Coverage-closure program — orchestrata-shaped (new tree)
The spec-conformance rebuild tree ~/workspace/orchestrata-shaped (plan of
record: ~/.kimi/plans/cloak-blue-marvel-spider-gwen.md, portal reviewed
through Round 23) runs a coverage-closure program: every checkpoint is a
single measured parallel pass, every test may run at most once per cycle,
and any file found under 90% gets lifted the same run.
Checkpoint trajectory
| cp | TOTAL | missed | notes |
|---|---|---|---|
| run1 | 71% | 7,010 | program start |
| cp2-cp5 | 79.4 → 84.9% | 5,371 → 4,336 | ported-test waves |
| cp6 | 86.4% | 3,922 | run 8: first full parallel pass (726s @ -j8) |
| cp7 | 90.07% | 2,858 | run 9: janus live, phantom-import gate (709 modules, 0 phantoms), live-path-hazard harvest, 20 commits |
| cp8 | 92.86% | 2,074 | run 10 (2026-09-18): 29,035 stmts, −784 vs cp7 |
| cp9 | 94.98% | 1,869 | run 11 (2026-09-20): 37,265 stmts, −205 vs cp8, +8,230 landed stmts |
| cp10 | 96.01% | 1,563 | run 12 (2026-09-20): 39,137 stmts, −306 vs cp9, REGISTERED_UNHARVESTED EMPTY |
Suite at cp10: 339 modules, 7,686P / 0F / 15S (skips 18 → 15). Gates 10/10
(978/978 import sweep, 0 phantoms, 0 deferred), ruff 1,434, mypy 1,027 green.
Tree clean @ f4beee0.
Run 12 (2026-09-20) — completing the missing coverage
- Lane P — pipeline-engine harvest + session/cli close: the old
orchestrator.pipeline_runnerengine harvested (60 src files / 5,357 LOC, 17 modules — run-11's "no corpus" census was scoped wrong; ~6,600 LOC of old tests existed). pipeline.py 109→0 (100%),__main__100%, overview 25→13 registered, requirement_resolution 65→100%, 3 capability_mesh skips deleted. REGISTERED_UNHARVESTED allowlist EMPTY (gate-10 0 deferred). 435P/0F/0S. - Lane Q — graph residuals: prolog_bridge 85→0, ingester 79→0
(+ adapters memory/rerankers harvested at 100%), retry.py 13→0
(deterministic mvcc corpus), absorber retrofit 60 raw test-side graph
sites / 8 modules green at -j12. Fixed a dead
__post_init__Pydantic conversion defect (validators silently never ran). 403P/0F/0S. - Close: completion 292 targets / 844s; 2 reds root-caused (CHANGEME env-poison paying its price a second run straight; a tracer-saturation timing race fixed with a conflict-confirmed handshake). cp10 authoritative after the F-R10-E1 pre-combine repair (third materialization — deposit-time re-root must become MECHANICAL).
Run 11 (2026-09-20)
- User ruling: "update all dependencies to latest" (
f0ba261) —uv lock --upgrade+ click 8.5.0 / typer 0.27.2 / psutil 7.2.2 (+types-psutil) landed; GIL off across every new import; vendored janus wheel intact. This resolved all three pending dep rulings at once. - Tools wave complete (slices 1+2): legacy-handler harvest (tool_executor 38.4→96.3%) + the full native closure (task_tools/search/lru/repl/worktree/ lsp/mcp×2/remote_trigger/schedule_cron/context/codebase×2 — all ≥94%) + ast_query pkg + typer CLI (wired via [project.scripts]). 129 registered skips retired on first-run evidence (ledger 136 → 8). 819 first-run tests, final 0F.
- session/cli harvest (click unblocked): prereqs archive/kg_context/
spec_parser/kg_sync + the 12-file cli package (55/55 @traces verbatim), 534
tests 530P→0F,
sessionconsole script wired. Found and fixed the bare-path-MERGE hijack defect class (project-qualify all graph keys). - U5d relocation closed: gen-skeleton API →
src/orchestrata_ast/skeleton/(24 files) + gate0.py; 85/85 defs ast-identical, 64/64 behavioral parity. - Close: completion pass 255 targets / 811s; 2 reds root-caused (raw-test-op transients — unabsorbed test-side graph ops under -j12 contention; sanctioned retry fix, assertions untouched); cp9 authoritative.
Run 10 lanes (2026-09-18)
- Lane A — B-C.4 Prolog governance corpus (
fb9fc03…): 12.plfiles / 3,158 lines landed byte-identical (zero atom translations) intosrc/orchestrata_shaped/prolog/; all 51 registered B-C.4 skips lifted; PrologGate issues real decisions (unknown-agent denial proven, not fail-closed). janus_bootstrap 48.5 → 99%. - Lane B1 — LLM drivers: deepseek 14.3→98, browser 29.4→100, qwen →100, claude_api →100, anthropic →100, aider →99. playwright==1.63.0 + greenlet==3.5.6 installed (user ruling — proven on old tree's 3.14t; additive-only sync, janus wheel intact, GIL off, real headless Chromium). 18 never-run browser tests un-gated → all green, 0 skips.
- Lane C — anomaly_detector harvest (
bdd2bf7…): 515-LOC old module → SC-11 split package (6 files) +SynchronizerProtocol; 8 files 100% (243/243 stmts);check_src_importableDEFERRED 11→10; synchronizer's 8 anomaly-branch misses closed; 42P/0F/0S; zero graph residue (bolt-verified). - Lane D — session/cli: DEP-BLOCKED on
click(registered, not hacked): 12 src files / 2,661 LOC + 212 tests / 3,465 LOC fully mapped; needs the click ruling (below). Corrected the old "2,217 LOC" undercount. - Lane E — completion + checkpoint 8 + format normalization (post-measurement): 196 files reformatted AFTER cp8 (line-shifts must never touch measured data); exemption line-refs re-anchored (34/34 pragma refs verified).
Mechanics (the run-once doctrine, now machine-enforced)
scripts/run_tests_parallel.py— the ONLY way tests run (threadpool, subprocess-per-module, LPT longest-first from a duration cache,-j 12, coverage data isolated incoverage-data/). xdist and bare pytest banned.reports/test-runs.ndjson— the ledger (607 + 280 records committed): one atomic O_APPEND record per target execution {run_id, cycle, target, verdict, P/F/S, cov, data_file, src_head}.scripts/test_ledger.pyderives unrun/stale/purge-list/cycle-tally mechanically — no agent memory.scripts/checkpoint_coverage.sh N— combine once →reports/ coverage-checkpoint-N.json→.coverage.jsonLAST; the floor ratchet (coverage-floor.json, sum-invariant) can never drop.- 10 static gates incl.
check_src_importable(709-module import sweep, PHANTOM class) andcheck_coverage_floor. - Coordinator pattern: plan/verify stays human-side;
deepseek-coderagents execute (thegeneralchannel is deny-ruled; model id isopenrouter/qwen/qwen3.8-max-0902after the Sep-18 config fix). Worktree lanes must re-root coverage paths at deposit (F-R10-E1 standing-rule).
Bug classes the program has caught (run 8-11)
Phantom split imports (2× twins at 0%), live-path unharvested-module ModuleNotFoundError, f-string-reflow prefix loss, silent ImportError swallowers (audited to zero), Memgraph DDL transient retry gap, teardown seam without transient absorption (only fails under 852s of -j12 load), a 32-node graph leak, LastFailed race corruption (cacheprovider now off), bare-path-MERGE hijack, SWI-json packaging, raw-test-op transients, external-base inheritance dominance, boundary-map collision.
PENDING RULING — pyright
7 registered skips + a 75%-covered backend retire with pyright added
(same shape as prior dep rulings: old tree pins it). Secondary: host
claude-CLI session pollution (blocks llm_integration marker-content
contracts — host fix or adapter-isolation ruling); CHANGEME placeholder
leak hygiene (~35 old-tree .env vars poison agent shells).
Run-13 queue
multiagent.pl + MultiagentPrologBridge · intent-DCG (generator + gate) ·
orphan tail (ast_pipeline, conversation_tools) · scripts/* harvest incl.
generated_orchestrator regeneration · llm_integration e2e per the cassette
ruling — standardize on the call_llm-scripted seam, do NOT port the
cassette fixture layer (retires pipeline_runner's 24) · packet_header.py:71
dead __post_init__ · tool_executor T-phase (16) · adapter tail (deepseek 31,
temporal 35, packer/_ops 34, usage_log 34, _autofix 31, canonical 39,
st_embedder 38) · untraceable-line audit · memgraph-adapter closure ·
kg.py deterministic drivers · pending rulings (pyright, volk_config A.5,
host claude-CLI pollution, CHANGEME hygiene, machine PYTHONPATH fix).