Opus 4.7 Migration Scanner

The scanner is intentionally narrow. It reports three categories of deterministic patterns plus two heuristic categories that require manual review.

Category A — Fixed-budget Extended Thinking (deterministic)

Opus 4.7 does not support Extended Thinking with a fixed budget_tokens value. Code that still passes thinking={"type": "enabled", "budget_tokens": N} fails or is ignored depending on SDK version.

Patterns flagged:

• Literal budget_tokens= in Python source
• "budget_tokens" keys inside thinking={...} dict literals
• .thinking.budget_tokens attribute references

Category B — Retired model ID aliases (deterministic)

Dated or superseded Claude model aliases that no longer map to the current model. Using them is not broken but defeats the platform's model routing.

Patterns flagged:

• claude-opus-4-5, claude-opus-4-4, claude-opus-4-3, claude-opus-4-2, claude-opus-4-1, claude-opus-4-0
• claude-sonnet-4-5, claude-sonnet-4-4, claude-sonnet-4-3, claude-sonnet-4-2, claude-sonnet-4-1, claude-sonnet-4-0
• Dated format aliases: claude-*-20250514, claude-*-20241022, etc.
• claude-3-* family (superseded)

Category C — Hardcoded model references outside config (deterministic)

Per the armory convention "model refs in config files only," any claude-* literal in .py / .ts / .js source code outside of config.toml, .env, or test-only sentinel strings is a candidate for extraction.

Category D — Opus-4.6-verbosity-assuming prompts (heuristic)

Opus 4.7 is less default-verbose than 4.6. Prompts that relied on Opus 4.6's eager output or never needed a depth cue may produce shorter-than-expected responses on 4.7.

Patterns flagged (review required):

• Prompts requesting "detailed" or "comprehensive" output without matching length directive
• Prompts that exceeded Opus 4.6's defaults by relying on implicit verbosity
• Absence of first-turn specification completeness in agent prompts

Category E — Non-explicit parallel sub-agent dispatch (heuristic)

Opus 4.7 delegates more judiciously than 4.6. Agent prompts instructing "spawn in parallel" without the phrase "in a single message" or stating sub-task independence often serialize on 4.7.

Patterns flagged (review required):

• Spawn + parallel language in agent prompts without "single message" or "independent"
• subagent_type dispatches in loops or sequential blocks where parallelism was intended

Workflow

Phase 1: Run the Scanner

Execute:

python3 scripts/scan.py /path/to/repo

Optional flags:

• --categories A,B,C — run only deterministic categories (skip heuristics)
• --exclude tests/,vendor/,node_modules/ — path exclusion
• --format json — machine-readable output for CI integration
• --exit-code — exit 1 if any findings (useful for pre-commit hooks)

Phase 2: Triage Findings

Group findings by category and sort by severity. Deterministic categories (A, B, C) are always actionable. Heuristic categories (D, E) require reading the surrounding prompt context before editing.

Priority order for a migration PR:

1. Category A first — these either error out or silently do the wrong thing on 4.7
2. Category B — rename to current logical aliases (claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5) or to the platform's auto-routing name
3. Category C — extract to config.toml per repo convention
4. Category D, E — manual review; fix those that materially change behavior

Phase 3: Write the Migration PR

For each category with findings, apply the actions documented in references/migration-map.md. Prefer small PRs grouped by category over one large PR.

Pair the migration with:

• Unit tests or integration tests that exercise the updated code paths on Opus 4.7
• A short CHANGELOG entry noting "Opus 4.7 compatibility: removed budget_tokens from …"
• A rollout plan if the service is production-critical (high/xhigh effort level on the targeted agents)

Output Format

The scanner produces a categorized report:

Opus 4.7 Migration Scan — /path/to/repo

Category A: Fixed-budget Extended Thinking       2 findings
  src/agent/reasoner.py:47   budget_tokens=8000
  src/agent/planner.py:112   thinking={"type": "enabled", "budget_tokens": 4096}

Category B: Retired model ID aliases             3 findings
  tests/test_agent.py:89     claude-opus-4-5
  config/dev.yaml:14         claude-sonnet-4-5
  src/llm/client.py:23       claude-sonnet-4-20250514

Category C: Hardcoded model refs outside config  1 finding
  scripts/bulk_process.py:8  DEFAULT_MODEL = "claude-sonnet-4-6"

Category D: Verbosity-assuming prompts          0 findings (heuristics disabled)
Category E: Non-explicit parallel dispatch      0 findings (heuristics disabled)

Total: 6 deterministic findings across 3 categories.

Error Handling

Limitations

• Categories D and E are heuristic and will produce false positives on carefully written prompts. Review before changing.
• The scanner does not exercise the code — it only does static pattern matching. A budget_tokens variable that is never passed to the Anthropic SDK will still be flagged.
• The scanner cannot detect runtime dynamic model selection (e.g., model = config["models"][stage]) — those need integration tests, not static analysis.
• Cross-language analysis is limited to Python, TypeScript, JavaScript, YAML, TOML, and Markdown. Other languages are skipped.

Related

• rules/adaptive-thinking-control/RULE.md — prompt-level controls that replace fixed thinking budgets
• skills/usage-audit/SKILL.md — broader context-bloat audit (complementary)
• skills/mcp-to-skill/SKILL.md — MCP-to-skill conversion (relevant when migration surfaces heavy MCP usage)