Escalation Governance

Escalation Governance

Overview

Model escalation (haiku→sonnet→opus) trades speed/cost for reasoning capability. This trade-off must be justified.

Core principle: Escalation is for tasks that genuinely require deeper reasoning, not for "maybe a smarter model will figure it out."

The Iron Law

Copy code
NO ESCALATION WITHOUT INVESTIGATION FIRST

Verification: Run the command with --help flag to verify availability.

Escalation is never a shortcut. If you haven't understood why the current model is insufficient, escalation is premature.

When to Escalate

Legitimate escalation triggers:

When NOT to Escalate

Illegitimate escalation triggers:

Decision Framework

Before escalating, answer these questions:

1. Have I understood the problem?

• Can I articulate why the current model is insufficient?
• Have I identified what specific reasoning capability is missing?
• Is this a capability gap or a knowledge gap?

If knowledge gap: Gather more information, don't escalate.

2. Have I investigated systematically?

• Did I read error messages/outputs carefully?
• Did I check for similar solved problems?
• Did I form and test a hypothesis?

If not investigated: Complete investigation first.

3. Is escalation the right solution?

• Would a different approach work at current model level?
• Is the task inherently complex, or am I making it complex?
• Would breaking the task into smaller pieces help?

If decomposable: Break down, don't escalate.

4. Can I justify the trade-off?

• What's the cost (latency, tokens, money) of escalation?
• What's the benefit (accuracy, safety, completeness)?
• Is the benefit proportional to the cost?

If not proportional: Don't escalate.

Escalation Protocol

When escalation IS justified:

1. Document the reason - State why current model is insufficient
2. Specify the scope - What specific subtask needs higher capability?
3. Define success - How will you know the escalated task succeeded?
4. Return promptly - Drop back to efficient model after reasoning task

Reasons That Do Not Justify Escalation

Escalation needs a reason tied to the task. These common ones fail the four steps above, and the constraint behind each says why:

• "This is complex" needs the attempt that showed the current model falls short. Complexity alone does not predict model failure.
• "Better safe than sorry" names no risk. A larger model spends more without reducing a risk nobody identified.
• "I tried and failed" needs the attempt count and why it failed. A failure not yet investigated is often a context problem that a larger model would repeat.
• "The user expects quality" is met by process: tests, review and evidence. Model size does not supply them.
• "Just this once" still needs the documented reason, since an exception repeated without one becomes the default.
• "Time is money" favors investigation. Thrashing on a larger model is slower than a systematic pass on the current one.

Agent Schema

Agents can declare escalation hints in frontmatter:

yamlCopy code
model: haiku
escalation:
  to: sonnet                 # Suggested escalation target
  hints:                     # Advisory triggers (orchestrator may override)
    - security_sensitive     # Touches auth, secrets, permissions
    - ambiguous_input        # Multiple valid interpretations
    - novel_pattern          # No existing patterns apply
    - high_stakes            # Error would be costly

Verification: Run the command with --help flag to verify availability.

Key points:

• Hints are advisory, not mandatory
• Orchestrator has final authority
• Orchestrator can escalate without hints (broader context)
• Orchestrator can ignore hints (task is actually simple)

Orchestrator Authority

The orchestrator (typically Opus) makes final escalation decisions:

Can follow hints: When hint matches observed conditions Can override to escalate: When context demands it (even without hints) Can override to stay: When task is simpler than hints suggest Can escalate beyond hint: Go to opus even if hint says sonnet

The orchestrator's judgment, informed by conversation context, supersedes static hints.

Red Flags - STOP and Investigate

If you catch yourself thinking:

• "Let me try with a better model"
• "This should be simple but isn't working"
• "I've tried everything" (but haven't investigated why)
• "The smarter model will know what to do"
• "I don't understand why this isn't working"

ALL of these mean: STOP. Investigate first.

Integration with Agent Workflow

Copy code
**Verification:** Run the command with `--help` flag to verify availability.
Agent starts task at assigned model
├── Task succeeds → Complete
└── Task struggles →
    ├── Investigate systematically
    │   ├── Root cause found → Fix at current model
    │   └── Genuine capability gap → Escalate with justification
    └── Don't investigate → WRONG PATH
        └── "Maybe escalate?" → NO. Investigate first.

Verification: Run the command with --help flag to verify availability.

Quick Reference

Model Capability Notes

MCP Tool Search (Claude Code 2.1.7+): Haiku models do not support MCP tool search. If a workflow uses many MCP tools (descriptions exceeding 10% of context), those tools load upfront on haiku instead of being deferred. This can consume significant context. Consider escalating to sonnet for MCP-heavy workflows or ensure haiku agents use only native tools (Read, Write, Bash, etc.).

Claude.ai MCP Connectors (Claude Code 2.1.46+): Users with claude.ai connectors configured may have additional MCP tools auto-loaded, increasing the total tool description footprint. This makes it more likely that haiku agents will exceed the 10% tool search threshold. When escalation decisions involve MCP-heavy workflows, factor in claude.ai connector tool count via /mcp.

Effort Controls as Escalation Alternative (Opus 4.6 / Claude Code 2.1.32+): Opus 4.6 introduces adaptive thinking with effort levels (low, medium, high). The max level was removed in 2.1.72 for Opus 4.6, and high became the ceiling on that model. Claude Code 2.1.111 reintroduced max and added xhigh (between high and max) for Opus 4.7 only; on other models xhigh falls back to high. Symbols: ○ (low) ◐ (medium) ● (high) ◉ (xhigh) ★ (max). Use /effort (interactive slider since 2.1.111) or /effort auto to reset. Before escalating between models, consider whether adjusting effort on the current model would suffice:

Default effort change (2.1.68+): Opus 4.6 now defaults to medium effort for Max and Team subscribers. Use /model to change effort level, or type "ultrathink" in your prompt to enable high effort for the next turn.

Opus 4/4.1 removed (2.1.68+): Opus 4 and 4.1 are no longer available on the first-party API. Users with these models pinned are automatically migrated to Opus 4.6. No action needed for agents using model frontmatter, as the migration is transparent.

Sonnet 4.5 → 4.6 migration (2.1.69+): Sonnet 4.5 users on Pro/Max/Team Premium are automatically migrated to Sonnet 4.6. Agent model frontmatter referencing Sonnet resolves transparently. The --model flags for claude-opus-4-0 and claude-opus-4-1 now correctly resolve to Opus 4.6 instead of deprecated versions.

Effort parameter fix (2.1.70+): Fixed API 400 error This model does not support the effort parameter when using custom Bedrock inference profiles or non-standard Claude model identifiers. Effort controls now work reliably across all deployment configurations.

Default Opus 4.6 on providers (2.1.73+): Bedrock, Vertex, and Microsoft Foundry now default to Opus 4.6 (was Opus 4.1). Subagent model: opus/sonnet/haiku aliases now resolve to the current version on all providers; previously they were silently downgraded to older versions (e.g., Opus 4.1 instead of 4.6). This fix means agent dispatch workflows on third-party providers now match first-party API behavior.

modelOverrides setting (2.1.73+): Maps model picker entries to provider-specific IDs (Bedrock inference profile ARNs, Vertex version names, Foundry deployment names). Use when routing model selections to specific inference profiles. See the model optimization guide for configuration details.

/output-style deprecated (2.1.73+): Use /config instead. Output style is now fixed at session start for better prompt caching.

Full model IDs in agent frontmatter (2.1.74+): Agent model: fields now accept full model IDs (e.g., claude-opus-4-6) in addition to aliases (opus, sonnet, haiku). Previously, full IDs were silently ignored. Agents now accept the same values as --model.

Effort controls do NOT replace the escalation governance framework: they provide an additional axis. The Iron Law still applies: investigate before changing either model or effort level.

Exit Criteria

• A decision (escalate / stay) is stated with a named trigger from the "When to Escalate" or "When NOT to Escalate" tables, not a vague claim of complexity.
• If escalation is recommended, the specific target model and the subtask scope are documented before the escalation occurs.
• Any instance of escalation triggered by "maybe a smarter model will figure it out" is flagged as an Iron Law violation and blocked.
• If the task is decomposable into smaller pieces that each fit the current model, that decomposition is proposed instead of escalation.