Scope Guard
Do what was planned. Nothing more.
Core Principle: Every line of code that wasn't in the plan is a line of code
that wasn't reviewed, wasn't tested against the spec, and wasn't approved by the
human. "While I'm here" is how technical debt and bugs are born.
The mechanical floor (opt-in)
This skill is the judgment layer. The MECHANICAL layer under it is the scope
gate: at plan approval, python3 sage/runtime/tools/manifest.py scope derive <manifest> turns the plan's per-task Files:/Output: lines into a machine
scope: block, and — with scope_gate: standard+ in .sage/config.yaml —
the sage-scope-gate.sh hook blocks any Edit/Write outside it. Two legal
exits from a block, both recorded: manifest.py scope add-collateral <path> --task TN --reason "…", or amend the plan and scope derive --refresh.
Scope changes go through the artifact, never silently through the diff.
Both knobs are OFF by default, and the 2026-08-02 campaign says leave
them there: bare and gated both held scope 3/3 at two model tiers, so the
gate prevented nothing (it is harmless and fast — arm it if you want the
floor, but do not claim it catches drift). This skill's judgment layer is
what carries scope discipline; see develop/evals/SCOPE-GUARD-CAMPAIGN.md.
When to Use
Active throughout implementation. Scope guard is a background discipline, not
a discrete step. Apply these rules continuously while working.
Rules — What's In Scope
In scope (MAY do without asking):
- Everything explicitly listed in the current task
- Infrastructure directly required by the task (imports, types, config)
- Tests for the functionality being implemented
- Constitution-mandated requirements (security, error handling) even if not in task
Rules — What's Out of Scope
MUST NOT do these without explicit human approval:
- "While I'm here" refactors: "This function is messy, let me clean it up" — NO.
If it's not in the task, don't touch it. File a note for future work.
- Unrequested features: "This would be better with caching" — NO.
If caching isn't in the spec, don't add it. YAGNI.
- Premature optimization: "This could be faster with..." — NO.
Make it work first. Optimize only when evidence shows it's needed.
- Style fixes in unchanged files: Don't reformat code you didn't write
and aren't changing. It pollutes the diff and hides the real changes.
- Dependency upgrades: Unless the task specifically requires it.
- Documentation for unrelated features: Stay focused.
SHOULD NOT do unless the constitution requires it:
- Additional error handling beyond what the spec requests. If the
constitution mandates error boundaries or input validation, it's in scope.
Otherwise, note it for future work.
Self-Check
Before committing any change, ask:
- "Is this in my current task?" — If no, don't commit it.
- "Was this in the plan?" — If no, don't commit it.
- "Did the human ask for this?" — If no, don't commit it.
- "Is this required by the constitution?" — If yes, it's in scope even if not in the task.
If the answer to all four is "no," you're scope creeping. Revert the change.
If you believe the change is truly important, note it in .sage/decisions.md
under a "Future Work" section and move on.
When Scope SHOULD Expand
Sometimes you discover mid-implementation that the plan is incomplete. This is
legitimate — but the response is NOT to silently expand scope:
- STOP implementation
- Describe what you discovered: "Task 3 requires X, which isn't in the plan"
- Ask the human: "Should I add this to the plan, or work around it?"
- If they approve, update the plan FIRST, then implement
- If they say "not now," note it as future work and proceed without it
Failure Modes
- Agent wants to "clean up" existing code: Redirect. "I noticed [X] could be
improved. I've noted it in decisions.md for future work. Continuing with current task."
- Dependency requires additional work: Legitimate. Flag it, ask human, update plan if approved.
- Constitution requires something not in plan: Legitimate. Constitution overrides
the plan. Implement the constitution requirement and note why.
1---2name: scope-guard3description: Monitors implementation scope and prevents drift beyond the plan. Detects unrequested refactors, gold-plating, and "while I'm here" additions. Use when implementing tasks from a plan, writing code for a feature, or when the agent starts modifying files not listed in the current task.4---56<!-- sage-metadata7cost-tier: haiku8activation: mandatory9tags: [discipline, scope, focus, yagni]10inputs: [spec, plan, implementation]11outputs: [scope-assessment]12-->1314# Scope Guard1516Do what was planned. Nothing more.1718**Core Principle:** Every line of code that wasn't in the plan is a line of code19that wasn't reviewed, wasn't tested against the spec, and wasn't approved by the20human. "While I'm here" is how technical debt and bugs are born.2122## The mechanical floor (opt-in)2324This skill is the judgment layer. The MECHANICAL layer under it is the scope25gate: at plan approval, `python3 sage/runtime/tools/manifest.py scope derive26<manifest>` turns the plan's per-task `Files:`/`Output:` lines into a machine27`scope:` block, and — with `scope_gate: standard+` in `.sage/config.yaml` —28the `sage-scope-gate.sh` hook blocks any Edit/Write outside it. Two legal29exits from a block, both recorded: `manifest.py scope add-collateral <path>30--task TN --reason "…"`, or amend the plan and `scope derive --refresh`.31Scope changes go through the artifact, never silently through the diff.32Both knobs are OFF by default, and the 2026-08-02 campaign says leave33them there: bare and gated both held scope 3/3 at two model tiers, so the34gate prevented nothing (it is harmless and fast — arm it if you want the35floor, but do not claim it catches drift). This skill's judgment layer is36what carries scope discipline; see develop/evals/SCOPE-GUARD-CAMPAIGN.md.3738## When to Use3940Active throughout implementation. Scope guard is a background discipline, not41a discrete step. Apply these rules continuously while working.4243## Rules — What's In Scope4445In scope (MAY do without asking):46- Everything explicitly listed in the current task47- Infrastructure directly required by the task (imports, types, config)48- Tests for the functionality being implemented49- Constitution-mandated requirements (security, error handling) even if not in task5051## Rules — What's Out of Scope5253MUST NOT do these without explicit human approval:54- **"While I'm here" refactors:** "This function is messy, let me clean it up" — NO.55 If it's not in the task, don't touch it. File a note for future work.56- **Unrequested features:** "This would be better with caching" — NO.57 If caching isn't in the spec, don't add it. YAGNI.58- **Premature optimization:** "This could be faster with..." — NO.59 Make it work first. Optimize only when evidence shows it's needed.60- **Style fixes in unchanged files:** Don't reformat code you didn't write61 and aren't changing. It pollutes the diff and hides the real changes.62- **Dependency upgrades:** Unless the task specifically requires it.63- **Documentation for unrelated features:** Stay focused.6465SHOULD NOT do unless the constitution requires it:66- **Additional error handling** beyond what the spec requests. If the67 constitution mandates error boundaries or input validation, it's in scope.68 Otherwise, note it for future work.6970## Self-Check7172Before committing any change, ask:73741. "Is this in my current task?" — If no, don't commit it.752. "Was this in the plan?" — If no, don't commit it.763. "Did the human ask for this?" — If no, don't commit it.774. "Is this required by the constitution?" — If yes, it's in scope even if not in the task.7879If the answer to all four is "no," you're scope creeping. Revert the change.80If you believe the change is truly important, note it in `.sage/decisions.md`81under a "Future Work" section and move on.8283## When Scope SHOULD Expand8485Sometimes you discover mid-implementation that the plan is incomplete. This is86legitimate — but the response is NOT to silently expand scope:87881. STOP implementation892. Describe what you discovered: "Task 3 requires X, which isn't in the plan"903. Ask the human: "Should I add this to the plan, or work around it?"914. If they approve, update the plan FIRST, then implement925. If they say "not now," note it as future work and proceed without it9394## Failure Modes9596- **Agent wants to "clean up" existing code:** Redirect. "I noticed [X] could be97 improved. I've noted it in decisions.md for future work. Continuing with current task."98- **Dependency requires additional work:** Legitimate. Flag it, ask human, update plan if approved.99- **Constitution requires something not in plan:** Legitimate. Constitution overrides100 the plan. Implement the constitution requirement and note why.
xoai