donaldfilimon/abi / .agents/skills/auto-improve
auto-improve
VerifiedRun 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
persistandadapt_routeroff. It recalls evidence and reportstask=,profile=,evidence_count=,audit_passed=,audit_vetoed=, andapply_allowed=, and it writes nothing. --applyre-runs the loop with the default config only whenaudit_passed=trueandaudit_vetoed=false. Otherwise it printsskipped: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_PATHat 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[review]are for a human: routes below 0.6 confidence, duplicate or low-confidence or superseded memories, and candidates missing provenance.--applynever touches them. - A dry run with no memory store does not create one.
Gotchas
- In abi, run tests and
./tools/check.shwith< /dev/null(an auth test reads stdin). - abi's
improveis the 14th frozen command. Its help,help.json, completion files, andsite/index.htmlchip 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/fixflags are unrelated and unreliable; do not use them as the improvement signal.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| 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: 471k | Updated | Claude CodeCodex | Verified | Install |
| 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: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| 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 available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |