DynosAI
Get started
Documentation navigation

DynosAI 1.0.0-rc.9

Package version: 1.0.0rc9 (PEP 440). Display: 1.0.0-rc.9. Public status: RC9 · certification pending. Schema remains v6.

This candidate is not stable 1.0. MATRIX_1.0 is not declared PASS. This candidate is not live-certified.

Why RC9 exists

RC8 verified provider-safe local runtime placement but revealed that Codex App Server still did not start or expose the managed DynosAI MCP server.

RC8 live codex.greenfield attempt 4 started Codex (provider_exit=0) with workspace and CODEX_HOME outside OS temp, then recorded mcp_calls=0 and no mcp-activity.jsonl. That is not a Fibonacci/model failure.

RC8 codex --version only proved the CLI could start with the managed home. It did not prove App Server used that home, loaded config.toml, saw mcp_servers.dynosai, started DynosAI, or could call a DynosAI tool.

RC9 adds deterministic App Server MCP introspection and a no-model-turn preflight so configuration, startup, tool discovery and direct MCP calls must succeed before provider tokens can be consumed.

Integrity work in this candidate

  • Codex App Server preflight: initialize, config/read, mcpServerStatus/list, mcpServer/startupStatus/updated, and a read-only mcpServer/tool/call (dynosai_get_next_action with execute=false).
  • Hard gate in scripts/run_matrix_1_0.py --live --cells codex.*: preflight FAIL returns exit 2, starts no turn/start, consumes 0 model tokens, and does not append a MATRIX trial / increment the live attempt counter.
  • Opaque client _meta (progress tokens / vendor keys) no longer aborts a negotiated MCP 2025-11-25 / 2025-06-18 session. MCP 2026 still requires per-request io.modelcontextprotocol/protocolVersion and capabilities.
  • Evidence files: codex-mcp-preflight.json, codex-mcp-preflight-app-server.jsonl, codex-mcp-preflight-stderr.log (secrets redacted; public/portable reports omit absolute paths).

Observed Codex App Server behavior

mcpServerStatus/list without threadId lists configured servers with runtimeStatus=null. DynosAI therefore starts a thread (never a turn) so startup status materializes.

Codex listed 15 DynosAI tools on the acceptance profile during the no-token preflight. The public contract remains 31 names. The preflight requires at least one unambiguous DynosAI tool, not a count of 31.

features.plugins, features.apps and features.recommended_plugins stayed false. Effective config/read still contained [mcp_servers.dynosai]. Those flags were not the RC8 MCP absence.

What this RC does not claim

  • Cursor and Codex are supported 1.0 target providers. They are not live-certified for 1.0 while MATRIX_1.0 is not PASS.
  • Passing the no-token App Server MCP preflight is a prerequisite for a later live Codex cell. It is not that live cell.
  • Roadmap 1.1–1.5 stay out of this candidate.
  • Remaining on 1.0.0rc9 until MATRIX_1.0 is green is an honest successful outcome.

Compatibility contract (unchanged)

  • Schema v6
  • 31 frozen public MCP names
  • Git source authority, knowledge.db workflow authority, human gates
  • Predictive Router remains shadow-only
  • No fake OS sandbox
  • MCP support for 2026-07-28, 2025-11-25 and 2025-06-18

Candidate identity

Canonical candidate:

dynosai_git_commit
+
certification_subject_sha256

candidate_version is a human label (1.0.0-rc.9). Release gates do not require the same attempt number across cells.

RC8 live evidence remains historical: commit 181624b8df6eab3224fbfc23c44ee27b4cacef0d, subject b433bd5252f6c0343e2b246e04d43b72b528871ad30f01b3886c089041e60120, codex.greenfield attempt 4 FAIL.