Controller-Agnostic Orchestration Audit

Current Runtime Surfaces

• bridges/base_bridge.py defines ModelType for codex, claude, and gemini, plus a shared BridgeResponse with success, model, session_id, content, error, and raw_messages. It also centralizes non-interactive execution, timeout handling, CLI availability checks, and optional auth env checks.
• bridges/codex_bridge.py wraps codex exec --json, supports sandbox selection, --cd, optional model, image inputs, and session resume via resume <session_id>.
• bridges/claude_bridge.py wraps claude -p --output-format stream-json, supports optional model, permission mode, and session resume via --resume.
• bridges/orchestrator.py exposes legacy runtime modes parallel, chain, role, and single, plus contract-aware task-run, team-run, task-plan, doctor, and strict Stage-I code-build routing.
• bridges/orchestrator.py can dispatch any runtime agent through _execute_runtime_agent() and preflight Codex, Claude, and Gemini through _runtime_preflight_error().

Current Contract Coverage

• standards/mcp-agent-capability-map.yaml has a task-level coordination_contract requiring task packet fields such as task_id, paper_type, topic, required_outputs, required_skills, required_skill_cards, and quality_gates.
• The same map has agent_registry entries for codex, claude, and gemini, functional role ownership through functional_agent_registry, and per-task primary_agent, review_agent, fallback_agent, required_mcp, and quality_gates.
• standards/agent-profiles.example.json defines runtime options per agent, including Codex sandbox, Claude permission mode, Gemini transport, non-interactive settings, and timeout values.
• bridges/orchestrator.py builds task packets, injects skill/MCP context, resolves draft/review/fallback runtimes, runs revision loops, optionally runs triad audit, and checks required output files with _validator_gate().
• This branch now has an initial standards/agent-run-contract.yaml plus run/review/handoff templates, but the orchestrator and strict validators do not yet consume that contract during real task execution.

Codex Bridge Gaps

• bridges/codex_bridge.py returns a generic BridgeResponse, but it does not emit controller-agnostic run metadata, artifact read/write lists, verification status, or declared write-set evidence.
• Codex sandbox configuration exists, but there is no bridge-level mapping from orchestration mode to sandbox/write permissions for solo_codex, duo, or triad runs.
• Codex parsing requires a thread_id and agent messages, but it does not normalize Codex events into a shared agent run packet or handoff/review packet.
• There is no Codex-specific solo role gate for writing tasks, citation risk, evidence ledger checks, or claim calibration.

Claude Bridge Gaps

• bridges/claude_bridge.py normalizes several Claude JSON output shapes into BridgeResponse, but it does not emit shared run metadata, artifacts touched, blocking issues, warnings, or verification status.
• Claude permission mode is profile-driven, but there is no mode-aware contract that distinguishes Claude as controller, primary draft agent, reviewer, fallback, or solo runtime.
• Claude output parsing accepts plain text fallback, which helps robustness, but there is no schema validation for review verdicts, required revisions, or handoff fields.
• There is no Claude-specific solo role gate for code tasks, implementation intent, declared write set, failing-test-first evidence, command evidence, or rollback notes.

Orchestrator Mode Gaps

• bridges/orchestrator.py has runtime modes and a task-run pipeline, but current CLI flags do not expose --execution-mode, --controller, --primary-agent, --review-agent, or --solo-role-gates.
• task-run chooses draft/review/fallback agents from standards/mcp-agent-capability-map.yaml; it cannot currently force solo_codex, solo_claude, duo, or triad through a canonical controller-agnostic enum.
• single mode executes one model directly, but it is not connected to the Task-ID contract, MCP/skill collection, validator gate, or solo self-review policy.
• parallel, chain, and role modes operate as collaboration helpers and do not produce structured handoff, review, disagreement, or adjudication artifacts.
• team-run has fanout_merge settings in standards/mcp-agent-capability-map.yaml, but the execution mode is scoped to team fanout rather than the planned controller-agnostic solo/duo/triad contract.

Solo Mode Gaps

• There is no standards/solo-role-policy.yaml or equivalent policy in the inspected files.
• Solo execution today means legacy single mode or degraded task-run fallback, not a first-class solo_codex, solo_claude, or solo_gemini mode.
• Solo runs do not require role-specific self-review templates, implementation intent, writing claim maps, quality gate reports, or handoff summaries.
• Current task-run excludes the draft runtime from review routing, but intentional single-runtime execution is still not modeled as a first-class solo mode with explicit controller, reviewer, and self-review semantics.

Validator Coverage Gaps

• _validator_gate() in bridges/orchestrator.py checks only whether required output paths exist under RESEARCH/[topic]/; it does not validate run packet fields, review packet fields, handoff metadata, controller identity, or execution mode enums.
• --mcp-strict and --skills-strict validate provider/skill availability, but not solo role gates or controller/mode compatibility.
• scripts/validate_research_standard.py is referenced by the broader standards workflow, but the inspected runtime files do not wire it into controller-agnostic run/review/handoff contract validation.
• The initial agent-run contract test enforces packet fields, but no runtime validator yet checks real Codex-only and Claude-only run artifacts for equivalent evidence, blocking issue reporting, confidence, and verification status.

P0 Implementation Recommendations

• Integrate standards/agent-run-contract.yaml as the source of truth for execution modes, runtime agents, required run/review/handoff fields, and verification statuses in orchestrator runtime output and strict validation.
• Add the remaining machine-readable and human-readable templates for disagreement/adjudication, solo task packets, solo self-review, implementation intent, writing claim maps, and quality gate reports.
• Extend bridges/orchestrator.py CLI and task packet construction to carry execution_mode, controller, primary_agent, selected reviewer, session IDs, artifact read/write declarations, warnings, blockers, confidence, and validator status.
• Promote solo execution to first-class task-run behavior: solo_codex, solo_claude, and solo_gemini should use the normal Task-ID contract, MCP/skill context, output validator, and role-specific self-review gates.
• Add offline validators and tests before behavior changes so bridge outputs can be exercised with fake responses without invoking real Codex, Claude, or Gemini CLIs.