v0.5.0

Added

• OMP (oh-my-pi) as provider #11. Reads and writes OMP sessions under ~/.omp/agent/sessions, including the fixed-width title line, and follows the active branch of OMP's session tree. Converts in both directions with all nine other providers and T3 Code (20 new directions). list omp, to-trace omp, search, and MCP support included.
• Thinking and tool calls for OMP. Assistant thinking and tool calls are read, and each tool result is attached as its output. Written sessions use OMP's own block shapes, with a toolResult entry after each call. Honors --no-thinking and --no-tool-calls.

Fixed

• Grok change detection compares rows in grok-cli's grok.db, not the Grok Build folder layout.
• OMP session lookup accepts a unique id prefix only, so a short prefix can't load an unrelated session.

Notes

• Directions: 110 total (90 between the ten file-based providers, plus 20 with T3 Code).
• Tests: 136 passing on Windows.
• Converted OMP files were checked with the installed omp binary (omp --export). Resuming them with a model was not tested.

Full changes: v0.4.0...v0.5.0

v0.4.0

Added

• Options to leave out thinking and tool calls. --no-thinking and --no-tool-calls on convert, bulk, and to-trace. Environment variables UNISESSIONS_NO_THINKING and UNISESSIONS_NO_TOOL_CALLS. The SDK has ConversionOptions and conversion_options(). Thinking and tool calls carry across by default.
• Thinking and tool calls in every direction. Each target writes its own native shape: Claude thinking blocks (unsigned), Codex reasoning and function calls, Factory thinking/tool_use/tool_result, Grok and Freebuff tool calls, and Windsurf Cascade steps.
• ML-ready traces. OpenAI fine-tuning format carries assistant tool_calls and role: tool messages. ShareGPT (LLaMA-Factory) carries function_call and observation. Reasoning is written as a <think> block. HuggingFace STS stays text-only.
• Grok. Reads grok-cli's grok.db as well as Grok Build folders. Writes go into grok.db, so grok-cli can open the converted session.
• Windsurf. Reads grep, view, list, find, MCP, and shell steps, code actions, todo lists, and step errors. Writes planner steps with thinking and tool calls.

Changed

• --on-conflict update rewrites a destination only when the source file is newer. Before, it compared message counts.
• export-all now writes to all nine targets. Before, it skipped every target except Pi and OpenCode.
• Codex output keeps summaries as reasoning and writes function calls.

Notes

• Directions: 90 total (72 direct conversions, plus 18 with T3).
• Tests: 96 passing.
• Converted files have not been opened in every target app. The round trips were checked with our own readers and the real-sample matrix.

Full changes: v0.3.0...v0.4.0

v0.3.0

Added

• T3 Code provider (issue #2). Read T3 threads from statev2.sqlite, falling back to the legacy state.sqlite. The store is read-only.
• Convert T3 threads to Pi, Codex, OpenCode, Claude Code, Devin, Factory, Windsurf, Grok Build, and Freebuff (t3-to-<provider>).
• Export T3 threads as traces and search them through the MCP chat index.
• Experimental: write sessions into a new, empty T3 home (<provider>-to-t3, with --t3-home and --t3-provider). T3 imports the thread on first start. The write refuses homes that already have statev2.sqlite, the live T3 data directory, and databases unisessions did not create.

Fixed

• Search index path keys no longer lowercase on case-sensitive filesystems, which had made every refresh re-index all sessions on Linux.

Notes

• Directions: 90 total (72 existing, plus 9 T3 reads and 9 experimental T3 writes).
• Tests: 73 passing.
• The T3 schema template matches T3 commit 9a3070b. Newer T3 releases may change the schema.

Full changes: v0.2.0...v0.3.0