Skip to content
Azhi FlowAzhi FlowAzhi Flow

Workers and OpenCode tools

The server image never runs scripts or harnesses. Workers do. A worker is a Linux host (in the alpha) that runs azhi worker start, with the tools your workflows need.

Script runtimes

Section titled “Script runtimes”

Script nodes need uv (Python 3.12) and/or Bun. The worker reports what it found, and the run plan marks anything missing as a blocker before a run starts.

Tools for OpenCode steps

Section titled “Tools for OpenCode steps”

OpenCode’s search tools (grep, glob) run ripgrep. Each step gets a fresh home folder, so without rg on the worker OpenCode downloads a copy on every step (and fails offline). Run npx azhi setup on each worker host: it checks git, OpenCode and ripgrep and prints the install command that fits the platform.

Terminal window
sudo apt install ripgrep

Without root, put rg in ~/.azhi/tools/bin (or $AZHI_TOOLS_DIR/bin, for a small system drive) or set AZHI_RG_BIN. azhi doctor flags OpenCode workers without ripgrep.

Optional: token saving

Section titled “Optional: token saving”

Off by default. Set AZHI_OPENCODE_TOKEN_SAVING=on on a worker, or token_saving: on|off under a profile’s harness.opencode, to add search-first reading rules to OpenCode steps’ prompts: search with grep and glob, read only the lines needed, never reread a file. The run’s Usage tab shows each turn’s cache share.

Optional: DCP context pruning

Section titled “Optional: DCP context pruning”

npx azhi setup --dcp installs the pinned DCP OpenCode plugin into the tools folder (about 180 MB). It is AGPL-3.0-or-later, so Azhi does not ship it.

  • When token saving is on and the worker has DCP, the step loads it with update checks, notifications and slash commands off.
  • DCP prunes only OpenCode’s own tool output. It never prunes the Azhi gateway tools (azhi_*, submit_output included), because their results carry the excerpt ids the model cites and the ledger records.
  • DCP’s settings come only from Azhi, so a checked-out repository cannot change how DCP runs.
  • DCP only asks the model to compress once context is large: from 50,000 tokens, and insistently past 100,000, by default. Change them with AZHI_DCP_MIN_CONTEXT and AZHI_DCP_MAX_CONTEXT (a token count such as 30000, or a share of the context window such as 40%).

Optional: OpenCode 2 (preview)

Section titled “Optional: OpenCode 2 (preview)”

Off by default. A worker started with AZHI_OPENCODE_MAJOR=2 runs OpenCode steps on OpenCode 2 through its own adapter. npx azhi setup --opencode2 installs the pinned @opencode/cli into the tools folder (about 250 MB), or set AZHI_OPENCODE2_BIN. It is not yet checked with real Copilot or ChatGPT accounts.

Concurrency

Section titled “Concurrency”

At most 2 runs work on one repository at once; the rest wait in order. Change it per workspace with azhi settings concurrency or the Workers page, or the server default with AZHI_MAX_RUNS_PER_REPO (0 means no limit).