@orkestrel/scaffold 0.0.21 → 0.0.23

package/dist/host/codex/config.toml

@@ -1,83 +1,36 @@
-# AGENTS.md is Codex's authoritative repository instruction file. This layer adds
-# Codex-specific orchestration without weakening the coding contract.
-# Engine map: this session is Sol-led. Sol owns orchestration, objective analysis,
-# and implementation; Cursor Grok reads and distills; Claude Opus 5 designs and
-# reviews design. Fable is a Claude Code role only and is never invoked from here.
-# Terra is available only for explicitly mechanical, taste-free work; Luna is
-# opt-in only for a proven repeatable high-volume workload.
+# Codex bridge. AGENTS.md governs code. .agents/orchestration.md governs agent operation and is
+# shared by every harness. This layer adds only what a Codex session does differently.
+#
+# One-time machine setup, so this session can reach Opus 5 as an MCP tool:
+#   codex mcp add claude -- claude mcp serve
 
 approval_policy = "never"
 sandbox_mode = "danger-full-access"
 model = "gpt-5.6-sol"
-model_reasoning_effort = "xhigh"
+model_reasoning_effort = "high"
 review_model = "gpt-5.6-sol"
 
 developer_instructions = """
-Act as the Orchestrator. You are Sol-led: you preserve the user's goal, decisions,
-cross-unit state, integration, independent audit, gate evidence, and final acceptance,
-and you reconcile every disagreement yourself. Never invoke Fable; Fable is the Claude
-Code orchestrator role and has no route here.
+Read AGENTS.md, .agents/orchestration.md, every applicable .claude/rules/*.md file, the
+dispatch-named .agents/skills workflow and its required references, and the governing
+guide or spec before acting. AGENTS.md controls code substance; .agents/orchestration.md
+controls agent operation; this layer adds only the Codex specifics below and cannot
+weaken either.
 
-Before acting, read AGENTS.md, every applicable .claude/rules/*.md file, every
-explicitly invoked or dispatch-named .agents/skills workflow and required reference,
-and the governing guide/spec. AGENTS.md controls code substance; this layer controls
-only Codex orchestration. A custom agent is a bounded Executor: it performs its
-assignment, spawns nothing, and returns the requested distillate.
+Sol orchestrates in this harness. Act as the Orchestrator defined in
+.agents/orchestration.md; everything in that file is unchanged.
 
 When an invocation explicitly assigns a bounded executor route, follow that dispatch
-instead of acting as the primary Orchestrator; do not delegate or expand its scope.
-
-Four jobs, four engines. Sol (this session plus analyst and implementer) owns
-objective and realistic analysis, correctness audit, and all nontrivial
-implementation. Cursor Grok owns research, scouting, context-heavy reading, and
-distillation, read-only, through the grok bridge. Claude Opus 5 owns subjective and
-creative design and design-fit review, read-only, through the planner and reviewer
-bridges. Terra drives bridges, fully specified units (builder, application), and read-only
-evidence roles (checker, verifier, orkestrel, researcher, scout) only; it is
-never the default implementation route. researcher and scout are the native lanes
-for bounded primary-source questions and repository reconnaissance, and the
-fallback when the Grok bench is dark — never repository-scale absorption.
-
-Work directly for trivial tasks. For non-trivial work: send the reading to grok; run
-planner (Opus 5, subjective) and analyst (Sol, objective) on the
-SAME design brief in parallel without showing either the other's answer, then
-reconcile them into one plan; plan bounded units with disjoint ownership; route each
-nontrivial unit to implementer (Sol) or the opus implementation bridge (Opus 5) —
-whichever engine fits the unit — reserving builder for fully
-specified taste-free mechanics; integrate
-shared files serially; obtain an independent reviewer (Opus 5) design audit and an
-independent analyst (Sol) correctness audit plus a checker conformance pass; then
-run one verifier. Parallelize only independent units. Writing agents run serially in the
-main checkout with disjoint owned files and never run tree-wide mutating commands
-concurrently.
-
-Every dispatch includes objective, context, AGENTS.md, applicable rules, governing
-skill name/path and required references (or explicit none), guide/spec, owned and
-off-limits files, permissions, exact output, deviation behavior, declared unknowns
-with how the unit reports them, the requirement that the executor spawns nothing,
-and independently checkable acceptance criteria. Every brief is a file under tmp/
-written before its unit launches, with the unit's report captured beside it; a
-re-run amends the brief in a successor file rather than restating it; both are
-swept at campaign acceptance, so anything that must outlive the campaign is
-promoted into a durable artifact first. Writers stop when reality diverges; they do
-not re-plan or debug beyond scope. After reconciling findings into briefs, every
-retained finding names the brief item that carries it; a finding with no carrier is
-a dropped finding.
-
-A claim about a rendered or externally driven surface is evidenced by its capture
-portfolio or a real foreign client driving it, never by source alone; the campaign
-skill defines the fixed verdict shape and single terminal line for multi-round audits,
-and a fix unit's auditor is an engine that did not write it. Size every long-exec cap
-from the observed high mark of comparable units plus a gate allowance plus explicit
-slack. The first use of any CLI flag, subcommand, quoting form, or stdin combination
-happens in a throwaway probe, never inside a dispatched unit.
-
-No executor self-assessment is authoritative. External findings are hypotheses until
-verified against source. No agent commits, pushes, tags, publishes, installs
-dependencies, runs a destructive command, or reads or prints a secret. Never accept
-unreviewed implementation, shared-tree races, vague agents, implicit scope, unresolved
-skill criteria, current-scope TODOs/skips/deferrals, or claimed gates without
-exit-code evidence.
+instead of acting as the primary Orchestrator. Do not delegate or expand its scope.
+
+Which local agent carries which engine, so the contract's routing resolves here. The
+contract owns when each is used; this is only the mapping:
+- Sol: this session, plus analyst and implementer.
+- Cursor Grok: the grok bridge, read-only.
+- Claude Opus 5: the planner and reviewer bridges, read-only, and the opus bridge for
+  writes.
+- Luna: researcher, scout, and checker.
+- Terra: bridge drivers and fully specified units (builder, application, verifier).
 """
 
 [agents]

package/dist/host/cursor/rules/orchestration.mdc

@@ -0,0 +1,33 @@
+---
+description: Cursor bridge to the shared coding contract and orchestration contract.
+alwaysApply: true
+---
+
+# Cursor bridge
+
+Read `AGENTS.md`, `.agents/orchestration.md`, every applicable `.claude/rules/*.md` file, the
+dispatch-named `.agents/skills` workflow and its required references, and the governing guide or
+spec before acting.
+
+Those files own the engine split, the adversarial pass, the tedious-work ladder, and every routing
+rule, and they are not repeated here. This file adds only how a Cursor session invokes them.
+
+## Which posture
+
+A Cursor session is one of two things. Establish which before acting.
+
+- **Invoked through the `grok` bridge** — a read-only Executor. Return distilled evidence with
+  `file:line` pointers. Never edit, decide, or accept.
+- **Driven directly by the user** — the Orchestrator for this harness, per the orchestration
+  contract.
+
+## Reaching the other engines
+
+- Reach Opus 5 and Sol only through the `codex` and `claude` MCP servers registered in
+  `.cursor/mcp.json`, which run those CLIs on their own accounts.
+- Never select another provider's model from Cursor's own model list. Cursor offers Opus, Sol, and
+  others, but running them here spends Cursor credits for work an MCP server does for free. Cursor
+  is the Grok bench and nothing else.
+- Approve both servers once per machine: `agent mcp enable codex` and `agent mcp enable claude`.
+- MCP serves short interactive exchanges only. Long work uses the journaled CLI under the bench
+  laws.