adk-improvelisted

Improve, learn, refresh-metadata, update-defaults, self-improve, train-skill, learn-from-session. The self-improvement loop. Always interactive (asks first: improve skill defaults from decision logs, metadata via MCP introspection, or both). For defaults: runs scripts/proposal_generator.py against `$ADK_DATA_HOME/improve/learning/decisions.jsonl`, drafts proposed updates to `$ADK_CONFIG_HOME/core.yaml.defaults.*`, presents each with ≥3 evidence lines, applies on confirm; rotates decisions.jsonl to learning/archive/ after each run; appends summary to learning/summary.md. For metadata: runs scripts/metadata_introspector.py to refresh `$ADK_DATA_HOME/improve/metadata/<source>.json` from every reachable MCP. Bounded: cannot change shared/constitution.md; cannot change Must-do/Must-not-do sections of any SKILL.md (those are constitution-grade). Each proposal requires per-item confirmation regardless of mode. Never auto-applies. Min-evidence default 3 (configurable). Manual-only by design (disable-model-invocation:
sujeet-pro/agents-devkit · ★ 0 · AI & Automation · score 76

Install: claude install-skill sujeet-pro/agents-devkit

# adk-improve Read accumulated decision logs + introspect MCPs; propose updates to `core.yaml`. **Global skill** — intermediate artifacts go to `$ADK_DATA_HOME/improve/<ts>/`. Mutates `$ADK_CONFIG_HOME/core.yaml` and `$ADK_DATA_HOME/improve/metadata/<source>.json` on confirm; never touches the cwd repo. `--detailed` inspects more decision evidence before proposing defaults. `--deep` selects the stronger model profile per `shared/model-depth.md`; use it for broad default rewrites or conflicting evidence, never to bypass per-item confirmation. ## Modes (mandatory interactive choice at start) The skill ALWAYS asks first, even under `--auto`: ``` What do you want to improve? [1] skill defaults — based on accumulated decision logs since <date> [2] metadata — re-introspect all configured data sources [3] both [4] custom — specify target ``` (`--target` flag pre-selects.) ## Workflow — defaults (option 1) ``` Phase 0 — read learning state - $ADK_DATA_HOME/improve/learning/decisions.jsonl (current cycle) - $ADK_DATA_HOME/improve/learning/summary.md (history) - $ADK_CONFIG_HOME/core.yaml.defaults (current state) Phase 1 — advise - Show count of decisions since last improve run - Ask: scope (all skills / one skill / custom), min-evidence threshold (default 3) - Show summary of detected patterns (programmatic via scripts/proposal_generator.py) Phase 2 — execute (proposal review loop) - For each proposal: 1. Show: skill, fork_id, current_default, prop