coder

README

examples/coder

A coding assistant: one agent wired to the four built-in execution tools from resolute-agent-core-go's tools package — read, write, edit, bash — rooted at a workspace directory, with a stdout observer that narrates tool activity (start / streamed partial output / done) as it runs.

Run it

go run ./examples/coder

runs keyless (a deterministic local provider echoes the last message — it does not call tools) with a SQLite store in ./harness-data and a workspace at ./harness-data/workspace. Set GEMINI_API_KEY to run against Gemini instead, which is what actually drives the tools.

Env var	Default	What it controls
GEMINI_API_KEY	unset	Present → Gemini provider; absent → keyless local echo provider.
MODEL	gemini/gemini-3.1-pro-preview	Model ref, only consulted when GEMINI_API_KEY is set.
DATA_DIR	./harness-data	SQLite store location.
WORKSPACE	$DATA_DIR/workspace	Root directory the read/write/edit/bash tools operate under. Created on first claim.
ADDR	:8490	HTTP listen address.

# dispatch asynchronously (202 + ids), or block on the result:
curl -s localhost:8490/agents/coder/demo -d '{"kind":"user","body":"what files are in your workspace?"}'
curl -s 'localhost:8490/agents/coder/demo?wait=true' -d '{"kind":"user","body":"hello there"}'

# watch the conversation live (replay + tail; reconnect with Last-Event-ID),
# including each tool_outcome record's full result content:
curl -N localhost:8490/agents/coder/demo

Live gate: three scenarios (needs GEMINI_API_KEY)

Start the server with GEMINI_API_KEY set, then place a fixture image for scenario 2:

cp <any small png> "${DATA_DIR:-./harness-data}/workspace/logo.png"

All three prompts use the blocking ?wait=true form so the reply comes back in the curl response; the stdout observer's [tool ...] lines appear in the server's own terminal as each scenario runs.

1. write + read + edit

curl -s 'localhost:8490/agents/coder/demo?wait=true' -d '{
  "kind": "user",
  "body": "Create hello.txt containing '\''hello world'\'', then change '\''world'\'' to '\''resolute'\'' using the edit tool, then read the file back."
}'

Expected: the stdout observer prints [tool write] start / done (ok), [tool edit] start / done (ok), [tool read] start / done (ok) in order; the settled reply's final assistant message contains hello resolute. Verify on disk:

cat "${DATA_DIR:-./harness-data}/workspace/hello.txt"
# hello resolute

2. image read

curl -s 'localhost:8490/agents/coder/demo?wait=true' -d '{
  "kind": "user",
  "body": "Read logo.png and describe what you see."
}'

Expected: [tool read] start / done (ok) on stdout; the tool_outcome record on the SSE stream (curl -N localhost:8490/agents/coder/demo) shows "content":"Read image file [image/png]"; the settled reply is a text description of the image (proof ImageContent reached Gemini and the turn completed).

3. streamed bash

curl -s 'localhost:8490/agents/coder/demo?wait=true' -d '{
  "kind": "user",
  "body": "Run: for i in 1 2 3 4 5; do echo step $i; sleep 1; done"
}'

Expected: while the command runs (about 5 seconds), the stdout observer prints a [tool bash] step N line as each step N reaches the tool's output buffer — proof ToolCallUpdatedEvent carries partial results end-to-end, not just the final result — followed by [tool bash] done (ok) and a settled reply that includes all five step N lines.

Documentation

Overview

Command coder is the coding-assistant example for resolute-harness-go: one agent wired to the four built-in execution tools (read, write, edit, bash) from resolute-agent-core-go's tools package, rooted at a workspace directory, with a stdout observer that narrates tool activity — including a running bash command's partial output — as it happens.

Run it keyless (a deterministic local provider stands in for the model):

go run ./examples/coder

or against Gemini by setting GEMINI_API_KEY (and optionally MODEL, e.g. "gemini/gemini-3.1-pro-preview") — see README.md for the three live-gate scenarios (write+read+edit, image read, streamed bash) this example is built to demonstrate.

Then, in another terminal:

# dispatch asynchronously (202 + ids), or block on the result:
curl -s localhost:8490/agents/coder/demo -d '{"kind":"user","body":"list the files in your workspace"}'
curl -s 'localhost:8490/agents/coder/demo?wait=true' -d '{"kind":"user","body":"hello there"}'

# watch the conversation live (replay + tail; reconnect with Last-Event-ID):
curl -N localhost:8490/agents/coder/demo

The workspace tools operate under WORKSPACE (default DATA_DIR/workspace), created on first claim.