
-
"""包 + 规则装配 → context 文件(**绝对路径**,工人执行期内有效)。
|
|
118
|
-
|
|
119
|
-
相对路径的解析基准是 spec.workdir,跨进程易错;绝对路径无歧义。
|
|
120
|
-
"""
|
|
121
|
-
os.makedirs(workdir, exist_ok=True)
|
|
122
|
-
base = _safe(pkg.get("bundle_id"))
|
|
123
|
-
bp = os.path.join(workdir, base + ".bundle.json")
|
|
124
|
-
rp = os.path.join(workdir, base + ".rules.json")
|
|
125
|
-
ents = [{k: e.get(k) for k in _ENTRY_KEYS}
|
|
126
|
-
for e in (pkg.get("entries") or [])[:int(entry_limit)]]
|
|
127
|
-
_dump_json(bp, {"bundle_id": pkg.get("bundle_id"),
|
|
128
|
-
"group_kind": pkg.get("group_kind"),
|
|
129
|
-
"group_key": pkg.get("group_key"),
|
|
130
|
-
"size": pkg.get("size"), "entries": ents})
|
|
131
|
-
_dump_json(rp, asm)
|
|
132
|
-
return [bp, rp]
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
# 生效条件:pkg/asm 给定,节点清单标题恒输出、节点项取 pkg.get("entries") or [];机械命中取 asm.get("mechanical") or [],非空列前 int(max_mech) 条否则输出「无」,llm 取 asm.get("llm") or [] 非空列前 40 条;返回 join(L);
|
|
136
|
-
def build_prompt(pkg: dict, asm: dict, *, max_mech=120) -> str:
|
|
137
|
-
"""user_prompt:节点清单 + 机械命中 + 待检问题 + 输出要求。"""
|
|
138
|
-
ents = pkg.get("entries") or []
|
|
139
|
-
L = ["【待评包】bundle_id=%s group=%s/%s 节点数=%d"
|
|
140
|
-
% (pkg.get("bundle_id"), pkg.get("group_kind"),
|
|
141
|
-
pkg.get("group_key"), len(ents)), ""]
|
|
142
|
-
L.append("【节点清单】(仅可裁决以下节点)")
|
|
143
|
-
for i, e in enumerate(ents, 1):
|
|
144
|
-
frag = (e.get("excerpt") or "").strip().replace("\n", " ")[:220]
|
|
145
|
-
L.append("%d) node_id=%s layer=%s basis=%s tags=%s"
|
|
146
|
-
% (i, e.get("node_id"), e.get("layer"),
|
|
147
|
-
e.get("verification_basis"),
|
|
148
|
-
",".join(e.get("tags") or []) or "-"))
|
|
149
|
-
L.append(" 正文摘录:%s" % (frag or "(无正文/正文缺失)"))
|
|
150
|
-
if e.get("issue_kinds"):
|
|
151
|
-
L.append(" 机械初筛命中:%s" % ",".join(e["issue_kinds"]))
|
|
152
|
-
L.append("")
|
|
153
|
-
mech = asm.get("mechanical") or []
|
|
154
|
-
if mech:
|
|
155
|
-
L.append("【确定性规则命中】共 %d 条(机械层已给出,只需判断其语义后果)"
|
|
156
|
-
% len(mech))
|
|
157
|
-
for m in mech[:int(max_mech)]:
|
|
158
|
-
L.append("- [%s] %s :: %s"
|
|
159
|
-
% (m.get("issue_kind"), m.get("ref") or m.get("node_id"),
|
|
160
|
-
str(m.get("detail") or m.get("message") or "")[:160]))
|
|
161
|
-
else:
|
|
162
|
-
L.append("【确定性规则命中】无(本包未触发任何机械规则)")
|
|
163
|
-
L.append("")
|
|
164
|
-
llm = asm.get("llm") or []
|
|
165
|
-
if llm:
|
|
166
|
-
L.append("【待你判断的问题】逐项判断,并在 reason 中体现结论:")
|
|
167
|
-
for q in llm[:40]:
|
|
168
|
-
L.append("- (规则 %s / 检查 %s) %s"
|
|
169
|
-
% (q.get("rule_id"), q.get("check"), q.get("question")))
|
|
170
|
-
L.append("")
|
|
171
|
-
L.append("【输出】严格按 system 中的 JSON 契约,对每个 node_id 给出一条裁决。")
|
|
172
|
-
return "\n".join(L)
|
|
173
|
-
|
|
174
|
-
|
|
175
|
-
# 生效条件:pkg/asm 给定,model 按 model or os.environ.get('HIVE_MODEL') or DEFAULT_MODEL 取值(空串/假值均回落),context_files 按 list(context_files or []),timeout_s 转 int、temperature 转 float、max_tokens 转 int,node_ids 取 pkg.entries 中 node_id 真值的 str,返回含 meta 的 spec 字典;
|
|
176
|
-
def build_spec(pkg: dict, asm: dict, *, context_files=None, model=None,
|
|
177
|
-
timeout_s=600, temperature=0.0, max_tokens=4096) -> dict:
|
|
178
|
-
"""(pkg, asm) → hive spec(文件协议的可投递单元)。"""
|
|
179
|
-
nids = [str(e.get("node_id")) for e in (pkg.get("entries") or [])
|
|
180
|
-
if e.get("node_id")]
|
|
181
|
-
return {"model": model or os.environ.get("HIVE_MODEL") or DEFAULT_MODEL,
|
|
182
|
-
"system_prompt": SYSTEM_PROMPT,
|
|
183
|
-
"user_prompt": build_prompt(pkg, asm),
|
|
184
|
-
"context_files": list(context_files or []),
|
|
185
|
-
"timeout_s": int(timeout_s),
|
|
186
|
-
"temperature": float(temperature),
|
|
187
|
-
"max_tokens": int(max_tokens),
|
|
188
|
-
# 归因元信息(hive 忽略;收卷侧用于配对与审计)
|
|
189
|
-
"meta": {"bundle_id": pkg.get("bundle_id"),
|
|
190
|
-
"group_kind": pkg.get("group_kind"),
|
|
191
|
-
"group_key": pkg.get("group_key"),
|
|
192
|
-
"size": pkg.get("size"), "node_ids": nids}}
|
|
193
|
-
|
|
194
|
-
|
|
195
|
-
# 生效条件:当模块级常量 CD.SOURCES 可用时,sources 为假值则回落 CD.SOURCES,调用 CD.generate 生成候选,再经 BD.bundle(c.get('candidates') or [], nodes, root, max_per_bundle) 与 RS.assemble_all(b, rules_dir) 组装,最终将 bundles 与 packages 按 zip 配对返回。
|
|
196
|
-
def build_packages(*, root=None, sources=None, limit=None, nodes=None,
|
|
197
|
-
max_per_bundle=50, rules_dir=None, now=None) -> dict:
|
|
198
|
-
"""级 1-3 串联:候选 → 捆包 → 规则装配,产出配对的 (pkg, asm) 列表。"""
|
|
199
|
-
c = CD.generate(root, sources=sources or CD.SOURCES, nodes=nodes, limit=limit,
|
|
200
|
-
with_report=False, now=now)
|
|
201
|
-
b = BD.bundle(c.get("candidates") or [], nodes=nodes, root=root,
|
|
202
|
-
max_per_bundle=max_per_bundle)
|
|
203
|
-
a = RS.assemble_all(b, rules_dir=rules_dir)
|
|
204
|
-
pairs = []
|
|
205
|
-
for pkg, asm in zip(b.get("bundles") or [], a.get("packages") or []):
|
|
206
|
-
pairs.append((pkg, asm))
|
|
207
|
-
return {"candidates": c, "bundles": b, "assembled": a, "pairs": pairs}
|
|
208
|
-
|
|
209
|
-
|
|
210
|
-
# ---- 级 4-B · 投递 / 收卷 -------------------------------------------------
|
|
211
|
-
|
|
212
|
-
# 生效条件:argv 给定,环境变量副本强制 PYTHONUTF8="1",env 真值时用其 str 值更新,随后 subprocess.run(argv, capture_output=True, text=True, encoding="utf-8", errors="replace", env=e, timeout=timeout, cwd=cwd, input=stdin_text, shell=False);
|
|
213
|
-
def _run(argv, *, env=None, timeout=120, cwd=None, stdin_text=None):
|
|
214
|
-
"""统一子进程入口:argv 列表 + 显式 UTF-8 + PYTHONUTF8=1,不经 shell(第15条)。"""
|
|
215
|
-
e = dict(os.environ)
|
|
216
|
-
e["PYTHONUTF8"] = "1"
|
|
217
|
-
if env:
|
|
218
|
-
e.update({k: str(v) for k, v in env.items()})
|
|
219
|
-
return subprocess.run(argv, capture_output=True, text=True, encoding="utf-8",
|
|
220
|
-
errors="replace", env=e, timeout=timeout, cwd=cwd,
|
|
221
|
-
input=stdin_text, shell=False)
|
|
222
|
-
|
|
223
|
-
|
|
224
|
-
# 生效条件:text 为假值时 (text or "").strip().splitlines() 为空、循环不执行并返回 None;否则从末尾行向前找以 "{" 开头且 json.loads 成功且 isinstance(obj, dict) 的行返回该 dict,未找到返回 None;
|
|
225
|
-
def _last_json(text: str):
|
|
226
|
-
"""stdout 逐行解析取最后一个 JSON 对象(容忍前导日志行)。"""
|
|
227
|
-
for line in reversed((text or "").strip().splitlines()):
|
|
228
|
-
s = line.strip()
|
|
229
|
-
if not s.startswith("{"):
|
|
230
|
-
continue
|
|
231
|
-
try:
|
|
232
|
-
obj = json.loads(s)
|
|
233
|
-
except ValueError:
|
|
234
|
-
continue
|
|
235
|
-
if isinstance(obj, dict):
|
|
236
|
-
return obj
|
|
237
|
-
return None
|
|
238
|
-
|
|
239
|
-
|
|
240
|
-
# 生效条件:exe/jobs_dir/spec 给定,创建 jobs_dir 并在临时目录写 spec.json 后执行 [exe,"submit","--spec",sp,"--jobs",jobs_dir];returncode 非 0 返回 ok False + error,否则 _last_json(p.stdout) 为假(含 None/空 dict)返回 ok False「submit 输出非 JSON」,否则 d.setdefault("ok", True) 后返回 d;
|
|
241
|
-
def submit(exe: str, jobs_dir: str, spec: dict, *, timeout=120) -> dict:
|
|
242
|
-
"""投递单包 spec → {"ok":True,"job_id":...};失败返回 ok=False + error。"""
|
|
243
|
-
os.makedirs(jobs_dir, exist_ok=True)
|
|
244
|
-
with tempfile.TemporaryDirectory(prefix="mrev_spec_") as td:
|
|
245
|
-
sp = _dump_json(os.path.join(td, "spec.json"), spec)
|
|
246
|
-
p = _run([exe, "submit", "--spec", sp, "--jobs", jobs_dir], timeout=timeout)
|
|
247
|
-
if p.returncode != 0:
|
|
248
|
-
return {"ok": False, "returncode": p.returncode,
|
|
249
|
-
"error": (p.stderr or p.stdout or "").strip()[:800]}
|
|
250
|
-
d = _last_json(p.stdout)
|
|
251
|
-
if not d:
|
|
252
|
-
return {"ok": False, "error": "submit 输出非 JSON",
|
|
253
|
-
"stdout": (p.stdout or "")[:800]}
|
|
254
|
-
d.setdefault("ok", True)
|
|
255
|
-
return d
|
|
256
|
-
|
|
257
|
-
|
|
258
|
-
# 生效条件:specs 为可迭代列表时逐项调用 submit(exe, jobs_dir, s, timeout=timeout) 并返回等长列表;specs 为空则返回空列表;
|
|
259
|
-
def submit_many(exe: str, jobs_dir: str, specs: list, *, timeout=120) -> list:
|
|
260
|
-
"""批量投递(hive 侧并发消费;本函数逐条投递并保序配对)。"""
|
|
261
|
-
return [submit(exe, jobs_dir, s, timeout=timeout) for s in specs]
|
|
262
|
-
|
|
263
|
-
|
|
264
|
-
# 生效条件:exe/jobs_dir 给定,job_id 为真值时 argv 追加 str(job_id),假值(None/"")时仅列全部;运行后 returncode 非 0 返回 ok False + error,否则 _last_json(p.stdout) 为真值(非空 dict)时返回该 dict,为假值时返回 ok False「poll 输出非 JSON」;
|
|
265
|
-
def poll(exe: str, jobs_dir: str, job_id=None, *, timeout=120) -> dict:
|
|
266
|
-
"""查询 job 状态(不传 job_id = 列出全部)。"""
|
|
267
|
-
argv = [exe, "poll"]
|
|
268
|
-
if job_id:
|
|
269
|
-
argv.append(str(job_id))
|
|
270
|
-
argv += ["--jobs", jobs_dir]
|
|
271
|
-
p = _run(argv, timeout=timeout)
|
|
272
|
-
if p.returncode != 0:
|
|
273
|
-
return {"ok": False, "returncode": p.returncode,
|
|
274
|
-
"error": (p.stderr or p.stdout or "").strip()[:800]}
|
|
275
|
-
d = _last_json(p.stdout)
|
|
276
|
-
return d or {"ok": False, "error": "poll 输出非 JSON",
|
|
277
|
-
"stdout": (p.stdout or "")[:800]}
|
|
278
|
-
|
|
279
|
-
|
|
280
|
-
# 生效条件:j.get("job_id") 为真值时返回该值,否则返回 j.get("id")(若 id 为 0/""/False 则原样返回该假值,若 id 缺失返回 None);
|
|
281
|
-
def _job_id(j: dict):
|
|
282
|
-
return j.get("job_id") or j.get("id")
|
|
283
|
-
|
|
284
|
-
|
|
285
|
-
# 生效条件:j.get("state") 为真值时返回 str(state);否则 j.get("status") 为 dict 时返回 str(status.get("state") or "");否则返回 str(status or "");
|
|
286
|
-
def _job_state(j: dict) -> str:
|
|
287
|
-
"""state 读取兼容两种形态(扁平 / status 嵌套);仅供轮询判终态。"""
|
|
288
|
-
s = j.get("state")
|
|
289
|
-
if s:
|
|
290
|
-
return str(s)
|
|
291
|
-
st = j.get("status")
|
|
292
|
-
if isinstance(st, dict):
|
|
293
|
-
return str(st.get("state") or "")
|
|
294
|
-
return str(st or "")
|
|
295
|
-
|
|
296
|
-
|
|
297
|
-
# 生效条件:exe/jobs_dir 给定,job_ids 真值集合限观测;循环 poll 直到 pending 为空且(want 为 None 或 want <= 已观测)返回 ok=not failed,若 pending 非空或 want 未全观测且到 max(1.0,float(timeout_s)) 截止则返回 ok False + error,每次间隔 sleep(max(0.05,float(interval_s)));
|
|
298
|
-
def wait_jobs(exe: str, jobs_dir: str, job_ids=None, *, timeout_s=900,
|
|
299
|
-
interval_s=1.0, on_tick=None) -> dict:
|
|
300
|
-
"""轮询至全部终态(或超时)→ done/failed/pending 三分类 + 原始快照。"""
|
|
301
|
-
want = set(str(x) for x in (job_ids or [])) or None
|
|
302
|
-
deadline = time.time() + max(1.0, float(timeout_s))
|
|
303
|
-
last = {}
|
|
304
|
-
while True:
|
|
305
|
-
r = poll(exe, jobs_dir)
|
|
306
|
-
states = {}
|
|
307
|
-
for j in (r.get("jobs") or []):
|
|
308
|
-
jid = _job_id(j)
|
|
309
|
-
if jid is None:
|
|
310
|
-
continue
|
|
311
|
-
jid = str(jid)
|
|
312
|
-
if want is not None and jid not in want:
|
|
313
|
-
continue
|
|
314
|
-
states[jid] = _job_state(j) or "pending"
|
|
315
|
-
last[jid] = j
|
|
316
|
-
done = sorted(k for k, v in states.items() if v == "done")
|
|
317
|
-
failed = sorted(k for k, v in states.items()
|
|
318
|
-
if v in TERMINAL_STATES and v != "done")
|
|
319
|
-
pending = sorted(k for k, v in states.items() if v not in TERMINAL_STATES)
|
|
320
|
-
if on_tick:
|
|
321
|
-
on_tick({"states": dict(states), "done": len(done),
|
|
322
|
-
"pending": len(pending), "failed": len(failed)})
|
|
323
|
-
if not pending:
|
|
324
|
-
if want is None or want <= set(states):
|
|
325
|
-
return {"ok": not failed, "done": done, "failed": failed,
|
|
326
|
-
"pending": [], "states": states, "jobs": last}
|
|
327
|
-
if time.time() >= deadline:
|
|
328
|
-
miss = sorted(want - set(states))
|
|
329
|
-
return {"ok": False, "done": done, "failed": failed,
|
|
330
|
-
"pending": miss, "states": states, "jobs": last,
|
|
331
|
-
"error": "未在超时内观测到 job:%s" % ",".join(miss)}
|
|
332
|
-
if time.time() >= deadline:
|
|
333
|
-
return {"ok": False, "done": done, "failed": failed,
|
|
334
|
-
"pending": pending, "states": states, "jobs": last,
|
|
335
|
-
"error": "等待终态超时(%ss)" % timeout_s}
|
|
336
|
-
time.sleep(max(0.05, float(interval_s)))
|
|
337
|
-
|
|
338
|
-
|
|
339
|
-
# 生效条件:jobs_dir/job_id 给定,join(jobs_dir,str(job_id),"result.json") 路径缺失时返回 ok False「result.json 不存在」,读取时 OSError/ValueError 返回 ok False「读取失败」,json 对象非 dict 返回 ok False「result 非对象」,否则返回该 dict;
|
|
340
|
-
def read_result(jobs_dir: str, job_id: str) -> dict:
|
|
341
|
-
"""读工人产出(job 目录的 result.json)。"""
|
|
342
|
-
p = os.path.join(jobs_dir, str(job_id), "result.json")
|
|
343
|
-
if not os.path.isfile(p):
|
|
344
|
-
return {"ok": False, "error": "result.json 不存在:%s" % p}
|
|
345
|
-
try:
|
|
346
|
-
with open(p, encoding="utf-8") as f:
|
|
347
|
-
obj = json.load(f)
|
|
348
|
-
except (OSError, ValueError) as exc:
|
|
349
|
-
return {"ok": False, "error": "result.json 读取失败:%s" % exc}
|
|
350
|
-
return obj if isinstance(obj, dict) else {"ok": False, "error": "result 非对象"}
|
|
351
|
-
|
|
352
|
-
|
|
353
|
-
# 生效条件:jobs_dir/job_ids 给定,job_ids 为假值(None/[])时返回空 dict,否则返回 {str(j): read_result(jobs_dir, j) for j in job_ids};
|
|
354
|
-
def collect(jobs_dir: str, job_ids) -> dict:
|
|
355
|
-
"""按 job_id 收卷(读 result.json),返回 {job_id: result} 保序映射。"""
|
|
356
|
-
return {str(j): read_result(jobs_dir, j) for j in (job_ids or [])}
|
|
357
|
-
|
|
358
|
-
|
|
359
|
-
# ---- 级 5-A · 意见解析与格式校验(校验率 100% 是验收口径)-----------------
|
|
360
|
-
|
|
361
|
-
# 生效条件:text 去空格后为空返回 None;否则先 try json.loads(整体),成功返回;否则找 ```(?:json)?...``` 围栏内 json.loads 成功返回;否则取 s.find("{") 与 s.rfind("}") 且 0<=i<k 的切片 json.loads 成功返回;均失败返回 None;
|
|
362
|
-
def _extract_json(text: str):
|
|
363
|
-
"""从模型文本抽 JSON 对象:整体 → ```围栏``` → 首尾大括号切片。"""
|
|
364
|
-
s = (text or "").strip()
|
|
365
|
-
if not s:
|
|
366
|
-
return None
|
|
367
|
-
try:
|
|
368
|
-
return json.loads(s)
|
|
369
|
-
except ValueError:
|
|
370
|
-
pass
|
|
371
|
-
m = re.search(r"```(?:json)?\s*(.+?)```", s, re.S)
|
|
372
|
-
if m:
|
|
373
|
-
try:
|
|
374
|
-
return json.loads(m.group(1).strip())
|
|
375
|
-
except ValueError:
|
|
376
|
-
pass
|
|
377
|
-
i, k = s.find("{"), s.rfind("}")
|
|
378
|
-
if 0 <= i < k:
|
|
379
|
-
try:
|
|
380
|
-
return json.loads(s[i:k + 1])
|
|
381
|
-
except ValueError:
|
|
382
|
-
pass
|
|
383
|
-
return None
|
|
384
|
-
|
|
385
|
-
|
|
386
|
-
# 生效条件:result 非 dict 返回 ok False;result.get("error") 真值返回 ok False「工人报错」;v=result.get("verdicts"),v is None 时从 result.get("content") or "" 经 _extract_json 取 dict,非 dict 返回 ok False,否则 v=obj.get("verdicts") 且 summary=obj.get("summary") or summary;v 非 list 返回 ok False;否则返回 ok True、verdicts=v、summary=str(summary);
|
|
387
|
-
def parse_opinions(result: dict) -> dict:
|
|
388
|
-
"""result.json → {"ok":True,"verdicts":[...],"summary":...}。
|
|
389
|
-
|
|
390
|
-
两形态都能收:
|
|
391
|
-
· 执行器直写 {"ok":true,"verdicts":[...],"summary":...}
|
|
392
|
-
· LLM 文本 {"ok":true,"content":"{...JSON...}"}(exec.py 默认形态)
|
|
393
|
-
"""
|
|
394
|
-
if not isinstance(result, dict):
|
|
395
|
-
return {"ok": False, "error": "result 非对象"}
|
|
396
|
-
if result.get("error"):
|
|
397
|
-
return {"ok": False, "error": "工人报错:%s" % result["error"]}
|
|
398
|
-
v = result.get("verdicts")
|
|
399
|
-
summary = result.get("summary") or ""
|
|
400
|
-
if v is None:
|
|
401
|
-
obj = _extract_json(result.get("content") or "")
|
|
402
|
-
if not isinstance(obj, dict):
|
|
403
|
-
return {"ok": False, "error": "未能从 content 中解析 JSON 意见块"}
|
|
404
|
-
v = obj.get("verdicts")
|
|
405
|
-
summary = obj.get("summary") or summary
|
|
406
|
-
if not isinstance(v, list):
|
|
407
|
-
return {"ok": False, "error": "verdicts 非数组(got %s)" % type(v).__name__}
|
|
408
|
-
return {"ok": True, "verdicts": v, "summary": str(summary)}
|
|
409
|
-
|
|
410
|
-
|
|
411
|
-
# 生效条件:v 经 str(v or "").strip().upper() 得 s,若 s 在模块级 VERDICTS 中则返回 s,否则返回 None;
|
|
412
|
-
def _norm_verdict(v):
|
|
413
|
-
s = str(v or "").strip().upper()
|
|
414
|
-
return s if s in VERDICTS else None
|
|
415
|
-
|
|
416
|
-
|
|
417
|
-
# 生效条件:op 非 dict 返回「意见非对象」;nid 为空返回「缺 node_id」;node_ids 非 None 且 nid 不在其中返回越界错误;_norm_verdict(op.get("verdict")) 为 None 返回「verdict 越界」;reason 去空格空返回「缺 reason」;vd=="REJECT" 且 evidence 去空格空返回「REJECT 缺 evidence」;否则返回 None;
|
|
418
|
-
def validate_opinion(op, node_ids=None):
|
|
419
|
-
"""单条意见格式校验:合法返回 None,非法返回错误串。"""
|
|
420
|
-
if not isinstance(op, dict):
|
|
421
|
-
return "意见非对象"
|
|
422
|
-
nid = str(op.get("node_id") or "").strip()
|
|
423
|
-
if not nid:
|
|
424
|
-
return "缺 node_id"
|
|
425
|
-
if node_ids is not None and nid not in node_ids:
|
|
426
|
-
return "node_id 不属于本包:%s" % nid
|
|
427
|
-
vd = _norm_verdict(op.get("verdict"))
|
|
428
|
-
if vd is None:
|
|
429
|
-
return "verdict 越界:%r" % (op.get("verdict"),)
|
|
430
|
-
if not str(op.get("reason") or "").strip():
|
|
431
|
-
return "缺 reason"
|
|
432
|
-
if vd == "REJECT" and not str(op.get("evidence") or "").strip():
|
|
433
|
-
return "REJECT 缺 evidence(有据才拒)"
|
|
434
|
-
return None
|
|
435
|
-
|
|
436
|
-
|
|
437
|
-
# 生效条件:verdicts 为假值(None/[])时 total=0、rate=1.0、ok=False;否则逐条 validate_opinion,无效记入 invalid,有效项复制并写入 _norm_verdict(op.get("verdict")),最后返回 total、valid、invalid、rate=len(valid)/total、ok=bool(valid) and not invalid;
|
|
438
|
-
def validate_opinions(verdicts, node_ids=None) -> dict:
|
|
439
|
-
"""批量校验:valid/invalid 明细 + 通过率(用于 100% 口径断言)。"""
|
|
440
|
-
total = len(verdicts or [])
|
|
441
|
-
valid, invalid = [], []
|
|
442
|
-
for i, op in enumerate(verdicts or []):
|
|
443
|
-
err = validate_opinion(op, node_ids)
|
|
444
|
-
if err:
|
|
445
|
-
invalid.append({"index": i, "error": err,
|
|
446
|
-
"node_id": (op.get("node_id")
|
|
447
|
-
if isinstance(op, dict) else None)})
|
|
448
|
-
continue
|
|
449
|
-
op = dict(op)
|
|
450
|
-
op["verdict"] = _norm_verdict(op.get("verdict"))
|
|
451
|
-
valid.append(op)
|
|
452
|
-
rate = 1.0 if total == 0 else len(valid) / float(total)
|
|
453
|
-
return {"total": total, "valid": valid, "invalid": invalid,
|
|
454
|
-
"pass_rate": rate, "ok": bool(valid) and not invalid}
|
|
455
|
-
|
|
456
|
-
|
|
457
|
-
# ---- 级 5-B · 落库(verify 令牌 + writepipe 六道闸;不自造通道)-----------
|
|
458
|
-
|
|
459
|
-
# 生效条件:当 actor 给出时,从 TK.role_spec('verify') 读取 clearance_cap/can_write/can_admin/layers_allow 构造 Principal;session 为假值时生成 'mrev_' + uuid 前 12 位;layers_allow 为假值时回落 spec.get('layers_allow') 或 [];ops_allow 原样传入。
|
|
460
|
-
def verifier_principal(*, actor="mreview-verifier", session=None,
|
|
461
|
-
layers_allow=None, ops_allow=None) -> Principal:
|
|
462
|
-
"""验证单元 Principal——角色规格取自 tokens 真源(避免手写漂移)。"""
|
|
463
|
-
spec = TK.role_spec("verify")
|
|
464
|
-
return Principal(actor=actor, clearance=spec.get("clearance_cap") or "internal",
|
|
465
|
-
can_write=bool(spec.get("can_write")),
|
|
466
|
-
can_admin=bool(spec.get("can_admin")),
|
|
467
|
-
role="verify", unit="verify",
|
|
468
|
-
session=session or ("mrev_" + uuid.uuid4().hex[:12]),
|
|
469
|
-
layers_allow=list(layers_allow or spec.get("layers_allow") or []),
|
|
470
|
-
ops_allow=ops_allow, auth_mode="mreview")
|
|
471
|
-
|
|
472
|
-
|
|
473
|
-
# 生效条件:pkg、opinions、reviewer 给定时恒返回 "\n".join(L);summary 为真值才追加「整包结论」行,invalid 为真值才追加「【未落库条目】」行,二者取默认 "" / None 时不追加;
|
|
474
|
-
def opinion_doc(pkg: dict, opinions: list, *, reviewer: str, summary="",
|
|
475
|
-
invalid=None) -> str:
|
|
476
|
-
"""评审意见节点正文(核心修改四要素:内容/原因/位置/验证)。"""
|
|
477
|
-
cnt = {k: 0 for k in VERDICTS}
|
|
478
|
-
for o in opinions:
|
|
479
|
-
cnt[_norm_verdict(o.get("verdict")) or "?"] = \
|
|
480
|
-
cnt.get(_norm_verdict(o.get("verdict")) or "?", 0) + 1
|
|
481
|
-
dist = " / ".join("%s %d" % (k, cnt.get(k, 0)) for k in VERDICTS)
|
|
482
|
-
L = ["# 记忆评审意见 · 包 %s(%s/%s,%d 节点)"
|
|
483
|
-
% (pkg.get("bundle_id"), pkg.get("group_kind"), pkg.get("group_key"),
|
|
484
|
-
len(pkg.get("entries") or [])),
|
|
485
|
-
"来源:级 4 蜂巢并发评审(评审者=%s)" % reviewer,
|
|
486
|
-
"裁决分布:%s" % dist]
|
|
487
|
-
if summary:
|
|
488
|
-
L.append("整包结论:%s" % summary)
|
|
489
|
-
L.append("")
|
|
490
|
-
for o in opinions:
|
|
491
|
-
seg = "- [%s] %s:%s" % (_norm_verdict(o.get("verdict")) or "?",
|
|
492
|
-
o.get("node_id"),
|
|
493
|
-
str(o.get("reason") or "").strip())
|
|
494
|
-
ev = str(o.get("evidence") or "").strip()
|
|
495
|
-
if ev:
|
|
496
|
-
seg += "(证据:%s)" % ev[:300]
|
|
497
|
-
L.append(seg)
|
|
498
|
-
if invalid:
|
|
499
|
-
L.append("")
|
|
500
|
-
L.append("【未落库条目】%s" % json.dumps(invalid, ensure_ascii=False))
|
|
501
|
-
L.append("")
|
|
502
|
-
L.append("【内容】级 4 并发评审对包 %s 的逐节点裁决意见。" % pkg.get("bundle_id"))
|
|
503
|
-
L.append("【原因】机械层已过滤确定性缺陷,语义判断留给验证单元;意见落 contextual "
|
|
504
|
-
"供治理层决策——评审不改知识层。")
|
|
505
|
-
L.append("【位置】源包 group=%s/%s(%d 节点)。"
|
|
506
|
-
% (pkg.get("group_kind"), pkg.get("group_key"),
|
|
507
|
-
len(pkg.get("entries") or [])))
|
|
508
|
-
L.append("【验证】机械规则命中 %d 条;意见格式校验 %d/%d 通过(越界条目已剔除)。"
|
|
509
|
-
% (len((pkg.get("_asm") or {}).get("mechanical") or []),
|
|
510
|
-
len(opinions), len(opinions) + len(invalid or [])))
|
|
511
|
-
return "\n".join(L)
|
|
512
|
-
|
|
513
|
-
|
|
514
|
-
# 生效条件:cg/pkg/opinions/reviewer/applier 给定,CC.detect_self_verify 判 reviewer 与 applier 同一→refused self_verify_disallowed;opinions 为空→ok True skipped no_opinions;dry_run 真→ok True committed False;否则 (pipe or WP.default_pipeline()).execute,AccessDenied→refused layer_denied,其他异常→write_error,成功→ok 取 out.ok、committed 取 out.committed、moved_to 取 out.moved_to、deferred=bool(out.moved_to) and not out.committed;
|
|
515
|
-
def apply_opinions(cg, pkg: dict, opinions: list, *, reviewer, applier,
|
|
516
|
-
dry_run=False, pipe=None, layer=OPINION_LAYER,
|
|
517
|
-
importance=0.4, tags=None, node_id=None,
|
|
518
|
-
summary="", invalid=None) -> dict:
|
|
519
|
-
"""落库单包意见(级 5 唯一的写入口)。
|
|
520
|
-
|
|
521
|
-
顺序(任一环不过即不落库):
|
|
522
|
-
1 自验违例检测(评审者 ≠ 落库裁决者,判据复用 crosscheck)
|
|
523
|
-
2 意见非空(全 DEFER 且空 → 不写,零噪声)
|
|
524
|
-
3 verify 令牌 + writepipe 六道闸 + 库层 require_layer_write
|
|
525
|
-
4 合规规则库(部署侧 MDCG_POLICY_FILE)——audit 的 text 验证器在**无规则**
|
|
526
|
-
时恒 DEFER(audit._rule_check「无规则不能假装合规」),意见将转审核队列
|
|
527
|
-
(`moved_to="review_queue"`)而非落盘。这是设计行为不是故障:本函数把
|
|
528
|
-
去向如实透出为 `moved_to` / `deferred`,**不把「入队」报成「落库」**。
|
|
529
|
-
"""
|
|
530
|
-
rows = [{"unit": CC.REFLECT_UNIT, "actor": str(reviewer or "")},
|
|
531
|
-
{"unit": CC.VERIFY_UNIT, "actor": str(applier or "")}]
|
|
532
|
-
if CC.detect_self_verify(rows):
|
|
533
|
-
return {"ok": False, "committed": False, "refused": True,
|
|
534
|
-
"reason": "self_verify_disallowed",
|
|
535
|
-
"detail": "评审者与落库裁决者为同一执行者(%s)——自验被拒(§7)"
|
|
536
|
-
% reviewer}
|
|
537
|
-
if not opinions:
|
|
538
|
-
return {"ok": True, "committed": False, "skipped": True,
|
|
539
|
-
"reason": "no_opinions(无有效意见,零噪声不落库)"}
|
|
540
|
-
nid = node_id or ("mr_opinion_" + _safe(pkg.get("bundle_id")))
|
|
541
|
-
doc = opinion_doc(pkg, opinions, reviewer=str(reviewer), summary=summary,
|
|
542
|
-
invalid=invalid)
|
|
543
|
-
if dry_run:
|
|
544
|
-
return {"ok": True, "committed": False, "dry_run": True, "node_id": nid,
|
|
545
|
-
"doc_len": len(doc)}
|
|
546
|
-
a = {"node_id": nid, "content": doc, "layer": layer,
|
|
547
|
-
"tags": list(tags or ["cap:记忆评审", "mreview", "opinion",
|
|
548
|
-
"bundle:" + _safe(pkg.get("bundle_id"))]),
|
|
549
|
-
"importance": float(importance), "verification_basis": OPINION_BASIS,
|
|
550
|
-
"content_kind": "text",
|
|
551
|
-
# 意见是评审产物(contextual),不参与知识层冲突判定:冲突闸若命中会
|
|
552
|
-
# 把意见当冲突源转入审核队列(语义不符),故显式跳过(闸层可关,
|
|
553
|
-
# 库层 require_layer_write 不可关——评审不改知识层的结构约束仍在)。
|
|
554
|
-
"consistency": False}
|
|
555
|
-
try:
|
|
556
|
-
out = (pipe or WP.default_pipeline()).execute(cg, a)
|
|
557
|
-
except AccessDenied as exc:
|
|
558
|
-
# 结构性拒绝(如 layer=knowledge 越出 verify 令牌的 layers_allow):
|
|
559
|
-
# 如实上报为拒写,不吞异常也不改写写入语义(§7 反面清单第 1 条)。
|
|
560
|
-
return {"ok": False, "committed": False, "refused": True,
|
|
561
|
-
"reason": "layer_denied", "node_id": nid, "layer": layer,
|
|
562
|
-
"detail": "%s: %s" % (type(exc).__name__, exc)}
|
|
563
|
-
except Exception as exc: # noqa: BLE001
|
|
564
|
-
return {"ok": False, "committed": False, "refused": False,
|
|
565
|
-
"reason": "write_error", "node_id": nid, "layer": layer,
|
|
566
|
-
"detail": "%s: %s" % (type(exc).__name__, exc)}
|
|
567
|
-
return {"ok": bool(out.get("ok")), "committed": bool(out.get("committed")),
|
|
568
|
-
"moved_to": out.get("moved_to"),
|
|
569
|
-
# 「改了去向但没落盘」(如合规/冲突闸转审核队列、gated 的 DROP/DEFER)
|
|
570
|
-
# 与「落盘」是两种结局:分开报,调用方不必读 response 才能分辨。
|
|
571
|
-
"deferred": bool(out.get("moved_to")) and not out.get("committed"),
|
|
572
|
-
"node_id": nid, "layer": layer, "response": out,
|
|
573
|
-
"doc_len": len(doc)}
|
|
574
|
-
|
|
575
|
-
|
|
576
|
-
# ---- 编排 ----------------------------------------------------------------
|
|
577
|
-
|
|
578
|
-
# 生效条件:pkg/asm/cg 给定,exe=hive_exe(exe),jobs_dir/ctx_dir 假值回落到 tempfile.gettempdir() 下路径;dump_context 后 build_spec、submit,若 submit 无 ok 或无 _job_id 返回 stage submit;wait_jobs 后若 failed 或 done 空返回 stage wait;read_result+parse_opinions 失败返回 stage parse;否则 validate_opinions 并 apply_opinions,返回 ok=app.ok 及计数;
|
|
579
|
-
def run_package(*, pkg, asm, cg, exe=None, jobs_dir=None, ctx_dir=None,
|
|
580
|
-
reviewer="mreview-worker", applier="mreview-applier",
|
|
581
|
-
model=None, timeout_s=900, interval_s=1.0, dry_run=False,
|
|
582
|
-
keep_jobs=True) -> dict:
|
|
583
|
-
"""单包全链路:dump context → submit → wait → parse → validate → apply。"""
|
|
584
|
-
exe = hive_exe(exe)
|
|
585
|
-
jobs_dir = jobs_dir or os.path.join(tempfile.gettempdir(), "mrev_jobs")
|
|
586
|
-
ctx_dir = ctx_dir or os.path.join(tempfile.gettempdir(), "mrev_ctx")
|
|
587
|
-
ctx_files = dump_context(pkg, asm, ctx_dir)
|
|
588
|
-
spec = build_spec(pkg, asm, context_files=ctx_files, model=model,
|
|
589
|
-
timeout_s=timeout_s)
|
|
590
|
-
sub = submit(exe, jobs_dir, spec)
|
|
591
|
-
if not sub.get("ok") or not _job_id(sub):
|
|
592
|
-
return {"ok": False, "stage": "submit", "error": sub.get("error"),
|
|
593
|
-
"submit": sub, "bundle_id": pkg.get("bundle_id")}
|
|
594
|
-
jid = str(_job_id(sub))
|
|
595
|
-
w = wait_jobs(exe, jobs_dir, [jid], timeout_s=timeout_s,
|
|
596
|
-
interval_s=interval_s)
|
|
597
|
-
if w.get("failed") or not w.get("done"):
|
|
598
|
-
return {"ok": False, "stage": "wait", "job_id": jid, "wait": w,
|
|
599
|
-
"bundle_id": pkg.get("bundle_id")}
|
|
600
|
-
res = read_result(jobs_dir, jid)
|
|
601
|
-
parsed = parse_opinions(res)
|
|
602
|
-
if not parsed.get("ok"):
|
|
603
|
-
return {"ok": False, "stage": "parse", "job_id": jid, "error":
|
|
604
|
-
parsed.get("error"), "bundle_id": pkg.get("bundle_id")}
|
|
605
|
-
nids = {str(e.get("node_id")) for e in (pkg.get("entries") or [])
|
|
606
|
-
if e.get("node_id")}
|
|
607
|
-
val = validate_opinions(parsed["verdicts"], nids or None)
|
|
608
|
-
pkg2 = dict(pkg, _asm=asm)
|
|
609
|
-
app = apply_opinions(cg, pkg2, val["valid"], reviewer=reviewer,
|
|
610
|
-
applier=applier, dry_run=dry_run,
|
|
611
|
-
summary=parsed.get("summary"), invalid=val["invalid"])
|
|
612
|
-
return {"ok": bool(app.get("ok")), "bundle_id": pkg.get("bundle_id"),
|
|
613
|
-
"job_id": jid, "opinions": len(parsed["verdicts"]),
|
|
614
|
-
"valid": len(val["valid"]), "invalid": len(val["invalid"]),
|
|
615
|
-
"pass_rate": val["pass_rate"], "deferred": bool(app.get("deferred")),
|
|
616
|
-
"apply": app, "stages": {"submit": True, "wait": True, "parse": True}}
|
|
617
|
-
|
|
618
|
-
|
|
619
|
-
# 生效条件:pairs/cg 给定,先对每对 pkg,asm 执行 dump_context/build_spec/submit,收集 ok 且有 job_id 的 jids;wait_jobs 后逐 jid 若 state!="done" 记 stage wait,否则 read_result/parse_opinions/validate_opinions/apply_opinions 并累计;返回 ok=all(results.ok) and bool(results) 及聚合计数;
|
|
620
|
-
def run_batch(*, pairs, cg, exe=None, jobs_dir=None, ctx_dir=None,
|
|
621
|
-
reviewer="mreview-worker", applier="mreview-applier", model=None,
|
|
622
|
-
timeout_s=1800, interval_s=1.0, dry_run=False, on_tick=None) -> dict:
|
|
623
|
-
"""批量并发:先全投递(hive worker 池并发消费),再统一等候,再逐包落库。
|
|
624
|
-
|
|
625
|
-
并发发生在**投递之后**——本函数投递不阻塞,故 N 包的墙钟≈最慢一包。
|
|
626
|
-
"""
|
|
627
|
-
exe = hive_exe(exe)
|
|
628
|
-
jobs_dir = jobs_dir or os.path.join(tempfile.gettempdir(), "mrev_jobs")
|
|
629
|
-
ctx_dir = ctx_dir or os.path.join(tempfile.gettempdir(), "mrev_ctx")
|
|
630
|
-
os.makedirs(jobs_dir, exist_ok=True)
|
|
631
|
-
staged, posted, jids = [], [], []
|
|
632
|
-
for pkg, asm in pairs:
|
|
633
|
-
ctx_files = dump_context(pkg, asm, ctx_dir)
|
|
634
|
-
spec = build_spec(pkg, asm, context_files=ctx_files, model=model,
|
|
635
|
-
timeout_s=timeout_s)
|
|
636
|
-
sub = submit(exe, jobs_dir, spec)
|
|
637
|
-
staged.append({"bundle_id": pkg.get("bundle_id"), "pkg": pkg,
|
|
638
|
-
"asm": asm, "submit": sub})
|
|
639
|
-
if sub.get("ok") and _job_id(sub):
|
|
640
|
-
jid = str(_job_id(sub))
|
|
641
|
-
jids.append(jid)
|
|
642
|
-
posted.append(staged[-1])
|
|
643
|
-
w = wait_jobs(exe, jobs_dir, jids, timeout_s=timeout_s,
|
|
644
|
-
interval_s=interval_s, on_tick=on_tick)
|
|
645
|
-
by_jid = {}
|
|
646
|
-
for i, s in enumerate(staged):
|
|
647
|
-
if s["submit"].get("ok") and _job_id(s["submit"]):
|
|
648
|
-
by_jid[str(_job_id(s["submit"]))] = s
|
|
649
|
-
results, ok_n, valid_n, invalid_n, applied_n, deferred_n = [], 0, 0, 0, 0, 0
|
|
650
|
-
for jid in jids:
|
|
651
|
-
s = by_jid.get(jid) or {}
|
|
652
|
-
pkg, asm = s.get("pkg") or {}, s.get("asm") or {}
|
|
653
|
-
state = (w.get("states") or {}).get(jid)
|
|
654
|
-
if state != "done":
|
|
655
|
-
results.append({"ok": False, "bundle_id": pkg.get("bundle_id"),
|
|
656
|
-
"job_id": jid, "stage": "wait", "state": state})
|
|
657
|
-
continue
|
|
658
|
-
res = read_result(jobs_dir, jid)
|
|
659
|
-
parsed = parse_opinions(res)
|
|
660
|
-
if not parsed.get("ok"):
|
|
661
|
-
results.append({"ok": False, "bundle_id": pkg.get("bundle_id"),
|
|
662
|
-
"job_id": jid, "stage": "parse",
|
|
663
|
-
"error": parsed.get("error")})
|
|
664
|
-
continue
|
|
665
|
-
nids = {str(e.get("node_id")) for e in (pkg.get("entries") or [])
|
|
666
|
-
if e.get("node_id")}
|
|
667
|
-
val = validate_opinions(parsed["verdicts"], nids or None)
|
|
668
|
-
app = apply_opinions(cg, dict(pkg, _asm=asm), val["valid"],
|
|
669
|
-
reviewer=reviewer, applier=applier,
|
|
670
|
-
dry_run=dry_run, summary=parsed.get("summary"),
|
|
671
|
-
invalid=val["invalid"])
|
|
672
|
-
ok_n += 1
|
|
673
|
-
valid_n += len(val["valid"])
|
|
674
|
-
invalid_n += len(val["invalid"])
|
|
675
|
-
applied_n += 1 if app.get("committed") else 0
|
|
676
|
-
deferred_n += 1 if app.get("deferred") else 0
|
|
677
|
-
results.append({"ok": bool(app.get("ok")),
|
|
678
|
-
"bundle_id": pkg.get("bundle_id"), "job_id": jid,
|
|
679
|
-
"opinions": len(parsed["verdicts"]),
|
|
680
|
-
"valid": len(val["valid"]), "invalid": len(val["invalid"]),
|
|
681
|
-
"pass_rate": val["pass_rate"],
|
|
682
|
-
"deferred": bool(app.get("deferred")), "apply": app})
|
|
683
|
-
total_op = valid_n + invalid_n
|
|
684
|
-
return {"ok": all(r.get("ok") for r in results) and bool(results),
|
|
685
|
-
"jobs": len(jids), "submitted": len(posted),
|
|
686
|
-
"done": len(w.get("done") or []), "failed": w.get("failed") or [],
|
|
687
|
-
"packages": len(results), "opinions": total_op,
|
|
688
|
-
"valid": valid_n, "invalid": invalid_n,
|
|
689
|
-
"pass_rate": (1.0 if total_op == 0 else valid_n / float(total_op)),
|
|
690
|
-
"committed": applied_n, "deferred": deferred_n,
|
|
691
|
-
"results": results, "wait": w}
|
|
692
|
-
|
|
693
|
-
|
|
694
|
-
# ---- CLI -----------------------------------------------------------------
|
|
695
|
-
|
|
696
|
-
# 生效条件:当 argv 给出时,argparse 解析命令行;root 取 a.root 或 os.environ.get('MDCG_ROOT'),若 root 为假值则 SystemExit;否则用 verifier_principal(actor=a.applier) 构造 MdCGSecure,build_packages(root=root, limit=None) 得到 pairs,按 a.packages(假值 0 则取全部)切片,run_batch 以 dry_run=(a.dry_run or not a.apply) 运行,打印 JSON 报告,返回 0(rep['ok'] 为真)或 1。
|
|
697
|
-
def main(argv=None): # pragma: no cover
|
|
698
|
-
import argparse
|
|
699
|
-
ap = argparse.ArgumentParser(prog="python -m md_cg.mreview.pipeline",
|
|
700
|
-
description="记忆评审 级4 蜂巢并发 + 级5 落库")
|
|
701
|
-
ap.add_argument("--root", default=None, help="认知图根(缺省 MDCG_ROOT)")
|
|
702
|
-
ap.add_argument("--packages", type=int, default=10, help="试跑包数(0=全部)")
|
|
703
|
-
ap.add_argument("--jobs", default=None, help="hive jobs 目录(隔离生产)")
|
|
704
|
-
ap.add_argument("--ctx", default=None, help="context 文件目录")
|
|
705
|
-
ap.add_argument("--hive", default=None, help="hive 可执行文件")
|
|
706
|
-
ap.add_argument("--model", default=None, help="模型名")
|
|
707
|
-
ap.add_argument("--timeout", type=float, default=1800.0)
|
|
708
|
-
ap.add_argument("--reviewer", default="mreview-worker")
|
|
709
|
-
ap.add_argument("--applier", default="mreview-applier")
|
|
710
|
-
ap.add_argument("--dry-run", action="store_true", help="只校验不落库")
|
|
711
|
-
ap.add_argument("--apply", action="store_true", help="实际落库(缺省不写)")
|
|
712
|
-
a = ap.parse_args(argv)
|
|
713
|
-
from .. import mdcos
|
|
714
|
-
root = a.root or os.environ.get("MDCG_ROOT")
|
|
715
|
-
if not root:
|
|
716
|
-
raise SystemExit("缺少 --root 或 MDCG_ROOT(拒绝在未知根上运行)")
|
|
717
|
-
cg = mdcos.MdCGSecure(root, principal=verifier_principal(actor=a.applier))
|
|
718
|
-
built = build_packages(root=root, limit=None)
|
|
719
|
-
pairs = built["pairs"][:a.packages] if a.packages else built["pairs"]
|
|
720
|
-
rep = run_batch(pairs=pairs, cg=cg, exe=a.hive, jobs_dir=a.jobs,
|
|
721
|
-
ctx_dir=a.ctx, reviewer=a.reviewer, applier=a.applier,
|
|
722
|
-
model=a.model, timeout_s=a.timeout,
|
|
723
|
-
dry_run=(a.dry_run or not a.apply))
|
|
724
|
-
print(json.dumps(rep, ensure_ascii=False, indent=2))
|
|
725
|
-
return 0 if rep.get("ok") else 1
|
|
726
|
-
|
|
727
|
-
|
|
728
|
-
if __name__ == "__main__": # pragma: no cover
|
|
1
|
+
# -*- coding: utf-8 -*-
|
|
2
|
+
"""记忆评审流水线 · 级 4 蜂巢并发对接(spec 构造/投递/收卷)+ 级 5 意见落库治理。
|
|
3
|
+
|
|
4
|
+
真源:docs/记忆评审系统_立项设计与施工交接_20260915.md §5(spec/result 契约)
|
|
5
|
+
§7(反面清单:评审不直接改知识层 / 意见必过 verify 令牌 / 不确定 DEFER)
|
|
6
|
+
|
|
7
|
+
五级分工:级 1-3(M1)=候选→捆包→规则装配,产出 (pkg, asm) 对;
|
|
8
|
+
级 4(本模块)=每对 → hive spec → submit → poll 收卷(真并发、文件协议留痕);
|
|
9
|
+
级 5(本模块)=意见格式校验 → verify 令牌过 writepipe 六道闸落 contextual。
|
|
10
|
+
|
|
11
|
+
纪律是结构性的,不靠自觉:
|
|
12
|
+
· 评审不改知识层:落库层固定 contextual,verify 令牌 layers_allow 不含 knowledge
|
|
13
|
+
——库层 require_layer_write 拦截,任何拦截器都绕不过(§7 反面清单第 1 条)。
|
|
14
|
+
· 自验禁止:产出意见者(reflect)与落库裁决者(verify)不得同一 actor,判据
|
|
15
|
+
复用 crosscheck.detect_self_verify(不另造判据,保持真源单一)。
|
|
16
|
+
· 不确定即 DEFER:解析失败/格式越界一律不落库(Precision over noise)。
|
|
17
|
+
· 零写入面:worker 只读 spec/context、只写自己的 result.json;认知图写入只发生在
|
|
18
|
+
收卷后的 apply_opinions(本模块,走 writepipe 通道)。
|
|
19
|
+
|
|
20
|
+
运行:python -m md_cg.mreview.pipeline --help
|
|
21
|
+
"""
|
|
22
|
+
from __future__ import annotations
|
|
23
|
+
|
|
24
|
+
import json
|
|
25
|
+
import os
|
|
26
|
+
import re
|
|
27
|
+
import subprocess
|
|
28
|
+
import tempfile
|
|
29
|
+
import time
|
|
30
|
+
import uuid
|
|
31
|
+
|
|
32
|
+
from .. import crosscheck as CC
|
|
33
|
+
from .. import tokens as TK
|
|
34
|
+
from .. import writepipe as WP
|
|
35
|
+
from ..security import AccessDenied, Principal
|
|
36
|
+
from . import candidates as CD
|
|
37
|
+
from . import bundle as BD
|
|
38
|
+
from . import ruleset as RS
|
|
39
|
+
|
|
40
|
+
__all__ = [
|
|
41
|
+
"VERDICTS", "OPINION_LAYER", "DEFAULT_MODEL", "TERMINAL_STATES",
|
|
42
|
+
"hive_exe", "build_packages", "build_prompt", "build_spec", "dump_context",
|
|
43
|
+
"submit", "submit_many", "poll", "wait_jobs", "read_result", "collect",
|
|
44
|
+
"parse_opinions", "validate_opinion", "validate_opinions",
|
|
45
|
+
"verifier_principal", "opinion_doc", "apply_opinions", "run_package",
|
|
46
|
+
"run_batch", "main",
|
|
47
|
+
]
|
|
48
|
+
|
|
49
|
+
# ---- 常量 ----------------------------------------------------------------
|
|
50
|
+
|
|
51
|
+
VERDICTS = ("ACCEPT", "DEFER", "REJECT", "BLINDSPOT") # 意见四态(资格裁决同源)
|
|
52
|
+
OPINION_LAYER = "contextual" # verify 令牌 layers_allow=rejected/contextual
|
|
53
|
+
OPINION_BASIS = "measurement" # 意见的验证基底:确定性规则命中 + 复核
|
|
54
|
+
TERMINAL_STATES = ("done", "error", "timeout", "killed")
|
|
55
|
+
DEFAULT_MODEL = "glm-4-flash"
|
|
56
|
+
_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
|
|
57
|
+
|
|
58
|
+
SYSTEM_PROMPT = """你是记忆评审流水线中的【验证单元】。
|
|
59
|
+
|
|
60
|
+
职责边界(违反即无效):
|
|
61
|
+
1. 只对给定节点清单逐条裁决,不得引入清单外的节点。
|
|
62
|
+
2. 只输出意见,不改写任何记忆——你没有写入通道,也不需要。
|
|
63
|
+
3. 不确定就 DEFER(硬纪律):宁可漏判,不可错判。
|
|
64
|
+
4. 只有拿得出证据时才 REJECT;说不出证据就 DEFER。
|
|
65
|
+
5. ACCEPT 表示「已确认适用」,不是「没发现问题」——正条件无法确认时用 DEFER。
|
|
66
|
+
|
|
67
|
+
四态语义:
|
|
68
|
+
- ACCEPT 已确认适用
|
|
69
|
+
- DEFER 条件不足/无法确认(默认落点)
|
|
70
|
+
- REJECT 有据否定(冲突/来源许可不足/重复且劣于既有)
|
|
71
|
+
- BLINDSPOT 当前观测位置看不见(不是"不知道",是"这个位置判断不了")
|
|
72
|
+
|
|
73
|
+
输出契约(严格 JSON,无额外文字、无 markdown 围栏):
|
|
74
|
+
{"verdicts": [{"node_id": "...", "verdict": "ACCEPT|DEFER|REJECT|BLINDSPOT",
|
|
75
|
+
"reason": "一句话理由", "evidence": "证据(REJECT 必填)"}],
|
|
76
|
+
"summary": "整包一句话结论"}
|
|
77
|
+
清单中每个节点都要有一条裁决,一条不多一条不少。"""
|
|
78
|
+
|
|
79
|
+
|
|
80
|
+
# ---- 级 4-A · spec 构造 ---------------------------------------------------
|
|
81
|
+
|
|
82
|
+
# 生效条件:exe 为真值时返回 str(exe),否则 HIVE_EXE 经 (os.environ.get("HIVE_EXE") or "").strip() 后非空则返回该值,否则按 os.name 返回 os.path.join(_ROOT, "hive", "target", "release", "hive.exe" 或 "hive");
|
|
83
|
+
def hive_exe(exe=None) -> str:
|
|
84
|
+
"""蜂巢可执行文件:显式参数 > HIVE_EXE > 仓内 release 构建。"""
|
|
85
|
+
if exe:
|
|
86
|
+
return str(exe)
|
|
87
|
+
env = (os.environ.get("HIVE_EXE") or "").strip()
|
|
88
|
+
if env:
|
|
89
|
+
return env
|
|
90
|
+
name = "hive.exe" if os.name == "nt" else "hive"
|
|
91
|
+
return os.path.join(_ROOT, "hive", "target", "release", name)
|
|
92
|
+
|
|
93
|
+
|
|
94
|
+
# 生效条件:name 为假值(None/""/0/False 等)时以 "pkg" 为输入,否则 str(name),结果经 re.sub(r"[^0-9A-Za-z_.\-]", "_", ...) 替换并截取前 120 字符;
|
|
95
|
+
def _safe(name) -> str:
|
|
96
|
+
return re.sub(r"[^0-9A-Za-z_.\-]", "_", str(name or "pkg"))[:120]
|
|
97
|
+
|
|
98
|
+
|
|
99
|
+
# 生效条件:path 与 obj 给定后,若 os.path.dirname(path) 非空则 os.makedirs 该目录,随后以 UTF-8、ensure_ascii=False、indent=2 将 obj 写入 path 并返回 path;
|
|
100
|
+
def _dump_json(path: str, obj) -> str:
|
|
101
|
+
d = os.path.dirname(path)
|
|
102
|
+
if d:
|
|
103
|
+
os.makedirs(d, exist_ok=True)
|
|
104
|
+
with open(path, "w", encoding="utf-8") as f:
|
|
105
|
+
json.dump(obj, f, ensure_ascii=False, indent=2)
|
|
106
|
+
return path
|
|
107
|
+
|
|
108
|
+
|
|
109
|
+
_ENTRY_KEYS = ("ref", "node_id", "proposal_id", "origin", "layer", "tags", "role",
|
|
110
|
+
"importance", "evidence_count", "verification_basis",
|
|
111
|
+
"lifecycle_state", "content_hash", "issue_kinds", "evidence",
|
|
112
|
+
"excerpt")
|
|
113
|
+
|
|
114
|
+
|
|
115
|
+
# 生效条件:pkg/asm/workdir 给定,base 取 _safe(pkg.get("bundle_id"))(bundle_id 假值→"pkg"),entries 取 pkg.get("entries") or [] 的前 int(entry_limit) 项且仅保留 _ENTRY_KEYS,写出 workdir/base.bundle.json 与 workdir/base.rules.json(asm 原样),返回这两个 join(workdir,...) 路径(workdir 相对时不保证绝对);
|
|
116
|
+
def dump_context(pkg: dict, asm: dict, workdir: str, *, entry_limit=64) -> list:
|
|
117
|
+
"""包 + 规则装配 → context 文件(**绝对路径**,工人执行期内有效)。
|
|
118
|
+
|
|
119
|
+
相对路径的解析基准是 spec.workdir,跨进程易错;绝对路径无歧义。
|
|
120
|
+
"""
|
|
121
|
+
os.makedirs(workdir, exist_ok=True)
|
|
122
|
+
base = _safe(pkg.get("bundle_id"))
|
|
123
|
+
bp = os.path.join(workdir, base + ".bundle.json")
|
|
124
|
+
rp = os.path.join(workdir, base + ".rules.json")
|
|
125
|
+
ents = [{k: e.get(k) for k in _ENTRY_KEYS}
|
|
126
|
+
for e in (pkg.get("entries") or [])[:int(entry_limit)]]
|
|
127
|
+
_dump_json(bp, {"bundle_id": pkg.get("bundle_id"),
|
|
128
|
+
"group_kind": pkg.get("group_kind"),
|
|
129
|
+
"group_key": pkg.get("group_key"),
|
|
130
|
+
"size": pkg.get("size"), "entries": ents})
|
|
131
|
+
_dump_json(rp, asm)
|
|
132
|
+
return [bp, rp]
|
|
133
|
+
|
|
134
|
+
|
|
135
|
+
# 生效条件:pkg/asm 给定,节点清单标题恒输出、节点项取 pkg.get("entries") or [];机械命中取 asm.get("mechanical") or [],非空列前 int(max_mech) 条否则输出「无」,llm 取 asm.get("llm") or [] 非空列前 40 条;返回 join(L);
|
|
136
|
+
def build_prompt(pkg: dict, asm: dict, *, max_mech=120) -> str:
|
|
137
|
+
"""user_prompt:节点清单 + 机械命中 + 待检问题 + 输出要求。"""
|
|
138
|
+
ents = pkg.get("entries") or []
|
|
139
|
+
L = ["【待评包】bundle_id=%s group=%s/%s 节点数=%d"
|
|
140
|
+
% (pkg.get("bundle_id"), pkg.get("group_kind"),
|
|
141
|
+
pkg.get("group_key"), len(ents)), ""]
|
|
142
|
+
L.append("【节点清单】(仅可裁决以下节点)")
|
|
143
|
+
for i, e in enumerate(ents, 1):
|
|
144
|
+
frag = (e.get("excerpt") or "").strip().replace("\n", " ")[:220]
|
|
145
|
+
L.append("%d) node_id=%s layer=%s basis=%s tags=%s"
|
|
146
|
+
% (i, e.get("node_id"), e.get("layer"),
|
|
147
|
+
e.get("verification_basis"),
|
|
148
|
+
",".join(e.get("tags") or []) or "-"))
|
|
149
|
+
L.append(" 正文摘录:%s" % (frag or "(无正文/正文缺失)"))
|
|
150
|
+
if e.get("issue_kinds"):
|
|
151
|
+
L.append(" 机械初筛命中:%s" % ",".join(e["issue_kinds"]))
|
|
152
|
+
L.append("")
|
|
153
|
+
mech = asm.get("mechanical") or []
|
|
154
|
+
if mech:
|
|
155
|
+
L.append("【确定性规则命中】共 %d 条(机械层已给出,只需判断其语义后果)"
|
|
156
|
+
% len(mech))
|
|
157
|
+
for m in mech[:int(max_mech)]:
|
|
158
|
+
L.append("- [%s] %s :: %s"
|
|
159
|
+
% (m.get("issue_kind"), m.get("ref") or m.get("node_id"),
|
|
160
|
+
str(m.get("detail") or m.get("message") or "")[:160]))
|
|
161
|
+
else:
|
|
162
|
+
L.append("【确定性规则命中】无(本包未触发任何机械规则)")
|
|
163
|
+
L.append("")
|
|
164
|
+
llm = asm.get("llm") or []
|
|
165
|
+
if llm:
|
|
166
|
+
L.append("【待你判断的问题】逐项判断,并在 reason 中体现结论:")
|
|
167
|
+
for q in llm[:40]:
|
|
168
|
+
L.append("- (规则 %s / 检查 %s) %s"
|
|
169
|
+
% (q.get("rule_id"), q.get("check"), q.get("question")))
|
|
170
|
+
L.append("")
|
|
171
|
+
L.append("【输出】严格按 system 中的 JSON 契约,对每个 node_id 给出一条裁决。")
|
|
172
|
+
return "\n".join(L)
|
|
173
|
+
|
|
174
|
+
|
|
175
|
+
# 生效条件:pkg/asm 给定,model 按 model or os.environ.get('HIVE_MODEL') or DEFAULT_MODEL 取值(空串/假值均回落),context_files 按 list(context_files or []),timeout_s 转 int、temperature 转 float、max_tokens 转 int,node_ids 取 pkg.entries 中 node_id 真值的 str,返回含 meta 的 spec 字典;
|
|
176
|
+
def build_spec(pkg: dict, asm: dict, *, context_files=None, model=None,
|
|
177
|
+
timeout_s=600, temperature=0.0, max_tokens=4096) -> dict:
|
|
178
|
+
"""(pkg, asm) → hive spec(文件协议的可投递单元)。"""
|
|
179
|
+
nids = [str(e.get("node_id")) for e in (pkg.get("entries") or [])
|
|
180
|
+
if e.get("node_id")]
|
|
181
|
+
return {"model": model or os.environ.get("HIVE_MODEL") or DEFAULT_MODEL,
|
|
182
|
+
"system_prompt": SYSTEM_PROMPT,
|
|
183
|
+
"user_prompt": build_prompt(pkg, asm),
|
|
184
|
+
"context_files": list(context_files or []),
|
|
185
|
+
"timeout_s": int(timeout_s),
|
|
186
|
+
"temperature": float(temperature),
|
|
187
|
+
"max_tokens": int(max_tokens),
|
|
188
|
+
# 归因元信息(hive 忽略;收卷侧用于配对与审计)
|
|
189
|
+
"meta": {"bundle_id": pkg.get("bundle_id"),
|
|
190
|
+
"group_kind": pkg.get("group_kind"),
|
|
191
|
+
"group_key": pkg.get("group_key"),
|
|
192
|
+
"size": pkg.get("size"), "node_ids": nids}}
|
|
193
|
+
|
|
194
|
+
|
|
195
|
+
# 生效条件:当模块级常量 CD.SOURCES 可用时,sources 为假值则回落 CD.SOURCES,调用 CD.generate 生成候选,再经 BD.bundle(c.get('candidates') or [], nodes, root, max_per_bundle) 与 RS.assemble_all(b, rules_dir) 组装,最终将 bundles 与 packages 按 zip 配对返回。
|
|
196
|
+
def build_packages(*, root=None, sources=None, limit=None, nodes=None,
|
|
197
|
+
max_per_bundle=50, rules_dir=None, now=None) -> dict:
|
|
198
|
+
"""级 1-3 串联:候选 → 捆包 → 规则装配,产出配对的 (pkg, asm) 列表。"""
|
|
199
|
+
c = CD.generate(root, sources=sources or CD.SOURCES, nodes=nodes, limit=limit,
|
|
200
|
+
with_report=False, now=now)
|
|
201
|
+
b = BD.bundle(c.get("candidates") or [], nodes=nodes, root=root,
|
|
202
|
+
max_per_bundle=max_per_bundle)
|
|
203
|
+
a = RS.assemble_all(b, rules_dir=rules_dir)
|
|
204
|
+
pairs = []
|
|
205
|
+
for pkg, asm in zip(b.get("bundles") or [], a.get("packages") or []):
|
|
206
|
+
pairs.append((pkg, asm))
|
|
207
|
+
return {"candidates": c, "bundles": b, "assembled": a, "pairs": pairs}
|
|
208
|
+
|
|
209
|
+
|
|
210
|
+
# ---- 级 4-B · 投递 / 收卷 -------------------------------------------------
|
|
211
|
+
|
|
212
|
+
# 生效条件:argv 给定,环境变量副本强制 PYTHONUTF8="1",env 真值时用其 str 值更新,随后 subprocess.run(argv, capture_output=True, text=True, encoding="utf-8", errors="replace", env=e, timeout=timeout, cwd=cwd, input=stdin_text, shell=False);
|
|
213
|
+
def _run(argv, *, env=None, timeout=120, cwd=None, stdin_text=None):
|
|
214
|
+
"""统一子进程入口:argv 列表 + 显式 UTF-8 + PYTHONUTF8=1,不经 shell(第15条)。"""
|
|
215
|
+
e = dict(os.environ)
|
|
216
|
+
e["PYTHONUTF8"] = "1"
|
|
217
|
+
if env:
|
|
218
|
+
e.update({k: str(v) for k, v in env.items()})
|
|
219
|
+
return subprocess.run(argv, capture_output=True, text=True, encoding="utf-8",
|
|
220
|
+
errors="replace", env=e, timeout=timeout, cwd=cwd,
|
|
221
|
+
input=stdin_text, shell=False)
|
|
222
|
+
|
|
223
|
+
|
|
224
|
+
# 生效条件:text 为假值时 (text or "").strip().splitlines() 为空、循环不执行并返回 None;否则从末尾行向前找以 "{" 开头且 json.loads 成功且 isinstance(obj, dict) 的行返回该 dict,未找到返回 None;
|
|
225
|
+
def _last_json(text: str):
|
|
226
|
+
"""stdout 逐行解析取最后一个 JSON 对象(容忍前导日志行)。"""
|
|
227
|
+
for line in reversed((text or "").strip().splitlines()):
|
|
228
|
+
s = line.strip()
|
|
229
|
+
if not s.startswith("{"):
|
|
230
|
+
continue
|
|
231
|
+
try:
|
|
232
|
+
obj = json.loads(s)
|
|
233
|
+
except ValueError:
|
|
234
|
+
continue
|
|
235
|
+
if isinstance(obj, dict):
|
|
236
|
+
return obj
|
|
237
|
+
return None
|
|
238
|
+
|
|
239
|
+
|
|
240
|
+
# 生效条件:exe/jobs_dir/spec 给定,创建 jobs_dir 并在临时目录写 spec.json 后执行 [exe,"submit","--spec",sp,"--jobs",jobs_dir];returncode 非 0 返回 ok False + error,否则 _last_json(p.stdout) 为假(含 None/空 dict)返回 ok False「submit 输出非 JSON」,否则 d.setdefault("ok", True) 后返回 d;
|
|
241
|
+
def submit(exe: str, jobs_dir: str, spec: dict, *, timeout=120) -> dict:
|
|
242
|
+
"""投递单包 spec → {"ok":True,"job_id":...};失败返回 ok=False + error。"""
|
|
243
|
+
os.makedirs(jobs_dir, exist_ok=True)
|
|
244
|
+
with tempfile.TemporaryDirectory(prefix="mrev_spec_") as td:
|
|
245
|
+
sp = _dump_json(os.path.join(td, "spec.json"), spec)
|
|
246
|
+
p = _run([exe, "submit", "--spec", sp, "--jobs", jobs_dir], timeout=timeout)
|
|
247
|
+
if p.returncode != 0:
|
|
248
|
+
return {"ok": False, "returncode": p.returncode,
|
|
249
|
+
"error": (p.stderr or p.stdout or "").strip()[:800]}
|
|
250
|
+
d = _last_json(p.stdout)
|
|
251
|
+
if not d:
|
|
252
|
+
return {"ok": False, "error": "submit 输出非 JSON",
|
|
253
|
+
"stdout": (p.stdout or "")[:800]}
|
|
254
|
+
d.setdefault("ok", True)
|
|
255
|
+
return d
|
|
256
|
+
|
|
257
|
+
|
|
258
|
+
# 生效条件:specs 为可迭代列表时逐项调用 submit(exe, jobs_dir, s, timeout=timeout) 并返回等长列表;specs 为空则返回空列表;
|
|
259
|
+
def submit_many(exe: str, jobs_dir: str, specs: list, *, timeout=120) -> list:
|
|
260
|
+
"""批量投递(hive 侧并发消费;本函数逐条投递并保序配对)。"""
|
|
261
|
+
return [submit(exe, jobs_dir, s, timeout=timeout) for s in specs]
|
|
262
|
+
|
|
263
|
+
|
|
264
|
+
# 生效条件:exe/jobs_dir 给定,job_id 为真值时 argv 追加 str(job_id),假值(None/"")时仅列全部;运行后 returncode 非 0 返回 ok False + error,否则 _last_json(p.stdout) 为真值(非空 dict)时返回该 dict,为假值时返回 ok False「poll 输出非 JSON」;
|
|
265
|
+
def poll(exe: str, jobs_dir: str, job_id=None, *, timeout=120) -> dict:
|
|
266
|
+
"""查询 job 状态(不传 job_id = 列出全部)。"""
|
|
267
|
+
argv = [exe, "poll"]
|
|
268
|
+
if job_id:
|
|
269
|
+
argv.append(str(job_id))
|
|
270
|
+
argv += ["--jobs", jobs_dir]
|
|
271
|
+
p = _run(argv, timeout=timeout)
|
|
272
|
+
if p.returncode != 0:
|
|
273
|
+
return {"ok": False, "returncode": p.returncode,
|
|
274
|
+
"error": (p.stderr or p.stdout or "").strip()[:800]}
|
|
275
|
+
d = _last_json(p.stdout)
|
|
276
|
+
return d or {"ok": False, "error": "poll 输出非 JSON",
|
|
277
|
+
"stdout": (p.stdout or "")[:800]}
|
|
278
|
+
|
|
279
|
+
|
|
280
|
+
# 生效条件:j.get("job_id") 为真值时返回该值,否则返回 j.get("id")(若 id 为 0/""/False 则原样返回该假值,若 id 缺失返回 None);
|
|
281
|
+
def _job_id(j: dict):
|
|
282
|
+
return j.get("job_id") or j.get("id")
|
|
283
|
+
|
|
284
|
+
|
|
285
|
+
# 生效条件:j.get("state") 为真值时返回 str(state);否则 j.get("status") 为 dict 时返回 str(status.get("state") or "");否则返回 str(status or "");
|
|
286
|
+
def _job_state(j: dict) -> str:
|
|
287
|
+
"""state 读取兼容两种形态(扁平 / status 嵌套);仅供轮询判终态。"""
|
|
288
|
+
s = j.get("state")
|
|
289
|
+
if s:
|
|
290
|
+
return str(s)
|
|
291
|
+
st = j.get("status")
|
|
292
|
+
if isinstance(st, dict):
|
|
293
|
+
return str(st.get("state") or "")
|
|
294
|
+
return str(st or "")
|
|
295
|
+
|
|
296
|
+
|
|
297
|
+
# 生效条件:exe/jobs_dir 给定,job_ids 真值集合限观测;循环 poll 直到 pending 为空且(want 为 None 或 want <= 已观测)返回 ok=not failed,若 pending 非空或 want 未全观测且到 max(1.0,float(timeout_s)) 截止则返回 ok False + error,每次间隔 sleep(max(0.05,float(interval_s)));
|
|
298
|
+
def wait_jobs(exe: str, jobs_dir: str, job_ids=None, *, timeout_s=900,
|
|
299
|
+
interval_s=1.0, on_tick=None) -> dict:
|
|
300
|
+
"""轮询至全部终态(或超时)→ done/failed/pending 三分类 + 原始快照。"""
|
|
301
|
+
want = set(str(x) for x in (job_ids or [])) or None
|
|
302
|
+
deadline = time.time() + max(1.0, float(timeout_s))
|
|
303
|
+
last = {}
|
|
304
|
+
while True:
|
|
305
|
+
r = poll(exe, jobs_dir)
|
|
306
|
+
states = {}
|
|
307
|
+
for j in (r.get("jobs") or []):
|
|
308
|
+
jid = _job_id(j)
|
|
309
|
+
if jid is None:
|
|
310
|
+
continue
|
|
311
|
+
jid = str(jid)
|
|
312
|
+
if want is not None and jid not in want:
|
|
313
|
+
continue
|
|
314
|
+
states[jid] = _job_state(j) or "pending"
|
|
315
|
+
last[jid] = j
|
|
316
|
+
done = sorted(k for k, v in states.items() if v == "done")
|
|
317
|
+
failed = sorted(k for k, v in states.items()
|
|
318
|
+
if v in TERMINAL_STATES and v != "done")
|
|
319
|
+
pending = sorted(k for k, v in states.items() if v not in TERMINAL_STATES)
|
|
320
|
+
if on_tick:
|
|
321
|
+
on_tick({"states": dict(states), "done": len(done),
|
|
322
|
+
"pending": len(pending), "failed": len(failed)})
|
|
323
|
+
if not pending:
|
|
324
|
+
if want is None or want <= set(states):
|
|
325
|
+
return {"ok": not failed, "done": done, "failed": failed,
|
|
326
|
+
"pending": [], "states": states, "jobs": last}
|
|
327
|
+
if time.time() >= deadline:
|
|
328
|
+
miss = sorted(want - set(states))
|
|
329
|
+
return {"ok": False, "done": done, "failed": failed,
|
|
330
|
+
"pending": miss, "states": states, "jobs": last,
|
|
331
|
+
"error": "未在超时内观测到 job:%s" % ",".join(miss)}
|
|
332
|
+
if time.time() >= deadline:
|
|
333
|
+
return {"ok": False, "done": done, "failed": failed,
|
|
334
|
+
"pending": pending, "states": states, "jobs": last,
|
|
335
|
+
"error": "等待终态超时(%ss)" % timeout_s}
|
|
336
|
+
time.sleep(max(0.05, float(interval_s)))
|
|
337
|
+
|
|
338
|
+
|
|
339
|
+
# 生效条件:jobs_dir/job_id 给定,join(jobs_dir,str(job_id),"result.json") 路径缺失时返回 ok False「result.json 不存在」,读取时 OSError/ValueError 返回 ok False「读取失败」,json 对象非 dict 返回 ok False「result 非对象」,否则返回该 dict;
|
|
340
|
+
def read_result(jobs_dir: str, job_id: str) -> dict:
|
|
341
|
+
"""读工人产出(job 目录的 result.json)。"""
|
|
342
|
+
p = os.path.join(jobs_dir, str(job_id), "result.json")
|
|
343
|
+
if not os.path.isfile(p):
|
|
344
|
+
return {"ok": False, "error": "result.json 不存在:%s" % p}
|
|
345
|
+
try:
|
|
346
|
+
with open(p, encoding="utf-8") as f:
|
|
347
|
+
obj = json.load(f)
|
|
348
|
+
except (OSError, ValueError) as exc:
|
|
349
|
+
return {"ok": False, "error": "result.json 读取失败:%s" % exc}
|
|
350
|
+
return obj if isinstance(obj, dict) else {"ok": False, "error": "result 非对象"}
|
|
351
|
+
|
|
352
|
+
|
|
353
|
+
# 生效条件:jobs_dir/job_ids 给定,job_ids 为假值(None/[])时返回空 dict,否则返回 {str(j): read_result(jobs_dir, j) for j in job_ids};
|
|
354
|
+
def collect(jobs_dir: str, job_ids) -> dict:
|
|
355
|
+
"""按 job_id 收卷(读 result.json),返回 {job_id: result} 保序映射。"""
|
|
356
|
+
return {str(j): read_result(jobs_dir, j) for j in (job_ids or [])}
|
|
357
|
+
|
|
358
|
+
|
|
359
|
+
# ---- 级 5-A · 意见解析与格式校验(校验率 100% 是验收口径)-----------------
|
|
360
|
+
|
|
361
|
+
# 生效条件:text 去空格后为空返回 None;否则先 try json.loads(整体),成功返回;否则找 ```(?:json)?...``` 围栏内 json.loads 成功返回;否则取 s.find("{") 与 s.rfind("}") 且 0<=i<k 的切片 json.loads 成功返回;均失败返回 None;
|
|
362
|
+
def _extract_json(text: str):
|
|
363
|
+
"""从模型文本抽 JSON 对象:整体 → ```围栏``` → 首尾大括号切片。"""
|
|
364
|
+
s = (text or "").strip()
|
|
365
|
+
if not s:
|
|
366
|
+
return None
|
|
367
|
+
try:
|
|
368
|
+
return json.loads(s)
|
|
369
|
+
except ValueError:
|
|
370
|
+
pass
|
|
371
|
+
m = re.search(r"```(?:json)?\s*(.+?)```", s, re.S)
|
|
372
|
+
if m:
|
|
373
|
+
try:
|
|
374
|
+
return json.loads(m.group(1).strip())
|
|
375
|
+
except ValueError:
|
|
376
|
+
pass
|
|
377
|
+
i, k = s.find("{"), s.rfind("}")
|
|
378
|
+
if 0 <= i < k:
|
|
379
|
+
try:
|
|
380
|
+
return json.loads(s[i:k + 1])
|
|
381
|
+
except ValueError:
|
|
382
|
+
pass
|
|
383
|
+
return None
|
|
384
|
+
|
|
385
|
+
|
|
386
|
+
# 生效条件:result 非 dict 返回 ok False;result.get("error") 真值返回 ok False「工人报错」;v=result.get("verdicts"),v is None 时从 result.get("content") or "" 经 _extract_json 取 dict,非 dict 返回 ok False,否则 v=obj.get("verdicts") 且 summary=obj.get("summary") or summary;v 非 list 返回 ok False;否则返回 ok True、verdicts=v、summary=str(summary);
|
|
387
|
+
def parse_opinions(result: dict) -> dict:
|
|
388
|
+
"""result.json → {"ok":True,"verdicts":[...],"summary":...}。
|
|
389
|
+
|
|
390
|
+
两形态都能收:
|
|
391
|
+
· 执行器直写 {"ok":true,"verdicts":[...],"summary":...}
|
|
392
|
+
· LLM 文本 {"ok":true,"content":"{...JSON...}"}(exec.py 默认形态)
|
|
393
|
+
"""
|
|
394
|
+
if not isinstance(result, dict):
|
|
395
|
+
return {"ok": False, "error": "result 非对象"}
|
|
396
|
+
if result.get("error"):
|
|
397
|
+
return {"ok": False, "error": "工人报错:%s" % result["error"]}
|
|
398
|
+
v = result.get("verdicts")
|
|
399
|
+
summary = result.get("summary") or ""
|
|
400
|
+
if v is None:
|
|
401
|
+
obj = _extract_json(result.get("content") or "")
|
|
402
|
+
if not isinstance(obj, dict):
|
|
403
|
+
return {"ok": False, "error": "未能从 content 中解析 JSON 意见块"}
|
|
404
|
+
v = obj.get("verdicts")
|
|
405
|
+
summary = obj.get("summary") or summary
|
|
406
|
+
if not isinstance(v, list):
|
|
407
|
+
return {"ok": False, "error": "verdicts 非数组(got %s)" % type(v).__name__}
|
|
408
|
+
return {"ok": True, "verdicts": v, "summary": str(summary)}
|
|
409
|
+
|
|
410
|
+
|
|
411
|
+
# 生效条件:v 经 str(v or "").strip().upper() 得 s,若 s 在模块级 VERDICTS 中则返回 s,否则返回 None;
|
|
412
|
+
def _norm_verdict(v):
|
|
413
|
+
s = str(v or "").strip().upper()
|
|
414
|
+
return s if s in VERDICTS else None
|
|
415
|
+
|
|
416
|
+
|
|
417
|
+
# 生效条件:op 非 dict 返回「意见非对象」;nid 为空返回「缺 node_id」;node_ids 非 None 且 nid 不在其中返回越界错误;_norm_verdict(op.get("verdict")) 为 None 返回「verdict 越界」;reason 去空格空返回「缺 reason」;vd=="REJECT" 且 evidence 去空格空返回「REJECT 缺 evidence」;否则返回 None;
|
|
418
|
+
def validate_opinion(op, node_ids=None):
|
|
419
|
+
"""单条意见格式校验:合法返回 None,非法返回错误串。"""
|
|
420
|
+
if not isinstance(op, dict):
|
|
421
|
+
return "意见非对象"
|
|
422
|
+
nid = str(op.get("node_id") or "").strip()
|
|
423
|
+
if not nid:
|
|
424
|
+
return "缺 node_id"
|
|
425
|
+
if node_ids is not None and nid not in node_ids:
|
|
426
|
+
return "node_id 不属于本包:%s" % nid
|
|
427
|
+
vd = _norm_verdict(op.get("verdict"))
|
|
428
|
+
if vd is None:
|
|
429
|
+
return "verdict 越界:%r" % (op.get("verdict"),)
|
|
430
|
+
if not str(op.get("reason") or "").strip():
|
|
431
|
+
return "缺 reason"
|
|
432
|
+
if vd == "REJECT" and not str(op.get("evidence") or "").strip():
|
|
433
|
+
return "REJECT 缺 evidence(有据才拒)"
|
|
434
|
+
return None
|
|
435
|
+
|
|
436
|
+
|
|
437
|
+
# 生效条件:verdicts 为假值(None/[])时 total=0、rate=1.0、ok=False;否则逐条 validate_opinion,无效记入 invalid,有效项复制并写入 _norm_verdict(op.get("verdict")),最后返回 total、valid、invalid、rate=len(valid)/total、ok=bool(valid) and not invalid;
|
|
438
|
+
def validate_opinions(verdicts, node_ids=None) -> dict:
|
|
439
|
+
"""批量校验:valid/invalid 明细 + 通过率(用于 100% 口径断言)。"""
|
|
440
|
+
total = len(verdicts or [])
|
|
441
|
+
valid, invalid = [], []
|
|
442
|
+
for i, op in enumerate(verdicts or []):
|
|
443
|
+
err = validate_opinion(op, node_ids)
|
|
444
|
+
if err:
|
|
445
|
+
invalid.append({"index": i, "error": err,
|
|
446
|
+
"node_id": (op.get("node_id")
|
|
447
|
+
if isinstance(op, dict) else None)})
|
|
448
|
+
continue
|
|
449
|
+
op = dict(op)
|
|
450
|
+
op["verdict"] = _norm_verdict(op.get("verdict"))
|
|
451
|
+
valid.append(op)
|
|
452
|
+
rate = 1.0 if total == 0 else len(valid) / float(total)
|
|
453
|
+
return {"total": total, "valid": valid, "invalid": invalid,
|
|
454
|
+
"pass_rate": rate, "ok": bool(valid) and not invalid}
|
|
455
|
+
|
|
456
|
+
|
|
457
|
+
# ---- 级 5-B · 落库(verify 令牌 + writepipe 六道闸;不自造通道)-----------
|
|
458
|
+
|
|
459
|
+
# 生效条件:当 actor 给出时,从 TK.role_spec('verify') 读取 clearance_cap/can_write/can_admin/layers_allow 构造 Principal;session 为假值时生成 'mrev_' + uuid 前 12 位;layers_allow 为假值时回落 spec.get('layers_allow') 或 [];ops_allow 原样传入。
|
|
460
|
+
def verifier_principal(*, actor="mreview-verifier", session=None,
|
|
461
|
+
layers_allow=None, ops_allow=None) -> Principal:
|
|
462
|
+
"""验证单元 Principal——角色规格取自 tokens 真源(避免手写漂移)。"""
|
|
463
|
+
spec = TK.role_spec("verify")
|
|
464
|
+
return Principal(actor=actor, clearance=spec.get("clearance_cap") or "internal",
|
|
465
|
+
can_write=bool(spec.get("can_write")),
|
|
466
|
+
can_admin=bool(spec.get("can_admin")),
|
|
467
|
+
role="verify", unit="verify",
|
|
468
|
+
session=session or ("mrev_" + uuid.uuid4().hex[:12]),
|
|
469
|
+
layers_allow=list(layers_allow or spec.get("layers_allow") or []),
|
|
470
|
+
ops_allow=ops_allow, auth_mode="mreview")
|
|
471
|
+
|
|
472
|
+
|
|
473
|
+
# 生效条件:pkg、opinions、reviewer 给定时恒返回 "\n".join(L);summary 为真值才追加「整包结论」行,invalid 为真值才追加「【未落库条目】」行,二者取默认 "" / None 时不追加;
|
|
474
|
+
def opinion_doc(pkg: dict, opinions: list, *, reviewer: str, summary="",
|
|
475
|
+
invalid=None) -> str:
|
|
476
|
+
"""评审意见节点正文(核心修改四要素:内容/原因/位置/验证)。"""
|
|
477
|
+
cnt = {k: 0 for k in VERDICTS}
|
|
478
|
+
for o in opinions:
|
|
479
|
+
cnt[_norm_verdict(o.get("verdict")) or "?"] = \
|
|
480
|
+
cnt.get(_norm_verdict(o.get("verdict")) or "?", 0) + 1
|
|
481
|
+
dist = " / ".join("%s %d" % (k, cnt.get(k, 0)) for k in VERDICTS)
|
|
482
|
+
L = ["# 记忆评审意见 · 包 %s(%s/%s,%d 节点)"
|
|
483
|
+
% (pkg.get("bundle_id"), pkg.get("group_kind"), pkg.get("group_key"),
|
|
484
|
+
len(pkg.get("entries") or [])),
|
|
485
|
+
"来源:级 4 蜂巢并发评审(评审者=%s)" % reviewer,
|
|
486
|
+
"裁决分布:%s" % dist]
|
|
487
|
+
if summary:
|
|
488
|
+
L.append("整包结论:%s" % summary)
|
|
489
|
+
L.append("")
|
|
490
|
+
for o in opinions:
|
|
491
|
+
seg = "- [%s] %s:%s" % (_norm_verdict(o.get("verdict")) or "?",
|
|
492
|
+
o.get("node_id"),
|
|
493
|
+
str(o.get("reason") or "").strip())
|
|
494
|
+
ev = str(o.get("evidence") or "").strip()
|
|
495
|
+
if ev:
|
|
496
|
+
seg += "(证据:%s)" % ev[:300]
|
|
497
|
+
L.append(seg)
|
|
498
|
+
if invalid:
|
|
499
|
+
L.append("")
|
|
500
|
+
L.append("【未落库条目】%s" % json.dumps(invalid, ensure_ascii=False))
|
|
501
|
+
L.append("")
|
|
502
|
+
L.append("【内容】级 4 并发评审对包 %s 的逐节点裁决意见。" % pkg.get("bundle_id"))
|
|
503
|
+
L.append("【原因】机械层已过滤确定性缺陷,语义判断留给验证单元;意见落 contextual "
|
|
504
|
+
"供治理层决策——评审不改知识层。")
|
|
505
|
+
L.append("【位置】源包 group=%s/%s(%d 节点)。"
|
|
506
|
+
% (pkg.get("group_kind"), pkg.get("group_key"),
|
|
507
|
+
len(pkg.get("entries") or [])))
|
|
508
|
+
L.append("【验证】机械规则命中 %d 条;意见格式校验 %d/%d 通过(越界条目已剔除)。"
|
|
509
|
+
% (len((pkg.get("_asm") or {}).get("mechanical") or []),
|
|
510
|
+
len(opinions), len(opinions) + len(invalid or [])))
|
|
511
|
+
return "\n".join(L)
|
|
512
|
+
|
|
513
|
+
|
|
514
|
+
# 生效条件:cg/pkg/opinions/reviewer/applier 给定,CC.detect_self_verify 判 reviewer 与 applier 同一→refused self_verify_disallowed;opinions 为空→ok True skipped no_opinions;dry_run 真→ok True committed False;否则 (pipe or WP.default_pipeline()).execute,AccessDenied→refused layer_denied,其他异常→write_error,成功→ok 取 out.ok、committed 取 out.committed、moved_to 取 out.moved_to、deferred=bool(out.moved_to) and not out.committed;
|
|
515
|
+
def apply_opinions(cg, pkg: dict, opinions: list, *, reviewer, applier,
|
|
516
|
+
dry_run=False, pipe=None, layer=OPINION_LAYER,
|
|
517
|
+
importance=0.4, tags=None, node_id=None,
|
|
518
|
+
summary="", invalid=None) -> dict:
|
|
519
|
+
"""落库单包意见(级 5 唯一的写入口)。
|
|
520
|
+
|
|
521
|
+
顺序(任一环不过即不落库):
|
|
522
|
+
1 自验违例检测(评审者 ≠ 落库裁决者,判据复用 crosscheck)
|
|
523
|
+
2 意见非空(全 DEFER 且空 → 不写,零噪声)
|
|
524
|
+
3 verify 令牌 + writepipe 六道闸 + 库层 require_layer_write
|
|
525
|
+
4 合规规则库(部署侧 MDCG_POLICY_FILE)——audit 的 text 验证器在**无规则**
|
|
526
|
+
时恒 DEFER(audit._rule_check「无规则不能假装合规」),意见将转审核队列
|
|
527
|
+
(`moved_to="review_queue"`)而非落盘。这是设计行为不是故障:本函数把
|
|
528
|
+
去向如实透出为 `moved_to` / `deferred`,**不把「入队」报成「落库」**。
|
|
529
|
+
"""
|
|
530
|
+
rows = [{"unit": CC.REFLECT_UNIT, "actor": str(reviewer or "")},
|
|
531
|
+
{"unit": CC.VERIFY_UNIT, "actor": str(applier or "")}]
|
|
532
|
+
if CC.detect_self_verify(rows):
|
|
533
|
+
return {"ok": False, "committed": False, "refused": True,
|
|
534
|
+
"reason": "self_verify_disallowed",
|
|
535
|
+
"detail": "评审者与落库裁决者为同一执行者(%s)——自验被拒(§7)"
|
|
536
|
+
% reviewer}
|
|
537
|
+
if not opinions:
|
|
538
|
+
return {"ok": True, "committed": False, "skipped": True,
|
|
539
|
+
"reason": "no_opinions(无有效意见,零噪声不落库)"}
|
|
540
|
+
nid = node_id or ("mr_opinion_" + _safe(pkg.get("bundle_id")))
|
|
541
|
+
doc = opinion_doc(pkg, opinions, reviewer=str(reviewer), summary=summary,
|
|
542
|
+
invalid=invalid)
|
|
543
|
+
if dry_run:
|
|
544
|
+
return {"ok": True, "committed": False, "dry_run": True, "node_id": nid,
|
|
545
|
+
"doc_len": len(doc)}
|
|
546
|
+
a = {"node_id": nid, "content": doc, "layer": layer,
|
|
547
|
+
"tags": list(tags or ["cap:记忆评审", "mreview", "opinion",
|
|
548
|
+
"bundle:" + _safe(pkg.get("bundle_id"))]),
|
|
549
|
+
"importance": float(importance), "verification_basis": OPINION_BASIS,
|
|
550
|
+
"content_kind": "text",
|
|
551
|
+
# 意见是评审产物(contextual),不参与知识层冲突判定:冲突闸若命中会
|
|
552
|
+
# 把意见当冲突源转入审核队列(语义不符),故显式跳过(闸层可关,
|
|
553
|
+
# 库层 require_layer_write 不可关——评审不改知识层的结构约束仍在)。
|
|
554
|
+
"consistency": False}
|
|
555
|
+
try:
|
|
556
|
+
out = (pipe or WP.default_pipeline()).execute(cg, a)
|
|
557
|
+
except AccessDenied as exc:
|
|
558
|
+
# 结构性拒绝(如 layer=knowledge 越出 verify 令牌的 layers_allow):
|
|
559
|
+
# 如实上报为拒写,不吞异常也不改写写入语义(§7 反面清单第 1 条)。
|
|
560
|
+
return {"ok": False, "committed": False, "refused": True,
|
|
561
|
+
"reason": "layer_denied", "node_id": nid, "layer": layer,
|
|
562
|
+
"detail": "%s: %s" % (type(exc).__name__, exc)}
|
|
563
|
+
except Exception as exc: # noqa: BLE001
|
|
564
|
+
return {"ok": False, "committed": False, "refused": False,
|
|
565
|
+
"reason": "write_error", "node_id": nid, "layer": layer,
|
|
566
|
+
"detail": "%s: %s" % (type(exc).__name__, exc)}
|
|
567
|
+
return {"ok": bool(out.get("ok")), "committed": bool(out.get("committed")),
|
|
568
|
+
"moved_to": out.get("moved_to"),
|
|
569
|
+
# 「改了去向但没落盘」(如合规/冲突闸转审核队列、gated 的 DROP/DEFER)
|
|
570
|
+
# 与「落盘」是两种结局:分开报,调用方不必读 response 才能分辨。
|
|
571
|
+
"deferred": bool(out.get("moved_to")) and not out.get("committed"),
|
|
572
|
+
"node_id": nid, "layer": layer, "response": out,
|
|
573
|
+
"doc_len": len(doc)}
|
|
574
|
+
|
|
575
|
+
|
|
576
|
+
# ---- 编排 ----------------------------------------------------------------
|
|
577
|
+
|
|
578
|
+
# 生效条件:pkg/asm/cg 给定,exe=hive_exe(exe),jobs_dir/ctx_dir 假值回落到 tempfile.gettempdir() 下路径;dump_context 后 build_spec、submit,若 submit 无 ok 或无 _job_id 返回 stage submit;wait_jobs 后若 failed 或 done 空返回 stage wait;read_result+parse_opinions 失败返回 stage parse;否则 validate_opinions 并 apply_opinions,返回 ok=app.ok 及计数;
|
|
579
|
+
def run_package(*, pkg, asm, cg, exe=None, jobs_dir=None, ctx_dir=None,
|
|
580
|
+
reviewer="mreview-worker", applier="mreview-applier",
|
|
581
|
+
model=None, timeout_s=900, interval_s=1.0, dry_run=False,
|
|
582
|
+
keep_jobs=True) -> dict:
|
|
583
|
+
"""单包全链路:dump context → submit → wait → parse → validate → apply。"""
|
|
584
|
+
exe = hive_exe(exe)
|
|
585
|
+
jobs_dir = jobs_dir or os.path.join(tempfile.gettempdir(), "mrev_jobs")
|
|
586
|
+
ctx_dir = ctx_dir or os.path.join(tempfile.gettempdir(), "mrev_ctx")
|
|
587
|
+
ctx_files = dump_context(pkg, asm, ctx_dir)
|
|
588
|
+
spec = build_spec(pkg, asm, context_files=ctx_files, model=model,
|
|
589
|
+
timeout_s=timeout_s)
|
|
590
|
+
sub = submit(exe, jobs_dir, spec)
|
|
591
|
+
if not sub.get("ok") or not _job_id(sub):
|
|
592
|
+
return {"ok": False, "stage": "submit", "error": sub.get("error"),
|
|
593
|
+
"submit": sub, "bundle_id": pkg.get("bundle_id")}
|
|
594
|
+
jid = str(_job_id(sub))
|
|
595
|
+
w = wait_jobs(exe, jobs_dir, [jid], timeout_s=timeout_s,
|
|
596
|
+
interval_s=interval_s)
|
|
597
|
+
if w.get("failed") or not w.get("done"):
|
|
598
|
+
return {"ok": False, "stage": "wait", "job_id": jid, "wait": w,
|
|
599
|
+
"bundle_id": pkg.get("bundle_id")}
|
|
600
|
+
res = read_result(jobs_dir, jid)
|
|
601
|
+
parsed = parse_opinions(res)
|
|
602
|
+
if not parsed.get("ok"):
|
|
603
|
+
return {"ok": False, "stage": "parse", "job_id": jid, "error":
|
|
604
|
+
parsed.get("error"), "bundle_id": pkg.get("bundle_id")}
|
|
605
|
+
nids = {str(e.get("node_id")) for e in (pkg.get("entries") or [])
|
|
606
|
+
if e.get("node_id")}
|
|
607
|
+
val = validate_opinions(parsed["verdicts"], nids or None)
|
|
608
|
+
pkg2 = dict(pkg, _asm=asm)
|
|
609
|
+
app = apply_opinions(cg, pkg2, val["valid"], reviewer=reviewer,
|
|
610
|
+
applier=applier, dry_run=dry_run,
|
|
611
|
+
summary=parsed.get("summary"), invalid=val["invalid"])
|
|
612
|
+
return {"ok": bool(app.get("ok")), "bundle_id": pkg.get("bundle_id"),
|
|
613
|
+
"job_id": jid, "opinions": len(parsed["verdicts"]),
|
|
614
|
+
"valid": len(val["valid"]), "invalid": len(val["invalid"]),
|
|
615
|
+
"pass_rate": val["pass_rate"], "deferred": bool(app.get("deferred")),
|
|
616
|
+
"apply": app, "stages": {"submit": True, "wait": True, "parse": True}}
|
|
617
|
+
|
|
618
|
+
|
|
619
|
+
# 生效条件:pairs/cg 给定,先对每对 pkg,asm 执行 dump_context/build_spec/submit,收集 ok 且有 job_id 的 jids;wait_jobs 后逐 jid 若 state!="done" 记 stage wait,否则 read_result/parse_opinions/validate_opinions/apply_opinions 并累计;返回 ok=all(results.ok) and bool(results) 及聚合计数;
|
|
620
|
+
def run_batch(*, pairs, cg, exe=None, jobs_dir=None, ctx_dir=None,
|
|
621
|
+
reviewer="mreview-worker", applier="mreview-applier", model=None,
|
|
622
|
+
timeout_s=1800, interval_s=1.0, dry_run=False, on_tick=None) -> dict:
|
|
623
|
+
"""批量并发:先全投递(hive worker 池并发消费),再统一等候,再逐包落库。
|
|
624
|
+
|
|
625
|
+
并发发生在**投递之后**——本函数投递不阻塞,故 N 包的墙钟≈最慢一包。
|
|
626
|
+
"""
|
|
627
|
+
exe = hive_exe(exe)
|
|
628
|
+
jobs_dir = jobs_dir or os.path.join(tempfile.gettempdir(), "mrev_jobs")
|
|
629
|
+
ctx_dir = ctx_dir or os.path.join(tempfile.gettempdir(), "mrev_ctx")
|
|
630
|
+
os.makedirs(jobs_dir, exist_ok=True)
|
|
631
|
+
staged, posted, jids = [], [], []
|
|
632
|
+
for pkg, asm in pairs:
|
|
633
|
+
ctx_files = dump_context(pkg, asm, ctx_dir)
|
|
634
|
+
spec = build_spec(pkg, asm, context_files=ctx_files, model=model,
|
|
635
|
+
timeout_s=timeout_s)
|
|
636
|
+
sub = submit(exe, jobs_dir, spec)
|
|
637
|
+
staged.append({"bundle_id": pkg.get("bundle_id"), "pkg": pkg,
|
|
638
|
+
"asm": asm, "submit": sub})
|
|
639
|
+
if sub.get("ok") and _job_id(sub):
|
|
640
|
+
jid = str(_job_id(sub))
|
|
641
|
+
jids.append(jid)
|
|
642
|
+
posted.append(staged[-1])
|
|
643
|
+
w = wait_jobs(exe, jobs_dir, jids, timeout_s=timeout_s,
|
|
644
|
+
interval_s=interval_s, on_tick=on_tick)
|
|
645
|
+
by_jid = {}
|
|
646
|
+
for i, s in enumerate(staged):
|
|
647
|
+
if s["submit"].get("ok") and _job_id(s["submit"]):
|
|
648
|
+
by_jid[str(_job_id(s["submit"]))] = s
|
|
649
|
+
results, ok_n, valid_n, invalid_n, applied_n, deferred_n = [], 0, 0, 0, 0, 0
|
|
650
|
+
for jid in jids:
|
|
651
|
+
s = by_jid.get(jid) or {}
|
|
652
|
+
pkg, asm = s.get("pkg") or {}, s.get("asm") or {}
|
|
653
|
+
state = (w.get("states") or {}).get(jid)
|
|
654
|
+
if state != "done":
|
|
655
|
+
results.append({"ok": False, "bundle_id": pkg.get("bundle_id"),
|
|
656
|
+
"job_id": jid, "stage": "wait", "state": state})
|
|
657
|
+
continue
|
|
658
|
+
res = read_result(jobs_dir, jid)
|
|
659
|
+
parsed = parse_opinions(res)
|
|
660
|
+
if not parsed.get("ok"):
|
|
661
|
+
results.append({"ok": False, "bundle_id": pkg.get("bundle_id"),
|
|
662
|
+
"job_id": jid, "stage": "parse",
|
|
663
|
+
"error": parsed.get("error")})
|
|
664
|
+
continue
|
|
665
|
+
nids = {str(e.get("node_id")) for e in (pkg.get("entries") or [])
|
|
666
|
+
if e.get("node_id")}
|
|
667
|
+
val = validate_opinions(parsed["verdicts"], nids or None)
|
|
668
|
+
app = apply_opinions(cg, dict(pkg, _asm=asm), val["valid"],
|
|
669
|
+
reviewer=reviewer, applier=applier,
|
|
670
|
+
dry_run=dry_run, summary=parsed.get("summary"),
|
|
671
|
+
invalid=val["invalid"])
|
|
672
|
+
ok_n += 1
|
|
673
|
+
valid_n += len(val["valid"])
|
|
674
|
+
invalid_n += len(val["invalid"])
|
|
675
|
+
applied_n += 1 if app.get("committed") else 0
|
|
676
|
+
deferred_n += 1 if app.get("deferred") else 0
|
|
677
|
+
results.append({"ok": bool(app.get("ok")),
|
|
678
|
+
"bundle_id": pkg.get("bundle_id"), "job_id": jid,
|
|
679
|
+
"opinions": len(parsed["verdicts"]),
|
|
680
|
+
"valid": len(val["valid"]), "invalid": len(val["invalid"]),
|
|
681
|
+
"pass_rate": val["pass_rate"],
|
|
682
|
+
"deferred": bool(app.get("deferred")), "apply": app})
|
|
683
|
+
total_op = valid_n + invalid_n
|
|
684
|
+
return {"ok": all(r.get("ok") for r in results) and bool(results),
|
|
685
|
+
"jobs": len(jids), "submitted": len(posted),
|
|
686
|
+
"done": len(w.get("done") or []), "failed": w.get("failed") or [],
|
|
687
|
+
"packages": len(results), "opinions": total_op,
|
|
688
|
+
"valid": valid_n, "invalid": invalid_n,
|
|
689
|
+
"pass_rate": (1.0 if total_op == 0 else valid_n / float(total_op)),
|
|
690
|
+
"committed": applied_n, "deferred": deferred_n,
|
|
691
|
+
"results": results, "wait": w}
|
|
692
|
+
|
|
693
|
+
|
|
694
|
+
# ---- CLI -----------------------------------------------------------------
|
|
695
|
+
|
|
696
|
+
# 生效条件:当 argv 给出时,argparse 解析命令行;root 取 a.root 或 os.environ.get('MDCG_ROOT'),若 root 为假值则 SystemExit;否则用 verifier_principal(actor=a.applier) 构造 MdCGSecure,build_packages(root=root, limit=None) 得到 pairs,按 a.packages(假值 0 则取全部)切片,run_batch 以 dry_run=(a.dry_run or not a.apply) 运行,打印 JSON 报告,返回 0(rep['ok'] 为真)或 1。
|
|
697
|
+
def main(argv=None): # pragma: no cover
|
|
698
|
+
import argparse
|
|
699
|
+
ap = argparse.ArgumentParser(prog="python -m md_cg.mreview.pipeline",
|
|
700
|
+
description="记忆评审 级4 蜂巢并发 + 级5 落库")
|
|
701
|
+
ap.add_argument("--root", default=None, help="认知图根(缺省 MDCG_ROOT)")
|
|
702
|
+
ap.add_argument("--packages", type=int, default=10, help="试跑包数(0=全部)")
|
|
703
|
+
ap.add_argument("--jobs", default=None, help="hive jobs 目录(隔离生产)")
|
|
704
|
+
ap.add_argument("--ctx", default=None, help="context 文件目录")
|
|
705
|
+
ap.add_argument("--hive", default=None, help="hive 可执行文件")
|
|
706
|
+
ap.add_argument("--model", default=None, help="模型名")
|
|
707
|
+
ap.add_argument("--timeout", type=float, default=1800.0)
|
|
708
|
+
ap.add_argument("--reviewer", default="mreview-worker")
|
|
709
|
+
ap.add_argument("--applier", default="mreview-applier")
|
|
710
|
+
ap.add_argument("--dry-run", action="store_true", help="只校验不落库")
|
|
711
|
+
ap.add_argument("--apply", action="store_true", help="实际落库(缺省不写)")
|
|
712
|
+
a = ap.parse_args(argv)
|
|
713
|
+
from .. import mdcos
|
|
714
|
+
root = a.root or os.environ.get("MDCG_ROOT")
|
|
715
|
+
if not root:
|
|
716
|
+
raise SystemExit("缺少 --root 或 MDCG_ROOT(拒绝在未知根上运行)")
|
|
717
|
+
cg = mdcos.MdCGSecure(root, principal=verifier_principal(actor=a.applier))
|
|
718
|
+
built = build_packages(root=root, limit=None)
|
|
719
|
+
pairs = built["pairs"][:a.packages] if a.packages else built["pairs"]
|
|
720
|
+
rep = run_batch(pairs=pairs, cg=cg, exe=a.hive, jobs_dir=a.jobs,
|
|
721
|
+
ctx_dir=a.ctx, reviewer=a.reviewer, applier=a.applier,
|
|
722
|
+
model=a.model, timeout_s=a.timeout,
|
|
723
|
+
dry_run=(a.dry_run or not a.apply))
|
|
724
|
+
print(json.dumps(rep, ensure_ascii=False, indent=2))
|
|
725
|
+
return 0 if rep.get("ok") else 1
|
|
726
|
+
|
|
727
|
+
|
|
728
|
+
if __name__ == "__main__": # pragma: no cover
|
|
729
729
|
raise SystemExit(main())
|