Vinix24/vnx-orchestration / .claude/skills/fabric-reference
fabric-reference
VerifiedRead-only runbook for how the VNX fabric actually works — state resolution, the single-entry dispatch door, gate invocation, the horizon planning layer, the plan-gate panel, and the hard gotchas that repeatedly bite (PATH-break, dual-CLI, codex-cert, classifier-protected actions). Use when you need to look up a fabric operation instead of rediscovering it, or when a fabric command behaves unexpectedly. Companion to the t0-orchestrator skill (judgment) and vnx-manager (infra maintenance); this one is pure reference.
VNX Fabric Reference
How the fabric works, as runbooks. This is lookup, not judgment — the t0-orchestrator skill owns orchestration decisions; vnx-manager owns infra maintenance. When a T0 rediscovers the same mechanism a third time, it belongs here.
Master-rule (see role-orchestrator.md): project files describe the project; the fabric describes itself. Fabric mechanism lives in the canonical role and here, never copied into a consumer's CLAUDE.md.
Runbook: state resolution
Tracks, objectives, receipts, and coordination state live in the central store, never repo-local.
- Central store:
~/.vnx-data/<project_id>/state/(resolved by the vnx runtime, not hardcoded). - Never pin
VNX_STATE_DIRto a repo-local.vnx-data/statepath in a role or script — a repo-local pin forks state from central = split-brain. - Resolution helpers: Python
scripts/lib/project_root.py; Bashscripts/lib/vnx_resolve_root.sh. project_idnever silently defaults tovnx-dev(ADR-007) — it resolves fromVNX_PROJECT_ID/.vnx-project-id/ git remote, else it rejects.
vnx status # situational awareness from central state
cat .vnx-data/state/t0_state.json | python3 -m json.tool # SessionStart projection
vnx fabric-audit # store-hygiene: split-brain stores, per-project ledgers, hash-chain
Runbook: the single-entry dispatch door
Every dispatch goes through the one door, which decides the lane. Calling a lane script directly is a side door.
vnx dispatch <pending-id> # the door; decides lane, runs phantom-guard
- Provider→lane (hard):
claude/Opus/Sonnet route via the headless lane (dispatch_envelope.run_envelope_headless_plan,claude -p, subscription-preserving); the tmux-spawn lane was removed on 2026-09-18 — NEVERprovider_dispatch.kimi/glm/deepseekroute viaprovider_dispatch.py. - Default build-worker model: sonnet (
workers-kimi-pinned, pin_semantics=default; operator decision 2026-09-23). kimi-k3 is an explicit per-dispatch choice, not the default. A spec that names its own model or provider wins.VNX_OVERRIDE_WORKERS_KIMI_PINNED=1only downgrades the warning, it does not change routing. - Rollback to legacy routing:
VNX_DISPATCH_LEGACY=1(per terminal). - Autonomous staging flow (no template): track → central
stage_spec_bundle→ dry-run → fire → post-mergelink-pr. Full rule:docs/core/DISPATCH_RULES.md(§12 for autonomous). - No Claude Code subagents (Task tool) for dispatch work — governed lanes only.
Runbook: gate invocation
Three gates decide what merges: codex + deepseek (adversarial review, seat 2 on API credit since 2026-11-11) + deterministic CI.
- Provider gates WRITE to the working tree. After a gate: stage the good changes,
git checkout --any stray edits. - codex = strict diff-mode; deepseek_gate (
deepseek-v4-pro) = seat 2 of the default stack. kimi (synthesis/operational angle, proven 3x parallel) is an explicit choice only: the subscription is cancelled and no default names it. - Phantom-guard (
scripts/lib/phantom_guard.py) rejects evidence-free GATE-GREEN receipts. Read-only review roles (REVIEW_ROLES) are exempt — a verdict, not a diff, is expected. - Required headless review gate is not complete until BOTH the result record and the normalized headless report exist.
- Self-merge rule: allowed when local CI is green AND the codex-gate passed — but confirm the FULL GitHub CI is green first, not just the gate (a literal in a CHANGELOG once tripped the state-pin gate and broke main).
Runbook: horizon (planning / future-state)
vnx horizon list # actionable-by-default: done hidden
vnx horizon list --all # include done tracks
vnx horizon list --horizon now --phase queued # the real open NOW work
vnx horizon show <track_id>
vnx horizon queue [--objective <track_id>] [--all] # queued deliverables: state, PR, deps
vnx horizon reconcile # git-grounded auto-close CHECK (no writes)
vnx horizon reconcile --apply # close CONFIRMED tracks (PR merge verified via gh)
vnx horizon close <track_id> --apply --approval-id <token> # human-gated single close
vnx horizon drift # advisory declared-vs-derived divergence
bin/vnxexposes the same verbs under the olderobjectivename (bin/vnx objective list). pipvnxuseshorizon.- A track is
donewhen reconcile confirms its PR merged. Done tracks stay in their band + the ledger;listhides them by default so a reconciled-but-unarchived track does not read as live drift. VNX_AUTO_CLOSEgates whether the SessionStart tick auto-closes; when unset, reconcile is manual.
Runbook: the plan-first gate (panel)
vnx horizon plan-gate ... # multi-model panel reviews a plan before any build
- Panel = Opus + Kimi + GLM-5.2-via-harness (three families → real disagreement).
- One flaky panelist (unparseable JSON, empty output) must not force a REVISE — quorum/abstain handles it (#910). Never treat a codex/glm flake as PASS; retry works.
- The panel has no closeout mode — Tier-2 closeouts can get category-error REVISEs; judge accordingly.
Runbook: staging and firing a dispatch from a script
dispatch_bridge.stage_spec_bundle (scripts/lib/dispatch_bridge.py:229) writes <data_dir>/dispatches/pending/<dispatch_id>/{instruction.md,dispatch-spec.json} and never fires; bridge_dispatch (scripts/lib/dispatch_bridge.py:498) stages and drives the door in one call.
data_diris aPath, never a string (scripts/lib/dispatch_bridge.py:253); the stage CLI's--data-diristype=Pathfor the same reason (scripts/lib/dispatch_bridge.py:710).- The stage-only CLI has no
--gate—--gate,--pr-id,--work-refand--base-refare not in its flag list at all (scripts/lib/dispatch_bridge.py:689-710). Stage from Python when the round needs any of those:gate=isscripts/lib/dispatch_bridge.py:238, validated againstREGISTERED_GATE_NAMES(scripts/lib/dispatch_spec.py:87). The legacy bridge CLI does have--gate/--pr-id/--work-ref(scripts/lib/dispatch_bridge.py:770-772) but it fires on staging, so it cannot be dry-run first. - A follow-up round stages
pr_id=,work_ref=<branch>andbase_ref="origin/<branch>"(scripts/lib/dispatch_bridge.py:241,248,251).pr_idis the PR join key: the obligation is written withpr_number/pr_id(scripts/lib/gate_obligations.py:346-347) andobligation_matches_prjoins on exactly those (scripts/lib/gate_obligations.py:224-239) — withoutpr_idthe round reads as owing no gate for that PR.work_ref=dispatch/<parent>also derivesparent_dispatch(scripts/lib/dispatch_bridge.py:178-200). - Fire through the door in its own tmux window. The door holds its caller for the lane's whole run (
deadline_seconds, default 3600 —scripts/lib/dispatch_bridge.py:240), so keep it out of the T0 pane. Dry-run in the pane first (vnx dispatch <dispatch-id> --dry-runprints the compiled lane/model/billing plan and spawns nothing), then:
tmux new-window -d -n "dlv-<dispatch-id>" \
"cd '${VNX_HOME:-$(git rev-parse --show-toplevel)}' && bin/vnx dispatch <dispatch-id>"
A bare dispatch-id whose bundle sits at pending/<id>/dispatch-spec.json routes to the door, never the legacy lane (docs/core/DISPATCH_RULES.md:258-259).
Runbook: context rotation
A T0 rotates at 500K context (scripts/hooks/t0_context_guard.py:2-5). Measurement, enforcement and execution are three pieces: scripts/lib/t0_context_budget.py measures, scripts/hooks/t0_context_guard.py enforces, scripts/t0_rotate_spawn.sh spawns the successor.
- The bands (
scripts/lib/t0_context_budget.py:39-44, operator decision 2026-09-25): WARN 400K = prepare the rotation; FORCE 500K = start nothing new, rotate at the next boundary; HARD 600K = rotate now, work in flight included. Read the current band withpython3 scripts/lib/t0_context_budget.py <transcript>; unmeasurable context is alwaysok(scripts/lib/t0_context_budget.py:207-210). - Enforcement: at FORCE and above, PreToolUse denies starting new work (a dispatch, a new
/goal,gh pr create) and UserPromptSubmit refuses a new/goal; at HARD the Stop hook demands the rotation now, work in flight included (scripts/hooks/t0_context_guard.py:33-43). - The rotation:
/build-log wrapwritesdaily-log/handoff.md— numbered next steps, plus a## Actief /goalsection with the directive and the remainder when a/goalis running (docs/operations/T0_CONTEXT_ROTATION.md:331-350; writerscripts/lib/context_rotation.py:208). Thenscripts/t0_rotate_spawn.shopens the successor in a new tmux window (scripts/t0_rotate_spawn.sh:460) in the T0 folder when that folder is healthy (scripts/t0_rotate_spawn.sh:418-442). - A
/goaldirective longer than about 650 characters arrives as a pasted block after a rotation and is not executed (OI-2314). The follow-up accepts only a real<command-name>/goal</command-name>user record; a prompt that carries the directive as pasted<pasted_content>text is no arrival (scripts/lib/t0_rotation_delivery.py:181-187;docs/operations/T0_CONTEXT_ROTATION.md:593-595). Keep the handed-over/goalunder that length or the successor starts without a goal.
Gotchas (the ones that repeatedly bite)
- Bash starts every call in the session's start folder, not where the last one ended: the cwd resets after each call, and a T0 starts in
.claude/terminals/T0, so write every command from${VNX_HOME:-$(git rev-parse --show-toplevel)}or an absolute path — never from a relative path. - Do not "fix" that by moving the T0's start folder to the project root: role-file discovery walks up from cwd, so a root start silently drops
role-orchestrator.mdwith no error (.claude/terminals/T0/role-orchestrator.md:137); the rotation's start-folder logic isscripts/t0_rotate_spawn.sh:418-442. - PATH-break in a bash loop: do not run
head/cmp/gitright aftervnxin the same loop iteration — split into separate calls. - Dual-CLI: the real
vnxis the pipvnx_clieditable install (vnx_cli/commands/*), NOT the bashbin/vnx(scripts/*). Verify operator/cutover claims againstvnx_cli, or you check the wrong code path. New commands go in BOTH. - codex-cert: a dead codex-gate is usually a macOS cert-revoke → Trash the binary; fix with
npm i -g @openai/codex@latest(node v22).glm-harnessneeds the local proxy on :4141 (restart per session, not persistent). - kimi-gate edits the working tree: stage the good, revert the stray.
- classifier-protected actions: "fully autonomous" does NOT cover
.claude/settings.jsonself-modification, git tag-push, or PyPI-publish — those stay human-gated. vnx start --helpruns start: it spawns the daemon set (footgun); SIGKILL the supervisor root if triggered.- Central-mode paths: embedded-layout path assumptions can break central-mode resolution (fleet burn-in) — resolve via the helpers, never hardcode
.vnx-data/literals.
Known issues (tracked, active)
- An empty provider completion is usually a TRANSIENT lane issue, not a structural door bug: if
bin/vnx dispatch <id>for a provider lane (kimi/glm/deepseek) lands an empty completion, check the lane's health FIRST — kimi OAuth/quota, glm proxy on:4141— not the envelope. The door's provider path works end-to-end; a blank completion is now downgraded to a loudfailure(embedding the raw spawn result) for ALL provider spawns as of #1107, and the door surfacesEnvelopeResult.errorinstead of a bare exit code, so the real cause is visible. - The plan-gate runs one seat (P2, 2026-10-17): in
plan_gate_panel.py(NOT the/panelskill) a run has the default seat (codex), plus opus only when codex returns BLOCK (configs/plan_gate_panel.yaml). A codex seat that abstains on unparseable verdict JSON or never delivers leaves zero readable verdicts, which isINFRA_FAIL, not a REVISE; the version stays unread and the next run gives it its seat round again. Read the scoring-seat count in the log before trusting a plan-gate PASS/REVISE.
Where the source of truth lives
- Dispatch mechanics, lanes, failure modes:
docs/core/DISPATCH_RULES.md - Architecture + data flow:
docs/core/00_VNX_ARCHITECTURE.md,docs/core/DISPATCH_AND_INTELLIGENCE_ARCHITECTURE.md - State fabric:
docs/core/STATE_FABRIC.md - Provider constraints (machine-readable SSOT):
scripts/lib/providers/provider_constraints.yaml - ADRs:
docs/governance/decisions/
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 469k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 321k | Updated | Claude CodeCodex | Verified | Install |