feat: add type: set step for binding computed values into context (#221)

[jrob5756]

Closes #221.

Summary

Adds a new agent type that evaluates one or more Jinja2 expressions and binds the results into the workflow context — no LLM call, no subprocess, no I/O.

Two surface forms (mutually exclusive):

# Single binding — output is the typed scalar / list / dict.
- name: compute_target
  type: set
  value: "{{ workflow.input.org }}/{{ workflow.input.repo }}"
  # accessible as: compute_target.output

# Multi-binding — output is a dict; each key accessible as step.output.<key>.
- name: derive_flags
  type: set
  values:
    is_breaking: "{{ research.output.severity in ['high', 'critical'] }}"
    target_branch: "{{ workflow.input.branch or 'main' }}"

See examples/set-step.yaml for a runnable demo with a route that branches on a derived boolean.

Design highlights

• Type detection defaults to ruamel safe YAML loading with a JSON-safety pass: datetime/date/time become ISO 8601 strings; non-string dict keys and any other non-JSON-safe value raise ExecutionError so checkpoint round-trips never silently change a stored type. Empty / whitespace renders become "", not None. Explicit output_type: (single-value only) supports string, number, integer, boolean, list, dict.
• Multi-values: renders every binding against the original pre-step context — no left-to-right ordering — per the issue's resolution. Users who want ordered dependencies chain multiple set steps.
• Composition with parallel groups and for-each: both are supported. The validator rejects same-group sibling references inside value / values.* templates for parallel groups so a set step can never silently miss a sibling's output via the pre-group snapshot.
• Routes attached to a set step evaluate against the raw bound value: dict outputs expose {{ output.<key> }} and (via simpleeval) bare <key> access; scalars expose {{ output }}.
• Output schema validation is permitted only when the rendered output is a dict; a scalar with a declared output: schema raises ValidationError with a suggestion to switch to values:. The same validation runs in linear, parallel-group, and for-each positions (shared WorkflowEngine._run_set_step helper).
• Events: dedicated set_started / set_completed / set_failed events mirror the script-step pattern and are emitted in all three positions (main loop, parallel group, for-each iteration). The CLI verbose log, the web dashboard's synthetic replay path, and the frontend event-type union / store reducers / log + activity builders all gain coverage. A new node type set is added to NodeType and NodeData carries set_output_type, set_output_keys, set_value_repr for the detail panel. SetCompletedData.output_type and SetOutput.output_type are typed with a shared SetOutputType literal union.
• Context relaxation: WorkflowContext.store now accepts any value; _add_agent_input handles non-dict outputs (deep-copy for dicts, shallow for scalars/lists; KeyError with a "is a <type>, not a dict" message when shorthand field access targets a scalar). The two trim strategies (truncate, summarize) skip / JSON-summarise non-dict outputs and log at debug level so a user debugging "context still too big" can see why.
• Shared helper: render_set_value_repr + SET_VALUE_REPR_MAX live in set_step.py and are reused by WorkflowEngine and the web server's synthetic replay path so the dashboard preview is consistent across live and resumed runs.

Files

• Schema/validator: src/conductor/config/schema.py, src/conductor/config/validator.py
• Executor: src/conductor/executor/set_step.py (new), src/conductor/engine/context.py, src/conductor/engine/workflow.py
• Event parity: src/conductor/cli/run.py, src/conductor/web/server.py, src/conductor/web/frontend/src/types/events.ts, src/conductor/web/frontend/src/stores/workflow-store.ts, src/conductor/web/frontend/src/lib/constants.ts
• Example: examples/set-step.yaml
• Tests: tests/test_config/test_set_schema.py, tests/test_executor/test_set_step.py, tests/test_engine/test_set_workflow.py, tests/test_web/test_server.py, plus regression cases in tests/test_engine/test_context.py
• Docs: AGENTS.md mentions the new executor module

Review pass

Ran the full pr-review skill (code, tests, comments, error handling, type design) on the initial commit. The follow-up commit (fix: address PR review findings on set step) lands the substantive findings:

• Parallel/for-each set branches now emit set_* events + run schema validation (was silently skipped).
• _to_json_safe now raises on non-string dict keys (was silently str()-coerced).
• Extracted shared render_set_value_repr + SET_VALUE_REPR_MAX (was duplicated 512-char literal in workflow.py + server.py).
• Synthetic replay now propagates declared output_type (was hard-coded "auto").
• SetOutput.output_type + frontend SetCompletedData.output_type typed with SetOutputType literal (was str).
• Several misleading comments rewritten; YAML/log/error paths now observable at debug level.
• Test additions: schema-path set_failed events, parallel/for-each event parity, web synthetic replay branch, render_set_value_repr truncation boundary, end-to-end date→ISO normalisation (single + multi).

Validation

• make lint clean
• make typecheck — single pre-existing warning in dialog_evaluator.py, no new ones
• Full unit suite: 2798 passed, 3 skipped
• Targeted pytest -k "set or synthetic": 184 passed
• make validate-examples: clean
• Frontend npm run build: clean
• Smoke runs of conductor run examples/set-step.yaml with both severity branches produce the expected outputs

Acceptance criteria

• type: set accepted by the YAML schema with mutually exclusive value: / values:
• Schema rejects both being present or both being absent
• Single-value: step binds to step.output as the scalar/object
• Multi-values: step binds each key to step.output.<key>
• Standard output: schema validation applies if specified (dict outputs; enforced in main loop, parallel groups, and for-each)
• Routes work as on any other step
• Example under examples/ showing derived flags + a route that branches on one
• Tests for: scalar value, multi-value, type detection, explicit output_type, schema rejection of invalid combinations, routing on a set output