plat 286

← Pulse platform issue index

PLAT-286 — No tool could change a step's type, so moving a message_sequence to scripted meant rebuilding it

Coordination	Value
Assigned agent	Claude Fable 5.1
Ticket state	fixed
Last synchronized	2026-09-05

• Priority: P2 — a builder-workflow gap, not a runtime failure, but one that made the platform's own recommended shape (deterministic work in a scripted step, judgment in a sequence around it) expensive to reach for an existing step, so agents either rebuilt steps by hand or left them conversational.
• Owner: agent_go/pkg/orchestrator/agents/workflow/step_based_workflow/change_step_type_tool.go (+ registration in planning_agent.go, allowlist in interactive_workshop_manager.go, IsPlanModificationTool in planning_management.go, cmd/server/toolset_invariant_test.go).
• Related: PLAT-280 (the drift this tool structurally prevents: plan type and declared mode disagreeing), PLAT-282 (same "tool surface has update but not the other operation" shape, for eval-plan deletion).

What was found

The trading workflow's agent, asked to make place-paper-trades scripted like it had just done for three other steps, reported that this one was different: those were already regular-type steps running agentic (a one-field flip); place-paper-trades is a message_sequence, and "scripted mode" is a step_config setting that exists only for the regular type. Checked on origin/main: correct. update_step_config refuses declared_execution_mode="scripted" on a message_sequence (the plan type and the mode would disagree and the scripted executor would never run it), and update_scripted_step's PLAT-280 conversion only fires for a sequence that was already declared scripted — a drift repair, not a converter. So the only path was add_scripted_step + delete_plan_steps + rewiring every context_dependencies/context_output/route by hand.

declared_execution_mode itself turned out to be the leftover of the old model where type and mode were independent: today it is implied on one type (regular is scripted by definition; without it a regular step is the legacy agentic shape that runs as a sequence) and forbidden on the other. A type change is the mode change; there was just no tool that did both.

What shipped

change_step_type(step_id, target_type: scripted | message_sequence, reason):

• Converts in place using the converters the runtime already used for its own compatibility paths (normalizeMessageSequenceStepToRegular, normalizeRegularStepToMessageSequence), so id, title, description, dependencies, outputs, validation_schema, next_step_id and position are kept, nested and orphan steps included.
• To scripted: plan type → regular, items dropped (a scripted step's work lives in learnings/<id>/main.py), step_config declares scripted with the reason as declared_execution_mode_reason, use_code_execution_mode synced; the result says whether main.py already exists or the step will fail until update_scripted_step(code=...) writes it. A legacy agentic regular step just gets the declaration (its type already matches).
• To message_sequence: plan type → message_sequence with one execute-and-verify item, declared mode and lock_code cleared, and a warning if a now-orphaned main.py remains.
• Plan and config are written as one operation, plan first; the tool is idempotent, so if the second write fails a re-run finishes the mode half — the plan type and the declared mode can never end up disagreeing, which is exactly PLAT-280's failure. Idempotent no-op when already the target (and a sequence carrying a stray scripted declaration gets it cleared).
• Full plan validation before writing; a revertable changelog entry carrying the full step JSON before and after (deleted_steps/added_steps, the same revert data delete_plan_steps records — snapshots are only hashed into before_ref/after_ref, never stored) plus per-field changes; the dependent-artifact review notice; registered in the Workshop allowlist and IsPlanModificationTool. update_step_config's rejection and update_scripted_step's description now point at this tool, and the workshop guidance says to use it rather than add + delete.

Verification

• change_step_type_tool_test.go: sequence → scripted (fields kept, items dropped, config declared with reason and synced, changelog with distinct before/after refs, the type change, and the old/new step JSON), scripted → sequence (one item, chain kept, mode and lock cleared, orphaned-script warning), legacy agentic regular → scripted (config entry created), no-op writes nothing, orphan steps convert, routing/unknown steps rejected, IsPlanModificationTool.
• PLAT-280 compat tests and TestToolSetInvariants (the guard that catches a registered-but-invisible tool) pass with it.
• Not yet done: a live conversion of a real step on a deployment; the first candidate is place-paper-trades on Dominion, which the user has not yet decided to convert (the agent's own recommendation is a split: scripted placement, judgment kept in the surrounding sequence).