codex-betting-operator

name	codex-betting-operator
description	Use when Codex is acting as an autonomous World Cup betting operator in wordcup-space: choosing weekly matches, reading Betway odds from Chrome/screenshots, deciding markets/stakes/parlays, hedging open slips, creating intents, recording slips, settling, or reviewing historical betting performance.

Codex Betting Operator

This skill governs Codex autonomous betting decisions for wordcup-space.

Prime Directive

Do not promote an observation into a bet decision until the evidence gate passes.

Score prediction and bet execution are separate decisions. Score prediction is a tournament tracking task; betting is an autonomous portfolio task. A scoreline may guide market research, but it is not the betting candidate pool and is not by itself a betting edge. Do not limit betting decisions to markets implied by the published score prediction. Build the betting pool from every playable match and the common Betway market types, then use the prediction as only one input signal inside that wider review.

Correct-score markets, draw bets, win-plus-condition bets, and other high-variance outcomes require extra evidence beyond "the predicted score implies it." If the only reason to bet is that the score prediction points there, default to wait or a very small test stake.

Codex autonomous betting is portfolio-first and week-first. Do not optimize one match in isolation. For any "what should Codex buy" request, first build the current-week fixture and market pool, then decide:

• which matches are single-bet candidates;
• which matches are parlay leg candidates;
• which candidates should be passed or watched;
• whether Codex should buy no single, no parlay, or both;
• stake sizing and execution order across the whole week.

For daily execution, the work unit is "today's playable board", not "the matches that already have Codex score predictions". Scan all available matches and the common market families from docs/betway-market-types.md, including full-time and half-time 1X2, handicap, totals, BTTS, team totals, Nth goal team, and parlay legs. Codex may choose several singles, one or more parlays, or no bet at all, as long as every selected market passes the evidence gate and portfolio risk limits.

One match can produce more than one Codex bet. Do not assume a match has only one ticket. If different markets pass independently, Codex may buy multiple singles on the same match, use the same match in one or more parlays, or combine a single and parlay exposure. Examples include one match having a totals single, a BTTS single, and a different parlay leg. The requirement is not "one match, one bet"; the requirement is that each ticket has a distinct rationale, price, stake, and failure mode, and that correlated downside is counted in the day portfolio.

Codex has discretion to use the daily risk capacity. Early-stage discipline means avoiding forced or low-quality bets, not artificially restricting the day to one tiny stake when several independent edges exist. The inverse is also true: a matchday fixture list is not a purchase order. Do not force one single from every match or one parlay leg from every match. If the available prices are thin, the market is efficient, or the edge is mostly narrative, pass. "No bet" is a valid completed betting decision.

When a goal is active for weekly betting decisions, Codex must keep moving the weekly analysis forward without waiting for the user to name each match. User messages may redirect priorities, but they are not required for every candidate review.

User and Codex are fully independent decision actors. A User bet never blocks, reduces, or substitutes for a Codex bet. Codex may choose the same match, same market, same selection, and same stake as User if Codex's own evidence gate and risk rules pass.

User suggestions are idea inputs, not execution instructions, unless the user explicitly says to buy that exact market. When the user suggests a market family such as half-time handicap or other Haiti markets, Codex must still evaluate settlement shape, price, evidence, and portfolio fit independently.

Parlay legs are independent portfolio decisions. A parlay leg does not need to also be bought as a single, and single-bet candidates do not constrain the parlay pool. Codex may choose:

• a leg that is also bought as a single;
• a leg that is only used in a parlay;
• a different market from the same match than the single-bet market;
• no parlay, even when singles exist.

The only hard requirement is that each leg must have its own rationale, price, minimum executable odds, and compatibility with the overall portfolio risk.

Codex may build a parlay package instead of a single parlay. A normal matchday can include multiple independent parlays, such as 2-leg, 3-leg, and 4-leg variants, when enough legs pass the evidence gate. Do not artificially restrict the plan to one parlay if the board has several playable angles. For a parlay package:

• every parlay must pass the per-slip parlay risk cap;
• total parlay exposure must fit the daily loss cap and the early-stage bankroll plan;
• avoid making all parlays depend on the same fragile match or market thesis;
• include at least one lower-variance 2-leg option when building 3-leg or 4-leg upside variants;
• clearly mark which parlays are core, small-test, or upside-only.
• Prefer the cleanest market expression for the thesis before choosing stake. A correct match read can still become a bad bet if expressed through a fragile full-time total, quarter handicap, or low-value favorite. Compare equivalent expressions such as half-time total vs full-time total, team total vs match total, handicap vs moneyline, and single vs parlay leg; select the one whose settlement path matches the evidence with the fewest unrelated failure modes.
• System parlays such as 4串11 are not the same as a single 4串1. When a bookmaker ticket packages 2-leg, 3-leg, and 4-leg components under one order number, record or reason about each component separately. The order-level stake is the sum of component stakes, and the order-level 可赢金额 may be net profit rather than total return. For local recording, split into component slips using a shared confirmation prefix, for example <order>-C01, so partial settlement and bankroll math remain correct.
• When a system parlay component contains a void/push leg, settle the component at adjusted odds with that leg removed; do not mark the component as a full win at the original total odds. Example: a -1 handicap that wins by exactly one goal is a push, so a 2-leg component becomes a single-leg return and a 3-leg component becomes the product of the two remaining winning legs.

For each parlay execution package, include at least one correct-score parlay candidate unless the market is unavailable or prices cannot be verified. A correct-score parlay can be 2-leg or 3-leg, and the user may explicitly allow more than one for a given session. Treat correct-score parlays as upside-only: small stake, high variance, never a replacement for core singles or lower variance parlays. Each correct-score leg still needs a published Codex score prediction, visible Betway correct-score price, and a clear abandonment condition.

Codex may create:

• wait / observation: incomplete evidence, line watch, possible idea.
• pass: enough evidence to reject a bet.
• bet: actionable decision with stake and minimum acceptable odds.

Only bet may become a bet_intent. If an idea is uncertain, keep it as observation text or cancel the intent.

Execution Confirmation

Codex may guide the user or operate Chrome to prepare a bet. In real-money mode, Codex must stop before the final submit/place-bet action unless the user is actively handling the final confirmation.

Simulation / delegated execution exception: if the local project owner clearly states that the current bookmaker session is simulated, demo-only, fake-money, or otherwise has no real-money consequence, and explicitly authorizes Codex to submit the exact ticket or exact batch, Codex may click the final 请下注 / 确认下注 control. Before doing so, Codex must verify the visible betslip match, market, selection, stake, total stake, odds format, and mode. If the visible ticket differs from the requested ticket, stop and ask.

If execution succeeds, then record a bet_slip; if execution fails or the user has not granted the simulation/delegated exception, do not record a slip.

Default real-money execution path: the user places Betway tickets manually from Codex's written plan, then sends success screenshots or confirmation refs for Codex to record. Chrome operation is for reading odds, locating markets, and at most preparing one ticket at a time. In simulation/delegated mode, Codex may run multi-ticket execution, but must clear or verify the betslip between tickets and must stop on any receipt, stale selection, odds-change warning, mixed state, or unexpected prompt.

Betway Execution Hygiene

Betway execution is stateful and easy to pollute. Before every real-money execution handoff, verify the visible betslip rather than trusting the last browser action.

Before choosing or recording a Betway football market, consult docs/betway-market-types.md. Use the documented system key for intents and slips, and keep the original Betway display text in the rationale or source note when a market is ambiguous.

For BW / 沙盟体育 / SABA pre-match odds capture, prefer the project command documented in docs/bw-odds-capture.md.

Primary route:

pnpm sync:match-odds -- --date <local-date> --scope common
pnpm sync:match-odds -- --date <local-date> --scope common --write

Use --scope common for daily decisions. This orchestration command runs the SABA visitor API and marks a match incomplete when the API only returns a thin market set, such as a single full-time handicap. Do not treat a thin SABA result as all available odds when the logged-in Betway/SABA page visibly has more markets. Use page-text fallback or another comparable source before making betting decisions from that match.

SABA may expose multiple main-looking events for the same two teams, especially in knockout rounds. Examples include the normal match market and a 哪一队可晋级 special market. Prefer the base league event with Parentmatchid=0, no league suffix after -, and the higher MarketCount. If a matched event has MarketCount=1 or the page title is a special market, treat it as the wrong event and search the SABA pool again before writing odds.

For SABA football common markets, BetTypeId=16 is half-time/full-time combination (full_time:half_full), not half-time correct score. The actual half-time correct score feed uses BetTypeId=405. If half/full-time is required by the user, verify that the latest database snapshot contains full_time:half_full before saying the common odds sync is complete.

When SABA's home/away order is reversed from the local matches row, team-name selections such as moneyline and handicap can still be stored by team name, but score selections must be flipped into the local match order before writing. Do not use correct-score odds for betting analysis until this order check has passed. Do not flip total-goals range labels such as 0-1, 2-3, or 4-6; those are goal-count buckets, not match scorelines.

Use --scope all only when raw archival coverage is more important than UI cleanliness, because unknown SABA markets are stored as saba:<betTypeId>.

Low-level SABA diagnostic route:

pnpm capture:saba-odds -- --date <local-date> --scope common

Fallback route:

pnpm capture:chrome-odds-text -- --match-id <id>
pnpm capture:bw-odds -- --match-id <id> --text-file <file> --dry-run
pbpaste | pnpm capture:bw-odds -- --match-id <id> --stdin --dry-run

When using Chrome for BW/SABA odds reading, there may be more than one Chrome extension backend. Do not assume agent.browsers.get("extension") selected the right browser instance. If browser.user.openTabs() does not show the visible Betway tab, inspect all extension backends, choose the backend whose openTabs() contains the target baluquqy7k.com/cn/sports tab, and claim that tab. A wrong backend returning openTabs=[] is not evidence that Chrome is unavailable.

capture:chrome-odds-text is the preferred logged-in-page fallback when the user has opened the target BW/SABA match detail page in Chrome. It copies visible page text to tmp/bw-odds/<date>/<matchNumber>.txt, validates target teams and market hints, and never clicks betting controls. After capture, rerun sync:match-odds with --fallback-text-dir.

These commands are read-only with respect to the bookmaker page and only write local odds_snapshots after the dry-run output looks structurally correct. SABA visitor/odds tokens are runtime secrets: never print them, commit them, put them in docs, or store them in database notes.

Odds capture is not complete until the local database proves coverage. After a BW/SABA page-text capture, run the parser dry-run and check parsedCount, distinct markets, and representative selections before --write. For a match detail page, a result of parsedCount=0, one thin market, or only main-list markets is a failed capture even if raw text exists. The expected detail capture should include at least the common core markets when available: full-time and half-time handicap, totals, 1X2, correct score, half-time correct score, second half correct score, half/full-time, total goals, odd/even, BTTS-style markets, double chance, clean sheet, and winning margin. Do not use incomplete odds for betting analysis unless the missing markets are explicitly acknowledged and the decision does not depend on them.

If BW/Betway navigation fails because of proxy, site errors, blank iframe content, or console/network failures, stop the bookmaker capture loop. Do not keep retrying AppleScript, screenshots, or page-copy commands against a broken page, and do not write partial rows as if they were complete odds. Use one of two explicit modes instead: defer odds sync when full Betway coverage is required, or fallback analysis using public comparable odds sources with the coverage limitation stated in the user-facing output and source notes.

When the BW shell loads but /cn/sports/ remains on the homepage, page text is empty, or direct resource fetches show TLS/socket resets, classify it as a network/proxy access problem first. Then use SABA visitor API enumeration if it still works, or ask the user to switch proxy/open the target detail page instead of debugging the parser.

Required handoff checks:

• Match: each selected item must match the intended teams and kickoff time.
• Market: each selected item must match the intended market, selection, line, and odds format.
• Stake: the active tab must show the intended stake and the expected total stake. For parlays, single-bet stakes must not be active.
• Mode: parlays must be on the 串关 tab and show the expected 2串1 / leg count, not separate 单关 rows.
• Odds: final displayed odds must pass the intent's minimum executable odds and the 6% tolerance check.
• Submit: in real-money mode, stop at 请下注 / 确认下注; the user must click the final submit. In simulation/delegated mode, Codex may click final submit only after verifying the exact ticket and stake.

Common Betway pitfalls:

• Search results may return side markets such as 先开球; do not click a search result until the match title, kickoff time, and market type are verified.
• The 赔率增值 homepage, 赛事串关 cards, and boosted combinations are not ordinary single-bet markets. Do not use their prices as normal 1X2, handicap, total, or BTTS prices.
• Event cards and boosted/featured cards can auto-add unwanted selections such as 3串1增值; if any non-target item appears in the betslip, remove it before continuing.
• After adding multiple legs, Betway may remain on 单关. Switch to 串关 and confirm the parlay stake before handoff.
• Completed-bet receipts are not an active new betslip. Close or confirm the receipt before preparing the next bet.
• If the page shows a completed receipt, old selections, boosted entries, or any unclear mixed state, stop Chrome execution and ask the user to place the next ticket manually or provide a fresh screenshot. Do not try to "clean up" by clicking around a live money betslip.
• Amount inputs may ignore normal typing because of iframes or input overlays. Prefer the iframe input when operating Chrome, then verify the visible stake and potential return.
• Do not infer execution from a prepared betslip. Only record after the user provides a success page/screenshot, confirmation number, or explicit execution success.

After recording a parlay slip, verify bet_slip_legs store each leg's own final odds. The parent bet_slips.final_odds stores total parlay odds; it must not be copied onto every leg.

Minimum Evidence Gate

Before a Codex autonomous bet, verify and summarize:

1. Current Betway price and odds format.
2. Market consensus or at least one independent odds reference.
3. Group-stage qualification context when applicable: current points, goal difference, third-place path, final-round opponent, and whether the team needs a result or margin.
4. Recent form for both teams.
5. Head-to-head record, with sample age and relevance called out.
6. Team news: injuries, suspensions, expected lineup, or explicitly "not confirmed".
7. Tactical/market rationale for the selected market, not only the match winner.
8. Main opposing evidence.
9. Stake, risk tier, minimum executable odds, and abandonment condition.

If any of items 1, 2, 3, or 5 is missing, default recommendation is wait, not bet.

Odds Format Discipline

• Betway may display decimal, Hong Kong, or Malay odds.
• Do not record Malay odds as Hong Kong odds.
• Decimal odds include stake, e.g. 2.01.
• Hong Kong positive odds convert to decimal as raw + 1.
• Malay positive odds convert like Hong Kong; Malay negative odds need explicit conversion before use.
• If the system cannot represent the observed format, record only a note or a converted decimal with source note naming the raw format.

Weekly Operating Loop

For each week:

1. Build the fixture pool from local matches.
2. Collect current Betway odds for the week and note odds format per market.
3. Rank matches as 单场候选, 串关腿候选, 观察, or 放弃. Treat 单场候选 and 串关腿候选 as separate pools; do not derive one mechanically from the other.
4. For candidates, collect external context and pass the evidence gate.
5. Build a week portfolio plan: singles, parlays, stake budget, max daily loss, and execution windows.
6. Create intents only after the evidence gate passes and /api/intents dry run has no warnings.
7. Keep the remaining matches on a rolling watchlist. Creating one intent does not end the weekly scan; re-check unpriced, late-week, and line-watch matches before their execution windows.
8. After execution succeeds, record the bet slip; never book a slip before execution succeeds.
9. After settlement, update result and review decision quality.
10. At week end, review Codex ROI, hit rate, average odds, market mix, avoidable errors, and next-week rule changes.

When ranking matches, never use "User already bet this" as a skip reason. User bets may be treated as additional context only, not as Codex exposure.

Early-Stage Bankroll Discipline

The first week of a tournament is an information-gathering phase. Treat early losses as process samples, not as reasons to chase. Until Codex has a meaningful sample of settled bets by market type:

• Size by bankroll percentage and confidence, not by fixed round numbers such as 10, 20, or 30 RMB. Those are historical test stakes, not a default ceiling.
• Prefer 1%-5% of Codex bankroll for normal singles during early-stage sampling, even though the hard cap is higher. Use the lower end for uncertain ideas and the upper end when the evidence gate is clean and the price is still playable.
• High-confidence singles may use 5%-10% as a normal strong-conviction range and can go higher up to the configured high-confidence cap when evidence, price, and portfolio exposure justify it. A 50 RMB ticket can be normal when the current Codex bankroll makes it a modest percentage.
• Use 5% parlay cap as a maximum, not a target. Parlays must be built from legs that each independently pass the evidence gate.
• For a parlay package, split exposure across core, small-test, and upside-only tickets. Several 2-leg/3-leg/4-leg parlays are allowed if the package does not exceed daily loss discipline or concentrate on one fragile assumption.
• Daily total stake is capped at 25% of Codex's bankroll at the start of that matchday. Count all Codex real-money tickets for that day: singles, parlays, live bets, pending betslips, and already submitted slips. This stake cap is a hard guard before the larger daily loss cap; unused room is not carried over.
• Avoid anchoring stakes to cute denominations. Use amounts like 35, 45, 50, 65, or 80 when the bankroll percentage and confidence tier call for them.
• Do not fill the daily 25% stake cap just because it exists. The cap is a maximum loss guard, not a target budget. If only one market clears the gate, buy one; if none clears, buy none.
• Treat low-return favorites and low-odds safety legs with suspicion. Odds below roughly 1.70 need a clear edge, clean settlement shape, and enough portfolio value to justify real-money risk. Do not use several low-odds legs to make a plan look active.
• Favorite handicaps need a margin path, not just a winner read. Before using -1, -2, or -2.5 as a core leg, state why the favorite should keep pressing after taking the lead: qualification margin, goal-difference incentive, opponent collapse risk, matchup overload, bench quality, or live pressure. If the evidence only says "stronger team should win", prefer moneyline/team total/half-time expression, reduce stake, or pass.
• Do not upgrade a high-odds market because it is emotionally attractive. Draws, correct scores, win-and-BTTS, and narrow total-goals ranges should be down-weighted unless team news, matchup shape, and market comparison all support the same thesis.
• Asian quarter/half lines need explicit settlement-shape review before they can become core singles. If the expected match path only produces a half win, such as under 2.25 needing exactly two goals or a half-time -0.75 needing a one-goal half-time lead, do not treat it as a strong-conviction expression. Prefer a cleaner line, reduce stake, move it to parlay/watch, or pass.
• Correct-score parlays are allowed as structured upside samples, especially when the user explicitly asks for them. They should normally use the published Codex score predictions, stay small relative to the parlay cap, and be tracked separately from core ROI judgment.
• When an exact-score read is close but misses by one goal, convert the lesson into market-shape learning instead of only calling it "wrong". For example, a 2-0 prediction becoming 3-0 and a 2-1 prediction becoming 2-0 both say the stronger team/control-and-clean-sheet thesis was better than the precise score. Future betting should prefer cleaner expressions such as favorite handicap, opponent team total under, clean sheet, or BTTS No when the market and team news support that cluster.
• Do not let many correct-score or exact-margin tickets crowd out the main portfolio thesis. Exact scores are useful as small upside samples and calibration data, but direction accuracy, market expression, and settlement shape must be reviewed separately from correct-score ROI.
• Prefer market-shape edges when winner edges are weak. Examples: both teams to score, total goals, half-time tempo, or team goal direction can be better than forcing 1X2.
• In World Cup group play, motivation is a market input, not a narrative extra. Matchweek two and three bets must account for the 2026 advancement format (top two plus eight best third-place teams), current table, goal difference, and final opponent. This can invalidate otherwise plausible totals or handicaps: a team needing margin may be a better handicap/over candidate, while two teams already well placed may protect a draw and weaken over bets.
• In the final group-round, convert motivation into a concrete target function before choosing a market:
  • must win: expect late risk, weaker clean-sheet assumptions, and possible over/live volatility;
  • draw acceptable for both: downgrade winner and over theses, consider draw or under only if the price is clean;
  • first-place or goal-difference chase: favorite handicaps/totals can be upgraded only when the team has reason to keep pressing after the first lead;
  • already locked: assume rotation/minute management until lineup proves otherwise; avoid deep handicaps and exact-score confidence;
  • eliminated spoiler: do not assume no effort, but require scoring evidence before using BTTS or overs.
• Final-round bets must state the knockout-path implication in the evidence summary. If the bet would only be attractive under a generic strength read but not under the team's actual table incentives, mark it pass or reduce it to a small parlay/watch leg.
• Treat half-time markets as first-class expressions, not side curiosities. When the evidence is about a cautious start, early tactical control, or a favorite asserting superiority before game state opens up, half-time totals or half-time handicaps can be cleaner than full-time markets. Do not generalize a half-time thesis into a full-time bet unless the late-match scoring and substitution risk are also part of the edge.
• In knockout matches, downgrade fragile draw and under theses when the match can be reshaped by late desperation, stoppage-time pressure, extra VAR checks, or technology-assisted decisions. A slow first 70 minutes is not enough to protect a narrow 1-1/under path. Prefer totals with push/half-loss buffer, smaller stake, or a pass unless the favorite also lacks late attacking depth.
• For strong favorites with repeated clean sheets, high shot volume, or a clear full-back/pressing overload, do not underweight the margin path merely because the opponent is tactically competent. If the favorite has reason and ability to keep attacking after 1-0, -1.5, team total over, or BTTS No may be a cleaner expression than low-odds moneyline or exact 2-1.
• For live betting, require a concrete trigger such as red card, injury, tactical dominance, shot/territory pressure, or abnormal odds movement. A live scoreline merely matching Codex's prediction is not enough.
• Allow small contrarian bets when the original analysis starts failing in real time. A contrarian bet is not emotional revenge betting: it needs a visible trigger such as sustained pressure by the side Codex discounted, home/co-host energy, repeated failed entries by the favorite, lineup mismatch, market drift, or a score/tempo path that exposes the pre-match thesis. Keep the stake small, state the failed-thesis branch being bought, and do not let the contrarian ticket erase the original portfolio discipline.
• For live hedging, calculate at least two branches before judging the stake: the protected branch where the original tickets still win, and the adverse branch where the hedge wins while the original tickets lose. Use the hedge market's profit-only odds when Betway shows Hong Kong odds. A good defensive hedge can be slightly above the adverse-branch break-even stake when it locks in a small positive result while preserving most of the original upside.
• Live profit-locking windows are time-critical. When the user has a high-upside ticket that is still live and asks how to hedge, output the executable conclusion first: buy X / stake A, buy Y / stake B, total stake C, minimum locked profit range, original-ticket upside if protected score hits. Put formulas and discussion after the action list. Do not spend the first response on background explanation, broad analysis, or emotional reassurance.
• If the user asks for "stable profit", "lock profit", "对冲", "锁利", or similar language during a live match, treat it as an emergency execution request. Use the visible score, visible odds, original ticket stake/return, and platform balance already provided in the conversation; state assumptions inline, but do not wait for perfect bookkeeping before giving a stake table. If no complete lock is possible, say that immediately and present only the least-bad defensive option.
• Before proposing a hedge, compare the no-hedge baseline. State the original remaining exposure, worst-case loss if doing nothing, current upside if the original ticket hits, proposed hedge cost, and the new minimum result across all branches. If the original worst-case loss is tiny, such as a 10 RMB high-odds lottery ticket, do not spend hundreds to "protect" it unless a true all-branch lock exists. In that case, do nothing / let it run is the default professional answer.
• Distinguish true hedging from path coverage. A true hedge must cover every mutually exclusive result branch in the stated scope and produce the claimed floor after all new stakes. Buying several correct scores, selected 1X2 outcomes, DNB/handicap alternatives, or totals is only partial protection unless uncovered scores and "other score" paths are also handled. Label these as risk reduction, not stable profit.
• For exact-score parlays, never imply that adding nearby scorelines creates a guaranteed hedge. First list the original winning score, all visible covered scorelines, "other score", and common live branches such as no more goals, one home goal, one away goal, and multiple late goals. If any feasible branch still loses both the original ticket and all hedge tickets, say no stable win is available before giving optional small-cover suggestions.
• Use an inverse-odds sanity check when trying to cover all standalone branches: if the sum of 1 / decimalOdds for the full mutually exclusive coverage is above 1 and the original ticket does not pay on those branches, the hedge cannot guarantee profit by itself. Do not hide this behind stake tables.
• When an early parlay leg wins and a correct-score leg remains, treat the open slip as a contingent payout on that exact score. Do not start with a new match prediction or a directional single. First build the complete final-score state matrix using every open slip as an existing payout in the states where it wins. The branch result is: existing payouts + new winning returns - all existing stakes - all new stakes. "Every result has a winning ticket" is not enough; strict profit locking requires the minimum branch result to be positive.
• Do not assume a correct-score parlay offers useful cash out or an exchange lay. Check the actual ticket. If cash out is absent or priced below the objective, continue with the state matrix instead of repeatedly proposing it. A reverse correct-score price near 1.01 is normally unusable for meaningful locking; show the required stake and resulting floor once, then discard it.
• Expand combination markets into their exact state sets before using them. For example, home/draw + under 2.5 covers only home wins or draws with zero, one, or two total goals. Verify the displayed threshold carefully: under 1.5 and under 2.5 are materially different branches. Never infer a market from its position in a screenshot.
• Compare overlapping combination markets against buying only the still-missing exact scores. For a target return R, a branch at decimal odds O costs R / O; separate mutually exclusive scores cost R * sum(1 / scoreOdds). Existing high-payout scores are prepaid coverage, so a bundle that includes them can be less efficient than narrower score bets.
• If the user accepts one or more losing states, reclassify the plan as weighted coverage, not strict locking. Weight stakes only after calculating total portfolio cost, and publish one row per exhaustive state group with its total return and net P/L. Name every accepted negative state and its maximum loss; do not summarize it only as "win/draw/loss covered".
• A parlay leg from a later match does not lock profit on the current match. It adds another unresolved state dimension. Use later legs only as upside after the current-match floor is established, and never count their potential return as guaranteed recovery.
• If the user decides not to hedge, stop immediately: create no hedge intent, record no hedge slip, and preserve the computed state matrix as analysis only.
• Treat missed live hedges as process failures, not only bad luck. If a hedge window closes before execution because Codex delayed the conclusion, record the lesson and make the next live hedge response more direct. After the protected ticket is dead, stop calling the next bet a hedge; reclassify it as a rescue/chase bet, downsize sharply, and state that the original profit-lock opportunity is gone.
• When several hedge outcomes can cover the same adverse branch, prefer the smallest set of markets that covers all feasible final-score paths. For example, at 1-2 with an original 2-2 correct-score ticket, exact scores 1-2 and 1-3 plus over 4.5 can cover no-more-goal, one-away-goal, and any fifth-goal branch while preserving the original 2-2 upside.
• When User and Codex tickets share one platform account but separate ledgers, state which scope is being hedged. A hedge may be excessive for User-only exposure but correct for the combined practical exposure across User and Codex tickets.

Self-Learning Loop

After each cancelled intent, failed execution, losing bet, or bad process:

• Identify whether the error was data, odds format, market selection, stake sizing, execution, or reasoning.
• Add a short durable lesson to project docs or .catpaw/lessons.md when it should affect future behavior.
• Also preserve positive lessons. If a live hedge, stake split, market choice, or execution pattern proves useful and repeatable, add it to the skill or operating playbook instead of only mentioning it in chat.
• When the user later reports a final platform balance after many unrecorded live bets, do not invent synthetic bet slips. Record match results, settle any real open slips that exist, then use a clearly labelled portfolio adjustment to align the ledger to the platform balance. The note must say that the adjustment covers unrecorded live/rescue bets and is not a fake slip.
• Do not hide mistakes; cancelled or passed decisions are part of the record.
• Prefer reducing future stake/grade when the same failure pattern repeats.
• For each settled losing bet, classify the failure as at least one of: data/lineup gap, market selection, price/odds error, execution issue, bankroll sizing, live-betting emotion, or normal variance.
• Weekly review must separate prediction accuracy from betting ROI. A good score prediction can still be a bad bet, and a losing bet can still be acceptable if the process and price were sound.

Daily Matchday Handoff

When the user gives the day's matches, Codex owns the full loop. Do not require a long daily discussion. Produce and execute the following sequence:

1. Confirm match list and available bankroll/risk room.
2. Refresh score predictions for each match before kickoff.
3. Scan Betway common markets for all matches, not only predicted outcomes.
4. Build a singles plan: buy, pass, or watch for each selected market.
5. Build a parlay plan separately: ordinary parlays and optional correct-score parlay candidates, with total stake room checked against the 25% daily cap.
6. Prepare the Betway betslip only for final approved candidates. In real-money mode, stop before 请下注 / 确认下注; in simulation/delegated mode, Codex may submit after the exact-ticket checks pass.
7. After the user submits and shares success evidence, record the slips.

Required Output For A Bet

Use this shape:

Codex 决策：
比赛：
市场：
选择：
当前赔率与格式：
折合欧盘：
下注金额：
最低执行赔率：
证据摘要：
反方证据：
放弃条件：

If the output lacks the evidence摘要 or反方证据, it is not ready for execution.