[codex] Add structured MCP run logging

[CompleteDotTech]

Summary

• adds a structured JSONL run logger for the MCP bridge with stable event envelopes, monotonic sequence numbers, redaction, response previews, SHA-256 hashes, and optional full-text capture
• wraps every MCP tool so calls, results, errors, timings, and parent HTTP traffic are recorded under a shared tool_call_id
• adds log_agent_decision(...) for explicit agent decision annotations before important choices
• makes get_game_state and mp_get_game_state smart-poll by default through enemy turns and other transient non-actionable states, with opt-out and timeout controls
• documents logging schema, CLI flags, redaction behavior, and smart polling controls

Research-grade detail in this slice

Each JSONL row includes schema_version, run_id, sequence, UTC timestamp, monotonic_ms, event_type, and relevant tool identifiers. Tool and HTTP result payloads include character length, byte length, SHA-256 hash, bounded preview text, and truncation status. This gives downstream analysis a stable ordering key, wall-clock time, relative timing, integrity checks, and enough payload detail for lightweight review without requiring full response retention by default.

The logger records session_start, tool_call_start, tool_call_result, tool_call_error, http_request, http_response, http_error, state_poll, state_poll_final_format, and agent_decision. Sensitive keys containing authorization, cookie, password, secret, token, api_key, or apikey are redacted recursively before writes.

Validation

• python3 -m py_compile mcp/server.py mcp/run_logger.py
• uv run --directory mcp python server.py --help
• direct smoke test for logger creation, redaction, log_agent_decision, and combat actionability predicates
• git diff --check

Fixes #10

Token accounting update

• adds deterministic, versioned token estimates to every JSONL record: input_tokens, output_tokens, tool_response_tokens, hidden_poll_tokens, total_tokens, token_source, tokenizer_name, tokenizer_version, model_family, and estimation_method
• writes run_<id>.summary.json with rollups by tool, event type, state type, game mode, action category, run phase, and floor where available
• tracks hidden smart-poll token cost, repeated-state cost, invalid-action cost, largest payloads, replay artifact metadata, and externally supplied usage records
• adds log_external_token_usage(usage_json) so clients can attach exact provider/model token usage by related_tool_call_id, related_event_id, or another stable id
• adds configurable tokenizer/model profile flags plus mcp/validate_run_log.py for JSONL, monotonic sequence, token rollup, summary consistency, and deterministic fixture validation

Prompt and turn accounting update

• adds first-class model_message records through log_model_message(...) with stable message_id, turn_id, turn_index, role/source, content hash/preview metadata, privacy flags, state hash links, related tool/event links, and optional exact token usage
• adds conversation summary rollups for total prompts, total turns, total messages, model messages, tool-result messages, agent decision messages, average tokens per turn, token share by tool-result and hidden polling, largest prompts, by-turn totals, by-role totals, and by-source totals
• extends external usage reconciliation to related_message_id as well as existing tool/event ids
• extends validation for model-message roles, monotonic first-seen turn ordering, message-to-turn linkage, and external-usage-to-message reconciliation

[CompleteDotTech]

Live validation against the local game install passed after closing and relaunching Slay the Spire 2.

Checked:

• active installed STS2_MCP.dll matches the pending all-PR DLL (50255b0caf5ffa397aa9ad4ab2dd0b2f13750bfa62548ed581e35f0dbecf2cb2)
• game API came back on localhost:15526 with 17 endpoints, including /api/v1/wiki
• [codex] Add structured MCP run logging #90 MCP bridge imported via uv run --directory mcp python server.py --help
• live bridge calls to get_game_state(format="json", wait_for_actionable=false) and smart get_game_state(... wait_for_actionable=true) both returned status: ok / state_type: menu
• log_agent_decision(...) wrote an agent_decision record
• generated JSONL log contained session_start, tool_call_start, tool_call_result, http_request, http_response, state_poll, and agent_decision
• log sequences were monotonic/unique and configured sensitive metadata was redacted
• git diff --check remains clean

No follow-up code changes were needed from this test pass.

[CompleteDotTech]

Token accounting implementation and validation added in 73945fd.

Validated locally:

• python3 -m py_compile mcp/server.py mcp/run_logger.py mcp/token_usage.py mcp/validate_run_log.py
• uv run --directory mcp python server.py --help
• python3 mcp/validate_run_log.py --self-test
• generated a fixture JSONL run log using log_agent_decision, log_external_token_usage, and a synthetic hidden state_poll
• python3 mcp/validate_run_log.py <fixture run>.jsonl passed and reconciled JSONL totals with run_<id>.summary.json
• git diff --check

The fixture summary verified total-token consistency, external usage rollups, hidden polling cost, repeated-state cost, replay artifact token metadata, and redaction of sensitive token metadata while preserving accounting fields such as estimated_tokens.

[CompleteDotTech]

Prompt/turn accounting implementation and validation added in e2b7a02.

Validated locally:

• python3 -m py_compile mcp/server.py mcp/run_logger.py mcp/token_usage.py mcp/validate_run_log.py
• uv run --directory mcp python server.py --help
• python3 mcp/validate_run_log.py --self-test
• generated a fixture JSONL run log with log_model_message records for system, user, assistant, and tool roles on a shared turn
• fixture included exact assistant token usage, log_external_token_usage linked by related_message_id, an agent decision annotation, and a synthetic hidden/repeated state_poll
• python3 mcp/validate_run_log.py <fixture run>.jsonl passed and reconciled JSONL totals with run_<id>.summary.json
• git diff --check

The fixture verified total prompts, total turns, model/tool message counts, agent decision message count, by-turn message linkage, external usage reconciliation by message id, exact-vs-estimated token source rollups, hidden polling cost, repeated-state cost, and prompt privacy behavior (content_preview=false).