fclt-evolution

name	fclt-evolution
description	Turn repeated fclt writebacks into reviewed capability changes.
tags	["fclt","evolution","proposals","capability"]

fclt-evolution

When To Use

Use this skill when repeated writebacks, stale canonical assets, or a clearly missing capability should become a concrete proposal.

Do not use it for a single weak preference or speculative idea.

Workflow

1. Review signal:

fclt ai review reconcile --since <window-start> --until <window-end> --json
fclt ai writeback group --by asset
fclt ai writeback summarize --by domain
fclt ai evolve list

For an enabled scheduled loop, inspect its durable queue and observed scheduler health with fclt ai loop status --json. Use fclt ai loop run --dry-run --json for a fresh incremental scan that does not advance cursors or write reconciliation or loop state. The full queue is authoritative; the notification delta intentionally suppresses unchanged items.

2. Assess proposal readiness before mutating state:

fclt ai evolve assess --asset <selector> --json

Use the assessment recommendation as the decision checkpoint:

• reconcile_sources: run the bounded read-only source review; writebacks alone cannot prove the window is empty.
• review_reconciled_signals: review correlated dispositions and linked work without creating one proposal per ticket.
• no_mutation: do not change capability state; ask for a target or evidence.
• record_more_writeback: explain what recurrence would justify evolution and record a new writeback only if there is fresh concrete evidence.
• propose: ask before running the proposal command, then create the smallest target-specific proposal.
• review_existing_proposal: inspect or revise the existing proposal instead of creating a duplicate.

3. Propose only when evidence is strong enough:

fclt ai evolve propose

4. Draft and inspect:

fclt ai evolve draft EV-00001
fclt ai evolve review EV-00001

5. Accept/apply only when scope, target, and evidence are correct:

fclt ai evolve accept EV-00001
fclt ai evolve apply EV-00001

6. Verify the outcome after the producing loop has had a real chance to run:

fclt ai writeback link WB-00001 --issue TEAM-123
fclt ai writeback disposition WB-00001 --type task --target TEAM-123
fclt ai evolve verify EV-00001 --effectiveness improved --evidence test:post-apply

Apply is not completion. Do not resolve source writebacks until post-apply evidence shows the intended behavior improved. Treat recurrence as unchanged or regressed evidence linked to the same evolution, not as an unrelated singleton.

Proposal Kinds

• update_asset
• create_asset
• extract_snippet
• add_skill
• promote_asset

Rules

• Prefer the smallest valid proposal kind.
• Keep project-specific behavior project-scoped until reuse is proven.
• Ask for approval before applying global instructions, global skills, plugin behavior, or other broad shared surfaces.
• Reject or park proposals that are stale, duplicated, vague, or unsupported.
• Use the operator's task system for executable implementation work that needs owner, priority, or state.
• A no-op answer must still be useful: include the evidence grade, missing signal, next writeback target, and exact approval boundary.
• State the problem, source evidence, reason for the selected target, risk, expected outcome, verification plan, assumptions, and undo path before any review-producing or mutating action.
• Preview and verify the exact scope. Never overwrite authored capability silently or treat generated/rendered output as the canonical target.
• Canonical apply and cross-scope promotion are unavailable through the plugin until a transaction-safe API can return tested rollback data. Do not bypass that boundary with shell or arbitrary CLI arguments.
• The plugin exposes only closed-schema loop status and preview actions. Scheduler enable/disable/run and external tracker mutation remain outside the MCP surface.

Output

• proposals reviewed
• repeated signal
• assessment recommendation
• proposal created or updated
• approvals needed
• apply/reject/no-op rationale
• actual changed records/artifacts, verification result, and recovery route