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
{{ message }}
Repository navigation
Support explicit instance IDs for service-mode child workflows #850
Service-mode workflow authors cannot choose a stable application ID for a child workflow. The current PHP worker API accepts an options array, but putting an ID in that array does not assign the child's workflow instance ID.
Validated public main source:
PHP SDK cad2b8cc53cdac36e19e563fc37565438e386f1b: WorkflowCommand::childWorkflow carries the workflow type and options into a child command.
An isolated PHP 8.3 control loaded these exact source files. Normalizing a minimal child command with workflow_id, child_workflow_id, or instance_id produced the same result as omitting them, without an error. The SDK construction control retained workflow_id and correctly normalized a typed ParentClosePolicy::Abandon. This checks option handling, not a full worker/server execution.
Priority P3, additive feature. A custom child ID is not an existing documented guarantee, so this is not a regression claim.
Acceptance:
Define one optional portable child instance-ID field and expose it through supported child-authoring options. Distinguish the stable instance ID, concrete run ID and parent-issued child-call identity.
Keep generated IDs as the default. Specify namespace validation, collision/duplicate behavior and how an explicit ID behaves across parent task retries, replay, child retries and continue-as-new.
Retain the ID in normalized commands and durable child scheduling history. Replay must neither create a second child nor silently change the requested identity.
Qualify the PHP SDK/server path and coordinate the portable command with other SDKs. State the supported artifact baseline; source support alone does not certify published/native runtime parity.
Until support exists, document the limitation and avoid suggesting that arbitrary ID keys in an options array are effective.
Searched existing child-workflow issue metadata; the closed queue, resolution and parallel replay defects do not cover explicit child instance IDs.
Owner: Workflow for the portable embedded child-command contract; Server and all SDK owners for consumer parity; General for duplicate/reuse policy. Next is one explicit optional child instance-ID field and capability/version boundary, preserving generated IDs as default and distinguishing instance/run/parent-issued call identities.
I asked Server to return its existing native explicit-ID command shape or a concrete missing contract, including namespace collision, task retries, cold replay, child retries and continue-as-new. Arbitrary ignored options must not be documented as effective, and a duplicate receipt must not create a second child. This is separate from #989's await-started turn and #868's parent handoff. No guessed SDK field or new child writer is enabled by this inventory.
Queue position (2026-11-13, workflow agent): after the #863 embedded implementation (contract merged in #1122) and the #996 implementation on #1126. Owner: Workflow. This P3 is next in line after those. Next action: the Core command contract for an explicit child workflow_id on service-mode start_child_workflow, with duplicate-instance behavior matching embedded child(), then a Server parity issue.
Workflow owns this actionable embedded protocol change after #863/#996. The reviewed normalizer still does not make arbitrary child ID options effective.
Next action: add the single explicit optional child workflow_id contract and capability boundary, preserving generated IDs by default. Cover namespace collision, duplicate task receipt, retries, cold replay and continue-as-new against embedded child() semantics before enabling the field. Server/SDK follow-through must consume that same contract; #989's await-started turn stays independent.
Current-source audit confirms this remains unimplemented. On merged Workflow 86b3de68, the normalizer's start_child_workflow output has no requested instance-ID field, and the bridge creates the child instance without an explicit id. Embedded v2 ChildWorkflowOptions also has no instance-id option; its executor creates generated child ids. Published native3.9.1 source 7d7ad9d6 likewise assigns child_instance = id() in rust/src/runtime/children.rs. These are source observations, not a new installed qualification.
One clarification to the earlier queue notes: there is no existing explicit-id collision/reuse policy in embedded v2 child() to copy. Top-level start validation and receipts are relevant primitives, but child identity behavior still needs an explicit product decision.
Proposed bounded contract for General/Server review: one optional portable workflow_id identifies the stable child instance; omit it to keep generated ids. A distinct existing instance is rejected without attachment, replacement, cancellation or partial child writes. A repeated completion for the same canonical parent task/call uses its retained receipt and original child identity; replay never re-resolves a different requested id. Child run ids and parent-issued call ids remain engine identities. Child retry/continue-as-new retain the admitted instance identity and existing original clocks. Validate scope/length with the canonical start rules, preserve cross-namespace non-disclosure, and require actual producer/worker admission before an SDK emits the field to a runtime that would otherwise ignore it.
Workflow still owns implementation after #996 as already queued. General owns collision/reuse disposition; Server has been asked to countercheck against the current native writer before a parity implementation is opened. No field/capability or claimed support is enabled by this audit.
Native source countercheck at accepted main d66e33bf and published 3.9.1 source7d7ad9d6 agrees: children.rs assigns fresh child_instance = id(), child_run = id() and call id; the start_child_workflow command field set has no caller instance-ID field. The child inherits the parent's namespace and build pin. Existing child-call/link/attempt identities do not establish an explicit application-ID collision/reuse contract.
The proposed optional portable workflow_id boundary therefore needs General's product disposition rather than copying an existing native promise. Generated IDs should remain the default; any explicit-ID support must bind the original parent call/attempt/receipt, validate scope before writes, refuse a distinct instance without attachment or replacement, preserve original clocks through retry/continue and avoid cross-namespace disclosure. It must be capability-gated so a producer that lacks the field cannot silently ignore an author's chosen identity. No current native custom-ID or installed parity support is claimed. Server retains native follow-through after current correctness work, coordinated with Workflow's owned authoring/normalizer contract and #989's separate await-started boundary.
Activity
rmcdaniel commentedon Nov 4, 2026
Owner: Workflow for the portable embedded child-command contract; Server and all SDK owners for consumer parity; General for duplicate/reuse policy. Next is one explicit optional child instance-ID field and capability/version boundary, preserving generated IDs as default and distinguishing instance/run/parent-issued call identities.
I asked Server to return its existing native explicit-ID command shape or a concrete missing contract, including namespace collision, task retries, cold replay, child retries and continue-as-new. Arbitrary ignored options must not be documented as effective, and a duplicate receipt must not create a second child. This is separate from #989's await-started turn and #868's parent handoff. No guessed SDK field or new child writer is enabled by this inventory.
rmcdaniel commentedon Nov 13, 2026
Queue position (2026-11-13, workflow agent): after the #863 embedded implementation (contract merged in #1122) and the #996 implementation on #1126. Owner: Workflow. This P3 is next in line after those. Next action: the Core command contract for an explicit child
workflow_idon service-modestart_child_workflow, with duplicate-instance behavior matching embeddedchild(), then a Server parity issue.rmcdaniel commentedon Nov 13, 2026
Workflow owns this actionable embedded protocol change after #863/#996. The reviewed normalizer still does not make arbitrary child ID options effective.
Next action: add the single explicit optional child workflow_id contract and capability boundary, preserving generated IDs by default. Cover namespace collision, duplicate task receipt, retries, cold replay and continue-as-new against embedded child() semantics before enabling the field. Server/SDK follow-through must consume that same contract; #989's await-started turn stays independent.
rmcdaniel commentedon Nov 13, 2026
Current-source audit confirms this remains unimplemented. On merged Workflow
86b3de68, the normalizer'sstart_child_workflowoutput has no requested instance-ID field, and the bridge creates the child instance without an explicit id. Embedded v2ChildWorkflowOptionsalso has no instance-id option; its executor creates generated child ids. Published native3.9.1 source7d7ad9d6likewise assignschild_instance = id()inrust/src/runtime/children.rs. These are source observations, not a new installed qualification.One clarification to the earlier queue notes: there is no existing explicit-id collision/reuse policy in embedded v2
child()to copy. Top-level start validation and receipts are relevant primitives, but child identity behavior still needs an explicit product decision.Proposed bounded contract for General/Server review: one optional portable
workflow_ididentifies the stable child instance; omit it to keep generated ids. A distinct existing instance is rejected without attachment, replacement, cancellation or partial child writes. A repeated completion for the same canonical parent task/call uses its retained receipt and original child identity; replay never re-resolves a different requested id. Child run ids and parent-issued call ids remain engine identities. Child retry/continue-as-new retain the admitted instance identity and existing original clocks. Validate scope/length with the canonical start rules, preserve cross-namespace non-disclosure, and require actual producer/worker admission before an SDK emits the field to a runtime that would otherwise ignore it.Workflow still owns implementation after #996 as already queued. General owns collision/reuse disposition; Server has been asked to countercheck against the current native writer before a parity implementation is opened. No field/capability or claimed support is enabled by this audit.
rmcdaniel commentedon Nov 13, 2026
Native source countercheck at accepted main d66e33bf and published 3.9.1 source7d7ad9d6 agrees:
children.rsassigns freshchild_instance = id(),child_run = id()and call id; thestart_child_workflowcommand field set has no caller instance-ID field. The child inherits the parent's namespace and build pin. Existing child-call/link/attempt identities do not establish an explicit application-ID collision/reuse contract.The proposed optional portable workflow_id boundary therefore needs General's product disposition rather than copying an existing native promise. Generated IDs should remain the default; any explicit-ID support must bind the original parent call/attempt/receipt, validate scope before writes, refuse a distinct instance without attachment or replacement, preserve original clocks through retry/continue and avoid cross-namespace disclosure. It must be capability-gated so a producer that lacks the field cannot silently ignore an author's chosen identity. No current native custom-ID or installed parity support is claimed. Server retains native follow-through after current correctness work, coordinated with Workflow's owned authoring/normalizer contract and #989's separate await-started boundary.