bug-evaluator-assigned-as-actor — Workgraph live mirror

Metadata

Status	open
Created	2026-05-02T03:04:55.321665603+00:00
Tags	`bug`, `agency`, `assignment`, `evaluator`

Description

Study why normal content/writing tasks can be assigned to evaluator-style roles and then fail before any actor output exists. Produce a Markdown bug report for WorkGraph maintainers, not a code fix.

Concrete observed failure from /home/erik/poietic.life:

task mission-copy-organizational failed at 2026-05-02T02:42:05Z.
It was a writing/copy task: rewrite mission body paragraphs, produce 2-3 alternatives, deliverable notes/mission-copy-organizational.md, no file edits.
Tags included landing, mission, org-design, copy, eval-scheduled.
WorkGraph assigned agent-96 with role/default identity 31847164, Default Evaluator.
The evaluator immediately inspected for completed output, found no notes/mission-copy-organizational.md, no .wg/output/mission-copy-organizational, and no artifacts, then failed the task with score 0.
There was an assignment helper .assign-mission-copy-organizational which completed immediately and appears to have selected the evaluator.

Useful control case:

Replacement task fix-mission-copy-organizational used clearer writing tags and full/context graph creation, but was then dispatched as Default Creator with exec_mode=bare/context_scope=clean and succeeded.
Fan-in task synthesize-mission-copy-alternatives was later assigned to Default Evaluator and succeeded because it was actually a synthesis/comparison task after upstream outputs existed.

Questions to answer:

Where in the assignment/agency pipeline can Default Evaluator be selected as the primary actor for a non-evaluation user task?
Why did eval-scheduled or copy/org-design tagging make the original task look evaluator-owned?
Should evaluator/flip/assignment roles be excluded from ordinary task actor selection unless the task is an internal dot-task or explicitly evaluation/synthesis?
Should tasks with deliverables be protected from failing as evaluation before any actor output exists?
Is context packaging involved? The Poietic investigation also saw task input index.html reported missing in one context path, but that seems secondary.

Acceptance criteria:

Write a standalone Markdown report with reproduction, observed logs, likely root-cause area, proposed fix options, and regression tests to add.
Include exact commands a developer can run to inspect analogous assignment records/logs if available.
Do not edit WorkGraph source code in this task; this is a bug report only.

Study why normal content/writing tasks can be assigned to evaluator-style roles and then fail before any actor output exists. Produce a Markdown bug report for WorkGraph maintainers, not a code fix.

Concrete observed failure from /home/erik/poietic.life:
- task mission-copy-organizational failed at 2026-05-02T02:42:05Z.
- It was a writing/copy task: rewrite mission body paragraphs, produce 2-3 alternatives, deliverable notes/mission-copy-organizational.md, no file edits.
- Tags included landing, mission, org-design, copy, eval-scheduled.
- WorkGraph assigned agent-96 with role/default identity 31847164, Default Evaluator.
- The evaluator immediately inspected for completed output, found no notes/mission-copy-organizational.md, no .wg/output/mission-copy-organizational, and no artifacts, then failed the task with score 0.
- There was an assignment helper .assign-mission-copy-organizational which completed immediately and appears to have selected the evaluator.

Useful control case:
- Replacement task fix-mission-copy-organizational used clearer writing tags and full/context graph creation, but was then dispatched as Default Creator with exec_mode=bare/context_scope=clean and succeeded.
- Fan-in task synthesize-mission-copy-alternatives was later assigned to Default Evaluator and succeeded because it was actually a synthesis/comparison task after upstream outputs existed.

Questions to answer:
1. Where in the assignment/agency pipeline can Default Evaluator be selected as the primary actor for a non-evaluation user task?
2. Why did eval-scheduled or copy/org-design tagging make the original task look evaluator-owned?
3. Should evaluator/flip/assignment roles be excluded from ordinary task actor selection unless the task is an internal dot-task or explicitly evaluation/synthesis?
4. Should tasks with deliverables be protected from failing as evaluation before any actor output exists?
5. Is context packaging involved? The Poietic investigation also saw task input index.html reported missing in one context path, but that seems secondary.

Acceptance criteria:
- Write a standalone Markdown report with reproduction, observed logs, likely root-cause area, proposed fix options, and regression tests to add.
- Include exact commands a developer can run to inspect analogous assignment records/logs if available.
- Do not edit WorkGraph source code in this task; this is a bug report only.

Depends on

(none)

Required by

(none)

Log

(no log entries)