You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: AGENTS.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -37,7 +37,7 @@ Git worktrees (`git_worktree`): **off** in a minimal `config.toml` until you set
37
37
38
38
Sub-agents (`subagent_spawn`, `task_*`, `subagent_plan_execute`, `task_history_list`): **off** in a minimal config until `[harness.subagents] enabled = true`. The onboarding template enables this for new workspaces. Sub-agents run a second `run_reasoning_loop` with a synthetic chat id (`subagent-…`) and optional tool allowlist. `cancel_children_on_parent_cancel` (default true) controls whether an **explicit** parent cancel (`BusMessage::Cancel` / API cancel / terminal **`/cancel`**) also cancels those child tasks (not triggered by queued follow-up user messages). Completed runs are recorded in the workspace SQLite table **`subagent_tasks`** (same DB as session memory). See `docs/harness-implementation-plan.md` Phase 5.
39
39
40
-
**Named agents** (Phase 5b): `subagent_spawn` now accepts an optional `agent` parameter to invoke specialized sub-agents defined in config. Each named agent has its own system prompt, tool allowlist, and optional model/temperature overrides (wired at spawn in `subagent.rs` via `ProviderCredentials`). Three built-in defaults are registered when no `[agents.<name>]` blocks exist in `config.toml`: **researcher** (web+read tools, temp 0.1), **coder** (full harness allowlist, temp 0.2), and **evaluator** (read-only review tools, temp 0.0). Tools: `agent_list` (list available agents), `task_dashboard` (unified active+history view). The coordinator system prompt is auto-injected with agent descriptions. `wake_on_completion` (default true) enqueues a synthetic inbound when a subagent finishes so the parent can consume the result without polling. Config keys (under `[harness.subagents]`): `wake_on_completion`, `task_history_retention`. Agent definitions use `[agents.<name>]` blocks or `[harness.agents.<name>]` (harness-level wins in merge): `description` (required), `mode` ("subagent"), `system_prompt`, `system_prompt_file`, `allowed_tools`, `model`, `temperature`, `max_iterations`, `hidden`, `color`. Implementation: `src/agent/registry.rs`, `src/agent/subagent.rs`. **`subagent_plan_execute`** accepts optional per-step `"agent"` in plan JSON (same schema as spawn).
40
+
**Named agents** (Phase 5b): `subagent_spawn` now accepts an optional `agent` parameter to invoke specialized agents defined in config. Four built-in defaults are registered when no `[agents.<name>]` blocks exist in `config.toml`: **Semble Scout** (modelsiz local code retrieval), **researcher** (web+read tools, temp 0.1), **coder** (full harness allowlist, temp 0.2), and **evaluator** (read-only review tools, temp 0.0). Semble Scout is a deterministic `mode = "semble_scout"` worker: it invokes `uvx --from semble[mcp]==0.5.1 semble search` against the host-supplied project root, never accepts a root from the LLM, and keeps its cache under `.isanagent/cache/semble`. Tools: `agent_list` (list available agents), `task_dashboard` (unified active+history view). The coordinator system prompt is auto-injected with agent descriptions. `wake_on_completion` (default true) enqueues a synthetic inbound when a subagent finishes so the parent can consume the result without polling. Config keys (under `[harness.subagents]`): `wake_on_completion`, `task_history_retention`. Agent definitions use `[agents.<name>]` blocks or `[harness.agents.<name>]` (harness-level wins in merge): `description` (required), `mode` (`"subagent"` or `"semble_scout"`), `system_prompt`, `system_prompt_file`, `allowed_tools`, `model`, `temperature`, `max_iterations`, `hidden`, `color`. Implementation: `src/agent/registry.rs`, `src/agent/subagent.rs`. **`subagent_plan_execute`** accepts optional per-step `"agent"` in plan JSON (same schema as spawn).
41
41
42
42
**MaxEvolve kernel porting** (`[harness.kernel_porting] enabled = true`): Triton/PyTorch→JAX/Pallas workflow via named agents (`kernel_orchestrator`, `gpu_to_jax`, `evolve_orchestrator`, …), `kernel-porting` skill, Python validators under `skills/kernel-porting/scripts/`, MAP-Elites tools (`kernel_db_init`, `kernel_db_sample`, `kernel_db_insert`, `kernel_db_status` in `src/tools/kernel_porting.rs`), and Colab/SSH hardware profiling. Operator guide: `docs/kernel-porting-user-guide.md`. Projects live under `kernels/projects/{id}/`. Onboarding copies prompts to `.agents/prompts/`, reference docs to `kernels/reference/`, and benchmarks to `benchmarks/`.
0 commit comments