benchmark_report.md now 7/7 direct + 4/4 transfer, all 0->1.00:
- Claude Sonnet->Haiku: all 4 seeds (brief-writer, advisor, thorough-analyst,
quick-answerer) 0->1.00
- Codex self-optimized: brief-writer, advisor, quick-answerer 0->1.00
- quick-answerer uses the real ./search tool loop on both runtimes.
This matches gbrain's own "4/4 skills 0->1.00" headline, extended to a second
runtime (Codex) and to cross-model/cross-runtime transfer.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
quick-answerer (judge: tool_called=search) reaches 0.00 -> 1.00 with Sonnet
optimizer -> Haiku target: the optimizer wrote an OVERRIDE of the "never use
tools" instruction and the Haiku target genuinely invoked the ./search shim.
All 4 gbrain skillopt-v1 seeds now at 0->1.00, matching gbrain's own headline.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
The 4th gbrain seed (quick-answerer) is judged by tool_called=search: the agent
must ACTUALLY call a search tool. Add an honest tool loop:
- Backend.attempt_with_tools(task, skill, memory, tools) -> (response, tools_called)
- Claude: exposes a real ./search shell shim, runs with --allowedTools Bash in a
clean cwd; detects the call from the shim's log (not a self-reported marker).
- Codex: same shim under `exec --sandbox workspace-write`.
- Mock: deterministic — "calls" a tool iff skill/memory instructs it (for CI).
- replay_one routes tasks with a tool_called check through the tool loop and
feeds detected calls to the rule judge; ReplayResult gains tools_called.
Verified live (Claude haiku): deficient skill -> tools_called=[] hard=0;
learned "must run ./search" rule -> tools_called=['search'] hard=1.0.
20 tests pass.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Machine-generated benchmark_report.md from a 9-config sweep:
- Direct (Sonnet->Haiku): brief-writer/advisor/thorough-analyst 0->1.00
- Direct (Codex): brief-writer/advisor 0->1.00
- Transfer (4/4 positive, incl. cross-runtime Codex<->Claude): all 0->1.00
Cross-model transfer confirms the price-difference value prop: a skill
optimized on a cheap model deploys for free on an expensive one, and skills
move between Codex and Claude. sweep.jsonl is the committed source data.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Strong-optimizer/weak-target (Sonnet -> Haiku), fully isolated:
brief-writer, advisor, thorough-analyst all 0.00 -> 1.00 on held-out.
thorough-analyst shows 2-night convergence (0.33 -> 1.00). Codex self-optimized
brief-writer also 0 -> 1.00.
Key finding answering the optimizer/target-split request: the OPTIMIZER MODEL is
decisive — weak Haiku-as-optimizer is flaky (0 or 1.0 across runs), strong
Sonnet-as-optimizer reliably hits 1.0 on every seed. Raw logs under docs/sleep/raw/.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
The default sweep direct plan now uses a DualBackend (Sonnet optimizer proposes
edits, Haiku target runs tasks) — the SkillOpt-faithful and more reliable setup,
since a weak self-optimizing model (Haiku-as-optimizer) produced flaky JSON.
report.py renders the optimizer->target pairing in the direct table.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
- reflect() now retries once with a firmer "JSON only" instruction when the
first reply doesn't parse to a non-empty array. A transient non-JSON reply
otherwise wastes a whole night (gate sees no edits -> reject), which made
weak optimizers (Haiku) flaky across runs.
- FINAL_REPORT.md: document the context-leak discovery honestly; Codex cells
stand (clean), Claude cells recomputed under strict isolation.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
The clean-cwd + --disallowedTools isolation was NOT enough: the user's GLOBAL
skills (~/.claude/skills) are injected regardless of cwd, so reflect/attempt
still sometimes replied with a list of installed skills instead of JSON edits
(advisor reflect returned 21KB of skill descriptions, n_edits=0 -> gate reject).
Add --bare (skip hooks/LSP/plugins) and --disable-slash-commands (disable all
skills). Verified: the optimizer now returns clean JSON. Re-validating all
seeds with the truly-isolated backend; prior Claude numbers are being recomputed
honestly (some earlier "successes" were partly leak-assisted).
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Critical correctness fix found by debugging the thorough-analyst failure:
* `claude -p` was running with the AMBIENT Claude Code project context (the
repo's CLAUDE.md, installed skills, tools). The optimizer/target calls were
polluted — reflect once replied with a list of the user's installed skills
instead of JSON edits. Now ClaudeCliBackend._call runs ISOLATED: a clean temp
cwd, --disallowedTools '*', --exclude-dynamic-system-prompt-sections. This is
essential for the backend to be trustworthy and reproducible.
* reflect prompt: translate failing rule-judge criteria into plain English
(max_chars=1200 -> "the ENTIRE response must be at most 1200 characters") and
require CONCRETE, verbatim thresholds in proposed rules (not "respect limits").
* attempt prompt: treat the Learned-preferences block as HARD CONSTRAINTS that
override earlier conflicting skill text.
Earlier Claude results predate this fix and are being re-validated clean; the
Codex backend was never affected (it runs in its own exec context).
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
- skillopt-sleep-plugin/.claude-plugin/marketplace.json so the plugin is
installable via `/plugin marketplace add ./skillopt-sleep-plugin`.
- README install section (clone -> add marketplace -> install -> /sleep status).
- docs/sleep/FINAL_REPORT.md: the consolidated presented results doc (real
Claude+Codex, transfer, and the honest thorough-analyst failure + fix).
- sweep.py flushes stdout for live monitoring.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
- sweep.py: run many (backend, model, seed, transfer-pair) configs sequentially,
append each result to JSONL incrementally (resumable, interrupt-safe).
- report.py: render the sweep JSONL into a presented Markdown scorecard with
direct-improvement and cross-model-transfer tables.
- reflect prompt now tells the optimizer its edits are APPENDED (can't delete the
base skill text), so on a conflict it must write a forceful OVERRIDE rule.
Diagnosed from a real failure: thorough-analyst (needs <=1200 chars) kept its
edits rejected because the base "be exhaustive" line won; a verified override
("HARD LIMIT ... supersedes") makes Haiku obey (1194/880 chars -> hard=1.0).
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Three additions driven by the goal of price-aware, model-flexible sleep:
1. DualBackend + build_backend(): route attempt->TARGET model and
reflect/judge->OPTIMIZER model (SkillOpt's target-vs-optimizer split).
gbrain runner gains --optimizer-backend/-model + --target-backend/-model.
2. run_transfer.py: sleep-scenario cross-model transfer. Optimize a skill on a
SOURCE model (e.g. cheap haiku), freeze it, evaluate held-out on a TARGET
model (e.g. expensive sonnet) with no further optimization — plus a direct
reference. Mirrors the SkillOpt paper's transfer table; quantifies the
"optimize cheap overnight, deploy anywhere" value prop.
3. llm_miner.py: turn real harvested transcripts into TaskRecords WITH checkable
rule/rubric judges, wired into the cycle for non-mock backends, so real-data
lift becomes measurable (heuristic miner remains the no-API fallback).
Fixed a str.format brace bug the new unit test caught.
19 tests pass.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Codex with the directive reflect prompt + 2 nights converges 0.00 -> 1.00
(up from 0.67 single-night); its night-2 edit diagnoses its own residual
failure ("preserve required sections even when keeping the brief short").
Claude (Haiku) reaches 1.00 in one night. Update plugin README + skill to
reference --backend claude|codex (was anthropic) and surface the benchmark.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Upgrade from mock-only to REAL multi-backend validation:
Backends (skillopt/sleep/backend.py):
- CliBackend base: shared attempt/judge/reflect prompts, response cache,
token accounting. Subclasses implement only _call().
- ClaudeCliBackend: drives `claude -p --output-format text`.
- CodexCliBackend: drives the REAL @openai/codex `exec -o <file>` for clean
output; resolve_codex_path() skips the hermes wrapper at ~/.local/bin/codex.
- reflect() now aggregates the exact failing judge criteria into the prompt
(gbrain's lesson: tell the optimizer what the scorer rewards).
Rule judges (skillopt/sleep/judges.py): gbrain-compatible local scorers
(section_present / regex / max_chars / contains / tool_called) — held-out
scoring with no judge-API spend. TaskRecord gains a `judge` field +
reference_kind="rule".
gbrain-evals adapter (experiments/gbrain_bench.py, run_gbrain.py): load
garrytan/gbrain-evals skillopt-v1 deficient skills + train/held-out task
sets and run our consolidate() loop against the SAME suite gbrain scores.
REAL results (docs/sleep/real_api_results.md), brief-writer seed, 1 night:
- Claude (Haiku): held-out 0.00 -> 1.00
- Codex: held-out 0.00 -> 0.67
Both proposed a correct, general format rule into the protected LEARNED block.
CLI: --backend {mock,claude,codex}, --codex-path, --model; experiment +
gbrain runners gain --limit-* cost controls. 17 tests pass.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Design for a nightly offline self-evolution plugin that synthesizes
SkillOpt (validation-gated bounded text optimizer), Claude Dreams
(offline memory consolidation), and the Agent-Sleep paper (short-term
to long-term experience). Harvests local ~/.claude transcripts, mines
recurring tasks, replays them offline, and consolidates memory+skills
behind a held-out gate.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Two bugs made local vLLM targets score acc=0.000: the router did not
forward 'timeout' to the Qwen backend (so runs used the 300s default),
and qwen_backend always injected chat_template_kwargs.enable_thinking,
which non-Qwen vLLM servers reject or answer with <think> output and no
<answer> tag. Forward timeout and only set the field when enabled.
Closes#28
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The shipped env_template.py and loader_template.py described the same
fictional async execute / evaluate / build_prompt API documented in
docs/reference/api.md. As a result TemplateBenchmarkEnv(cfg) raised
'TypeError: Can't instantiate abstract class' for every copy-and-paste
user who followed the in-tree scaffold.
Rewrite the template so it's a working starting point:
- env_template.py: TemplateBenchmarkEnv(EnvAdapter) now implements all
five real abstract methods (build_train_env, build_eval_env, rollout,
reflect, get_task_types) with no-op defaults documented as TODO.
Instantiable today; pytest 60/60 still passes.
- loader_template.py: TemplateBenchmarkLoader(SplitDataLoader)
implements load_split_items for .json / .jsonl input and explains the
optional load_raw_items override for split_mode="ratio".
- README.md: usage steps now point at scripts/train.py's _ENV_REGISTRY
(the real registry) instead of a non-existent BENCHMARK_REGISTRY in
skillopt/envs/__init__.py, and link to the rewritten new-benchmark
guide.
- config_template.yaml: _base_ is a string path (not a list, which the
loader rejects); skill_init is commented out with a note so the
template config doesn't reference a file the user hasn't created.
Verified locally: 'from skillopt.envs._template.env_template import
TemplateBenchmarkEnv; TemplateBenchmarkEnv()' succeeds. Refs
microsoft/SkillOpt#30.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
docs/reference/api.md previously documented a fictional EnvAdapter API
(execute / evaluate / build_prompt + DataItem / TaskResult) and a
BENCHMARK_REGISTRY that never existed in code. Anyone following the
documented contract would hit ImportError or TypeError on the first
instantiation.
Replace both pages with the real shape from skillopt/envs/base.py and
skillopt/datasets/base.py:
- EnvAdapter: build_train_env, build_eval_env, rollout, reflect,
get_task_types (the 5 actual abstract methods).
- Rollout dicts: id / hard / soft required; everything else preserved
into RolloutResult.extras.
- Reflect dicts: {patch, source_type} schema as consumed by
run_minibatch_reflect.
- BatchSpec: slotted-but-mutable dataclass matching the actual
definition (payload defaults to None, metadata to dict()).
- SplitDataLoader.load_split_items as the one mandatory loader method.
- Registry: _ENV_REGISTRY in scripts/train.py (lazy try/except
ImportError block), not a non-existent BENCHMARK_REGISTRY in
skillopt/envs/__init__.py.
- _base_: documented as a string path, since the current YAML loader
only accepts strings.
The new-benchmark.md guide now walks through a docfaithful worked
example with a real rollout helper (chat_target + scorer) instead of
hand-waving over the rollout step. Refs microsoft/SkillOpt#30.
Co-Authored-By: Claude Opus 4 <noreply@anthropic.com>
Add initial test infrastructure covering:
- skillopt/utils/scoring.py (compute_score, skill_hash)
- skillopt/utils/json_utils.py (extract_json, extract_json_array)
- skillopt/types.py (Edit, Patch dataclass serialization)
All tested functions are pure/deterministic with no LLM dependencies.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add docs/guideline.html, a single self-contained documentation guide
(left-nav + content + on-this-page TOC) covering installation, data
preparation, training/eval, full configuration reference, framework
internals, and an API reference. Link it from the README with local,
htmlpreview, and GitHub Pages access instructions.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
PR #26 added a MiniMax chat backend but left three loose ends that
silently dropped any YAML / CLI configuration of minimax_* keys: only
the environment-variable path worked.
- skillopt/config.py: add 6 model.minimax_* entries to _FLATTEN_MAP so
the keys declared in configs/_base_/default.yaml actually survive
flatten_config() (mirroring the existing model.qwen_chat_* block).
- skillopt/engine/trainer.py: import configure_minimax_chat and call
it alongside configure_qwen_chat, so cfg-supplied credentials,
temperature, max_tokens, and enable_thinking reach the backend. Also
apply cfg["minimax_model"] via set_target_deployment when the active
target backend is minimax_chat.
- scripts/train.py: add 6 --minimax_* CLI flags + the corresponding
_CLI_TO_YAML entries, add 'minimax' / 'minimax_chat' to the --backend
choices, auto-route to target_backend=minimax_chat, and pick the
right default target_model for the new backend.
Default behavior on existing backends (openai, claude, qwen, codex,
claude_code_exec) is unchanged; all 8 shipped configs continue to load
with gate_metric falling back to 'hard' for paper reproduction.
feat: add MiniMax as first-class chat backend
Adds skillopt/model/minimax_backend.py (clean port of qwen_backend.py
targeting MiniMax-M2.7 via https://api.minimax.io/v1) and registers it
in the router, backend_config, and common defaults. Existing backends
(openai_chat, claude_chat, qwen_chat, codex_exec, claude_code_exec)
remain bit-for-bit unchanged.
Verified via 10 import / routing / parity subtests; backward-compat
sweep across the 8 shipped configs passes with no regression.
A follow-up commit completes the YAML / CLI plumbing that this PR left
half-wired (FLATTEN_MAP entries, trainer-level configure_minimax_chat
call, and --minimax_* CLI args).
Add optimizer.slow_update_gate_with_selection to control how epoch-boundary
slow-update guidance is applied:
- false (default): force-injected - inject guidance into current & best
unconditionally (unchanged behavior).
- true: gated - evaluate the slow-update candidate on the selection set and
accept/reject via the same validation gate as step-level updates
(logic follows the SkillReflection ablation).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Port qwen_backend.py pattern to minimax_backend.py as a new
OpenAI-compatible urllib-based backend. Includes:
- BASE_URL defaulting to https://api.minimax.chat/v1
- API_KEY, TIMEOUT_SECONDS, MAX_TOKENS, TEMPERATURE env vars
- ENABLE_THINKING support (MiniMax thinking mode)
- configure_minimax_chat() runtime configurator
- chat_target() and chat_target_messages() functions
- TokenTracker integration and get_token_summary()
- set_target_deployment() support
- Default model: MiniMax/MiniMax-Text-01
Move the soft/mixed gate-metric configuration introduced in PR #25 out of
the base default config and into a standalone example config so that
default SkillOpt runs (and paper reproduction) remain bit-for-bit on the
original hard gate.
- configs/_base_/default.yaml: drop gate_metric / gate_mixed_weight keys.
The trainer's cfg.get("gate_metric", "hard") fallback preserves the
original behavior unchanged.
- configs/examples/soft_gate.yaml: new standalone reference config with
a header explaining when to consider it (small selection split with
continuous rewards) and when not to (paper reproduction, large or
binary-reward settings).
- README.md: add a short "Community-contributed configs" section that
clearly flags this as user-contributed and non-default.
Add configurable gate metric (hard / soft / mixed) for skill validation
Default is `hard`, preserving exact pre-PR behavior — verified by 22 unit
assertions on the gate module plus an end-to-end 8-step trainer-trajectory
test that produces a bit-for-bit identical accept/reject sequence between
the pre-PR and post-PR code paths under `gate_metric: hard`. Paper-
reproduction results are unaffected.
`soft` and `mixed` are opt-in via `evaluation.gate_metric` in the config
and address small-selection-set runs where discrete hard accuracy is too
coarse to distinguish candidate skills.
