Orchestrata Ultimate › Blockers

Orchestrata Ultimate Blockers

Refreshed on 2026-10-23 against:

• the session 271 handoff;
• the host inventories of 2026-10-22;
• the tracker's backlog (context.json in wt-MAP-PRECOMMIT, read with jq), which holds 142 open items, 18 of them P1 or high.

The generated tasks page lists every open item.

Active blockers (2026-10-23)

Risks

• Stale AST data from hidden spawners. A PostToolUse hook runs serialize --incremental from the primary checkout, which is 1,102 commits behind its origin/master ref. ~/.venv holds an editable ast-sync install of that checkout. See hidden AST spawners.
• The T9 btrfs pool. The October stall came from full qgroups and was fixed by switching to simple quotas. Incus re-enables full quotas if quotas are ever turned off. See T9 Incus btrfs qgroup stall.
• Coverage numbers are not trustworthy yet. The baseline is a 2026-07-08 snapshot, and parallel waves may have cross-contaminated .coverage.* files (tasks 109, 448, 451 below).
• Hook defects that look like guards. Several hooks do nothing or cannot block: the codetools plugin's PreToolUse hooks, infra_check.py, and PostToolUse "BLOCK" hooks. See host inventories.

Open P1 and high tasks (18)

From context.json deferred_work, which has 564 items: 422 done and 142 open. The number is the item's index in that array. Notes in italics were added on this refresh.

Earlier P1 list (2026-07-08 backlog, kept as history)

This list predates the 18 measured items above and was not re-measured on 2026-10-23.

1. Communication: volk↔volk messaging + orchestrator↔volk with real-time status/heartbeat awareness.
2. Context injection: wire SPEC-009 + memory (SPEC-010) / KG (SPEC-008) retrieval into volk execution.
3. KG schema: audit + add semantic node/edge types, reusable queries, constraints (SPEC-008/105); remediation batch per plans/kg-schema-gaps.md.
4. Spec-gap remediation: verify implementation against all specs 000–105; close real gaps.
5. AST module evaluation + improvement (serialize_codebase_ast.py + orchestrata_tools) per the audit's ranked list; plus new deterministic codegen templates.
6. Coverage back-fill: ~95 src files still <90% (see cov_baseline.json); then verify coverage in isolation (parallel .coverage.* cross-contamination risk).
7. Escalation absorption design pass: graft kernel PipelineEscalationMixin capabilities onto the orchestrator runner — design fork (capability-mesh routing vs fixed StageSpec) needs sign-off; touches live gateway control flow.
8. D12 unified persistence layer: funnel ~30 direct GraphDatabase.driver/redis.Redis call sites through core factories + architecture-checker gate; converge 6 KG writers into one canonical writer.

Earlier P2 list (2026-07 to 2026-09, not re-measured)

• Test-purity remediation: 9 files / 11 violations (mocks/skipif) breaking the no-mock/no-skip doctrine. DONE — audit_test_purity.py reports 0 violations repo-wide (verified session 253).
• MemoryEngineProtocol divergence: agent_id vs volk_id naming across protocol/interface/impl — needs a canonical-name decision.
• Volk MERGE-key divergence in KG: pipeline_kg keys on alias, harness on name, kernel audit on volk_id → duplicate Volk nodes; converge on alias.
• Provider-protocol + core interfaces-vs-protocols split-brain: duplicate protocol/Message trees (core.protocols vs core.interfaces) across ~20 import sites; reconcile per plans/provider-protocol-reconciliation.md.
• AnthropicAdapter tool forwarding: does not translate OpenAI-shaped tools to Anthropic schema, so tool_use can't trigger in production (SPEC-011/103).
• Pipeline resume is not real resume: PipelineRunner never calls checkpoint.save_stage(); no store records original intent/work_type.
• (TESTING, ARCHITECTURE) routing-subject gap (surfaced 2026-09-21): CAPABILITY_SUBJECTS has no entry for the performance route's U:Qa→U:Ar edge, so U:Ar receives fallback subject "process" (unanswered) and the route dies at stage 2 — the same defect class 9fbed221 fixed for intake, its guard doesn't cover this edge. Choosing U:Ar's subject (it answers implement/scope_validated/ux_complete/design_feedback/review_edit) is a design decision for Mohamed; the rewritten test_performance_work_type_unhandled_subject_fails_fast documents the gap.
• Per-worker Neo4j test pool on Incus: unlocks -j 8–10, full suite ~6–9 min (clone containers with unique IPs, then worker-scaling ladder). DONE — NEO4J_URI_GW1..GW7 live in .env; suite runs 8 workers in ~7 min. Gotcha (session 253): ORCHESTRATA_WORKER_SLOT must be gwN-format — bare N silently falls back to the shared primary (see ADR-024 in decisions).
• Memory Phase-2: 5 volk call sites bypass base recall by calling module-level call_llm directly.
• Async-plugin convergence: 56 test files use @pytest.mark.anyio under asyncio_mode=auto — latent dual-loop hazard.

Earlier risks (2026-07, not re-measured)

• 2026-07-29 gate check (working tree, feat/graph-rename, uncommitted): the most recent run_tests.py invocation aborted before the test phase — parallel-safety gate found 3 blocking violations (2 unscoped Neo4j DETACH DELETE/CREATE in test files, one of them in a new test_migrate_kg_to_orchgraph.py) and the test-purity gate found 4 [STUB] violations in 2 files (scripts/test_affected_tests.py, scripts/test_rulings_lint.py); see reports/run-tests-latest.json. Not yet re-verified against a clean master checkout — the last confirmed-GREEN gated run remains 2026-07-27 (reports/run-tests-2026-07-27-green.json, 29,200 passed / 1455 modules). This does not overturn the session-253 "0 violations repo-wide" test-purity verdict below, which was true at that commit; it flags fresh violations introduced since.
• HA controller (Memgraph) single point of failure RESOLVED — superseded by the orchgraph Raft cluster (3 coordinators, routing 10.53.1.160:7691); the legacy controller container (.153) is deleted. Residual risk: losing 2 of 3 coordinators halts routing.
• 19 pre-existing test_postman_interface_compliance failures (MessageStore/ReceiptStore/ReplayEngine) — confirmed pre-existing, uninvestigated.
• Coverage numbers from parallel waves may be contamination-inflated until the clean sequential verification pass runs.

Recently Resolved

• 2026-09-21 evening — the "order-dependent suite" mirage (see reports/isolation-bisect-2026-09-21.md in the repo): ~79 failures in the gateway/cli/claude modules looked batch-order-dependent (bisect-proven pre-existing at 9fbed221) but were deterministic environment/staleness failures — solo and batch identical, no polluter exists. Four clusters, all fixed: GATEWAY_HOST=CHANGEME_* template sentinel (the .env on this machine was never substituted; gateway bind died with gaierror → "did not start within 15s"); kimi ghost-session residue in ~/.local/state/claude-kimi/sessions.json (per-test registry isolation now); intercept tests stale vs the silent-allow contract 559c6920; three more unguarded CHANGEME_ consumers (kernel/cli, cxp/endpoint_config, harness — convention now uniform in all 7). Root enablers: env/test.env now pins the endpoints and neutralizes sentinels/real keys, and conftest fast-fails at bootstrap on proven-dangerous CHANGEME_ values (ADR-0053 applied to configuration — a template .env can never masquerade as flakiness again). Also closed: a latent prod leak of the conversation-graph endpoint, and 2 stale pins from 9fbed221's subject-routing correction.
• 2026-09-21 migration wave (see architecture migration 2026-09): Prolog query layer repaired (library(json)→library(http/json) fallback for Guix swipl 9.3.8; code_intel reindex + trace_intel report exit 0); vendored janus-swi build fixed for GCC≥14 (scripts/_janus_patched/BUILDING.md — the suite was un-runnable without it); KG REL_CONTAINS/CALLS/IMPORTS canonicalized with identity drift-lock; 3 PEP 649/749 lazy-annotation NameErrors (ADR-0057); CHANGEME_ config-sentinel crash; stale pipeline-CLI fail-fast test (9fbed221's routing correction proven intentional, test rewritten); all 5 generated_* suites regenerated with backward-compatible signature pins. Session suite 2405/2405 green; static architecture debt 100→86.
• Governance wired into PipelineRunner (P0.1) + Prolog atom escaping (P0.2) — see Decisions.
• Stale vector index entries in semantic_graph tests; Memgraph replica-name quoting; GitLab push auth (GITLAB_TOKEN shadowing glab).

Related

• Tasks (generated)
• Upcoming work
• Open decisions
• Session 271
• Research
• Deployment
• Tests