Skip to content

donaldfilimon/abi / .agents/skills/auto-improve

auto-improve

Verified

Run the propose-then-apply self-improvement loops in abi (`abi improve`) and abbey (`abbey learn improve`). Use when asked to "self-improve", "auto-improve", or preview/apply what abi or abbey would learn from recent activity. Both default to a dry run; applying is an explicit, gated step. Not a trainer, not LoRA, and not skill-loop.

auto-improve — preview, then apply, what abi and abbey would learn

Two local CLIs, one contract: a dry run is the default, and --apply performs only the steps its gate allows. Neither command edits code, skills, or prompts.

abi — abi improve [--apply] [--model <id>] <input>

Run from /Users/donald/sources/repos/abi. It wraps the SEA learn loop (abi-sea run_learn_loop).

  • Dry run runs the loop with persist and adapt_router off. It recalls evidence and reports task=, profile=, evidence_count=, audit_passed=, audit_vetoed=, and apply_allowed=, and it writes nothing.
  • --apply re-runs the loop with the default config only when audit_passed=true and audit_vetoed=false. Otherwise it prints skipped: and persists nothing, so a rejected completion never becomes evidence.
  • It needs a persistent WDBX store and exits 1 without one. For experiments, point ABI_WDBX_PATH at a scratch directory, never the live ~/.abi/wdbx.
./tools/cargo.sh build -p abi-cli
ABI_WDBX_PATH=target/skill-scratch/wdbx target/debug/abi improve "plan next repair"
ABI_WDBX_PATH=target/skill-scratch/wdbx target/debug/abi improve --apply "plan next repair"

abbey — abbey learn improve [n] [--apply]

Run from ~/dev/active/abbey. It reads the last n route records (default 20), the memory reflect report, and train_candidate curation counts.

  • Output rows tagged [auto] are applied by --apply. Today the only auto step is promoting recent routes into the activity layer.
  • Rows tagged [review] are for a human: routes below 0.6 confidence, duplicate or low-confidence or superseded memories, and candidates missing provenance. --apply never touches them.
  • A dry run with no memory store does not create one.

Gotchas

  • In abi, run tests and ./tools/check.sh with < /dev/null (an auth test reads stdin).
  • abi's improve is the 14th frozen command. Its help, help.json, completion files, and site/index.html chip list are golden-tested, so a change to its wording means updating those files together.
  • The abbey help text calls LoRA "Proposed". Neither loop trains weights beyond abi's adaptive router, so do not describe either as fine-tuning.
  • skill-loop's review/fix flags are unrelated and unreliable; do not use them as the improvement signal.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall