feat: Support for logging session request-response pairs as OpenAI messages

[titu1994]

What does this PR do?

Fixes #7535

This pull request introduces a new trace logging feature for LLM request-response pairs, allowing users to save detailed traces of model interactions for debugging or auditing purposes. It adds a --trace-dir CLI option (also configurable via the OPENCODE_TRACE_DIR environment variable), implements a TraceLogger utility, and integrates trace logging into the LLM streaming pipeline.

Additionally, it has some workarounds for project directory handling for the run command due to bun dev issues. I can revert those.

The most important changes are:

Trace Logging Infrastructure:

• Added a new TraceLogger utility (trace-logger.ts) to create, update, and persist detailed trace logs of LLM requests and responses, including errors and system info. Traces are saved as JSON files in a configurable directory.
• Introduced a --trace-dir CLI option and OPENCODE_TRACE_DIR env variable to enable trace logging, and initialized the logger during CLI startup.
• Integrated trace logging into the LLM streaming pipeline: traces are created for each request, updated with streamed response data or errors, and written to disk upon completion.
• Add a env flag to disable plugin installation of node modules simply because a .opencode directory exists.
  • This is necessary when large scale data generation is being done. Each execution of opencode run is done on a separate directory with specific .opencode configuration of different permissions. This causes repeated network IO to download and install 6.4MB of opencode package in node modules and blows up disk space as well as network bandwidth.

CLI and Project Directory Handling:

• Added a --project-dir option to the run command, allowing users to specify the working directory for project execution. All relevant paths and server initialization now respect this directory.
  • I needed this because running opencode run from the directory I wanted to wasnt working. Maybe this isnt strictly necessary and i just didnt use it properly with the bun dev setup.

Model Data Loading Fix:

• Replaced a macro import for model metadata with a runtime function that fetches model data from disk or the network, ensuring compatibility with browser conditions and bun run.
  • Again, this seems to be an issue with bun dev / bun run setup. Maybe not necessary and i can revert in that case.

How did you verify your code works?

I ran the following command

opencode run {problem_text}  --project-dir {task_dir} --trace-dir {trace_dir}

And here are some traces -
2026-01-09T22-43-08-873Z_ses_45b128da9ffevGAq3ugKoAeozK_trace_mk7gphvn_1smwbk5h.json
2026-01-09T22-43-40-672Z_ses_45b128da9ffevGAq3ugKoAeozK_trace_mk7gpm2h_8cwrj23j.json
2026-01-09T22-43-45-843Z_ses_45b128da9ffevGAq3ugKoAeozK_trace_mk7gqam5_fubms5zt.json

[github-actions]

Hey! Your PR title Support for logging session request-response pairs as OpenAI messages doesn't follow conventional commit format.

Please update it to start with one of:

• feat: or feat(scope): new feature
• fix: or fix(scope): bug fix
• docs: or docs(scope): documentation changes
• chore: or chore(scope): maintenance tasks
• refactor: or refactor(scope): code refactoring
• test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

No duplicate PRs found