Delphi

/delphi — Codex Adapter for NexusOS Delphi

This skill gives Codex direct access to Delphi by wrapping the existing NexusOS/Claude Delphi implementation. It does not duplicate Delphi Pro logic.

The adapter itself inherits the Codex host default (gpt-5.6-sol, high). The --model option applies only to the legacy Claude backend; the Codex/Python research backends have their own deterministic provider routing.

Decision

Use a thin adapter, not a rewrite:

• Delphi Pro source of truth: ~/.claude/plugins/delphi/
• NexusOS agent identity: ~/.nexus/agents/delphi/
• V2 wrapper/iron laws: ~/.nexus/v2/agents/delphi/
• Codex entrypoint: scripts/codex-delphi-dispatch.sh

Exact in-process replication of Claude Delphi Pro is not technically correct because Delphi Pro relies on Claude Code plugin semantics, Claude CLI model routing, Agent-tool scout dispatch, and plugin-local skills/resources. Codex uses those files as specifications, then runs its own codex backend without Claude runtime dependencies.

The dispatcher passes Delphi's full D3/D4 tool allowlist to Claude CLI, including Agent, Skill, Cortex, Tavily, Brave, Exa, ArXiv, Wikipedia, OpenAlex, and YouTube transcript MCP tool names. Actual availability still depends on Claude CLI quota and MCP configuration.

Usage

~/.codex/skills/delphi/scripts/codex-delphi-dispatch.sh \
  --topic "research topic" \
  --budget 0.15 \
  --backend codex

Optional:

--timeout 900
--model claude-sonnet-4-6
--depth D1|D2|D3|D4
--backend auto|codex|claude|python
--store-cortex
--dry-run

Depth Rules

• If --depth is omitted, auto-classify before dispatch using Delphi Pro intake rules.
• D1: quick lookup, low budget, no critic.
• D2: standard brief, multi-source, default for normal research.
• D3: deep research, critic required by Delphi rules.
• D4: exhaustive research. In the Codex backend, an explicit Pafi-requested D4 dispatch is treated as owner-authorized and writes d4-approved.json plus d4-approval-used.json for auditability before running. Use only when Pafi explicitly asks.

Auto-Classification

Codex mirrors Delphi Pro's Claude-side routing:

• D1: single factual lookup, quick check, what is X, short public-knowledge answer.
• D2: normal research, comparison, multiple perspectives, exploratory technology landscape.
• D3: deep research, strategic synthesis, competitive intelligence, trend analysis, complex/high-stakes/academic topics with complexity score ≥3.
• D4: explicit D4, exhaustive, investment-grade, Pafi-grade, due diligence, final/production-lock decisions. Codex dispatch auto-writes an owner approval marker for explicit D4 requests; direct low-level runner calls still require an approval artifact or flag.

The dispatcher records depth_source: auto|explicit in DISPATCH.md so downstream gates know whether depth was inferred or forced.

Output Contract

The adapter creates a task workspace under:

~/.nexus/workspace/codex-delphi/codex-delphi-YYYYMMDD-HHMMSS-PID/

It writes:

• DISPATCH.md
• PROGRESS.md
• EXECUTION.log
• output.md
• credential-matrix.json for the Codex backend
• delphi-result.json and critic.json for Codex D2/D3/D4 runs
• d4-approved.json and d4-approval-used.json for owner-authorized Codex D4 runs

If --store-cortex is set and Cortex is reachable, it stores output.md to Cortex collection research with genie_visible: true.

Backends

• auto: use the Codex-native Delphi backend.
• codex: run Delphi through local Python/Codex code with no Claude CLI, Claude Code plugin, Anthropic Claude model, Claude agent, or Claude quota dependency.
• claude: legacy backend that uses the real Delphi Pro plugin through Claude CLI.
• python: use the local deterministic Delphi research engine only.

The Codex backend reads Delphi Pro files as specifications only and produces Delphi-compatible workspace artifacts. The Python backend is useful for low-level smoke tests but does not write the full Codex workspace contract.

VERIFY

Run these checks after changing the adapter:

bash -n ~/.codex/skills/delphi/scripts/codex-delphi-dispatch.sh
cd ~/repos/delphi && pytest -q tests/test_codex_runner.py && pytest -q
~/.codex/skills/delphi/scripts/codex-delphi-dispatch.sh --topic "D1 smoke" --depth D1 --backend codex --dry-run
~/.codex/skills/delphi/scripts/codex-delphi-dispatch.sh --topic "D4 owner-approved smoke" --depth D4 --backend codex --timeout 60

Expected results:

• Syntax and pytest commands exit 0.
• D1 dry-run workspace has budget_usd: 0.01.
• D4 Codex workspace has d4-approved.json; it should not stop at BLOCKED_BY_D4_GATE. It may still finish INCOMPLETE if the D4 quality contract fails.
• Codex backend output always includes claude_runtime: false.

Safety

• Do not copy secrets.
• Do not edit Delphi plugin files.
• Do not claim exact replication unless an E2E Delphi job succeeds.
• Prefer D1 or D2 smoke tests before D3/D4.
• If Claude quota or plugin runtime fails, preserve the workspace and report the failure.

When Blocked

If Claude CLI, quota, or Agent-tool access is unavailable, report:

• task workspace path
• exact failing command class, without secrets
• whether DISPATCH.md, PROGRESS.md, and EXECUTION.log were created
• recommended fallback: Codex-native research or Genie/Delphi dispatch

Changelog

• 1.1.0 — Added Codex-native backend contract, no-Claude runtime guarantee, workspace artifact list, and verification checklist.
• 1.2.0 — Added Delphi Pro-compatible auto-classification when --depth is omitted and records depth_source.
• 1.2.1 — Explicit Pafi-requested Codex D4 dispatches now auto-write owner approval artifacts instead of stopping at the D4 gate.