Context Pattern — Problem

Pattern: Context
Component: problem.md
Version: 1.2 | Updated: 2026-07-29

Statement

Agents receive unranked dumps, stale docs, or untrusted retrieval in the system channel — then “confidently” implement against poisoned or oversized packs.

Measurable symptoms

Symptom	How you detect it
Token blowups	Packs routinely exceed budget with low signal
Instruction smuggling	Retrieved text alters tool policy
Stale truth	Answers cite deleted APIs or old ADRs
Hidden task	More context added instead of clarifying AC

Root cause

Retrieval without trust labels and budgets treats all tokens as equal. System-channel contamination turns data into unintended instructions.

What breaks when this pattern is skipped

Planner/coder thrash on contradictory sources
Security review finds tool allowlist changed by retrieved text
Cost spikes with no quality gain

Non-goals of this pattern

Not a substitute for clear acceptance criteria
Not infinite memory
Not a license to paste production secrets

Failures — recovery table
Workflow — ordered steps
Checklist — exit criteria

Changelog

1.2 (2026-07-29): Context problem page expanded with skip-impact and operator symptoms.