surface voice_config in Agent.get_config and fail loudly on codegen n… by berges99 · Pull Request #113 · timbal-ai/timbal

berges99 · 2026-08-04T13:02:21Z

…o-ops

voice_config rode extra="allow" and was write-only: set-config saved it but get-flow never returned it, so Studio fell back to stale values on every refresh. Agent.get_config now emits it JSON-safely (dict passthrough, callable as "", VoiceConfig dumped).

Also fixes the write-side silent failures:

annotated assignments (agent: Agent = Agent(...)) are now transformed (AnnAssign support in cst_utils + AgentConfigSetter)
factory entry points (agent = build()) error instead of corrupting the call
unmatched shapes (agent = _agent) exit non-zero instead of phantom-saving; transformers report matched=True so idempotent saves still succeed, and idempotent ops (re-add tool/mcp, remove absent tool/step, dup edge) opt out via allow_noop

…o-ops voice_config rode extra="allow" and was write-only: set-config saved it but get-flow never returned it, so Studio fell back to stale values on every refresh. Agent.get_config now emits it JSON-safely (dict passthrough, callable as "<name>", VoiceConfig dumped). Also fixes the write-side silent failures: - annotated assignments (agent: Agent = Agent(...)) are now transformed (AnnAssign support in cst_utils + AgentConfigSetter) - factory entry points (agent = build()) error instead of corrupting the call - unmatched shapes (agent = _agent) exit non-zero instead of phantom-saving; transformers report matched=True so idempotent saves still succeed, and idempotent ops (re-add tool/mcp, remove absent tool/step, dup edge) opt out via allow_noop

cursor · 2026-08-04T13:02:27Z

PR Summary

Medium Risk
Touches all codegen transformers and Studio config round-trip; behavior changes (errors on previously silent no-ops) could break workflows that relied on wrong FQNs or aliased agents, but reduces risk of corrupt or unchanged source files.

Overview
voice_config now round-trips through Studio: Agent.get_config() exposes voice_config (dict, callable placeholder, or VoiceConfig dump) so get-flow matches what set-config writes, fixing stale voice settings on refresh.

Codegen CST handling is broader and stricter: Entry resolution and transformers now support annotated assignments (agent: Agent = …), import aliases, and module-qualified constructors; factory entry points (agent = build()) error instead of getting kwargs injected; aliased assignments (agent = _agent) fail with a non-zero exit instead of phantom success. Shared apply_operation logic treats unchanged output as failure unless the transformer set matched=True (idempotent save) or allow_noop=True (re-add tool/MCP, remove absent step/edge, etc.). Workflow ops validate the FQN variable via has_step_expr where aliases still use .step() under the expected name.

Minor: Eval discovery skips evalconf.yaml; extensive tests cover the new shapes and behaviors.

^{Reviewed by Cursor Bugbot for commit 0d128a4. Bugbot is set up for automated code reviews on this repo. Configure here.}

…lconf from eval discovery voice_config rode extra="allow" and was write-only: set-config saved it but get-flow never returned it, so Studio fell back to stale values on every refresh. Agent.get_config now emits it JSON-safely (dict passthrough, callable as "<name>" with a callable schema variant, VoiceConfig dumped). Codegen write-side hardening: - annotated assignments (agent: Agent = Agent(...)) now work for agent- and tool-level set-config (AnnAssign support in cst_utils, AgentConfigSetter, ToolConfigSetter, and the inline-tool migration insert point) - factory entry points (agent = build(), local or imported) error instead of corrupting the call; aliased (Agent as A) and module-qualified (timbal.Agent) constructors are canonicalized so they keep working - unmatched shapes (agent = _agent) exit non-zero instead of phantom-saving; transformers report matched=True so idempotent saves still succeed, and idempotent ops (re-add tool/mcp, remove absent tool/step, dup edge) opt out via allow_noop evals: evalconf.yaml matches the eval*.yaml glob, so discover_eval_files collected the shared config file as an eval suite and the CLI crashed with "Invalid eval file" before running anything. Discovery now skips it.

The unchanged-output guard only exempted transformers with allow_noop or matched=True, which remove-edge, set-param, set-position, and add-step never set — so idempotent calls (removing an already-absent edge, setting a param/position to its current value, re-adding an identical step) raised "produced no changes" instead of succeeding. Deletes/adds opt out via allow_noop; the setters track matched so unmatched shapes still fail loudly.

… and set-position collect_assignments now resolves `agent: Agent = Agent(...)`, but these three transformers still only matched plain Assign nodes when updating the constructor: - add-tool passed validation, emitted the tool variable + import, but never appended to tools=[...] — and inserted the variable after the agent - remove-tool left the tool in place and reported success via allow_noop - set-position raised "produced no changes" instead of updating metadata Add leave_AnnAssign to ToolAdder, ToolRemover, and ConstructorPositionSetter (mirroring their leave_Assign logic), and teach ToolAdder's leave_Module insertion scan to account for AnnAssign statements. Regression tests exec the generated source for each operation against the annotated shape.

…osition Workflow steps declared as `agent_a: Agent = Agent(...)` resolve through collect_assignments, so validation passed, but StepConstructorConfigSetter and StepPositionSetter only transformed plain Assign nodes — matched stayed false and the save failed with "produced no changes". Add leave_AnnAssign to both (extracting the kwarg merge into a shared helper in StepConstructorConfigSetter) and cover the annotated shape with regression tests for both operations.

cursor

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit b3a6851. Configure here.}

Systematic audit of all transformers against annotated assignments, aliased entry points, and the allow_noop/matched no-op guard: - add-mcp: MCPAdder now handles annotated agents (tools list, in-place server replacement, collision check) and inserts server variables before them - add-step: validate the entry point exists; insert new step variables before annotated/aliased entry points instead of after their .step() call (NameError); update annotated step variables in place on re-add - remove-step: validate the entry point so a wrong fqn fails loudly instead of riding allow_noop to a phantom success - cst_utils: _BareFunctionWrapper inserts the Tool wrapper before the first entry-point reference; collect_chained_step_names sees annotated workflows; new has_step_expr helper keeps alias workflows (workflow = _wf) working - remove_unused_code: clean up annotated dead variables too Regression tests for each fix; add-edge/remove-edge/set-param verified safe (Workflow type check + require_step).

Conflicts were codegen-only: main's transformer hardening (PR #113 — annotated assignments, aliased/module-qualified entry points, loud no-op failures) landed on the same files our dedupe refactor consolidated. Resolution keeps the shared-helper structure and re-expresses main's behavior through it: - insert_before_assignments grew AnnAssign anchors and an optional step_calls_of anchor — one change instead of main's five copies - assignment_resolves_to accepts AnnAssign; StepCallRewriter sets matched - allow_noop / matched / leave_AnnAssign / has_step_expr validation applied per transformer exactly as on main All 505 codegen tests (incl. main's new coverage and the full suite (3080) pass. EOF ) Co-authored-by: Cursor <cursoragent@cursor.com>