---
name: goal-anchor
description: This skill should be used before spawning any agent expected to run more than 15 turns, before resuming an agent after /compact or session interruption, or before launching the next phase in a multi-phase chain where prior phases have accumulated significant context. Prepends an OBJECTIVE block to the subagent prompt to prevent gradual goal drift. Also trigger when user asks to "prevent agent drift" or "anchor the objective".
---

Inject this block at the start of the subagent prompt. Replace bracketed placeholders: OBJECTIVE from the user's original request / pipeline goal, SCOPE from the `/plan-tasks` execution plan.

```
OBJECTIVE (fixed — does not change during this session):
[Original objective — copy verbatim from the user's original request / pipeline goal]

SCOPE (fixed):
[File list and constraints — copy from execution plan]

Every 5 turns, check: "Is my current output directly addressing the OBJECTIVE above?"
If not → stop, explain why drift occurred, return to the objective.
```

**maxTurns guidance** (set in the agent definition, not here):
- Parallel debate agents (multiple reviewers running simultaneously): `maxTurns: 7` — echoing risk increases after 7 turns
- General task agents: `maxTurns: 15` to `maxTurns: 20` — goal drift risk rises with turn count
- These are two distinct thresholds for two distinct failure modes

**Compact + re-inject pattern for phase transitions.**

Instead of letting an agent run continuously across multiple phases in one session:

1. At the end of each phase, run `/compact` with the instruction: "Preserve objective and key decisions; drop implementation details."
2. Re-inject the original objective verbatim into the next phase's context.
3. Spawn the next phase with a clean context that includes the objective explicitly.

Reason: longer context causes models to pattern-match against accumulated text rather than reason from the objective, increasing drift with each turn. Re-injecting the objective after compaction resets this tendency.

**When to use this skill:**
- Any agent expected to run more than 15 turns
- Any agent resuming after a `/compact` or session interruption
- Any agent in a chain where prior phases have accumulated significant context

---

## Echo Prevention (for multi-reviewer / debate setups — issue 6.1)

When spawning multiple agents to review the same artifact in parallel, apply these constraints to prevent echoing (agents abandoning assigned roles and mirroring each other's outputs):

**Spawn rules:**
- Spawn review agents in parallel and in complete isolation — do NOT pass Agent B's output to Agent A before both have completed their independent assessment
- Set `maxTurns: 7` in review/debate agent definitions (echoing begins after 7+ turns)
- Require each agent's output to include a `DISAGREEMENT:` field stating explicitly what it disagrees with from any prior context

**Runtime echo detection:**
After receiving outputs from two or more review agents — if their conclusions appear substantially identical in both reasoning and evidence, spawn a fresh LLM check:
> "Compare these two outputs: [A] and [B]. Are they independently reasoned, or does B appear to echo A's reasoning? Answer: INDEPENDENT or ECHO with one sentence of explanation."

If result is ECHO: discard the echoing agent's output and re-spawn with a stronger role isolation prompt (agent must not see A's output at all).