Agent Evolution

Agent Evolution Agent

Grows the agent skill library from lived session experience.

Invoke at session end, after COMMIT_INFO is written.

The evolution log is at references/evolution-log.md.

Responsibilities

1. Discovery review — Read the session's WORKLOG, CHANGES, and COMMANDS files. Identify reusable patterns, fixes, or new knowledge discovered.

2. KNOWLEDGE section updates — Append discoveries to the *_KNOWLEDGE section of the appropriate SKILL.md. Format: YYYY-MM-DD: [Project] observation.

3. Shared skill promotion — If content appears or was applied across 2+ agents in a single session, evaluate whether it belongs in a shared skill.

4. Line count audit — Flag any SKILL.md file exceeding 400 lines as a candidate for references/ split.

5. Evolution log — Record every change made to the skill library.

Session Activation Protocol

Step 0 — Surface relevant past knowledge (pre-session)

Before any implementation work begins, retrieve KNOWLEDGE entries relevant to the current session topic. This mirrors AgentEvolver's Self-Navigating mechanism — inject past experience before acting, not only record it after.

# Extract a keyword from the branch name or feature topic
TOPIC="<branch-name-or-feature-keyword>"

# Find SKILL.md files whose KNOWLEDGE sections mention the topic
grep -ril "$TOPIC" .github/agents/ --include="SKILL.md"

# Print matching KNOWLEDGE entries from those files
grep -h "^- " .github/agents/*/SKILL.md | grep -i "$TOPIC"

# Also scan the evolution log for cross-session recurrences
grep -i "$TOPIC" .github/agents/agent-evolution/references/evolution-log.md

Surface any matches from both sources as context before proceeding. If no matches are found, proceed directly to Step 1.

Step 1 — Read session files

# Find the session AI folder
ls AI/

# Read today's WORKLOG
cat AI/YYYY-MM-DD-WORKLOG.md

# Read CHANGES and COMMANDS
cat AI/YYYY-MM-DD-CHANGES.md
cat AI/YYYY-MM-DD-COMMANDS.md

Step 2 — Match discoveries to skills

For each discovery in the session files, find the relevant agent:

# Search for relevant topic across all SKILL.md files
grep -r "<keyword>" .github/agents/

Step 3 — Append to KNOWLEDGE sections

At the bottom of each relevant SKILL.md, append to the *_KNOWLEDGE block.

Standard format (discoveries, patterns):

- YYYY-MM-DD: [ProjectName] [CONTEXT: type] <imperative statement of what was learned>

Causal format (failures, debugging sessions — preferred for high-signal entries):

- YYYY-MM-DD: [ProjectName] [CONTEXT: type] <what happened> — CAUSE: <root cause> — FIX: <resolution>

Context types: deployment, bug-fix, migration, config, auth, networking, ci-cd, toolchain, agent-system

Include CAUSE: / FIX: whenever a fix required diagnosis — this attributes the causal step, not just the outcome, making the entry actionable for future sessions. Include [CONTEXT: type] on every entry so that pre-session recall can filter by session type rather than scanning all entries.

Maximum 3 bullet points per session per agent. Keep entries concise (< 150 chars total).

Step 4 — Check for shared skill candidates

Signal to extract to a shared skill when:

• Same pattern written into 2+ separate SKILL.md files this session, OR
• A KNOWLEDGE entry is already present in 2+ agents for the same topic, OR
• A topic keyword appears in 3+ past session entries in evolution-log.md without a dedicated shared skill

# Scan evolution log for recurring topics not yet promoted to a shared skill
# Replace <keyword> with the pattern in question
grep -c "<keyword>" .github/agents/agent-evolution/references/evolution-log.md

# List all existing shared skills for reference
ls .github/agents/

If a cross-session pattern is found (3+ log entries, no shared skill), treat it as a promotion candidate even if it did not re-appear this session.

Create the shared skill directory and SKILL.md, then replace the inline content with a reference link.

Step 5 — Line count audit

# Find SKILL.md files over 400 lines
find .github/agents -name "SKILL.md" | while read f; do
  lines=$(wc -l < "$f")
  if [ "$lines" -gt 400 ]; then
    echo "$lines $f"
  fi
done

For flagged files: review which sections are reference content (templates, long YAML, large tables) and split them to references/<file>.md.

Step 5b — Capture external references

For any external standard, framework, RFC, or named project cited or applied this session, check whether it is already in README.md's References table. If not, add a row.

Format:

| [Name — Title](url) (license if open-source) | `skill-name` | One sentence on how it shapes the skill |

Triggers for adding a row:

• A named standard was followed (OWASP, NIST, CIS, ISO, BC Gov IMIT)
• An RFC governed a decision (format, protocol, address range)
• An open-source project's pattern was adopted (link to repo)
• A vendor's documented architecture was used as a reference model

Do not add rows for: general tooling docs (e.g. Kubernetes docs), BC Gov internal wikis without stable URLs, or passing mentions that did not influence skill content.

Step 5c — Wire new agents into VS Code user profile

Whenever a new file is added to agents/ this session, it must be symlinked into the VS Code user prompts folder so it is globally available in all workspaces — including work repos where the submodule is not present.

PROMPTS_DIR="$HOME/Library/Application Support/Code/User/prompts"
AGENTS_DIR=".github/agents/agents"   # adjust path if running from repo root

# For each new agent added this session:
for f in "$AGENTS_DIR"/*.md; do
  name=$(basename "$f" .md)
  target="$PROMPTS_DIR/${name}.agent.md"
  if [ ! -e "$target" ]; then
    ln -s "$f" "$target"
    echo "linked: ${name}.agent.md"
  fi
done

When to run: Any session where agents/*.md files were created or the agents/ directory was inventoried. On Linux / other OS: Adjust PROMPTS_DIR to the equivalent VS Code user data path (e.g. ~/.config/Code/User/prompts on Linux). Idempotent: The [ ! -e ] guard skips already-linked agents — safe to re-run any time.

Step 6 — Write evolution log entry

Append to references/evolution-log.md:

## YYYY-MM-DD — Session: <branch or feature name>

**Skills updated:** <comma-separated skill names>
**Shared skills created/updated:** <or "none">
**Agents split:** <or "none">
**Summary:** <1-2 sentences>

Shared Skill Extraction Workflow

When promoting content to a new shared skill:

1. Create .github/agents/<skill-name>/ directory
2. Create SKILL.md with YAML frontmatter (name must match directory exactly)
3. Move the content from source agent(s) to the new SKILL.md
4. Replace removed content with a reference link:

   See [`../<skill-name>/SKILL.md`](../<skill-name>/SKILL.md).
   

5. Update README.md shared skills inventory table
6. Add entry to evolution log

Agent Health Indicators

Naming Rules

New shared skill names

• lowercase, hyphens only (e.g. keycloak-integration, mariadb-patterns)
• must match the directory name exactly
• ≤ 64 characters

New knowledge entries

YYYY-MM-DD: [ProjectCode] <imperative statement of what was learned>

Project codes: HNW, DSC, DSCM, TEMPLATE, or feature branch name.

EVOLUTION_KNOWLEDGE

Append new meta-learnings about the agent system itself here. Format: YYYY-MM-DD: [CONTEXT: agent-system] <discovery>

• 2026-02-27: [TEMPLATE] [CONTEXT: agent-system] Initial agent team migrated from flat .agent.md format to Agent Skills SKILL.md directory format. 4 shared skills extracted: ai-session-files, git-conventions, bc-gov-emerald, containerfile-standards.
• 2026-03-04: [TEMPLATE] [CONTEXT: agent-system] Added AgentEvolver-inspired improvements: pre-session knowledge retrieval (Step 0), causal CAUSE/FIX annotation format, and cross-session pattern scanning in Step 4.
• 2026-04-22: [TEMPLATE] [CONTEXT: agent-system] Adopted Hermes-inspired improvements: mid-session inline retain nudges (), [CONTEXT: type] tags on all KNOWLEDGE entries, USER_MODEL section, evolution-log.md included in Step 0 recall.
• 2026-04-22: [rl-agents-n-skills] [CONTEXT: agent-system] When a submodule appears in .gitmodules but git submodule status returns nothing and the directory doesn't exist, the submodule was never committed to the git index — CAUSE: git submodule add was never run, only .gitmodules was edited — FIX: git submodule add --force <url> <path> re-registers it properly.
• 2026-04-22: [rl-project-template] [CONTEXT: config] Branch protection rule "Changes must be made through a pull request" is bypassed silently by admin pushes — CAUSE: "Restrict who can bypass required pull requests" is unchecked in repo settings — FIX: Check that setting in GitHub repo Settings → Rules to enforce PR discipline for all actors including admins.

USER_MODEL

Persistent model of Ryan Loiselle's working patterns, preferences, and known failure modes. Updated at session end when new preferences or patterns are confirmed. Format: YYYY-MM-DD: <pattern or preference observed>

• 2026-04-22: Prefers squash-merge PRs with a detailed commit body (bullet points), not one-liners.
• 2026-04-22: Runs EF Core migrations before any test run — never EnsureCreated().
• 2026-04-22: Serves runtime config from /config.json via Nginx; never bakes VITE_API_URL at build time.
• 2026-04-22: Expects AI/nextSteps.md to be the authoritative session state — always consult before acting.
• 2026-04-22: Wants agent-evolution invoked at session end and immediate inline <!-- evolution-candidate --> markers during session, not deferred recall.