v1.0.0 - Low-Impact Runtime and Verification Evidence

v1.0.0 - Low-Impact Runtime and Verification Evidence

This release focuses on making codex-agent-mem cheaper to run, easier to audit, and safer for long-lived Codex Desktop sessions.

Highlights

• Low-impact MCP profiles: minimal, standard, and full.
• Real --read-only mode for MCP usage.
• Lazy SQLite initialization so unused MCP connections stay cheap.
• Compact MCP response text with full payloads retained in structuredContent.
• known_pack_hash / not_modified support to avoid resending unchanged continuity packs.
• Runtime health diagnostics with cache, lazy-init, heartbeat, and spawn-storm fields.
• Optional local telemetry and optional daemon/stdio bridge.

Verifiable results

This release includes reproducible public verification data under docs/verification/.

The published verification run was executed with Codex Desktop, model GPT-5.4, reasoning effort xhigh, using synthetic fixtures only.

Measured scenarios:

• Small project continuity: 1,841 -> 216 tokens, ~88.27% saved.
• Medium agent workflow: 4,855 -> 233 tokens, ~95.20% saved.
• Large repeated audit: 9,731 -> 232 tokens, ~97.62% saved.
• Sub-agent handoff example: 6,523 -> 239 tokens, ~96.34% saved.

These numbers are not a universal guarantee. They are reproducible measurements showing the potential when repeated project context can be compacted.

Validation

• pytest: 32 passed.
• ruff: all checks passed.
• staging build: wheel and sdist generated.
• public verification runner reproduced successfully.
• staged content scanned for internal/private paths before publication.