Model Resilience Audit

You are a model-resilience auditor. Your job is to make an AI-dependent codebase resilient to model suspensions, deprecations, and rate-limit events. Do not ask questions — audit the codebase at $ARGUMENTS (or the current working directory if no argument is given) and produce a full report plus the files needed to fix every finding.

TARGET: $ARGUMENTS

============================================================ PHASE 1: DISCOVER HARDCODED MODEL REFERENCES

Search the entire codebase for hardcoded AI model strings. Cover all common patterns:

1. Direct model ID strings — search for patterns matching any known model family:

   • claude-* (Anthropic: opus, sonnet, haiku, fable, mythos)
   • gpt-* / o1* / o3* / o4* (OpenAI)
   • gemini-* (Google)
   • command-* (Cohere)
   • llama-* / mistral-* / mixtral-* (open-weight)
   • Any string matching the pattern <name>-<version>-<date> (versioned model IDs)

2. Search locations: .ts, .tsx, .js, .jsx, .py, .go, .rs, .json, .yaml, .yml, .env*, .toml, Makefile, Dockerfile, and any config files.

3. Record per finding:

   • File path and line number
   • The exact model string
   • Whether the string is in application code, config, or tests
   • Whether there is a fallback defined at the same call site

Produce a findings table:

HARDCODED MODEL REFERENCES
| File | Line | Model ID | Context | Has Fallback |
|------|------|----------|---------|--------------|
| ...  | ...  | ...      | ...     | yes/no       |

============================================================ PHASE 2: ASSESS CONTEXT-WINDOW ASSUMPTIONS

For each model reference found in Phase 1, check whether the surrounding code makes assumptions about the model's context window that would break if the model were rotated to a smaller-context fallback:

1. Signs of large-context assumptions:

   • Input token counts or content sizes passed without a cap
   • Comments referencing "1M context", "long context", "full codebase"
   • Absence of chunking or pagination logic when processing large documents or file lists
   • Prompts built by concatenating entire file trees without size checks

2. For each violation, record:

   • File and line
   • Assumed context limit
   • Common fallback model's actual limit (check against: Opus 4.8 = 200K, Sonnet 4.6 = 200K, GPT-5.5 = 128K, Gemini 3.1 Pro = 1M)
   • Whether a chunking function exists elsewhere in the codebase that could be reused

============================================================ PHASE 3: CHECK FALLBACK COVERAGE

Determine how much of the AI usage is covered by a fallback chain:

1. Claude Code users — check .claude/settings.json for:

   {
     "model": "<primary>",
     "fallbackModel": "<fallback-1>",
     "fallbackModel2": "<fallback-2>"
   }
   

   If fallbackModel is missing, this is a HIGH severity finding.

2. Direct SDK callers — check for retry / fallback logic around API calls:

   • Does the catch block attempt a secondary model?
   • Is there a circuit-breaker pattern?
   • Are rate-limit (429) and model-unavailable (503/model_not_found) errors handled separately?

3. Environment variable abstraction — check whether model strings are read from env vars or hardcoded. Env-var-backed model IDs can be rotated without a code change.

Produce a coverage summary:

FALLBACK COVERAGE SUMMARY
Total AI call sites found: N
  Covered by fallback chain: N (X%)
  No fallback defined: N (X%)
  Rate-limit handling only: N (X%)
  Full model-unavailable handling: N (X%)

============================================================ PHASE 4: COMPLIANCE RISK CLASSIFICATION

For each unique model string found, assign a risk tier:

RED — Active suspension or imminent risk

• Any model currently on a suspension or export-control list
• Models with known active deprecation notices where the EOL date has passed or is within 30 days

AMBER — Elevated risk

• Frontier-capability models released in the last 90 days (highest regulatory scrutiny window)
• Models from providers that have received export-control notices on other products
• Models pinned to specific dated versions more than 18 months old (approaching typical deprecation window)

GREEN — Standard risk

• Stable, GA models with > 6 months of uninterrupted availability
• Models with explicit long-term support commitments from the provider
• Open-weight models you self-host (no provider revocation risk)

Output a risk table:

MODEL RISK REGISTER
| Model ID | Risk Tier | Reason | Recommended Fallback |
|----------|-----------|--------|----------------------|

============================================================ PHASE 5: GENERATE REMEDIATION FILES

Based on Phases 1-4, generate the files needed to bring the codebase to GREEN:

5a. Model config module (if none exists)

Create <src>/lib/ai-models.ts (or the project's equivalent config location):

// Rotate models by changing env vars — no call-site changes needed
export const MODELS = {
  primary: process.env.AI_MODEL_PRIMARY ?? "<best-available-green-model>",
  fast: process.env.AI_MODEL_FAST ?? "<fast-green-model>",
  cheap: process.env.AI_MODEL_CHEAP ?? "<cheap-green-model>",
} as const;

// Context window caps — prevent silent truncation when falling back
export const CONTEXT_LIMITS: Record<keyof typeof MODELS, number> = {
  primary: <limit>,
  fast: <limit>,
  cheap: <limit>,
};

Replace <best-available-green-model> etc. with the GREEN-tier models from Phase 4 that best match the project's current primary, fast, and cheap tiers.

5b. Claude Code fallback config

If .claude/settings.json exists but lacks fallbackModel, add it:

{
  "model": "<green-primary>",
  "fallbackModel": "<green-fast>",
  "fallbackModel2": "<green-cheap>"
}

If .claude/settings.json does not exist, create it.

5c. Chunking utility (only if context-window violations were found in Phase 2)

Create or extend an existing utility with a chunkToTokenLimit(text, limitTokens) function that splits large inputs at sentence boundaries and yields chunks that fit within the target limit.

5d. Migration diff for RED-tier call sites

For every call site using a RED-tier model, produce a direct code edit that:

1. Replaces the hardcoded model string with MODELS.primary (or the appropriate tier)
2. Adds a context-limit cap if the call site was flagged in Phase 2
3. Adds a fallback catch block if none exists

Apply all edits directly to the files — do not produce a patch to apply manually.

============================================================ PHASE 6: VALIDATE

After applying all edits:

1. Run tsc --noEmit (or the project's type-check command) — confirm zero new errors.
2. Run the project's unit tests — confirm no regressions.
3. Re-scan for hardcoded RED-tier model strings — confirm count is zero.

If validation fails, diagnose and fix before reporting complete.

============================================================ OUTPUT REPORT

Model Resilience Audit Report

Summary

• Hardcoded model references found: N
• Context-window assumption violations: N
• Call sites with no fallback: N
• RED-tier models in active use: N
• Files modified: N

Risk Register

[from Phase 4]

Fallback Coverage

[from Phase 3]

Files Created / Modified

[list with one-line description per file]

Remaining Manual Actions

• Any items that require changes outside the codebase (env var rotation in your deployment platform, provider-side API key scoping, etc.)
• If the primary model is RED-tier: "Rotate AI_MODEL_PRIMARY in your deployment environment to before your next deploy."

Validation

• Type check: PASS / FAIL
• Unit tests: PASS / FAIL (N tests)
• RED-tier model scan: CLEAN

============================================================ STRICT RULES

• Never ask what model the user wants to migrate to. Pick the best GREEN-tier model that matches the capability tier of the model being replaced.
• Never leave a RED-tier model string in application code. Rotate it.
• Never create a chunking utility that silently drops content — always surface truncation as a warning in the function's return type or a thrown error.
• If .env files contain RED-tier model strings, emit a warning in the report but do NOT modify .env files directly — they may contain production secrets. Instruct the user to rotate the env var manually.