Orchestrata Ultimate › Plans › handoff-2026-10-20-supervisor

Handoff — supervisor session, 2026-10-20 (compaction-safe entry point)

Source: plans/handoff-2026-10-20-supervisor.md at 8a2cb42ad · last changed 2026-10-20 · first committed 2026-10-20 · 114 lines.

Prose pending.

Plan text

Handoff — supervisor session, 2026-10-20 (compaction-safe entry point)

Status: handoff for the next context (post-compaction or new session). Session 272, branch components, worktree /home/mohamed/workspace/wt-COMPONENTS. Written 2026-10-20 after the component-fixes handoff's Steps A/B/C all completed and the P-CAP/P-ARCHSYNC/INT-01 packets landed.

Read order: this file → docs/supervisor-workflow.md (the operating manual — verified tool invocations §3/§3.1, packet lane §4, categorization §5–6, context injection §8, automation ledger §9, starter list §10) → docs/lessons/README.md (25 lessons with detail pages) → reports/errors-and-lessons-2026-10-19.md. The older plans/handoff-component-fixes-2026-10-20.md is HISTORICAL (its Steps A/B/C are done — keep for provenance, do not restart from it).

0. Role and routing (standing for this workstream)

Coordinator/supervisor/PM/architect. You brief, observe, verify, commit, register — you write no component code. Workers: openrouter/qwen/qwen3-coder-next, passed as explicit model= on every dispatch (Mohamed's session ruling; stated override of RUL-037's sonnet default until M2 lands — compensating control is your full verification pass before commit). Audit/analysis panels: minimax-m3.1-flash + glm-5.3-flashx (cross-vendor). After every dispatch verify the served model: workflow §3.1 E's session_v2 query. Session priority ruling (Mohamed): finishing packets outranks perfect 90% coverage — honest measured numbers + named second-pass gaps, never padding.

1. What landed today (all committed; evidence in the commit messages)

Memory fix (Step C) applied: components-replaces-ultimate.md now says tools ARE a package.

2. IN FLIGHT at handoff time — finish this first

T17 env-key drift rule (worker ses_ea85f10e7ffeH8o0qxAEqgDGJ6, qwen3-coder-next, fence: scripts/test-review.py + tests/scripts/test_test_review_tool.py; brief: plans/briefs/t17-env-key-drift.md, committed). On its completion (or if already done — check git status on the fence files and the session_v2 row):

1. Verify per the brief's VERIFY block YOURSELF (pytest its test file, test-review --stdout on it, test-review.py scripts/run_tests.py --stdout must be T17-clean, ruff/black).
2. Read the rule: registry scan of *_ENV constants, token-run ≥2 similarity → ERROR, exact-literal- outside-declaring-module → WARNING; RUL-003 infra literals must stay clean; own tests must not self-trip.
3. Check its routing measurement report (which staged files reach test-review via gate.py review-staged + commit_review.pl) — the wiring change is Mohamed's ruling, do not implement it.
4. Commit (explicit list) with evidence; log verdict via session work.

3. Queue (in order; tasks are in session task list)

1. P-RUL039 — promotion gate = functionality tests only; rehearsal-class (real deploy/ship/incus/ terraform/tofu/load) tests exist but NEVER in the gate. Mechanism design: docs/ast-checkable-rules.md §4 + §3#13 (category derivation feeds the exclusion). Touches run_tests.py/promoter.py (test verb = promoter.py:223 → run_tests.py --json). Register RUL-039 wiki page WITH the landed mechanism (rulings_lint fails declared-but-missing).
2. P-RUL040 — reuse ledger + blast radius accurate: replace the run_tests.py:1446 full-run escalation with fail-closed "cannot bound, named set, explicit decision"; symbol-precise expansion (rts.py + affected-symbols). The dead run's 181s spec_012 affected-scope failure belongs here. Register RUL-040 with the mechanism.
3. Build-order packets (Mohamed-ruled, tasks queued): GATE-RUL038 (#3, ADR-0111's owed gate) → T17-GEN (#1) → RUL036-DERIVE (#13) → QUERY-INJECTION (#2) → DAG-BOUNDARY (#4). Each ships with a founding-incident regression fixture + clean-pass proof + the ruling's Enforcement line updated.
4. INT-02+ — integration LIFT increments. BLOCKED on ruling gates for most rows: E-1 (18 increments, Mohamed via AD-33/AD-35 — recommendation stands, needs his clear), N4 (INT-11), N-I1 (INT-04). INT-02/INT-03 (wave-1, no ruling gate beyond C-00 family deps) could dispatch first — write their briefs from packages/integration/increments/README.md rows + the §7.8 shape.
5. Second-pass coverage gaps (named, per the priority ruling): reader L402-403 ProcessLookupError arm; capped-exit end-to-end test pin; sync_architecture_to_graph.py at 65%.
6. Debts: M-ledger items (workflow §9) incl. M5 rename .opencode/agent/deepseek-coder.md → implementer; precommit-walls reason field (A3 UNKNOWN); stale swarm watches (B1-B3, dr-hannah×4, adr0111×2 — owners should unwatch); cut-over-era vocab unification (capitalised service rows vs lowercase DAG vs per-package Components — decision recorded, not due).

4. Decisions owed to Mohamed (ask, don't assume)

1. T17 routing: should staged SOURCE .py files enter test-review (a commit_review.pl policy change)?
2. E-1 ruling clear (unblocks 18 INT increments)?
3. Push components to origin? (Nothing pushed today — doctrine: no push unless asked. Branch is ahead of origin/master by ~13 commits now.)

5. Alerts (the expensive knowledge — do not rediscover)

• Prod graph verification is PER-INSTANCE direct bolt (main bolt://10.53.1.123:7687, replica .128); the routing URI serves divergent read views (measured). Write flake on first attempt is known (driver auto-retries). Memgraph rejects IN-list/UNWIND + OPTIONAL MATCH shapes; never 2>/dev/null a new query shape. SHOW INSTANCES unreachable from this host's port map.
• Neo4j is NOT used — orchgraph (Memgraph fork); neo4j:// + the neo4j driver are Bolt-client naming. Kernel transport = cxpt direct (env key CXP_GRPC_PORT keeps its name). Both clarified everywhere models read (AGENTS.md, SESSION_START.md, wiki banners).
• Workers fabricate under pressure (L11): measure every claim; REWORK-corpus after any reject (L12); verify contract CONSTANTS not shape (L13); run review gates yourself before commit (L14); producer before deletion (L15); fences name every file VERIFY creates (L20); sweep authority explicit (L21).
• Do NOT git add -A: remaining dirty files are NOT this workstream's (context.*, docs/CODEBASE_MAP*, plans/ast-*.json, reports/*.ndjson ledgers — leave them). No push unless asked. No --no-verify.
• The dead run 906142's 20 failures are triaged (tasks #566/#567 closed; rehearsal-class → RUL-039, the 181s outlier → P-RUL040). Do not re-triage.
• RUL-032: never re-run a recorded green to decorate anything; RUL-038-cap semantics now: 300s, exit non-zero with CAPPED, ledger resume, multi-file refusal without the sanctioned stamp.

6. Key files

docs/supervisor-workflow.md · docs/lessons/README.md · docs/ast-checkable-rules.md · reports/errors-and-lessons-2026-10-19.md · plans/briefs/{t17-env-key-drift,component-dag-seed, int-01-release-scaffold,pcap-cap-tests,p-archsync-table-correction,history-audit-2026-10-19}.md · plans/component-fixes-2026-10-19.md (the ruled plan) · packages/integration/{design,increments} · wiki: projects/orchestrata-ultimate/{component-dag,architecture-ontology,lessons-supervisor-2026-10-20}.md, rulings/RUL-038-* (amended).

Related

• Plans
• Orchestrata Ultimate — project index
• Workflow
• Tasks