Appendix B - 2. Dynamic context injection

The judgment — Push the constraints into context; don't hope the actor pulls them.

The Structure of Dynamic context injection — its shape at a glance:

One slicing operator runs in two directions. Forward maps target files to the constraints that will govern them and pushes them into the brief; reverse maps a diff's line ranges to the findings it introduced and feeds a self-heal gate.

Full description → Dynamic context injection.

Intent — Map the files an agent is about to touch to the exact constraints that govern those files — lints, conventions, component boundaries, tests — and inject that subset into the agent's brief before it writes code, moving detection left of the cheapest CI gate.

Problem

The recurring failure is constraint under-specification at dispatch time. A coding agent lacks the tacit knowledge an experienced engineer has of which rules apply to a given change, so it makes plausible edits that violate them, then spends rounds discovering and repairing the violations — "pinball." A layered validation hierarchy makes it worse: context is lost between where the agent authored the change and where the failure surfaces. Across many concurrent agents, the wasted work multiplies.

Mechanism

One slicing operator runs in two directions. Forward (files → constraints): given the target files, predict which constraints will apply and pull their declarations plus fix-hints — this powers pre-briefing, either as brief-time auto-injection or an agent-discovery-time pull once the agent knows what it will touch. Reverse (diff → findings): intersect checker output with the change's diff line ranges to attribute which findings this change introduced — this powers CI self-heal, asking an agent to fix only what it broke.

Engineering consequences

The applicable rules land in context whether or not the agent would have gone looking — push, not pull. It is a change of status, relocating a rule from available to binding, because a brief is mandatory reading by construction where a doc is optional reference. Forward injection stays advisory, though: it raises salience and shifts the odds, but a downstream gate is still what guarantees the rule.

Implementation seam

A constraint-extraction tool for the forward slice and the diff-line-range attribution machinery for the reverse, plus one adapter per file-addressable registry — the lint fleet's scope tags, the component model, the banned-API list, the test corpus, the doc index. Each constraint must declare a file scope and carry an actionable fix-hint, or it cannot be selected or acted on.

Known limitations

Garbage-in: a rule with no scope tag can't be sliced, and one with no fix-hint can't be acted on — the mechanism depends on that discipline fleet-wide and does not create it. The relevance operator is itself fallible: over-injection floods the brief with noise and lowers the salience it trades on, while under-injection silently omits a governing rule. Each adapter must stay in sync with its registry, or the sliced set drifts from truth.