Knowledge Wiki Orchestrata Ultimate

Orchestrata Ultimate › Plans › spec-audit › reverify-2026-10-18 › L-newest

L-newest — Spec-gap re-verification (SPEC-106 … SPEC-114)

Source: plans/spec-audit/reverify-2026-10-18/L-newest.md at 8a2cb42ad · last changed 2026-10-18 · first committed 2026-10-18 · 252 lines.

Prose pending.

Plan text

L-newest — Spec-gap re-verification (SPEC-106 … SPEC-114)

Cluster L — newest specs, bonus beyond the 000-105 task scope. Verified against CURRENT code on branch components (HEAD de9e996f2, 337 commits ahead of the 3-month-old audits). No prior audits exist for this cluster; every verdict below is fresh, evidence-grade (file:line), judged by code only — never by checklists/requirements.md boxes.

AST toolkit used exclusively for Python structure (symbols, refs, source); .pl/.sh/.json/.tf/.yml inspected directly (non-Python).

SPEC-106 context-semantic-layer — DONE

Specifies: Deterministic context assembly: canonical KG identity/schema, request analysis → selection → budgeted packing, Prolog (symbolic) relevance refinement, production wiring.

Verified:

Remaining work / plan: None. Scope —, target —, approach —, deps —, effort —.

Evidence log:

SPEC-107 context-injection-improvements — DONE

Specifies: Fill/refresh the layered assembler (session/plan, bounded memory, semantic-graph, AST, artifact store, on-demand tools, caching, expanded graph context, runtime signals, validator loop, AST-snapshot freshness, deployed claude/ config, ship).

Verified:

Remaining work / plan: None. Two cosmetic divergences only — spec names PlanParser/SemanticGraphContextProvider classes; code implements render_plan_context/render_semantic_context functions (same behavior, different shape). Effort —.

Evidence log:

SPEC-108 environment-installer — DIVERGENCE

Specifies: Tracked idempotent installers (claude-env/install.sh, install-infra.sh) replicating the global Claude layer (delegating to deploy/install-claude-config.sh) and the infra fleet (Incus/Terraform or Docker), with neo4j fleet (FR-007/008), license boundary, session-state replication.

Verified:

Remaining work / plan: Narrow — update spec FR-007/FR-008 (and "Complete" status) to the neo4j→orchgraph/memgraph retirement; the installer machinery itself is built and functional. Scope: spec text + verification table. Target: specs/108-environment-installer/spec.md. Approach: amend FR-007 (4-node redis+3 prolog public tier, orchgraph via private release) and FR-008 (redis+memgraph community+data). Dependencies: none. Effort S, low risk.

Evidence log:

SPEC-109 rename-engine — DONE

Specifies: Type-resolved (basedpyright LSP) symbol rename with same-name disambiguation, atomic module move, declarative rename plans, residue report, volk tool.

Verified:

Remaining work / plan: None. Note: spec explicitly scopes agents/tool_registry.py wiring OUT of SPEC-109 (later step), so its absence is not a gap. Effort —.

Evidence log:

SPEC-110 codemod-engine — DONE

Specifies: In-memory lossless CST-forest platform: load/index (CIG+CSR), composable transforms, review-before-write, transactional write-back, graph persistence, observability, analysis passes, 40-check in-memory review, warm daemon.

Verified:

Remaining work / plan: None. Two spec-text nuances: ReviewPass is a Protocol not ABC (review.py:135), and _cig_builder.py shipped as _cig_builder_ast.py (ADR-0068). Effort —.

Evidence log:

SPEC-111 live-code-substrate — GAP

Specifies: Resident multi-agent substrate: warm forest daemon, subtree checkouts + MVCC delta/merge queue, streaming per-delta gates, basedpyright oracle over in-memory buffers, crash-safe checkpoints, agent steering, deterministic Prolog write-gate + escalation loop.

Verified:

Remaining work / plan: Scope: 5 new packages (substrate, delta, oracle, stream, checkpoint) + steering + write-gate. Target files: src/orchestrata_ultimate/refactor/codemod/{substrate,delta,oracle,stream,checkpoint}/. Approach: build FR-001 wiring (watcher→incremental reparse already noted resolved 2026-08-03), then FR-002/003 delta+merge, then FR-004..009 gates/oracle/checkpoint/steering/write-gate. Dependencies: SPEC-110 (forest/CIG/txn), SPEC-109 (lsp_client), SPEC-033 (MergeQueue model). Effort XL, high risk (P0 multi-week substrate; concurrency + optimistic-concurrency core).

Evidence log:

SPEC-112 scripts-to-modules — GAP

Specifies: Consolidate scripts/ to thin CLI shims; extract 6 category-A scripts (flagship serialize_codebase_ast.py → refactor.codemod), 52 category-B, audit 79 category-C, enforce a thin-shim gate.

Verified:

Remaining work / plan: Scope: category-A first (flagship AST-toolkit decomposition from a 12K-line script), then B by package, then C audit + FR-007 gate. Target files: scripts/serialize_codebase_ast.py → refactor/codemod/*, 5 more A-scripts, 52 B-scripts; new standards package. Approach: behavior-preserving extraction (shim --help/outputs byte-compatible), @traces + gates on extracted modules. Dependencies: SPEC-110/111 (substrate needs the toolkit importable). Effort XL, high risk (12K-line decomposition + 52-package migration + gate).

Evidence log:

SPEC-113 measurement-retention — PARTIAL

Specifies: Retain test/runtime measurements with joinable run identity (never statistical prediction): run/batch identity, per-phase durations, per-fixture cost, freshness guard, container-CPU floor, Prolog-boundary prohibition.

Verified:

Remaining work / plan: One small reader. Scope: scripts/ fixture-durations reader (or extend scripts/test_durations.py) answering "top-N fixtures by total setup seconds for run R" from retained rows. Target: scripts/test_durations.py (or new scripts/fixture_durations.py). Approach: glob fixture-durations-*.ndjson, group by fixture argname/scope, join (run_id,batch_id), report top-N. Dependencies: none. Effort S, low risk. (The spec's own 2026-08-20 status already flags this as the sole open item in FR-001..FR-005.)

Evidence log:

SPEC-114 governance-rule-learning — PARTIAL

Specifies: Induce governance rules from open-vocabulary defect corpora (LEARN), quarantine as pending (QUARANTINE), promote by dual human review (PROMOTE), why-not sidekick (EXPLAIN).

Verified:

Remaining work / plan: Scope: (1) LEARN miner over rulings/adjudications/postmortems → provenance'd .pl candidates + tombstones; (2) quarantine rulebase with candidate_violation/4 + :- module + pending severity value in gate_severity.pl (one channel); (3) PROMOTE .pl-diff + dual-review + demotion. Target files: new src/orchestrata_ultimate/governance/rule_learning/ (or scripts/mine_governance_rules.py), prolog/gate_severity.pl, prolog/quarantine.pl. Approach: FR-002 quarantine isolation (head+module, not file) first (prerequisites already landed), then FR-001 miner, then FR-003 promotion. Dependencies: SPEC-113 FR-006 (statistical-fact boundary is the shared prohibition); explain.pl/ADR-0031; factory/promotion/store.py. Effort L, high risk (correctness of gate isolation is load-bearing — the spec documents a measured 5→1 clobber).

Evidence log:

Cluster summary

Spec Class Effort Hard deps
106 context-semantic-layer DONE — —
107 context-injection-improvements DONE — —
108 environment-installer DIVERGENCE S — (spec-text amend for neo4j retirement)
109 rename-engine DONE — —
110 codemod-engine DONE — —
111 live-code-substrate GAP XL SPEC-110, SPEC-109, SPEC-033
112 scripts-to-modules GAP XL SPEC-110/111 (toolkit importable)
113 measurement-retention PARTIAL S — (FR-003 reader)
114 governance-rule-learning PARTIAL L SPEC-113 FR-006, explain.pl/ADR-0031, factory/promotion

Cross-cluster dependencies