Obstacles survey

The Augmented Coding Patterns catalog lists 15 obstacles — inherent limitations of LLMs that no prompt or tool can fully eliminate. The review skill cannot solve these obstacles (they are properties of the model, not the workflow), but its architecture is designed to work around them. This survey marks each obstacle as addressed (the architecture mitigates it, with links to evidence), partial (mitigated in some cases but not structurally), or not applicable (the obstacle doesn't arise in a read-only batch review).

Obstacle	Status	How the architecture addresses it
Black Box AI — AI's reasoning is hidden; misalignment goes undetected	Addressed	The scripts-first protocol makes reasoning visible: scanner output is deterministic and inspectable, and the judgment pass must cite it verbatim. Offload Deterministic; Perfect Recall Fallacy audit
Cannot Learn — fixed weights, no memory between sessions	Addressed	settings.json is the project's persistent memory — every run re-reads it. Ground Rules; Context Management
Compliance Bias — AI says "sure thing, boss" even for nonsensical requests	Not applicable	The review is not conversational — there is no user request to comply with. Subcommand scopes are fixed (Focused Agent); the reviewer cannot be talked into narrowing scope. See Answer Injection audit
Context Rot — instructions lose influence as context grows	Addressed	Each subagent starts fresh with a small, focused context — one SKILL.md, one cache file, one report. Focused Agent; External Context; Obsess Over Rules audit
Degrades Under Complexity — multi-step tasks accumulate errors	Addressed	The three-phase protocol breaks work into small verifiable steps; each subagent has one scope. Chain of Small Steps; Sunk Cost audit
Excess Verbosity — AI overwhelms with unnecessary detail	Addressed	Reports are sliced by subcommand; SUMMARY/TODO/specialist-report split; "Do Not Overcomplicate Responses" rule. Noise Cancellation; Cognitive Overload audit
Hallucinations — AI makes up APIs, methods, or syntax	Addressed	Scanner output is deterministic — the LLM cannot report a finding the scripts didn't surface. Offload Deterministic; Perfect Recall Fallacy audit
Keeping up — AI generates faster than humans can review	Addressed	The review is the human's catch-up mechanism — it audits AI-generated code after the fact. Slice for Review; Flying Blind audit
Limited Context Window — fixed context size; everything competes for space	Addressed	Context is externalized to files (_context.md, cache, settings) and loaded on demand per subagent. External Context; Context Management
Limited Focus — attention dilutes when too much is loaded	Addressed	One subagent per scope, each loading only its own SKILL.md and cache. Focused Agent; Distracted Agent audit
Negative Bleedthrough — negation activates what you want to avoid	Not applicable	The review does not use negative instructions to suppress behavior — it uses positive scope definitions ("review all source code in src/") and deterministic scanner output. Answer Injection audit
Non-Determinism — same input yields different results across runs	Partial	Deterministic scanner output anchors the judgment, but the LLM judgment pass itself is non-deterministic. Offload Deterministic; Determinism boundary
Obedient Contractor — quick fixes over maintainability, no pushback	Addressed	The review is the pushback — it flags quick fixes and technical debt. ADR directives enforce decided tradeoffs. Decision Guards; Silent Misalignment audit
Selective Hearing — AI ignores instructions mid-task	Addressed	Rules are loaded fresh per subagent from files, not relied on from conversation history. Ground Rules; Reference Docs; Obsess Over Rules audit
Solution Fixation — AI latches onto the first plausible solution	Addressed	The reason-to-change test forces exhaustive enumeration before verdict; scanner evidence prevents declaring "found it" without verification. Constrained Tests; Unvalidated Leaps audit