GH-600 gap notes
These are the six boundaries where plausible-sounding answers go wrong, plus the "choose two" traps. The consistent theme: the safe answer names a boundary, an artifact, or a gate.
1. MCP and tool boundary
MCP questions are usually permission questions wearing an integration costume.
| Scenario signal | Correct move |
|---|---|
| Agent needs external API/data | Add an MCP server only for that source |
| MCP server has many tools | Allow only required tools |
| MCP tool can write external state | Gate or disable the write tool |
| Third-party MCP server | Review trust, data exposure, auth, logging, and owner |
| Agent cannot see expected tool | Check repository settings, agent profile, allow list, and auth |
| Agent made bad choices after adding MCP | Remove noisy tools or improve routing instructions |
| External doc source is stale | Verify source-of-truth and cache behavior |
Eliminate options that:
- enable all MCP tools
- treat MCP as inherently trusted
- fix every failure by broadening permissions
- ignore auth, token scope, or write capability
- skip logging and review for external systems
Fast drill: explain why a read-only GitHub MCP tool and a ticketing-system write tool require different review.
2. Environment and execution boundaries
GH-600 can make several GitHub controls sound interchangeable. They are not.
| Control | Use it for |
|---|---|
| Branch | Isolate code mutation |
| Pull request | Review and discuss code changes |
| Ruleset/branch protection | Enforce merge conditions |
| Required checks | Block merge until validation passes |
| CODEOWNERS | Route review for owned paths |
| Actions workflow | Run automated jobs |
| Runner | Provide execution environment |
| Environment | Gate deployment and secrets |
| Secrets/variables | Store sensitive or configurable values |
| Firewall/network | Restrict reachable endpoints |
| MCP allow list | Restrict external agent tools |
Eliminate options that:
- use PR review as if it controls runtime network access
- use runner configuration as if it approves deployment
- release production secrets before approval
- let the agent mutate default branch directly
- bypass checks for agent-generated code
Fast drill: for any deployment question, say which control blocks the code, which control blocks the secret, and which artifact proves the run.
3. Memory, state, and secrets
This is the one actual miss. Lock it down.
| Need | Correct place |
|---|---|
| Stable coding preference | Scoped persistent memory |
| Temporary task details | Session context |
| Long-running progress | Issue, PR, plan, task board, session log |
| Shared multi-agent state | External store or explicit handoff artifact |
| Current source truth | Repository files, branch state, checks, logs |
| Production token | Approved secret storage |
Eliminate options that:
- store tokens, credentials, or production secrets in memory
- rely on chat history as the only handoff
- keep stale memory after architecture changes
- let memory override repository instructions
- share state informally across agents
Fast drill: if a memory answer sounds convenient, ask whether it is safe, scoped, current, inspectable, and resettable.
4. Evaluation and root cause
Do not tune before classification.
| Failure | Evidence first | Likely fix |
|---|---|---|
| Wrong API choice | Plan/session log | Instruction or context fix |
| Missing tool | Tool trace/settings | Narrow tool grant |
| Permission denied | Token/repo/env logs | Correct scope |
| Secret unavailable | Environment/fork/secret logs | Secret boundary fix |
| Runner-only failure | Runner logs | Environment fix |
| Repeated work | Handoff/session state | Durable state fix |
| Security alert | CodeQL/secret/dependency scan | Remediate code or dependency |
| Conflicting patches | Branch diffs | Isolation and aggregation fix |
Eliminate options that:
- start by disabling scans
- restart without reading artifacts
- add permissions without naming the missing permission
- call every failure a reasoning error
- call every failure a tool problem
Fast drill: classify ten failures without proposing a fix. Then add the smallest fix.
5. Multi-agent orchestration
Multi-agent is not a badge of seriousness. It is a coordination cost.
| Need | Pattern |
|---|---|
| One bounded task | Single agent |
| Ordered dependency | Sequential |
| Independent slices | Concurrent |
| Specialist routing | Handoff |
| Shared critique | Group chat |
| Broad open-ended program | Manager-led |
Eliminate options that:
- add agents without a decomposition reason
- let agents edit overlapping files
- skip branch or artifact isolation
- let handoffs loop without routing metadata
- aggregate by first response instead of evidence
- restart all work when one lane fails
Fast drill: for any multi-agent answer, say the pattern, ownership boundary, state artifact, and recovery path.
6. Guardrails and accountability
Guardrails should match risk.
| Action | Gate |
|---|---|
| Draft plan or summarize issue | Log, no heavy gate |
| Create branch and PR | Checks and PR review |
| Merge | Required reviews and required checks |
| Deploy | Environment required reviewer |
| Change secrets or permissions | Explicit human approval and audit |
| Write external production system | Narrow tool plus approval |
| Delete or disable security controls | High-risk review, rollback, audit |
Eliminate options that:
- gate every tiny action equally
- gate after the irreversible action
- rely only on agent self-report
- use broad tokens because the agent is trusted
- hide the human initiator or session log
Fast drill: list the audit chain for a high-risk agent change without looking: human, session, branch, PR, checks, review, workflow logs, audit event, rollback.
Choose-two traps
When the question says "Choose two," expect one answer to be operationally useful but unsafe.
| If one option says... | Treatment |
|---|---|
| Store token in memory | Eliminate |
| Disable scan/check | Eliminate unless the task is explicitly to test scan config in a safe sandbox |
| Give org-wide access | Eliminate unless the scenario truly needs org-wide read and has approval |
| Use durable artifact | Usually strong |
| Narrow tool list | Usually strong |
| Require approval before risky action | Usually strong |
| Classify failure from logs | Usually strong |
| Reset/prune stale memory | Usually strong |