Add full Codex provider support across MAP workflows

.agents/references/map-json-output-contracts.md

@@ -0,0 +1,83 @@
+# MAP JSON Output Contracts
+
+Use these contracts when a MAP skill prompt asks an agent to return JSON that is not already covered by evidence-first output examples.
+
+Every `Output JSON with:` prompt section must be either:
+
+- Evidence-first: include `evidence` or `quotes` before verdict, risk, score, root-cause, or decomposition judgment fields, and link to `map-output-examples.md` when the prompt is high-risk.
+- Reference-backed: cite one of the compact workflow contracts below before listing fields.
+
+## Decomposition Output
+
+Use for TaskDecomposer prompts that split a user request into ordered work units.
+
+Required shape:
+
+```json
+{
+  "subtasks": [
+    {
+      "id": "string",
+      "description": "string",
+      "acceptance_criteria": "string | array",
+      "depends_on": []
+    }
+  ],
+  "total_subtasks": 1
+}
+```
+
+The prompt may add workflow-specific fields such as `debug_type`, `estimated_complexity`, or `estimated_duration`, but it must keep subtasks atomic, testable, and dependency-aware.
+
+## Actor Change Summary
+
+Use for Actor prompts that edit files directly and return a compact status summary instead of serialized file contents.
+
+Required shape:
+
+```json
+{
+  "approach": "string",
+  "files_changed": ["path/to/file"],
+  "tests_run": [],
+  "remaining_risks": []
+}
+```
+
+The prompt may add workflow-specific fields such as `trade_offs`, `why_this_fixes_it`, or `potential_side_effects`. `tests_run` is an array of command strings and should be empty when no tests were run. The prompt must still say that files were edited directly with Edit/Write tools and that full file contents must not be serialized in the response.
+
+## Monitor Verdict
+
+Use for Monitor prompts that validate written repository state.
+
+Required shape:
+
+```json
+{
+  "valid": true,
+  "issues": [],
+  "verdict": "approved | needs_revision | rejected",
+  "feedback": "string"
+}
+```
+
+If the Monitor prompt can reject, block, or materially change workflow direction based on code, test output, or artifacts, prefer the evidence-first review finding contract from `map-output-examples.md` and include evidence before verdict fields.
+
+## Learning Summary
+
+Use for Reflector or learning prompts that extract durable rules from a completed workflow.
+
+Required shape:
+
+```json
+{
+  "key_insight": "string",
+  "patterns_used": [],
+  "patterns_discovered": [],
+  "suggested_new_bullets": [],
+  "workflow_efficiency": {}
+}
+```
+
+The prompt must also tell the agent not to repeat existing learned rules already shown in context.
+

.agents/references/map-output-examples.md

@@ -0,0 +1,75 @@
+# Evidence-First Output Examples
+
+Use these compact examples when a MAP prompt asks an agent to return JSON after reviewing code, specs, logs, or workflow artifacts.
+
+## Review Finding
+
+```json
+{
+  "evidence": [
+    {
+      "file_path": "src/service.py",
+      "line_range": "42-47",
+      "quote": "user_id = request.args['user_id']",
+      "relevance": "The value is trusted before authorization checks."
+    }
+  ],
+  "valid": false,
+  "verdict": "needs_revision",
+  "issues": [
+    {
+      "severity": "HIGH",
+      "category": "security",
+      "description": "The endpoint trusts a caller-controlled user id before authz.",
+      "file_path": "src/service.py",
+      "line_range": "42-47",
+      "suggestion": "Resolve the authenticated principal first and compare it to the requested account."
+    }
+  ]
+}
+```
+
+## Debug Root Cause
+
+```json
+{
+  "quotes": [
+    {
+      "source": "test output",
+      "locator": "pytest tests/test_service.py::test_retry",
+      "quote": "AssertionError: expected 3 attempts, got 1",
+      "relevance": "Confirms the retry loop exits after the first failure."
+    },
+    {
+      "source": "src/retry.py",
+      "locator": "lines 18-23",
+      "quote": "except TimeoutError: raise",
+      "relevance": "The handler re-raises instead of continuing the retry loop."
+    }
+  ],
+  "root_cause": "TimeoutError is re-raised before the retry counter can advance.",
+  "next_steps": ["Change the TimeoutError branch to continue until attempts are exhausted."]
+}
+```
+
+## Spec Review Finding
+
+```json
+{
+  "evidence": [
+    {
+      "file_path": ".map/feature/spec_feature.md",
+      "line_range": "31-39",
+      "quote": "Background sync runs every 5 minutes",
+      "relevance": "The spec omits conflict handling for overlapping sync runs."
+    }
+  ],
+  "finding": {
+    "severity": "HIGH",
+    "category": "concurrency",
+    "description": "The spec schedules repeated background work but does not define locking or idempotency.",
+    "suggested_fix": "Add an invariant for single active sync per account and define stale-lock recovery."
+  }
+}
+```
+

.agents/references/map-xml-prompt-envelopes.md

@@ -0,0 +1,65 @@
+# MAP XML Prompt Envelopes
+
+Use this reference when a MAP skill builds a long subagent prompt that mixes user
+requirements, persisted artifacts, workflow policy, and an output contract.
+
+## Purpose
+
+MAP prompts should preserve the user's requirements and branch artifacts before
+asking an agent to reason over them. For long-context prompts, put the documents
+or artifacts first, then the task and instructions, then the expected output.
+
+This follows Anthropic's prompt engineering guidance, accessed 2026-05-19:
+
+- `https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/use-xml-tags`
+- `https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/long-context-tips`
+
+The relevant constraints are:
+
+- use consistent, descriptive XML tags when prompts mix instructions, context,
+  examples, and variable inputs
+- put long documents and data above the query/instructions for long-context work
+- wrap multi-document inputs in nested document tags with source metadata
+- keep output contracts in their own section so schema requirements are not
+  confused with task context
+
+## Standard Envelope
+
+```xml
+<documents>
+  <document source="path-or-origin">
+    <document_content>
+    ...long artifact, diff, spec, finding, or request...
+    </document_content>
+  </document>
+</documents>
+
+<task>
+The one-sentence job for this agent.
+</task>
+
+<workflow_policy>
+The MAP phase rules, ordering constraints, and hard stops that apply to this
+agent call.
+</workflow_policy>
+
+<instructions>
+The concrete checks or actions the agent should perform, in order.
+</instructions>
+
+<expected_output>
+The response schema, evidence requirements, and any formatting constraints.
+</expected_output>
+```
+
+## Rules
+
+- Keep artifact text inside `<documents>` or `<artifacts>` before instructions.
+- Use `<task>` for the user's goal or current subtask, not markdown `**Task:**`
+  inside generated subagent prompts.
+- Use `<workflow_policy>` for MAP sequencing rules and hard stops.
+- Use `<constraints>` when the agent must obey scope, file, or phase limits.
+- Use `<expected_output>` for JSON fields and evidence-first requirements.
+- Keep existing MAP semantic tags such as `<MAP_Contract>` and `<map_context>`;
+  they may live inside `<documents>` or `<artifacts>` when they are input data.
+