Self-Evolving Context Protocol

Basics

Activation

'POST_TASK' (default) — execute post-task protocol when task warrants it
'ALWAYS_ON' (opt-in) — monitor every interaction; use only when explicitly requested
'ON_REQUEST' — full context lifecycle audit when user asks

Paths

'SKILL_DIR' — the directory containing this SKILL.md file (resolved by the agent at runtime)
'Scripts': ${SKILL_DIR}/scripts/ — skill automation scripts
'Detailed docs': ${SKILL_DIR}/docs/protocols.md — full protocol specifications

Managed Targets

'Root README': README.md — human entry point
'Index File': One of AGENTS.md | CLAUDE.md | CODEX.md | GEMINI.md | CONTEXT.md
'Docs Index': /docs/README.md — living index of /docs
'Project Docs': /docs/' — documentation files

Purpose

Maintain a living knowledge system that captures insights, prevents repeated mistakes, and evolves project documentation through structured feedback loops.

'Target format': GitHub-Flavored Markdown. 'Sweet spot': Projects with extensive documentation. 'Core guarantee': No insight is lost. No mistake is repeated. No documentation goes stale.

Axioms

A1: Reflexive Integrity

The protocol applies its own rules to itself. Process context (index file) ≠ project docs (/docs).

A2: Constraint-Driven Evolution

Complexity is earned through discovered constraints, not invented upfront.

A3: Single Source of Truth

Every piece of knowledge lives in exactly one place. Navigate via hierarchy, never duplicate.

Three-Layer Architecture

Layer 0: README.md            ← "what is this & where to go"
Layer 1: AGENTS.md (index)    ← "how we work" (conventions, insights, mistakes)
Layer 2: /docs/README.md      ← "what docs exist"
Layer 3: /docs/'              ← "what we build"

Routing Decision Tree

About how we work?         → Layer 1 (index file)
About what we build?       → Layer 3 (/docs), then update Layer 2
About project identity?    → Layer 0 (README.md)

Connectivity Invariant

README.md → /docs/README.md → /docs/'. README.md → AGENTS.md.

Every doc reachable within 2 clicks from root. No dead ends.

Activation Modes

Mode 1: POST_TASK (Default)

'Trigger decision tree':

Task touched any *.md file? → YES → run POST_TASK
                              NO  → Changed public APIs/architecture? → YES → run POST_TASK
                                                                        NO  → skip

'Farmville guard': Before writing any update, ask: "Does this prevent a future mistake or capture a non-obvious constraint?" If no → skip. Silence is a valid output. Measuring documentation activity instead of prevented errors is the Farmville trap.

'Light post-task' (default — ≤3 files changed, no architectural impact):

EVALUATE — anything worth capturing? If no → skip
ROUTE & WRITE — update appropriate layer
SYNC — ensure /docs/README.md is current

'Full post-task' (>3 files changed OR architectural impact detected): 1–7. Full protocol steps (see protocols.md) 8. VALIDATE — bash "${SKILL_DIR}/scripts/validate-context"

Mode 2: ALWAYS_ON (Opt-in)

Only when explicitly requested by the user. Silently track during any task: surprises, wrong assumptions, recurring patterns, stale/missing/duplicated docs. Flag for post-task capture.

Cost: agent attention on every turn. Use only when task density justifies it.

Mode 3: ON_REQUEST (Explicit)

Full context lifecycle audit + validation. See protocols.md.

Pre-Task Protocol

1. READ index file — scan for relevant conventions and insights
2. READ /docs/README.md — understand documentation landscape
3. CHECK if similar task was done before → review past decisions
4. SCAN relevant /docs/' files
5. PROCEED with task execution

Entry Templates

Defaults for new files. If a file already has a format — match THAT format. Full templates: docs/templates.md

Quick Reference

'Index file entry':

- '[label]': [description] | Trigger: [cause] | Action: [what to do]

'Change History entry':

- `[Current]` [Task]. Impact: [what changed]. Insight: [lesson learned].

'Change History rotation': [Current] → [Previous] → [Legacy-0] → [Legacy-1] → [Legacy-2] → CHANGELOG.md

Tooling: `validate-context`

Automated documentation health checker. Detects index file, validates links (skipping code blocks), analyzes bloat heuristically, checks docs index coverage.

'Bootstrap': "${SKILL_DIR}/scripts/_bootstrap" installs validate-context into ~/.local/bin
'Design': docs/validation-design.md
'Usage': bash "${SKILL_DIR}/scripts/validate-context" from project working directory
'Override root': VALIDATE_CONTEXT_ROOT=/path/to/project bash ...
'Exit 0' = passed, 'Exit 1' = errors found

Lifecycle & Maintenance

'Growth control': Discovery → Active → Consolidation → Strategic Pattern → Archive
'GC triggers': 3+ bloat signals from validator → mandatory garbage collection
'Change History': Max 5 entries, rotate older to CHANGELOG.md
'Mistake escalation': Insight → Convention → Hard rule → Structural change
'Details': protocols.md

evolve-context