Reflect

Review the conversation and persist learnings — to .flux/brain/, as new skills, to existing skill files, or as structural enforcement.

Adapted from brainmaxxing by @poteto, with skill extraction adapted from Claudeception.

Session Phase Tracking

On entry, set the session phase:

PLUGIN_ROOT="${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)}}"
[ ! -d "$PLUGIN_ROOT/scripts" ] && PLUGIN_ROOT=$(ls -td ~/.claude/plugins/cache/nairon-flux/flux/*/ 2>/dev/null | head -1)
FLUXCTL="${PLUGIN_ROOT}/scripts/fluxctl"
$FLUXCTL session-phase set reflect

On completion, reset:

$FLUXCTL session-phase set idle

External Memory Awareness

Before processing learnings, check if an external memory provider is configured:

PLUGIN_ROOT="${DROID_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)}}"
[ ! -d "$PLUGIN_ROOT/scripts" ] && PLUGIN_ROOT=$(ls -td ~/.claude/plugins/cache/nairon-flux/flux/*/ 2>/dev/null | head -1)
FLUXCTL="${PLUGIN_ROOT}/scripts/fluxctl"
EXTERNAL_MEMORY_PROVIDER=$($FLUXCTL config get externalMemory.provider --json 2>/dev/null | jq -r '.value // empty')
EXTERNAL_MEMORY_TOOL=$($FLUXCTL config get externalMemory.tool --json 2>/dev/null | jq -r '.value // empty')

If EXTERNAL_MEMORY_PROVIDER is set, two additional behaviors activate:

1. Dedup on write — before writing a learning to .flux/brain/, query the external memory MCP to check if it already exists externally. If a close match is found, skip the brain write and log: "Already in {provider} — skipping brain write for: {topic}". This prevents the same insight living in both places.

2. Personal-scope routing — when a learning is personal/cross-project (not repo-specific), offer to store it in the external provider instead of brain. Use AskUserQuestion:

{
  "questions": [{
    "question": "This learning seems personal/cross-project. Store in {EXTERNAL_MEMORY_PROVIDER} instead of brain vault?",
    "header": "Learning: {one-line summary}",
    "multiSelect": false,
    "options": [
      {"label": "Yes — store in {EXTERNAL_MEMORY_PROVIDER}", "description": "Cross-project, personal, not git-tracked"},
      {"label": "No — store in brain vault", "description": "Repo-specific, git-tracked, shared with team"},
      {"label": "Skip", "description": "Don't store this learning anywhere"}
    ]
  }]
}

Heuristic for personal vs repo-scoped: A learning is personal if it's about the user's preferences, workflow habits, tool choices, or general engineering patterns that aren't specific to this codebase. A learning is repo-scoped if it references specific files, APIs, architecture, team members, or project decisions.

If EXTERNAL_MEMORY_PROVIDER is NOT set, skip both behaviors — reflect works exactly as before.

Process

1. Read .flux/brain/index.md to understand what notes already exist
2. Scan the conversation for:
   • Mistakes made and corrections received
   • User preferences and workflow patterns
   • Codebase knowledge gained (architecture, gotchas, patterns)
   • Tool/library quirks discovered
   • Decisions made and their rationale
   • Friction in skill execution, orchestration, or delegation
   • Repeated manual steps that could be automated or encoded
   • Human review findings where the developer confirmed "we do not want this pattern in this repo"
   • Non-obvious debugging solutions (>10 min investigation, not in docs)
   • Error resolutions where the error message was misleading
   • Workarounds discovered through trial-and-error
3. Skip anything trivial or already captured in existing brain files or skills
4. Route each learning to the right destination (see Routing below)
5. Update .flux/brain/index.md if any brain files were added or removed

Routing

Not everything belongs in the brain. Route each learning to where it will have the most impact.

Decision flow

For each learning, ask in order:

1. Can this be a lint rule, script, metadata flag, or runtime check? → Encode it structurally. See .flux/brain/principles/encode-lessons-in-structure.md and docs/review-structuralization.md.
2. Is this a reusable, non-obvious solution with clear trigger conditions? → Extract as a new skill (see Skill Extraction below).
3. Is this about how an existing skill works? → Update that skill directly.
4. Is this codebase knowledge, a principle, or a gotcha? → Write to .flux/brain/.
5. Is this follow-up work? → File as a backlog item via fluxctl.

Brain files (.flux/brain/)

Codebase knowledge, principles, gotchas — anything that informs future sessions. This is the default destination. Use the flux-brain skill for writing conventions.

• One topic per file. File name = topic slug.
• Group in directories with index files using [[wikilinks]].
• No inlined content in index files.

Skill extraction (.secureskills/ preferred)

When a learning meets ALL of these criteria, extract it as a standalone skill:

Criteria:

• Reusable — will help with future tasks, not just this instance
• Non-trivial — requires discovery, not just documentation lookup
• Specific — has clear trigger conditions (error messages, symptoms, scenarios)
• Verified — the solution actually worked in this session

When to extract:

• Non-obvious debugging that took significant investigation
• Error resolution where the root cause wasn't what the error message suggested
• Workaround for a tool/framework limitation found through experimentation
• Tool integration knowledge that documentation doesn't cover well

Process:

Invoke the flux-skill-builder skill autonomously. Pass it the learning context — what was discovered, why it's reusable, and the verified solution from this session. The skill builder handles research, drafting, validation (validate_skills.py), and installation to project-local .secureskills/ through PlaTo when available, with loose .codex/skills/<name>/ mirrors only as fallback.

Do NOT manually create skill files. The skill builder encodes all best practices from docs/skills-best-practices.md and produces validated, trigger-optimized skills with proper progressive disclosure.

Don't extract when:

• The solution is a standard documentation lookup
• It's project-specific knowledge (use brain vault instead)
• It duplicates existing documentation
• The solution hasn't been verified

Review-derived structuralization

When the session includes a manual review or engineering-taste review, look specifically for developer-confirmed patterns:

• "Yes, we do not want this abstraction/pattern in this repo."
• "This should be blocked going forward."
• "This is acceptable here, but not as a general pattern."

For those, run the structuralization checkpoint from docs/review-structuralization.md:

• if the pattern is repeatable and machine-detectable, prefer a lintcn rule
• if it is important but too nuanced to lint, write a brain principle or pitfall instead
• if it is not worth encoding, do not create structural churn just because it came up in review

Default new taste-oriented lintcn rules to warn, and escalate to error only for clear anti-patterns or after the codebase has been cleaned up.

Gotchas

• Do not create a new skill when updating an existing one would cover the same trigger.
• Do not dump project-specific knowledge into a reusable skill. Route that to .flux/brain/.
• Do not restate generic engineering advice the model already knows. Capture non-obvious triggers, constraints, and failure modes.
• If the lesson can become a script, lint rule, config flag, or runtime guard, prefer that over more prose.
• Do not turn a subjective review opinion into a lintcn rule without developer confirmation that the pattern should be enforced repo-wide.

Skill improvements (skills/<skill>/)

If a learning is about how a specific Flux skill works — its process, prompts, or edge cases — update the skill directly.

Backlog items

Follow-up work that can't be done during reflection — bugs, non-trivial rewrites, tooling gaps. File as a task via fluxctl.

Summary

## Reflect Summary
- Brain: [files created/updated, one-line each]
- External memory: [learnings routed to {provider}, one-line each] (only if external memory is configured)
- Dedup skipped: [learnings already in {provider}, one-line each] (only if any were deduped)
- Skills extracted: [new skill files created, one-line each]
- Skills updated: [existing skill files modified, one-line each]
- Structural: [rules/scripts/checks added]
- Todos: [follow-up items filed]

Auto-Meditate (conditional)

After reflection is complete, check if the brain vault needs maintenance:

PITFALL_COUNT=$(find .flux/brain/pitfalls -name "*.md" 2>/dev/null | wc -l | xargs)

If PITFALL_COUNT >= 20, automatically run /flux:meditate to prune stale content and promote recurring patterns to principles. Do not ask — just run it. If meditate fails, report the error but do not fail the overall reflect. Tell the user:

Brain vault has {PITFALL_COUNT} pitfalls — running meditate to prune and promote...