o :Liue@stdZddlZddlZddlZddlZddlZddlZddlZddlZddl m Z ddl m Z ddl m Z mZmZmZmZejejddeeZe dZed Ze d Ze d Zed Zed ZedZedZddddddddZ GdddZ!GdddZ"GdddZ#GdddZ$d efd!d"Z%d#efd$d%Z&d&d'Z'd(d)Z(d*d+Z)d,d-Z*ed.kre*dSdS)/ue Rook Hyperagent — Phase 1: Skill Self-Improvement Analyzes agent session data to identify underperforming skills, proposes improvements, tests them against historical sessions, and promotes successful variants. This is the entry point for self-improving behavior. The agent doesn't just use skills — it makes them better over time. Architecture: Training DBs → Identify weak skills → LLM proposes fix → Test against saved sessions → Promote or archive Runs inside Citadel sandbox. Uses the same free-tier models the agent uses for chat — no additional cost. Usage: python3 skill_improver.py run # One improvement cycle python3 skill_improver.py status # Show skill performance python3 skill_improver.py history # Show improvement history python3 skill_improver.py rollback # Revert a skill change N)datetime)Path)AnyDictListOptionalTuplez[hyperagent] %(message)s)levelformatz.hermesskillsz2/opt/hermes-data/training/default/user-sessions.dbz//opt/hermes-data/training/default/tool-calls.dbz .hyperagentarchivez history.jsonlz config.jsonzqwen/qwen3.6-plus:free openrouterg333333? T)max_improvements_per_cyclemin_sessions_for_evaluationimprovement_modelimprovement_providertest_threshold daily_budgetenabledc@sleZdZdZddZdeeeffddZdeefddZ deefd d Z deefd d Z d dZ dS) SkillAnalyzerz3Analyzes skill performance from training databases.cCs<d|_d|_trtt|_trtt|_dSdSN)conv_dbtool_db TRAINING_DBexistssqlite3connectTOOL_DBselfr#6/home/geodesix/Jetson/src/hyperagent/skill_improver.py__init__Hs zSkillAnalyzer.__init__returncCs(|jsiS|jd}dd|DS)z3Get usage stats for each skill from tool call data.a SELECT tool_name, COUNT(*) as uses, SUM(CASE WHEN tool_response IS NOT NULL AND tool_response != '' THEN 1 ELSE 0 END) as successes FROM tool_calls GROUP BY tool_name ORDER BY uses DESC cSsBi|]}|d|d|d|ddkr|d|dnddqS)r)uses successes success_rater#).0rowr#r#r$ ]sz1SkillAnalyzer.get_skill_usage..)rexecutefetchall)r"rowsr#r#r$get_skill_usagePs zSkillAnalyzer.get_skill_usagec Cs|jsgS|jd}i}|D]\}}}||vr!gid||<||d||dqg}|D]H\}}|t|dtdd|dDtdd|dDtdd|dDtd d|dDtd td d|dDd }||q4|S) z/Extract quality signals from conversation data.z SELECT session_id, role, content FROM conversations ORDER BY session_id, turn_index )messagessignalsr3rolecontentcs0|]}|ddkrd|dpdvVqdS)r6userretryr7Nlowerr,mr#r#r$ }.z.css:|]}|ddkrdD] }||dpdvVq qdS)r6r9)thanksz thank youperfectgreatawesomer7r;Nr<)r,r?wr#r#r$r@~s8csr8)r6 assistanterrorr7r;Nr<r>r#r#r$r@rAcss,|]}|ddkrt|dpdVqdS)r6rGr7r;N)lenr>r#r#r$r@s*r'css |] }|ddkrdVqdS)r6rGr'Nr#r>r#r#r$r@s) session_id turn_count has_retry has_thanks has_erroravg_response_len) rr/r0appenditemsrIanysummax) r"r1sessionsrJr6r7resultsdatar4r#r#r$get_session_quality_signalsfs*  0 z)SkillAnalyzer.get_session_quality_signalsc Csg}ts|StD]R}|sq |jt||d|ddd|dDtdd|dDd}|d}|rYz|}|d d |d <Wn t yXYnw| |q |S) z*List installed skills with their metadata.SKILL.mdzDESCRIPTION.mdcSsg|] }|r|jqSr#)is_filenamer,fr#r#r$ sz6SkillAnalyzer.get_installed_skills..*cs"|] }|r|jVqdSrrZstatst_sizer\r#r#r$r@ z5SkillAnalyzer.get_installed_skills..)r[path has_skill_mdhas_descriptionfiles size_bytesNicontent_preview) SKILLS_DIRriterdiris_dirr[strrglobrS read_text ExceptionrP)r"r skill_dir skill_infoskill_mdr7r#r#r$get_installed_skillss.      z"SkillAnalyzer.get_installed_skillsc Cs@|}|}|}g}|D]3\}}|ddkrE|ddkrE|d|d|d|ddd |d d |dd d |d|dqdd|D}|ro|ddt|dt|td t|t|t|dddd|D}|r|ddt|dt|td t|dt|id|jdddd|S)z4Identify skills that could benefit from improvement.r)r r+gffffff?low_success_ratezTool 'z' has .0%z success rate (r*/ calls)r')typetargetreasonpriorityrWcSg|]}|dr|qS)rLr#r,sr#r#r$r^z>SkillAnalyzer.identify_improvement_targets.. user_retryresponse_qualityz3 sessions had user retries (dissatisfaction signal)) retry_counttotal_sessionscSr~)rNr#rr#r#r$r^rerror_responseserror_handlingz# sessions contained error responses error_countcSs|dS)Nr}r#)tr#r#r$sz.T)keyreverse)r2rXrurQrPrIrTsort) r" tool_usagesession_signalsinstalled_skillstargets tool_namestatsretry_sessionserror_sessionsr#r#r$identify_improvement_targetssD*       z*SkillAnalyzer.identify_improvement_targetscCs(|jr|j|jr|jdSdSr)rcloserr!r#r#r$rs  zSkillAnalyzer.closeN) __name__ __module__ __qualname____doc__r%rrnr2rrXrurrr#r#r#r$rEs  -rc @sXeZdZdZddededefddZdefd d Zd ed ed eede efddZ dS) SkillModifierz*Uses an LLM to propose skill improvements.Nmodelproviderapi_keycCs||_||_|p ||_dSr)rr _load_api_keyr)r"rrrr#r#r$r%szSkillModifier.__init__r&cCsxtd|jd}|r|Stdd}|r:|D]}|dr9| dddSq&dS) zLoad API key from secrets.secretsz.keyz hermes-agentz.envzOPENROUTER_API_KEY==r'r;) HERMES_HOMErrrpstriprhome splitlines startswithsplit)r"key_fileenv_fileliner#r#r$rs  zSkillModifier._load_api_key skill_contentr{session_examplesc CsJ|js tddSd|dd}d|ddd|dd |ddd }z`d dl}d dl}d }d d|jd} t|jd|dgdd } |j j || | d} | } |j j | d| d} t| }|dd ddWdWS1swYWdSty}ztd|WYd}~dSd}~ww)z6Ask the LLM to propose an improved version of a skill.z*No API key available for improvement modelNz r zWYou are a meta-agent improving an AI assistant's skills. ## Current Skill Content ``` iz ``` ## Problem Identified r|z ## Relevant Session Examples u ## Your Task Propose an improved version of this skill that addresses the identified problem. Return ONLY the improved skill content — no explanation, no markdown wrapping. The output should be a drop-in replacement for the current skill file. Focus on: - Clearer instructions that reduce errors - Better handling of edge cases - More specific guidance for the agent - Removing ambiguity that might cause retries Keep the same format and structure. Make targeted improvements, not a complete rewrite.r-https://openrouter.ai/api/v1/chat/completionsapplication/jsonBearer z Content-Type Authorizationr9r5rr3 max_tokensrWheaderstimeoutcontextchoicesmessager7zLLM call failed: )rloggerrHjoinurllib.requestssljsondumpsrencoderequestRequestcreate_default_contexturlopenloadsreadrrq)r"rr{r examples_textprompturllibrurlrbodyreqctxresprWer#r#r$propose_improvementsH     (z!SkillModifier.propose_improvementr) rrrrrnr%rrrrrr#r#r#r$rs & rc@seZdZdZddZdededefddZd edefd d Zde e fd d Z de fddZ dde de e fddZde fddZdS) SkillArchivez/Manages skill versions and rollback capability.cCs tjdddtjddddS)NTparentsexist_ok) ARCHIVE_DIRmkdirHYPERAGENT_DIRr!r#r#r$r%7szSkillArchive.__init__ skill_name skill_pathr&cCsR|dtd}t|}t|r't||t d|d||S)z.r_)idskill timestampsize) rrsortedrlrmrPr[rrbst_mtimerSro)r"backupsdr#r#r$ list_backupsVszSkillArchive.list_backupsentrycCsTt|d<ttd}|t|dWddS1s#wYdS)zAppend to improvement history.ra N)rr isoformatopen HISTORY_FILEwriterr)r"rr]r#r#r$record_improvementds "zSkillArchive.record_improvementlimitc CsbtsgStd}g}|| dD]}z |t|Wqtjy.Yqw|S)zRead improvement history.rN) rrrprrrPrrJSONDecodeError)r"rlinesentriesrr#r#r$ get_historyjszSkillArchive.get_historycs*tdtfdd|dDS)zCount improvements made today.z%Y-%m-%dc3s&|]}|ddrdVqdS)rr;r'N)getr)r,rtodayr#r$r@zs z/SkillArchive.get_daily_count..d)rrrrSrr!r#rr$get_daily_countws zSkillArchive.get_daily_countN)r)rrrrr%rnrboolrrrrrintrrr#r#r#r$r4s  rc @s@eZdZdZdefddZdedededeed ef d d Z d S) SkillEvaluatorzBEvaluates whether a proposed skill improvement is actually better.modifiercCs ||_dSr)r)r"rr#r#r$r%s zSkillEvaluator.__init__original_contentimproved_contentr{rr&c Csd|ddd|ddd|dd}zddl}ddl}d }d d |jjd } t|jjd |dgdd} |jj || | d} | } |jj | d| dJ} t | }|dddd}d|vr|dd}|dr|dd}t |}t|dd|dddWdWS1swYWdSty}ztd|dd |dWYd}~Sd}~ww)!z5Ask the LLM to evaluate if the improvement is better.zQYou are evaluating a proposed improvement to an AI skill. ## Original Skill ``` Niz" ``` ## Proposed Improvement ``` z! ``` ## Problem Being Addressed r|a ## Evaluate Rate the improvement on a scale of 0.0 to 1.0: - 0.0 = worse than original - 0.5 = about the same - 1.0 = significantly better Consider: 1. Does it address the identified problem? 2. Could it introduce new problems? 3. Is it clear and unambiguous? 4. Is it a targeted fix, not an unnecessary rewrite? Respond with ONLY a JSON object: {"score": 0.X, "reason": "brief explanation"}rrrrrr9r5rrrrrrr7z```r'rscorer;)r r|zEvaluation failed: g?zEvaluation error: )rrrrrrrrrrrrrrrrrfloatrrqrrH)r"r r r{rrrrrrrrrrrWresponseresultrr#r#r$evaluate_improvementsP        (z#SkillEvaluator.evaluate_improvementN) rrrrrr%rnrrrr#r#r#r$rsrr&cCsptr4z#tt}itt|WdWS1swYWtSty3YtSwtS)zLoad hyperagent configuration.N) CONFIG_FILErrDEFAULT_CONFIGrloadrqcopy)r]r#r#r$ load_configs $ rconfigcCsNtjdddttd}tj||ddWddS1s wYdS)zSave hyperagent configuration.TrrFr()indentN)rrrrrdump)rr]r#r#r$ save_configs "rc Cst}|ddstddSt}|}||dkr-td|d|dddStd t}|}|sFtd |dStd t |d |dd D]}td|ddd|dqWt |d|d}t |}g}|j r|j d} dd| D}d} t|d|d|} |d| D]} td| d|} | stdntdd| D| d}t|d}|d}|sq|}td |d!||| |}|r||krtd"||d!| dd#d$qtd%|||| |}td&|d'dd(|d|d'|d)kr||d!t|}||td*|d!d+|d,z$dd-lm}|d.|d!d/| ddd0d1|d'dd2|d3 Wn tysYnw||d!| dd4|d'|d|d5| d67} qtd7|d'dd8|d)||d!| dd9|d'|dd:q|td;| d<dS)=z(Run one cycle of skill self-improvement.rTz)Hyperagent is disabled. Enable in config.NrzDaily budget reached (rxz ). Skipping.zAnalyzing skill performance...z9No improvement targets identified. Everything looks good.zFound z improvement targets:r [r}.2f] r|rrzISELECT role, content FROM conversations ORDER BY turn_index DESC LIMIT 20cSs&g|]\}}|d|ddqS)z: Nr r#)r,r6r7r#r#r$r^s&z)run_improvement_cycle..rrz Processing: zNo installed skills to improve.css|] }|dr|VqdS)rfNr#rr#r#r$r@%sz(run_improvement_cycle..rerYz# Proposing improvement for skill: r[z$ No improvement proposed. Skipping. no_change)rr{statusz$ Evaluating proposed improvement...z Score: r  — ru ✓ PROMOTED — skill 'z' improved (backup: )) notify_ownerzImproved skill z Reason: rz Score: z* Rollback: rook hyperagent rollback zpromoted)rr{rr r|rr'u ✗ REJECTED — score z below threshold rejected)rr{rr r|z Cycle complete: z improvements promoted.)rrrrrrrrrrIrrrr/r0minrunextrrrprrrrrn write_textnotifyr"rq)rr daily_countanalyzerrrr evaluatorrr1improvements_mademax_improvementsr{r rr skill_filer r  evaluationrr"r#r#r$run_improvement_cycles     "           r0c Cst}t}t}tdtd|ddtd|dtd|d|d d td |d |}td t||ddD]}td|dddt|ddd|dddqO|}|rtdt | dddddD]1\}}dt |ddd dt |dd}td|d!d|d|dd"d#|d$d% q| } | rtd&t| | dd'D]} td(| d)d*d+| d,q| } | rtd-t| | dd'D] } td| d.q|dS)/z5Show current skill performance and hyperagent status.u ⚔ Rook Hyperagent — Status zEnabled: rTzModel: rzDaily budget: rxrz iterations used todayzThreshold: rz Installed skills: Nr r[30 rh3z files riz>8z bytesz Tool performance:cSs|dd S)Nr'r)r#)xr#r#r$rszshow_status..)ru█r+u░25rwz (r)ryz Improvement targets: rrr}rrr|z Archived versions: r)rrrprintrrrurIr2rrQrrrr) rr r*r rrr[rbarrrrbr#r#r$ show_statusms: 4$,0  r:c Cst}|d}|stddStdt|D]R}|ddkr#dn |ddkr,d nd }|d d }t|ttfrBd |ddnd }td|d|ddddd|ddd|dddd| qdS)zShow improvement history.rzNo improvement history yet.Nu- ⚔ Rook Hyperagent — Improvement History rr#u✓r$u✗u—r r;z (rr!r1z [r?rrr r{<)rrr7reversedr isinstancerr)r historyr status_iconr  score_strr#r#r$ show_historys  (  NrCcCsJddl}|jdd}|jdd}|jddd|jd d d|jd d d|jd dd}|jddd|jddd|jddd|}|jdkrOtdS|jd krYtdS|jd krct dS|jd krst }| |j dS|jdkrt }d|d<t|tddS|jdkrt }d|d<t|tddS|dS)Nru*Rook Hyperagent — Skill Self-Improvement) descriptioncommand)destrunzRun one improvement cycle)helprz,Show skill performance and hyperagent statusr@zShow improvement historyrzRevert a skill changerzBackup ID to restoreenablezEnable hyperagentdisablezDisable hyperagentTrzHyperagent enabled.FzHyperagent disabled.)argparseArgumentParseradd_subparsers add_parser add_argument parse_argsrEr0r:rCrrrrrr7 print_help)rKparsersub rollback_cmdargsr rr#r#r$mains>              rV__main__)+rhashlibrloggingosrrsystimerpathlibrtypingrrrrr basicConfigINFO getLoggerrrrrrkrr rrrrrrrrrrrr0r:rCrVr#r#r#r$sZ    VPW ')