Orchestrata Ultimate › Sessions › Session 243

Session 243

‹ previous · next › · timeline

Prose pending.

Purpose

DSL benefit investigation — evidence-based research report (reports/research-dsl-benefit.md)

Work log

1. Precedent study: SPEC-104 adopts external OpenSpec markdown (stdlib parser); SPEC-105 adopts OWL/TTL+SHACL standard (Python IR + Jinja2); org-mode research = REJECT new format; build-factory Mermaid-contract mini-DSL already live. 3 auditor agents fanned out over 8 candidate surfaces.
2. Audits 1+2 returned. Harness: N=3 YAML, pydantic-validated, no DSL. Prompts: drift bugs real (PM brace-escape, false-confidence handoff tests) but fix = PersonaSpec contract + template engine. Governance: LIVE BUG found — prolog_client.py:163 may_execute/3 vs governance.pl:690 may_execute/2, silent fallback masks real rules; intent_grammar.pl 16-keyword drift vs intent_vocabulary.py, 9/10 categories. pipeline_router: 2 commits ever, well-tested, leave alone.
3. Context-injection iteration 2 committed (b905bc53): BUILD template gains DEFINES, ambiguous candidates hop-0 only; SearchEngine build pack 4 sections/148 tokens -> 10 sections/516 tokens (class + all 6 methods); 63 tests green
4. Prolog impact proposal delivered by analyst (reports/proposal-prolog-next-impact.md): top pick = finish CodebaseTools volk wiring + reindex lifecycle (asset at 0% utilization, no volk can see the 4 tools)
5. Module-versioning + plugin-installer research delivered (reports/research-module-versioning-plugin-installer.md): COMPONENTS.json manifest + extras recommended (uv-workspace split rejected: 12-package entangled cluster = 73% of modules); orchgraph-convention installer via GitLab registry, fail-closed verification, ~4 sessions after the MCP server; 6 user decisions listed
6. Ontology-skill consumption research delivered (reports/research-ontology-skill-consumption.md): BOTH consumers over one shared ArchitectureQuery read layer; verified the read side is missing (selector selects zero architecture labels, ArchitectureChunk twin invisible to volks, design_patterns catalog-only); 12-query catalog + 5-phase plan + 5 decisions
7. Killed pgrep-self-match-deadlocked wiki index chain wrapper; final wiki index + orchestrata source re-ingest now running sequentially

Artifacts

• reports/research-dsl-benefit.md (120 lines)

Decisions

• DSL verdict: NO new DSL for any surface: All 10 surveyed surfaces show the same failure mode — N hand-maintained representations of one truth; every evidenced defect closes via single-source + codegen/parity at ~1 week total vs a multi-week grammar project; precedent (SPEC-104/105 adopt existing formats, org-mode REJECT) already ratifies reuse-over-invention

Upcoming

• Fix may_execute arity bug (decision D1 first) + Prolog boundary gate + intent lexicon parity, per reports/research-dsl-benefit.md section 4

Handoff

• Key decisions: D1-D4 pending user: may_execute arity direction, intent lexicon source of truth, trace-semantics program, schema.py codegen-vs-delete (report section 5)
• Blockers: None
• In-progress: None
• Important files: reports/research-dsl-benefit.md, src/orchestrata_ultimate/governance/prolog_client.py
• Commands: None
• Warnings: LIVE BUG (found, NOT fixed — research task was read-only): prolog_client.py:163 queries may_execute/3 but governance.pl:690 declares may_execute/2; production silently falls back so per-role rules governance.pl:707-713 are dead; masked by test_prolog_client.py:239 self-asserted fact, intent_grammar.pl missing 16 keywords vs intent_vocabulary.py (9/10 categories) — no parity test

Related

• Session log
• Session timeline
• Orchestrata Ultimate — project index
• Workflow