fix: prevent agents from duplicating delegated subagent work

[code-yeongyu]

Problem

Users reported that Sisyphus/Hephaestus spawn subagents (explore, librarian, sis-junior) but then immediately perform the same task themselves, making subagents pointless.

User Reports (Discord #get-help)

1. seelosc: "there's no point in firing off subagents if the main agent is going to do the task themselves" — Sisyphus delegates design to sis-junior but does the design itself; fires explore agents but explores on its own then cancels them
2. moneyput: "sisyphus + opus 4.6 keeps reading the explore agents checking if it's done then it gets impatient and reads everything which bloats the context"
3. swiggityswerve: "GPT 5.4 as Atlas keeps stopping to ask about continuation at each step of a plan instead of just doing it"
4. boxer01: "got the same issue right now"

Root Cause

Agent prompts said "keep working" after firing background agents, with no clear rule against duplicating the delegated research. Atlas had no auto-continue policy, causing confirmation loops.

Changes

Anti-Duplication (Sisyphus, Hephaestus, Sis-Junior)

• New buildAntiDuplicationSection() in dynamic-agent-prompt-builder.ts with explicit rules
• FORBIDDEN: After delegating to explore/librarian, manually grep/search for the same information
• ALLOWED: Continue with non-overlapping work that does not depend on the delegated research
• Updated all "keep working" → "continue only with non-overlapping work"
• Added "Delegation Trust Rule" to explore section
• Added "Delegation Duplication" to anti-patterns list

Atlas Auto-Continue

• New <auto_continue> section in all 3 Atlas variants (default, gemini, gpt)
• NEVER ask "should I continue?" between plan steps
• Auto-continue after verification passes
• Only pause for: missing info, external dependency, critical failure
• Fixed plan file path bug (.sisyphus/tasks/ → .sisyphus/plans/)

Testing

• 3553 tests pass, 0 failures
• New test files: anti-duplication.test.ts, delegation-trust-prompt.test.ts, atlas-prompt.test.ts

---

Summary by cubic

Stops main agents from redoing work after delegating to subagents and removes Atlas “should I continue?” pauses between plan steps. Also fixes the Atlas plan path to .sisyphus/plans/.

• Bug Fixes
  • Add anti-duplication rules and examples; integrate into Sisyphus, Hephaestus, and Sisyphus‑Junior prompts; change “keep working” to “continue only with non-overlapping work”.
  • Add “Delegation Trust Rule” to the explore section and flag “Delegation Duplication” in anti‑patterns.
  • Introduce Atlas auto‑continue in all variants (default/gpt/gemini): never ask for approval; auto‑continue after verification; pause only when truly blocked.
  • Fix Atlas plan path to .sisyphus/plans/{plan-name}.md.
  • Update background_output: default full_session is now false while a task is active and true once completed to reduce context bloat.
  • Add tests for anti‑duplication, delegation trust, and Atlas prompts; suite passes.

^{Written for commit 21ddd55. Summary will update on new commits.}

[cubic-dev-ai]

2 issues found across 14 files

Confidence score: 3/5

• There is moderate merge risk: a high-confidence compatibility issue indicates plans should be written to .opencode/plans/, and keeping .sisyphus/plans/ can break Opencode-native plan tooling/permissions expectations.
• src/tools/background-task/create-background-output.ts introduces a dynamic full_session default that conflicts with the schema description, creating concrete behavior-vs-doc mismatch risk for callers and integrations.
• Pay close attention to src/agents/atlas/atlas-prompt.test.ts, src/tools/background-task/create-background-output.ts - align plan storage path with Opencode conventions and reconcile full_session runtime defaults with the declared schema.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="src/agents/atlas/atlas-prompt.test.ts">

<violation number="1" location="src/agents/atlas/atlas-prompt.test.ts:86">
P1: Custom agent: **Opencode Compatibility**

To maintain compatibility with Opencode's native plan tooling and permissions, plans should be stored in `.opencode/plans/` rather than `.sisyphus/plans/`. Opencode explicitly whitelists `.opencode/plans/*.md` for plan-related operations.</violation>
</file>

<file name="src/tools/background-task/create-background-output.ts">

<violation number="1" location="src/tools/background-task/create-background-output.ts:125">
P2: The new dynamic default for `full_session` contradicts the tool schema description.

This line changes the default behavior to `false` when the task is active. However, the tool schema description (around line 53) still states `(default: true)`. This discrepancy will mislead the LLM into expecting full session output by default, potentially causing it to incorrectly conclude the subagent hasn't produced any output when it only receives a status update. Update the schema description to document the dynamic default (e.g., `default: true if completed, false if active`).</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P1: Custom agent: Opencode Compatibility

To maintain compatibility with Opencode's native plan tooling and permissions, plans should be stored in .opencode/plans/ rather than .sisyphus/plans/. Opencode explicitly whitelists .opencode/plans/*.md for plan-related operations.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/agents/atlas/atlas-prompt.test.ts, line 86:

<comment>To maintain compatibility with Opencode's native plan tooling and permissions, plans should be stored in `.opencode/plans/` rather than `.sisyphus/plans/`. Opencode explicitly whitelists `.opencode/plans/*.md` for plan-related operations.</comment>

<file context>
@@ -0,0 +1,118 @@
+    const prompt = ATLAS_SYSTEM_PROMPT
+
+    // when / then
+    expect(prompt).toContain(".sisyphus/plans/{plan-name}.md")
+    expect(prompt).not.toContain(".sisyphus/tasks/{plan-name}.yaml")
+    expect(prompt).not.toContain(".sisyphus/tasks/")
</file context>

[cubic-dev-ai]

P2: The new dynamic default for full_session contradicts the tool schema description.

This line changes the default behavior to false when the task is active. However, the tool schema description (around line 53) still states (default: true). This discrepancy will mislead the LLM into expecting full session output by default, potentially causing it to incorrectly conclude the subagent hasn't produced any output when it only receives a status update. Update the schema description to document the dynamic default (e.g., default: true if completed, false if active).

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At src/tools/background-task/create-background-output.ts, line 125:

<comment>The new dynamic default for `full_session` contradicts the tool schema description.

This line changes the default behavior to `false` when the task is active. However, the tool schema description (around line 53) still states `(default: true)`. This discrepancy will mislead the LLM into expecting full session output by default, potentially causing it to incorrectly conclude the subagent hasn't produced any output when it only receives a status update. Update the schema description to document the dynamic default (e.g., `default: true if completed, false if active`).</comment>

<file context>
@@ -123,6 +122,10 @@ export function createBackgroundOutput(manager: BackgroundOutputManager, client:
         }
 
         const isActive = isTaskActiveStatus(resolvedTask.status)
+        const fullSessionProvided = args.full_session !== undefined
+        const fullSession = fullSessionProvided
+          ? (args.full_session ?? true)
</file context>