f463e726dd
## Why Evaluating buzz agent memory retrieval by seeding a memory then asking the buzz agent a question it needs that memory. **Bug Found**: System prompt had no inclusion of retrieving cold memories and suggested looking in a mem/*.md directory that does not exist. Updated `system-prompt.md` to include memory CLI tools and usage. Eval Before System Prompt Change: 0/3 Eval After System Prompt Change: 3/3 ## What - Add a `memory-retrieval` benchmark that seeds agent memory with `buzz mem set` before asking a direct question. - Grade the observable threaded answer without inspecting tool calls or exposing the answer in channel history. - Teach agents to use `buzz mem set`, `buzz mem ls`, and `buzz mem get` for cold memory. - Add a wire-debug endpoint configuration for diagnosing ACP tool calls in local runs. - Add fixture, seeding, verifier, and prompt coverage. ## Risk Assessment Low. The runtime changes are limited to the benchmark harness. The production-facing change clarifies existing memory commands in the base prompt; it does not change memory storage, relay behavior, or authorization. ## References - Before the system-prompt changes, 0/3 attempts passed because agents never invoked the `buzz mem` CLI and instead searched a non existent filesystem - After the changes, 3/3 attempts passed. ACP wire logs confirmed that every agent ran `buzz mem ls` followed by `buzz mem get` and returned `net_gpv`. --------- Signed-off-by: Philip Azar <pazar@squareup.com>
Endpoint launch configs
Deployment-time mapping from manifest endpoint names to
EndpointLaunchConfig (provider / api_key_env / env), passed to
harbor run as --agent-kwarg endpoint_config=<path>. This is deployment
config, deliberately OUTSIDE the immutable condition manifest — the manifest
endpoint string remains the join key.
Every key in these files must be a manifest endpoint name; the loader treats all entries as endpoint configs (no comment keys).
openai-live-wire-debug.json
Diagnostic variant of openai-live.json for local runs. It enables
acp::wire=debug, so retained agent stdout logs include full ACP messages,
including tool-call arguments and results. These logs may contain prompt or
command content; keep them local. The verifier and reward do not read them.
m1-local.json
M1 wiring proof: both placeholder endpoints resolve to one local llama-server
(OpenAI-compatible, http://127.0.0.1:8091/v1, no cloud keys).
buzz-agent env contract (crates/buzz-agent/src/config.rs, pinned at the M1
binary SHA): provider=openai reads OPENAI_COMPAT_API_KEY +
OPENAI_COMPAT_BASE_URL; the runtime sets BUZZ_AGENT_MODEL from the
manifest endpoint name, which overrides OPENAI_COMPAT_MODEL — llama-server
ignores the model name, so the placeholder value is harmless there.
llama-server needs no real key; the provisioner's per-endpoint
llm_api_keys map supplies a dummy value.
The Databricks pilot config is the same file shape with real serving endpoint hosts/keys.