fireworks2bedrock-migrate

One-shot migration of Kimi K3 inference from Fireworks to Amazon Bedrock, for Python repositories. One command scans the repository, rewrites the inference calls, and opens a pull request. The model does not change — only the backend does.

JavaScript / TypeScript repositories use the npm package instead: npx -y fireworks2bedrock-migrate migrate --pr.

For coding agents (use this README as a skill)

This package is meant to be driven by a coding agent. When asked to "migrate this repository from Fireworks Kimi K3 to Bedrock", follow this workflow exactly and do not improvise on the parts it covers.

Step 1 — Scan

uvx fireworks2bedrock-migrate scan <repo> --json

Read the findings and bucket them:

Step 2 — Migrate and open the pull request

uvx fireworks2bedrock-migrate migrate <repo> --pr --json

This creates the branch f2b/kimi-k3-bedrock, rewrites the MIGRATE files, drops in the llm_client.py helper and f2b.json routing config, adds the runtime dependencies, writes BEDROCK_MIGRATION.md, and opens the pull request.

Step 3 — Act on the exit code

Step 4 — Verify

The tool does not run the repository's own tests yet (checks in the JSON is reserved for that), so the agent verifies:

1. Read review_items from the JSON output — each one needs a decision before deploy.
2. Run the repository's own checks: pytest -q, plus lint/type-check/build from .github/workflows/*.yml or pyproject.toml.
3. Confirm the diff touches only MIGRATE files plus the helper and config.
4. Fix only failures in the files the tool listed, then push to the branch.

Step 5 — Report

Report the pull request URL, the BLOCK items that were deliberately left alone, and any REVIEW items still needing attention.

Rules — never break these

• Never edit a file or finding the tool classified BLOCK.
• Never hand-edit a model ID; the tool and f2b.json own that mapping.
• Never touch infrastructure files unless the tool lists them as REVIEW.
• Do not add setup steps for the developer — one command is the whole experience.

What the migration changes

• Model ID accounts/fireworks/models/kimi-k3 → global.moonshotai.kimi-k3 (us.moonshotai.kimi-k3 with --us-only)
• OpenAI(...) / Fireworks(...) / AsyncFireworks(...) / ChatFireworks(...) → get_llm_client() / get_async_llm_client() / ChatOpenAI(model=KIMI_K3, ...)
• Adds the llm_client.py helper, f2b.json routing config, and the runtime deps openai + aws-bedrock-token-generator
• Writes BEDROCK_MIGRATION.md: full change log, everything left untouched, the IAM policy to attach, and rollback steps

Rollback

Set F2B_BACKEND=fireworks to route back to Fireworks with no code change.

CLI reference

fireworks2bedrock-migrate scan    [path] [--json]
fireworks2bedrock-migrate migrate [path] [--pr] [--us-only] [--region <r>] [--json] [--no-git]

• --pr pushes the branch and opens a pull request (without it the commit stays local).
• --us-only targets the us.moonshotai.kimi-k3 profile instead of global..
• --no-git skips branch/commit/push (for testing).

Scope