@dreb/coding-agent 2.23.0 → 2.25.0

package/CHANGELOG.md

@@ -34,10 +34,14 @@
 - `/copy` now opens a multi-select message picker instead of copying only the last assistant message. All message types are listed (user, assistant, tool results, bash executions, branch/compaction summaries, custom). Navigate with arrow keys, toggle selection with Space, select-all with Ctrl+A, confirm with Enter. Copies source text without wrap-induced line breaks. Status message distinguishes between native clipboard and OSC 52 (SSH/mosh) delivery. New `app.clipboard.copyMessages` keybinding (Ctrl+Shift+C) opens the picker directly. ([#217](https://github.com/aebrer/dreb/issues/217))
 - Added TUI handling for `stream_retry` events: stale streaming components are removed, a retry spinner with attempt count is shown, and `agent_end` unconditionally clears retry state. ([#230](https://github.com/aebrer/dreb/issues/230))
 - Added `stream_retry` to the extension event surface so extensions can observe and react to provider stream retries. ([#230](https://github.com/aebrer/dreb/issues/230))
+- Added TUI handling for `length_retry` events: when a response is truncated at the model's token limit, stale streaming components are removed and a retry spinner shows "Response truncated, retrying with larger token budget" with attempt count. Reuses the same retry-state cleanup as `stream_retry`. ([#240](https://github.com/aebrer/dreb/issues/240))
+- Added `length_retry` to the extension event surface so extensions can observe and react to token-budget truncation retries. Carries `attempt`, `maxAttempts`, `previousMaxTokens`, `nextMaxTokens`, and an optional `discardedPartial`. ([#240](https://github.com/aebrer/dreb/issues/240))
 - Added `/dream` memory consolidation command — backs up all memory directories, merges duplicates, scans session history for unrecorded patterns, prunes stale entries, and validates links. Uses tar.gz archives with retention policy (keep last 10), lockfile-based concurrency protection, and a 10-step LLM pipeline with explicit backup verification. Configurable archive path via `dream.archivePath` setting. ([#99](https://github.com/aebrer/dreb/issues/99))
 
 ### Fixed
 
+- Fixed subagents silently returning empty results when a child's final response was truncated at the model's token limit. The subagent spawn handler now inspects the last assistant message's `stopReason`: a clean exit with no output now surfaces a descriptive error (truncation, loud truncation failure, or "completed with no output") instead of a silent empty string, and a clean exit with partial output now flags the truncation via `errorMessage` while preserving the partial text — covering both raw `length` truncation and the loud `error` stopReason produced when the core agent loop exhausts its length retries. The parent renders these errors even on a clean (exit code 0) exit. ([#240](https://github.com/aebrer/dreb/issues/240))
+
 - Fixed ghost whitespace appearing after the TUI content shrinks (e.g., closing the copy selector, filtering slash-command autocomplete, editor line deletion). The differential renderer now triggers a full screen redraw when content shrinks instead of clearing extra lines individually, which some terminals would not collapse. ([#217](https://github.com/aebrer/dreb/issues/217))
 
 - Added missing `ajv` direct dependency; previously relied on transitive install via `@mariozechner/pi-ai` which broke standalone installs ([#2252](https://github.com/badlogic/pi-mono/issues/2252))

package/README.md

@@ -359,6 +359,8 @@ The `subagent` tool delegates tasks to independent child agent processes. Each s
 
 **Model availability probes:** When an agent definition specifies a fallback list (comma-separated models), each model is verified with a lightweight API call via the same `streamSimple` path the agent loop uses before the subagent is spawned. The probe uses normal coding-agent thinking defaults and does not pass a synthetic `maxTokens` override, which keeps the request shape representative for reasoning models as well as non-reasoning models. Models that fail the probe (rate limit, quota exhaustion, auth failure, timeout) are skipped with a loud log line, and the next fallback is tried. If all configured models fail, the parent session's model is used as a last resort. Per-invocation model overrides and single-model configs skip probing entirely.
 
+**Per-agent model overrides:** The model used by each agent type can be overridden via the `agentModels.models` setting (a map of agent name → ordered fallback list) without copying or editing the agent definition `.md` files. Configure it in `settings.json` or via `/settings` → **Agent Models**. Resolution order: per-invocation `model` override → `agentModels` setting → agent definition `model` → parent session model. See [docs/agent-models.md](docs/agent-models.md).
+
 **Session metadata:** Each child process records its agent type in the session JSONL header (`agentType` field), providing an audit trail of which agent definition executed the work.
 
 ---

package/dist/core/agent-session.d.ts

@@ -246,6 +246,17 @@ export declare class AgentSession {
      * Used after both initial load (sdk.ts) and user-initiated reload.
      */
     warnResourceDiagnostics(resourceLoader: ResourceLoader): void;
+    /**
+     * Detect agentModels settings keys that reference agents which no longer exist
+     * (e.g. an upstream agent was renamed or removed) and surface a LOUD warning.
+     *
+     * Such keys are silently ignored during resolution — getAgentModelsForAgent
+     * never matches them — so without this check a stale override would vanish
+     * with no signal to the user. The agent-name keys must match exactly how
+     * discoverAgentTypes keys agents (case-sensitive, e.g. "Explore"), which is
+     * the same lookup getAgentModelsForAgent uses.
+     */
+    warnStaleAgentModelKeys(): void;
     /**
      * Surface a warning in the session so both the human and the AI agent can see it.
      * During streaming: steers the warning as a user message into the conversation.