Workflow: fake-strategist-task.yaml prompt diverged from TaskSpawner after PR #556 merge

[kelos-bot]

🤖 Kelos Agent @gjkim42

Area: Task Template Maintenance

Summary

PR #556 (merged 2026-03-06) made substantive prompt changes to kelos-fake-strategist.yaml — removing the "Workflow Improvements" focus area, widening the dedup search window, and adding cross-agent boundaries. These changes were not propagated to self-development/tasks/fake-strategist-task.yaml, creating significant prompt drift that affects manual strategist runs via the run-fake-strategist.yaml workflow.

The open PR #551 (which addresses issue #544 about task template drift) only covers cosmetic changes identified before #556 merged — it does not include these newer, more impactful divergences.

Findings

1. "Workflow Improvements" focus area still present in task template (removed from TaskSpawner)

PR #556 deliberately removed focus area #2 ("Workflow Improvements") from kelos-fake-strategist.yaml to reduce inter-agent overlap — this area is handled by kelos-self-update and kelos-config-update. The TaskSpawner now has 3 focus areas.

TaskSpawner (kelos-fake-strategist.yaml, lines 39-55) — 3 focus areas:

1. New Use Cases
2. Integration Opportunities
3. New CRDs & API Extensions

Task template (tasks/fake-strategist-task.yaml, lines 34-58) — still has 4 focus areas:

1. New Use Cases
2. Workflow Improvements ← removed from TaskSpawner
3. Integration Opportunities
4. New CRDs & API Extensions

Impact: Manual strategist runs via run-fake-strategist.yaml may choose "Workflow Improvements" as their focus area, producing issues that overlap with kelos-self-update and kelos-config-update output — exactly the problem #556 was designed to prevent.

2. Dedup search window not widened in task template

PR #556 changed the dedup search from --limit 10 to --limit 50 --json number,title across all cron agents.

TaskSpawner (kelos-fake-strategist.yaml, line 70):

gh issue list --label generated-by-kelos --state open --limit 50 --json number,title

Task template (tasks/fake-strategist-task.yaml, line 71):

gh issue list --label generated-by-kelos --limit 10

Impact: Manual strategist runs only check the 10 most recent issues for duplicates, missing 40+ existing issues. With 40+ open generated-by-kelos issues, this makes duplicate creation likely.

3. Cross-agent awareness constraints missing from task template

PR #556 added explicit cross-agent boundaries to the TaskSpawner (lines 71-74):

- The following areas are handled by other agents — do not create issues about them:
  - Self-development workflow improvements, prompt tuning, config alignment → kelos-self-update
  - Documentation, CLI UX, error messages → kelos-fake-user
  - Agent configuration based on PR reviews → kelos-config-update

The task template has none of these constraints (lines 66-72). Manual runs lack any awareness of other agents' responsibilities.

4. "Exit without output" constraint missing from task template

The TaskSpawner includes (line 75):

- If after reviewing existing issues you cannot identify something genuinely new and actionable, exit without creating an issue or PR. Not every run needs to produce output

The task template only says (line 72):

- If after reviewing existing issues you cannot identify something genuinely new and actionable, exit without creating an issue or PR. Not every run needs to produce output

Wait — actually the task template does have a similar line. Let me check... The task template's version (line 72) is weaker:

- If after reviewing existing issues you cannot identify something genuinely new and actionable, exit without creating an issue or PR. Not every run needs to produce output

This line was present before #556 but is equivalent. No gap here — correction: this finding is not applicable.

Proposed Changes

Sync fake-strategist-task.yaml prompt with kelos-fake-strategist.yaml

Replace the task template's prompt section (lines 29-72) with the current TaskSpawner prompt content. Specifically:

1. Remove focus area Add badges to README #2 ("Workflow Improvements") — lines 42-46
2. Renumber remaining focus areas — current Add release workflow to build and push Docker image #3 becomes Add badges to README #2, Use cheapest model in e2e test #4 becomes Add release workflow to build and push Docker image #3
3. Update dedup search — replace --limit 10 with --state open --limit 50 --json number,title (line 71)
4. Add cross-agent awareness — add the "areas handled by other agents" constraint block
5. Add backtick fencing around gh issue create command (PR Sync task templates with TaskSpawners and document tasks/ directory #551 already proposes this)

Update PR #551 or create a new PR

Since PR #551 is still open and addresses #544, the simplest approach is to update #551 to include these additional sync changes. Alternatively, a new PR can address the #556-related drift separately.

Relationship to existing issues

Issue/PR	What it covers	What it doesn't cover
#544 (open issue)	General task template drift problem	Not aware of #556 changes
#551 (open PR, fixes #544)	Backtick fencing, YAML quoting, README docs	Does not sync #556 prompt changes (focus area removal, dedup widening, cross-agent constraints)
#553 (closed issue)	Cross-spawner dedup widening	Only covered TaskSpawner files, not task templates

This issue specifically tracks the new drift introduced by #556 that is not covered by the existing #544/#551 pair.

Evidence

• PR self-development: widen cross-spawner dedup and reduce focus area overlap #556 diff: removed "Workflow Improvements" from kelos-fake-strategist.yaml, widened dedup to --limit 50, added cross-agent constraints
• Current tasks/fake-strategist-task.yaml on main: still contains all 4 focus areas, --limit 10, no cross-agent constraints
• run-fake-strategist.yaml workflow: directly applies tasks/fake-strategist-task.yaml, meaning manual runs use the outdated prompt

/kind cleanup

[kelos-bot]

🤖 Kelos Agent @gjkim42

Kelos Triage Report

Kind: kind/cleanup
Priority: priority/backlog — Config drift causes suboptimal manual strategist runs (focus area overlap, weaker dedup) but is not blocking, user-facing, or a security concern
Actor: actor/kelos — Well-scoped config sync with clear acceptance criteria; no architectural decisions needed

Status: STILL VALID

Evidence:

• Verified self-development/tasks/fake-strategist-task.yaml on main still has all 3 divergences from self-development/kelos-fake-strategist.yaml:
  1. Focus area Add badges to README #2 "Workflow Improvements" still present in task template (lines 42-46) — removed from TaskSpawner by PR self-development: widen cross-spawner dedup and reduce focus area overlap #556
  2. Dedup search uses --limit 10 (line 71) instead of --state open --limit 50 --json number,title as in TaskSpawner (line 70)
  3. Cross-agent awareness constraints (TaskSpawner lines 71-74) are completely absent from task template
• PR self-development: widen cross-spawner dedup and reduce focus area overlap #556 was merged 2026-03-06 and introduced these changes to the TaskSpawner without propagating to the task template
• No merged PRs reference issue Workflow: fake-strategist-task.yaml prompt diverged from TaskSpawner after PR #556 merge #575; no relevant commits found on main

Duplicates found: #544 (general task template drift — predates PR #556 and does not cover the focus area removal, dedup widening, or cross-agent constraints introduced by #556). PR #551 (fixes #544) is still open and only addresses cosmetic changes (backtick fencing, YAML quoting, README docs). This issue tracks genuinely new drift that #544/#551 do not cover.

Recommendation: KEEP OPEN — The simplest resolution path would be to either expand PR #551 to include these sync changes, or create a separate PR. The acceptance criteria are straightforward: task template prompt should match TaskSpawner prompt (3 focus areas, widened dedup, cross-agent constraints).