Skip to content

fairgradient-ai/aegis-desk

v1.0.0MIT

Aegis Desk: a read-only options position-monitoring and exit-alert desk for Grok Bot and Grok Build, run by the Aegis Desk Lead, the Aegis Exit Sentinel, Aegis Position Control, the Aegis Book & Risk Analyst, which shows exposures and stress views, the Aegis Events and Calendar Analyst, which keeps the event calendar, the Aegis Rules Steward, which keeps rulebook change control, the Aegis Review Desk, which posts the weekly review, counts only, and, only when you ask, the Aegis Structuring Desk, which prices choices for positions you hold by your own criteria. Stdlib tools compute every number; the desk never places, modifies or cancels an order.

Changelog

Unreleased

  • The 1.0.0 release record (docs/RELEASE.md): the release commit, the gate, the pin check, CI and the oracle on it, the Claude Code checks on the tag (validation, both denials, the read, the marketplace install), the tag, the GitHub Release, and the decision for main until the next tag (the bump waits until just before it).
  • The Grok Bot template is ready-to-publish (template/grok-bot.json): it claims every skill and role skill and exports the bootstrap skill. docs/GROK_BOT_TEMPLATE.md and docs/FAQ.md say so, and the template page names release 1.0.0 and the eight role skills.

1.0.0 (2026-11-17)

  • Release preparation: Stages 2 and 3 complete. Eight Bots: six on the Floor and two by direct message.
    • Present tense. Every line that called a landed item planned is rewritten, and the roadmap's mention table is empty. The roadmap's introduction, the manifests' descriptions, the runtime pages and the README count eight role profiles.
    • docs/UPDATING.md gains the step that creates a new Bot during an update, and the notes for a desk moving from 0.2.1: four new Bots, nine new routines, the new Allow text, the American model, the new optional keys.
    • docs/RELEASE.md is the 1.0.0 checklist, with ten new counsel rows for the texts the roadmap added. The 0.1.0 receipt is kept as docs/RELEASE-0.1.0.md.
  • The Aegis Review Desk (roadmap step 3.4): a direct-message seat (seat: off-floor) with its role skill and the skill aegis-signal-journal-and-review.
    • weekly_review.py (new) posts the week's signal facts, the counts per rule and a waiting candidate's replay over the week, counts only.
    • The alert lint's review mode refuses a currency amount outside a per-signal fact line and the grading phrases.
    • The routine. weekly-review (Saturday 11:00 ET). Lean 66, active 86.
  • The Aegis Rules Steward (roadmap step 3.3): a direct-message seat (seat: off-floor) with its role skill. It holds rulebook change control (candidate file, dry run, replay, lint, adoption on the owner's phrase and tap); the Desk Lead runs the same steps on request.
    • rules_sanity.py (new) verifies the adopted rulebook and lints it against the structures held today, or prints the weekly memo.
    • The routines. rules-sanity (16:45 ET; posts only a finding) and rulebook-memo (Friday 17:20). Lean 65, active 85.
  • Rules replay (roadmap step 3.2): rules_replay.py (new) replays a candidate rulebook over the desk's end-of-day snapshots beside the adopted one. It lists the days and groups whose verdict would have differed, then one count line. It states its limits (end of day only, no rule state carried across days, no profit or loss) and writes nothing.
  • The signal journal (roadmap step 3.1; counts only, the owner's answer to open question 6).
    • signal_journal.py (new) prints each signal's facts: when it was posted, its rules, the owner's ack/acted/snooze mark, and the group's end-of-day P/L at the mid, UNAVAILABLE for a day without a usable snapshot. It then prints the counts per rule in fixed words, never a money total.
    • journal-close. The Exit Sentinel's 17:30 ET routine closes each day into journal/outcomes/<ET date>.json and posts nothing.
  • The Aegis Events and Calendar Analyst (roadmap step 2.12): the sixth Floor seat, so the Floor is full. It comes with its role skill and the skill aegis-event-calendar-sourcing, and owns two Monday routines:
    • calendar-build (07:40): calendar_check reminders, then dividend_source proposals for the owner's tap. dividend_source.py now exits 10 when it proposes a dividend.
    • event-sheet (07:50): the decision calendar of the next two weeks.
    • calendar_check, dividend_source and decision_calendar become routine step tools with a postable message. Both profiles gain 2 runs a week (lean 54, active 74).
  • The Aegis Book & Risk Analyst (roadmap step 2.10): the fifth Floor seat, with its role skill and the skills aegis-exposures-limits-and-stress and aegis-vol-context, claimed by the template. It owns three routines:
    • opening-marks (10:20 ET, trading weekdays): the daily beta fit, then YOUR BOOK. YOUR BOOK now prints the beta-weighted delta and its r2 flag when the day's beta file holds them.
    • pre-event-stress (14:50 ET): scenario.py --upcoming 2, an event stress for each held symbol with an earnings or ex-dividend date within two trading days; quiet otherwise.
    • weekend-stress (Saturday 10:00): scenario.py --from-snapshot, the grid of the last end-of-day snapshot.
    • scenario.py gains --upcoming, --from-snapshot and a postable message, and beta_refresh and scenario become routine step tools. Both profiles gain 11 runs a week (lean 52, active 72).
  • Sourced cash dividends and the decision calendar (roadmap step 2.11; the owner's answer to open question 1).
    • One new read path. The market-data client gains /v1/corporate-actions (types=cash_dividend), the eighth allow-listed path (SECURITY.md).
    • dividend_source.py (new, read-only) compares the held symbols' declared dividends with research/calendar.toml and proposes the new or different ones.
    • The owner's tap. desk_init.py gains sourced_dividends answers. Run with --replace --owner-tapped, it rewrites only those ex_dividend lines and keeps a stated yield under the new yield_source key.
    • decision_calendar.py (new, read-only) lists the act-by instants, last trading days, ex-dividend and earnings dates in date order, and an unknown date as UNAVAILABLE.
  • Scenario and event stress (roadmap step 2.8). scenario.py, a new read-only tool, marks the book as book.py does and re-prices every open group through book.leg_price, the book's own leg pricer.
    • The grid. Spot -10 to +10 % crossed with volatility -5 to +5 points, 21 cells (the desk's proposal).
    • Event stresses. --event earnings|ex_dividend --symbol SYM prices the close of the first trading day after the event, without cutting spot by the dividend.
    • Output. A zero move equals the book. Each group's change is split into stock and options, so a hedge's offset shows. A group without usable inputs is excluded with its reason (exit 20). Nothing is written.
  • Exposures and two limit metrics (roadmap step 2.7; the owner's answer to open question 4).
    • Beta. aegis_desk.beta fits the OLS beta of daily simple returns against SPY over the 252 split-adjusted closes ending the previous trading day. Under 126 paired returns, or past a 40 % daily return in either series, it is UNAVAILABLE; r2_flag is ok at r2 >= 0.30 and low below.
    • beta_refresh.py (new, read-only) fetches those bars once a day with adjustment=split on the existing bars path and writes data/beta/<date>.json; exit 20 when a fit is UNAVAILABLE.
    • The book fills each underlying's net_delta, beta and beta_r2_flag and the portfolio's beta_weighted_delta in dollars from that file, and prints it on the portfolio line. It sends no extra request.
    • The exit rulebook gains beta_weighted_delta_pct and delta_exposure_pct, with the example rules BWD and UDELTA in rules/examples/limits-1.0.0.toml (both owner_must_set).
  • Stated cash and per-account approval levels (roadmap step 2.6; the owner's answers to open questions 5 and 7).
    • New account keys. Each [[owner.accounts]] entry may state its own approval_level, which the structuring tools read before [owner] approval_level. It may also state cash with cash_as_of, the two given together.
    • The cash check. aegis_desk.cash_check compares a ticket's or proposal's peak cash out with that stated cash. The peak is taken over the legs in the desk's fixed order and includes the cash a short put opened secures, less what a short put closed releases.
      • Tickets print it under the cash block and carry cash.check in JSON; proposals print it after the approval block.
      • A shortfall is a failed check item, never a refusal.
      • It reads UNAVAILABLE when no cash is stated, the figure is dated ahead or older than 5 trading days, or a leg has no quote side.
    • Retired values. engagement, chatty_or_quiet and attention_thresholds are optional and read by nothing. desk_init.py no longer requires them but still renders them when given, so an existing desk keeps its hash. It accepts the new account keys.
  • Weekly expirations (roadmap step 2.5; the owner's answer to open question 8). Every structure section of structuring.toml may set expirations = "weekly" (schema version 2; default monthly). Such a section adds every listed week-end expiration (the last trading day of its week: a Friday, or the Thursday when that Friday is a holiday) to the standard monthlies, and its dte window may span at most 21 days, so it holds at most four expirations and the four-chain cap stays as it is. The notes print expirations: weekly and monthly in your N to M day window, and desk_init.py takes the key. Daily expirations are never listed.
  • Step 2.3 landed: the structuring prices and the early-exercise line (the owner's answers to open questions 2 and 3).
    • Structuring prices. The Structuring Desk prices every candidate on the American model of record. Without a sourced yield an answer is UNAVAILABLE dividend_unknown, which replaces the q = 0 warning. The parity-carry estimate alone keeps European volatilities, the input of its own early-exercise bound.
    • Notes and model wording. The long put's american-style note: is retired. A short call's ex-dividend note and a short put's note read early-assignment note:. MODEL_LINE and MODEL_NOTE name the American model. The structuring texts say "dividend rate", because the structuring lint refuses the word yield.
    • The early-exercise line. New aegis_desk.early_exercise gives assignment_watch.py a line under each leg it lists, read by side: a long leg's exercise note or near-boundary line, a short leg's early-assignment warning, a long put's wait line when a stated dividend exceeds the interest on its strike, or why the line was not evaluated. It also gives the dividend note. A leg no tier lists is listed as EARLY EXERCISE <leg> when the lattice flags it, and --json adds early_exercise. The tiers, verdicts and exit codes do not change.
    • Doctor. The selftest gains an american model check: Hull's five-step put and an American volatility round trip, so the selftest now has six sub-checks.
    • Journeys. Four new journeys run on the new fixture root tests/fixtures/desk_american (the evening before an ex-date) and on the main root. They cover a dividend tier that stays HIGH where the lattice flags nothing, the wait line, the early-assignment warning, the band in the book, the exit rules' band guard, and dividend_unknown with NO_YIELD.
    • Oracle. The structuring-legs check compares with QuantLib's American engine.
    • Docs. Every "European-only" and "American pricing is planned" line is rewritten, and the roadmap marks 2.3 landed.
  • The book on the American model of record (roadmap step 2.3, third part; the owner's answers to open questions 2 and 3). Every option leg's IV inverts the American lattice price from its mid, and its Greeks come from the lattice (aegis_desk.american). Each leg gains model (american_crr_200), early_exercise_premium, greeks_quality, exercise_now and near_boundary. A leg in the boundary band keeps an approximate delta and has null gamma and theta. A group's net_greeks go Greek by Greek and gain a greeks_quality, so one band leg no longer blanks a net delta. A leg without a sourced yield is iv_status = "dividend_unknown" (no IV, no Greeks), which replaces the q = 0 warning in the book. An abs_delta rule threshold within 0.05 of a band delta reads UNAVAILABLE. Trigger levels' model levels solve the lattice delta with the sourced yield, and vol_context counts only days priced on the American model. The exit rules still read snapshots written before this change. The fixture book's American values move a few goldens: SPY's put IV 24.68 % to 24.59 %, net delta 278.84 to 278.67, and the MONETIZE level 572.79 to 575.48. The structuring tools still price European until the last part lands.
  • The dividend yield as a sourced fact (roadmap step 2.3, second part; the owner's answer to open question 3). research/calendar.toml is schema version 2: the yield is the yield of the symbol's ex_dividend entry (ex_dividend = { date = ..., amount = ..., yield = 0.0090, status = ..., source = ..., checked_utc = ... }), so it shares that entry's source, time and STALE check. Version 2 refuses a top-level dividend_yield; status none takes no yield and means zero. Version 1 files still load. calendar_events.sourced_yield says whether a yield is sourced (a version-2 yield, or status none); a bare version-1 yield is not. calendar_check.py gains the finding NO_YIELD for every held symbol without a sourced yield. desk_init.py writes version 2 and takes an optional yield with each dated ex-dividend. The fixture calendars are version 2. The book still prices with whichever yield the file holds until the wiring part lands.
  • The American leg pricer (roadmap step 2.3, first part; engine only, not wired). aegis_desk.american prices one leg on the owner's model of record: the lattice's price, delta, gamma and theta in desk units, vega and rho as central differences of the lattice price, the early-exercise premium and model = "american_crr_200". A leg in the boundary band keeps its price, vega, rho and an approximate delta (lattice.BAND_DELTA_TOL = 0.05); its gamma and theta are None. american.implied_vol inverts the lattice price (bounds: the undiscounted intrinsic value and the spot or strike; a quote on the intrinsic value has no single volatility), starting from the European volatility, which it never exceeds. The QuantLib oracle now asserts band delta within the tolerance, and the new pricer's vega, rho and implied volatility against QuantLib's finite-difference engine.
  • The owner's answers to the roadmap's open questions 1 to 8 (docs/ROADMAP.md). Step 2.3 comes next and 2.8 after it. The model of record is American everywhere, one value per leg; in the early-exercise band delta is marked approx within a tested 0.05, and gamma and theta are UNAVAILABLE. The dividend stays a continuous yield, moved into the sourced ex_dividend entry, with a note on every leg spanning an ex-date and a long put's wait line; step 2.2 (discrete dividends) is dropped. Weeklies narrow the dte window to 21 days under the same four-chain cap. Cash is optional per account, checked at the peak cash out and UNAVAILABLE after 5 trading days. Limits are two rule metrics in the exit rulebook. One GET-only path reads cash dividends. Journal and review totals are counts only. The planned t1-call and midday-limits routines are dropped (tests/engine/test_routines.py now fits fifteen ids), and three recorded desk.toml values are retired with step 2.6.
  • Volatility context (roadmap step 2.9). The new read-only tool vol_context.py (aegis_desk.vol_context) reads the desk's own end-of-day book snapshots: a symbol's value for a day is the median implied volatility of its open option legs with iv_status ok (never the vendor's Greeks), and the context places the latest day's value among the earlier recorded days (VOL CONTEXT | SPY | implied volatility 17.0 % on ... | 2 of 6 earlier days lower (33 %) | range ...). Fewer than 5 earlier days, no value on the latest day or no snapshot is UNAVAILABLE (exit 20). It is information only: a test proves no module but the tool and the doc generator imports it, so no rule, alert, ticket or ranking can read it. No network, nothing written; a journey and docs/TOOLS.md cover it. docs/ARCHITECTURE.md no longer lists the collar's roll analysis as planned (it landed in step 2.4).
  • A vertical's roll and call-spread coverage (roadmap step 2.4, the owner's answer to open question 9). exit_ticket.py --action ROLL on a vertical moves both legs to a later expiration at the same width: --new-strike names the new short leg and the long leg keeps the held distance. The four legs open the new long, close the held short, close the held long and open the new short. The approval convention gains the step open_spread_call (level 3): open a short call covered by a long call of the same expiry (a defined-risk spread). approval.Holdings.spread_calls counts short calls a same-expiry long call covers after the shares have covered what they can (holdings_of pairs them per expiry), approval.Leg.spread marks a spread's short call (opening, or closing so that its spread goes with it), and closing a spread's long call first is close_uncovers (level 4). The kind is replacement_vertical, judged under no structuring.toml section: the ticket's **The new vertical** block gives the width, the new vertical's net and the roll's net per share in place of checks, and the payload's replacement.held_other carries a vertical's held long leg (or a collar's held call). A vertical has no roll analysis; its NOT_APPLICABLE reason says that a vertical's roll is a ROLL ticket you name.
  • A collar's two legs replaced as a unit (roadmap step 2.4). roll_analysis.py --trade on a collar group ranks every pair of a later standard monthly inside [collar] dte and the floor and cap bands, judged as a fresh collar (its own net debit against max_net_debit, the strike floor, the call's events beyond the held call's life) and ranked by the section's objective; exit_ticket.py --action ROLL takes the new --new-call-strike beside --new-strike (the new put's). The kind is replacement_collar; the answer carries the held call as held_call beside held (a new key of the roll analysis and proposal payloads, null elsewhere). The four legs open the new put, close the held put, close the held call and open the new call (open_long 2, close_option 0 twice, open_covered_call 1), so the shares stay protected and covered at every step. The checks add net of the roll per share and name the new pair's own new collar net per share; with no best candidate both held legs' closes are quoted (the roll's net closes the held ...); the pointer asks for a ticket with its expiration and both strikes. A collar whose put and call differ in size is NOT_APPLICABLE. after_state closes more than one held leg, and its realized amount sums them. A vertical's other choices wait on the owner (the roadmap's open question 9: the approval convention counts only shares as a short call's cover).
  • A cash-secured put's short put: its roll analysis, a ROLL ticket and the early-assignment note (roadmap step 2.4). roll_analysis.py --trade on a cash_secured_put group ranks every put of a later standard monthly inside [roll] dte, and exit_ticket.py --action ROLL prices the one the owner names; both judge it under [roll] as a short call's roll is (the bands, min_credit, the roll test), with a put's time value bid - max(0, K - S) and the lower strike first among ties. strike_floor and ex_dividend are call rules and are not read; an earnings release inside the new put's life still counts. The held put closes at its ask first, then the new one opens at its bid (close_option level 0, open_short_put level 1), and the cash note says what each put secures (cash: the new put is secured by ... and the held put by ...; the desk does not see your cash balance). The kind is replacement_short_put, the header the roll analysis of the short <date> <strike> put and the ticket's a roll of the short put. structuring.short_put_caveat reads a long put's two tests (absolute delta 0.80 or more, or in the money with a time value at the mid below the interest on its strike) and prints early-assignment note: ... can be assigned before expiry, and the European model does not price that. An analysis of a group that holds no shares drops the your cost segment, and a ROLL ticket's checks caption names the new leg's right. Collars and verticals stay NOT_APPLICABLE.
  • The roll analysis of a protective put's long put (roadmap step 2.4, its first part). roll_analysis.py --trade on a protective_put group ranks every put of a later standard monthly inside [protective_put] dte, labelled against that section's criteria and ranked by its objective (stress_cost_per_day by default), with the same pricing as a ROLL ticket for the long put. The text reads the roll analysis of the long <date> <strike> put, the best candidate opens the new put at its ask first and then closes the held put at its bid (S7), each check carries the net cost per day added, and with no best candidate the held put's close is quoted (the net costs close the held ... put | bid ...). The approval-level check runs on those two legs before any market data (opening a long put is level 2 by the desk's convention). [protective_put] enabled = false makes it NOT_APPLICABLE disabled, naming that section. A short put, a collar and any other structure stay NOT_APPLICABLE, now with the roll analysis covers the short call of a covered call and the long put of a protective put. "over" joins the structuring text's template words (the replacement put's net-cost row already printed it). The roadmap marks step 2.4 in progress.
  • Release 1.0.0 (owner's decision): release_bump.py 1.0.0 moves the manifests, pins and template hashes from 0.3.0, which was never tagged. It is the first major release: the Stage 1 desk with the Stage 2 and Stage 3 roadmap and the roadmap's first engine step (below). No answer of the desk changes from 0.3.0.
  • Version bump to 0.3.0 (release_bump.py 0.3.0): v0.2.1 is tagged, so the skills this release changes are claimed at 0.3.0 (docs/RELEASE.md, step 8). It merges the desk-view work of the 0.2.2 branch (one message per answer, the book view, the brief's totals) with the structuring quant review.
  • One answer, one message. A chat client that draws a markdown table as a card of its own split every structuring answer into several bubbles. The Structuring Desk's desk view now prints no table: the context is one italic line under the title (*account 2 of desk.toml (ira) | SPY mid 648.12 | your cost 632.40 | quotes 19:31 UTC, 15-min delayed*), and the checks, the group now and after the legs, the cash, the approval steps, the ranked candidates and a ticket's key-value blocks are list items under capitalised captions (**Checks for this candidate**, - <green> **floor** 4.34 % below | your criteria: 3 to 10 %; **The group, now and after both legs**, - **max loss** 63,240.00 now, 1,165.00 after; **Candidates by rank**, - <red> **2** 2026-12-18 | new call 112.5 | debit 0.70 | extrinsic_per_day 0.0597 | fails credit_only). The reasons none passed have their own caption (**Why each one fails**), and the ticket pointer and the collar's outside-bands count are plain lines, so no list runs into another. The wording, the numbers, the JSON and the compliance lines are unchanged. approval.step_row is approval.step_item; structuring_text.table, status_cell, context_table, checks_table, group_table, candidates_table and cash_table are replaced by strip_line, status_mark, item_line, labelled_items, context_line, checks_block, check_item, group_block, candidate_items and cash_block.
  • The alert lint admits the three status circles at the start of a list item (- at the very start of a line, then the circle and a space) of a structuring text, beside the start of a table cell; an indented or * item, a quoted item, a circle inside a word or a number, and any circle under the rules or snapshot footer stay non_ascii.
  • The Morning Desk Sheet and the EOD Desk Brief are a desk view too: a ### title, bold section titles (**P/L**, **By trade**, **Desk status**, **To do**), items with their details nested under them (- **WMT covered call** +$1,145.00 at the mids (+$1,124.00 at liquidation) over - *AG-20261020-01 | today +$84.00 (row 3), -$50.00 (row 4)*; - **Routines** 1 of 7 ran today over the late, missed and pending slots), commands in code spans, the records read in one italic line at the end, and a blank line before the fixed footer so a chat never reads the footer as part of the last item. A gain carries its +. The numbers are the same; nothing is summed.
  • One total, one today, percentages and colours in the EOD brief and the book view. **Total** is the open positions' unrealized P/L plus what was closed this year, with its percentage of the open positions' cost and the liquidation figure, and an italic line of what it is made of and which groups it leaves out; **Today** is the book's day P/L with its percentage of the value at the start of the day (or the book's own reason it is unavailable). Each position reads - <colour> **WMT collar** +$7,634.14 (+17.99 %) | today ... with since entry, on a $42,436.22 cost | +$4,963.13 at liquidation under it, so its P/L says what it is measured from, and one day P/L per position instead of one per row. An unavailable position says which row had no usable quote and why (row 2 had no usable quote (grade E, stale)). The new aegis_desk.pl_figures computes every total, per-group day P/L and percentage from the book document in Decimal (a group's day P/L only when every open leg has one, never partially; no percentage on a credit or on a start value at or below zero); the book and its snapshot are unchanged. A gain is green, a loss red and an unavailable figure yellow: the alert lint now admits its three status circles at a list item's start under the rules footer too (never in a table cell, a nested item, the footer line or under the snapshot footer); fail_structuring_under_rules_footer now fails footer_missing alone.
  • book.py --view: the book of record as one linted markdown message for "how's my book?" and "analyze my positions". A ### YOUR BOOK <date> title with an italic line of the open positions and the quotes' source and time; **The whole book** (unrealized at the mids and at liquidation, today, realized, net delta, the groups the totals leave out); a #### <UNDERLYING> <structure> section per open group with its id and account in italics, its P/L and net cost, its payoff and one item per open leg in words (**2 short WMT 105 calls, 2026-11-20** cost $3.42, now $3.425 | P/L -$1.00 | today -$50.00, the quote grade, liquidation mark, IV and days left in an italic sub-item); then the session line and the fixed footer. An unbounded max profit names the shares the short calls leave uncovered (21.6301 of the 321.6301 shares carry no call), so no one subtracts. The account prints as account <n> of desk.toml (<type>) and never an OCC symbol. A view that fails its lint prints only WITHHELD: ... (exit 1 lint_failed); --json --view carries the message, its emitted document and the lint. The plain book.py output is unchanged.
  • The Aegis Desk Lead answers "how's my book?" with book.py --view and may post it, like the sheet, the brief and the opening bell, as one message, as it is, never in a code block, with at most one plain sentence before it and no number of its own. The Structuring Desk posts its answers as one message, never split. The agents, the role skills, the structuring, roll, proposals, ticket, threshold-level and incident skills, docs/STRUCTURING.md, docs/MONITORING.md, SECURITY.md, docs/TOOLS.md, the alert fixtures pass_structuring_* and fail_structuring_*, the CLI journeys (one more, for book.py --view) and the template hashes follow.
  • Multi-leg answers show what the quotes' spreads cost: a ticket of two legs or more adds spread cost against the mids to its cash block, and a collar proposal adds spread cost (half of both legs' spreads per share) to its checks. Each leg's JSON carries spot_used, the stock price its volatility was solved against, and a call's time value is now measured against that same price. The per-day objectives' JSON formulas say that T_c counts calendar days to the last trading day.
  • Threshold levels: the N sigma distance of a model level is now the lognormal measure |ln(level / S)| / (sigma x sqrt(tau)) instead of the arithmetic |level - S| / (sigma x sqrt(tau) x S), so a rise and a fall of the same ratio read the same. The fixture's SPY MONETIZE level reads 1.42 sigma (was 1.34); the NVDA DELTA_SIGNAL level is unchanged.
  • A ROLL ticket of a protective put's long put adds the row net cost per day added: (the new put's ask - the held put's bid) / S / (days added) x 100, in percent of the share price, so the cost reads net of what the held put still fetches. Candidates carry incremental_cost_per_day and days_added.
  • [roll] roll_test (schema_version = 2, in place of credit_only): "credit", "none" or "net_time_value", which fails a roll whose new call holds less time value than the held call (N = E_new - E_held < 0) and passes an in-the-money roll that credit_only would fail. Its checks row shows N per share and says when N lies within the held call's own quote spread. Candidates carry net_time_value in the JSON.
  • structuring.toml schema_version = 2: a section that ranks by band_centre may set its own centre in any band (dte_centre, abs_delta_centre, floor_pct_centre, cap_pct_centre, each inside its band) and band_weights (an inline table of numbers above 0 and at most 10 per band and for quote). The interview asks only where the owner chose band_centre; desk_init.py writes version 2 only when an answer uses these keys, so every version-1 file and its sha are unchanged. The answer names the owner's settings under the objective's words.
  • Structuring Desk: each expiration is priced at the carry its own put-call parity implies when those quotes rule out the calendar's dividend rate (with a quote-width bound and a one-sided early-exercise bound from a binomial tree); otherwise the calendar stands, with the reason in the JSON's new carry list. The text adds one carry: line per switched expiration after the model line. On the fixture the SPY March chain switches (call and put volatilities now agree within a few hundredths of a point instead of about 1.9 points apart), so the SPY group deltas read 70.17 and 23.75 instead of 69.69 and 24.14; the WMT and NVDA chains keep the calendar, and no best candidate changes. The book, the levels and the exit rules keep the calendar's rate.
  • Setup asks which objective ranks each structuring section. The interview (SETUP step 9) shows every objective a section accepts, with its words and the documented default first, and records the owner's choice; with no choice the default stays. An update (docs/UPDATING.md step 7) asks the same for an existing structuring.toml: keeping is the default and rewrites nothing; a switch is the owner's tap on desk_init.py --replace --owner-tapped and the acknowledgement.
  • Structuring Desk: two optional objectives. band_centre ([covered_call], [protective_put], [collar]) ranks by the distance from the centre of your bands, each band counted in its own half-width, plus each leg's quote spread counted in your max_spread_pct, all summed with equal weight; net_extrinsic_per_added_day ([roll]) ranks a replacement call by the time value it sells less the time value the roll buys back, per calendar day added when that is at least zero and as the whole amount given up otherwise. Each section still defaults to its release-1 objective, and a file that names it ranks exactly as before.
  • Quote precision, information only: an answer notes the ELIGIBLE candidates whose score lies within the best's own quote spread (rank 2 lies within the best's own quote spread on stress_cost_per_day: ...). The order and the best never change. The SPY protective-put example gains that line.
  • band_centre answers name the share prices, above and below, at which another candidate would be best, against one feed delay's move at the best's volatility. Information only.
  • The American caveat of a long put also applies to an in-the-money put whose time value at the mid is below the interest on its strike to expiry, K x (1 - e^(-r x tau)); the absolute-delta rule (0.80) is kept.
  • aegis_desk.carry (not yet wired): the carry an expiration's quotes imply by put-call parity, with a quote-resolution bound and a one-sided bound on the American early-exercise bias; the calendar yield is replaced only when the quotes exclude it.
  • Candidates carry objective_scores (every objective the section accepts, at the conservative side and at the mid) and a ranking carries near_best in JSON.
  • The roadmap (docs/ROADMAP.md, linked from the README's table of docs and from docs/ARCHITECTURE.md): every item the tree marks "planned for Stage 2" or "planned for Stage 3", ordered into steps with their prerequisites, the gates each must pass, the open owner questions and a ledger of the files each step rewrites when it lands. Work the roadmap adds on its own judgement is marked "proposed here". The ledger counts, per file, how many times each file names Stage 2 and Stage 3 (spaced, hyphenated or joined, even across a line break), and tests/gates/test_roadmap_ledger.py fails when the counts differ from the tree, when a counted file is in no item row, or when a listed path does not exist; a stage is complete when its column is empty. The comment in tests/engine/test_routines.py about the dropped roll-window routine no longer names a stage.
  • aegis_desk.lattice (not yet wired; step 2.1 of the roadmap): the American early-exercise lattice, a Cox-Ross-Rubinstein tree with a continuous dividend yield. It reports the American value with the closed form as a control variate, the early-exercise premium, delta, gamma and theta, exercise_now (only when the root and every node of the first two steps exercise, so a spot just inside the tree's exercise boundary, where a finer tree holds, is never an exercise signal) and near_boundary (the band where the tree cannot tell, and its Greeks are coarse). Tested against Hull's printed American put, the exact tree properties, carry.early_exercise_gap and the band; the QuantLib oracle compares it with QuantLib's finite-difference American engine on a grid and across the exercise boundary. No answer of the desk changes. The lines that said the lattice or an early-exercise model was planned now say what is still planned (American pricing, American Greeks, or early-exercise analysis beside the expiry tiers), and those that name the lattice say its engine is in the tree: docs/FAQ.md, docs/PROVENANCE.md, docs/STRUCTURING.md, docs/ARCHITECTURE.md (which also lists the lattice among the engine's modules), agents/position-control.md and its generated role skill, the covered-call, expiry-and-assignment, book-marking and roll-decision skills (their template hashes move with them) and the structuring model_note, whose new wording counsel signs before the tag; "lattice" leaves the structuring text's template word list with it.

0.2.1 (2026-11-13)

  • Version bump to 0.2.1 (release_bump.py 0.2.1).
  • The alert lint's fixed leg grammar takes one leading - list mark, and a leg's quote segment may end at its grade (| bid P, ask P, mid P | quote grade G) when the text names its quote source once in its header: the groundwork for the Structuring Desk's easier-to-read layout. Any other list mark, an indent or a free tail after a leg is still order_field.
  • The Structuring Desk's answers are now a desk view in markdown, built to read in a chat on a phone instead of long |-joined lines. A roll analysis, a proposal and a ticket each print a ### title; a context strip (the account, the underlying's mid, your cost or the ticket's times, the quotes' time); the legs as list items in the fixed leg grammar (never a table with order columns); the caveat as a quote; a checks table with each criterion's value, your band and a status column; the group now and after the legs; the approval-level check with a table of the steps; the ranked candidates with their status; and the notes as a quoted list (the objective's words, the quote source named once, the model line, the criteria sha, the limit and act lines), then a blank line and the footer. Amounts carry thousands separators, the best candidate's rank and score are bold, and a net prints in words (credit 0.05 (10.00 in all), debit 0.70) instead of a negative net credit. The Aegis Structuring Desk posts the text as it is, never in a code block.
  • The status column: a green circle (U+1F7E2) for a criterion that passes, a yellow circle (U+1F7E1) for an earnings or ex-dividend label you accepted (accepted by #assign-ok, #earnings-ok or your structuring.toml), a red circle (U+1F534) with the keys a candidate fails. An answer with no best candidate lists under its candidates table why each shown candidate fails (1, 2, 3 fail ex_dividend: ..., 4 not priced: ...).
  • A proposal whose contracts cover fewer shares than the group holds says so under the group table (3 contracts cover 300 of your 321.6301 shares; the other 21.6301 are not part of the collar), which explains an unbounded max profit on an odd lot.
  • Tickets: the cash at the conservative prices and at the mids, the realized P/L of a closing leg and the most the steps need at once form one table; a LET_ASSIGN ticket's short options and a MONETIZE_PUT ticket's puts are key-value tables; the act line reads if you decide to act, you type it into your brokerage account yourself, beside the unchanged you supply the limit; the desk prints none. approval.approval_lines returns the heading line and a steps table (leg_order_line and step_line are gone).
  • The alert lint's one exception to plain ASCII: the three status circles, only at the start of a table cell (| before, a space after) of a structuring text and never in its footer; anywhere else they are non_ascii. The structuring-content check also finds a first line behind ### or **. check.sh step 3 exempts the three circles from the docs emoji gate only inside a closed fenced block that quotes the desk's output.
  • The alert fixtures pass_structuring_* are regenerated and the fail_structuring_* fixtures re-derived from the desk view (fail_structuring_under_rules_footer now also fails non_ascii); the CLI journeys match the new text. The guide's examples (docs/STRUCTURING.md), the agent's "Reading the output", the roll, proposals, ticket and role skills, SECURITY.md, CONTRIBUTING.md, docs/POSITIONS_CSV.md and the counsel items in docs/RELEASE.md follow the desk view; the template's skill hashes are updated.
  • 4 more tests, 3108 in all (floor 3108).
  • One front door. The owner talks to one Bot, the Aegis Desk Lead, in its 1:1 chat, for setup and everything after it: questions, position changes, plan cards, tickets, roll analyses, proposals, the criteria interview and re-acknowledgements. The Lead runs each seat's tools by that seat's agent file and enables the skills they need; no Bot sends the owner to another chat. The Ask-first card is the confirmation (no typed "approve", "re-ack" or "run it" before a card), the owner-authority commands of one request run as one &&-joined command line so one tap covers them, a trade id is found from the symbol instead of copied by the owner, one Bot answers the owner in the Floor, and Bots other than the Lead stay quiet until desk.toml exists (no tool on the empty desk, no routine before SETUP section 9.3). Saving positions.csv joins the owner-authority Auto-review rule (391 of 1000 characters), so the save's card replaces the typed approval. Ticket texts now say after a fill, tell the Aegis Desk Lead the fill and ask the Aegis Desk Lead for a ROLL ticket.
  • Options quoted minutes before the stock are priced, not dropped. On a live paper account the indicative feed updated thinly traded WMT options minutes before the IEX stock quote, and every collar candidate came back UNPRICED spot_unaligned. The Structuring Desk now solves such an option's implied volatility against the stock's one-minute IEX bar of the minute the option was quoted in (alpaca_client.minute_bars, timeframe=1Min on the allow-listed bars path, cached by quote_cache.get_minute_bars, one request), and reads its Greeks at the current spot; the text adds spot: <n> option legs quoted more than 60 seconds from the stock quote were priced against the stock's one-minute IEX bar of the minute each was quoted in, and the leg JSON carries spot_bar_utc. An option with no bar for its minute stays unpriced. A stock quote reused from an earlier run's cache but stamped more than 60 seconds before the chains is refetched once first (structuring.align_spot). PROVENANCE's open item is closed.
  • Earnings dates through the intake. desk_init.py takes an earnings answer ({date, time pre | post | unknown, status confirmed | estimated} or {status none, source}) beside ex_dividends and writes both into research/calendar.toml; SETUP 9.1 asks for each held symbol's next earnings date, and the new phrase "update the calendar" (SETUP 9.5) adds a date later through one card (--replace --owner-tapped joined to ack.py config).
  • Criteria that can never pass are said up front. structuring_config.never_passing_notes names each call-writing section whose dte window starts past 100 calendar days with earnings or ex_dividend = exclude (the documented collar defaults): every such call spans a quarterly report or ex-date, so no candidate on a reporting or dividend-paying stock passes. desk_init.py prints them as note: lines and in its payload's new notes; the interview asks the owner to choose. Past such a failure the Bots offer allow in structuring.toml (structuring only, no alert silenced) before the row tags, whose alert silencing they now state.
  • Plainer first-run lines. A new desk's first adoption posts rulebook adopted: your first rulebook, no earlier verdicts to compare | rules X.Y.Z, sha ... instead of rulebook changed: digest unavailable (no snapshot); the morning sheet says last night's levels: none yet (the end-of-day level sheet has not run on this desk) instead of an UNAVAILABLE line while no level sheet exists; positions_check.py prints a group's net cost to the cent (the JSON keeps every digit). The charter adds two rules: no date or dividend from memory (UNAVAILABLE, never "has not declared one") and one clock (ET as the tools print it, never a converted time beside it).
  • Updating a desk already set up. docs/UPDATING.md, the README's "Update a desk you already set up" prompt and SETUP 9.6: the changelog in plain words, the checkout moved to the new tag after the owner's tap, the gate, each Bot's skills refreshed with a readback, any changed Auto-review text re-pasted, the routines re-checked and an update receipt; the desk files are never rewritten, and the page lists what 0.2.1 changes for an existing desk.
  • 12 more tests, 3120 in all (floor 3120).

0.2.0 (2026-11-13)

  • Version bump to 0.2.0 (release_bump.py 0.2.0).
  • Market data now comes from Alpaca, read with a paper key; Tradier support is removed. The owner's one first-run action is a free Alpaca paper key pair (a key ID and its secret, from a paper-only account opened with an email address), added through the secure secret card as APCA_API_KEY_ID and APCA_API_SECRET_KEY, never pasted in chat (SETUP section 7, the bootstrap skill, the FAQ). The fallback is the 0600 file /workspace/.aegis-secrets/alpaca.key (key ID on line 1, secret on line 2); opening_bell.py --token-file becomes --credentials-file.
  • The one network module is now scripts/aegis_desk/alpaca_client.py (check_no_write_paths.NETWORK_CLIENT), with every earlier safety property kept: one literal GET with no body, redirects refused, no proxy handler, TLS 1.2 or later, a response-size cap, redaction of both the key ID and the secret, and refusal of an answer that echoes either (a non-200 body that echoes either in JSON escapes is withheld whole). Two hosts are pinned: data.alpaca.markets (/v2/stocks/quotes/latest, /v2/stocks/bars, /v1beta1/options/quotes/latest, /v1beta1/options/snapshots/<UNDERLYING>) and paper-api.alpaca.markets (only /v2/clock, /v2/calendar and /v2/options/contracts). The live trading host api.alpaca.markets is refused. Answers are paginated (next_page_token, at most 10 pages, each counted as a request); new statuses no_credential, credential_file_mode, subscription (a request the free plan does not cover) and incomplete replace no_token and token_file_mode.
  • The free plan's feeds, labelled: stocks from IEX; options from Alpaca's indicative feed, delayed 15 minutes and not exchange (OPRA) quotes. Every quote segment, footer, alert, structuring text and the Opening Bell's snapshot line now name the source from one constant (aegis_desk.data_source): Alpaca paper (IEX stocks, indicative options: approximate). options_feed = "opra" is accepted for a subscriber. feed_delay_min stays 15, the conservative bound, and the indicative feed requires at least 15. The Structuring Desk's disclosure text now says the quotes are delayed Alpaca paper quotes, the option quotes approximate (its id is unchanged).
  • desk.toml [market_data]: vendor = "alpaca", host = "paper" (any other host, live included, fails at load), key_env, secret_env, optional credentials_file, options_feed, stock_feed = "iex"; token_env and token_file are gone. key_env and secret_env must be upper-case variable names with an underscore (APCA_API_KEY_ID), one rule shared with desk_init.py, so a pasted key ID or secret is refused and never shown. desk_init.py writes the new block and refuses a live host (live_host).
  • The desk doctor: check 4 is key_source (names the two variable names or the file, never a value); check 5 is paper_key, one GET of /v2/clock on the paper trading host (which refuses a key from another environment): paper key verified, or a 401/403 FAIL not a paper key: remove it and add a paper key. Check 7 exempts exactly the configured key and secret names from the broker-name rule (any other ALPACA_* name still fails; a configured APCA_LIVE_* name fails too) and fails on an APCA_API_BASE_URL that names the live trading host.
  • The live Opening Bell reads its contract listing on the paper trading host, so a key that host refuses after the market-data host answered it is UNAVAILABLE: not a paper key: ... (exit 20, not_paper_key): a live key fails the first run's check.
  • Quote grading: an ask of 0 under a positive bid (Alpaca's "no active offer") is a missing ask, ask missing (no offer), grade F, never a crossed market.
  • The quote cache counts every request, not every call: a quotes batch with stocks and options is 2 requests, a chain at least 2 (the contract listing, then the snapshots). The desk-wide window is 200 requests a minute (the free plan's market-data limit) minus a headroom of 20; no new fetch starts once a run has made 40 requests (a chain only while its minimum of 2 still fits; a paginated answer can finish past 40), and a run fetches at most 4 chains. A symbol the vendor did not quote is fetch_failed:not_returned. Index options stay unquotable, now worded "the data source cannot quote index options".
  • The offline fixtures moved to tests/fixtures/desk*/fixtures/alpaca/ in Alpaca's response shapes, generated from the same synthetic values, so every price, size and stamp golden holds; request counts and the desk.toml hash in the journeys changed.
  • The write-side Ask-first Auto-review rule now names Alpaca: any request to a brokerage's order, position or account endpoints, including any request to api.alpaca.markets, and any request to paper-api.alpaca.markets other than /v2/clock, /v2/calendar and /v2/options/contracts (487 characters; the Allow rule says "GET-only Alpaca paper client"). Paste the reprinted rules (routine_audit.py --rules-text) again. The Grok Build deny rules name the Alpaca hosts.
  • The broker-write hook no longer exempts any market-data read on a broker host: an HTTP call that names a broker host is denied, reads included (the desk's tools read market data from Python, never through a shell or an MCP tool). The replaced vendor's host left the hook's broker-host list; a write to any account or order path is still denied on every host.
  • 3103 tests in all (floor 3103).
  • Every document, skill, agent and role skill is rewritten for Alpaca, including the FAQ's paper-key answer, the free plan's limits and the personal-use data licence note.

0.1.1 (2026-11-13)

  • Onboarding on Grok Bot is staged. The first run (SETUP sections 1 to 8) asks the owner for one thing: a free Tradier sandbox market-data token, added through the secure secret card, never pasted in chat. It shows a demo on a copy of the sample desk before anything is asked, then returns a receipt that opens with three plain lines and ends with Setup stayed read-only: one sandbox market-data token was added by you; no order was created or sent. Everything else waits for a phrase the owner says, each a part of SETUP section 9:

    • "add my positions" (9.1): first the two Ask-first Auto-review rules and the connected-apps readback (ack.py connectors --owner-tapped), because desk_init.py follows; then four desk questions with the other settings proposed for one "yes"; Aegis Position Control drafts the positions.csv rows from the owner's own words and saves them once the owner approves; dividends for held symbols; the full doctor;
    • "adopt my exit rules" (9.2): one answer (keep the examples, or name those to change or drop), proposed defaults for DAILY and CONC, and --allow-rth offered during regular trading hours;
    • "turn on monitoring" (9.3): the routine capability probe, Notifications and Timezone, the Allow rule for routine.py, the safe.directory fix if the doctor asks, routines created paused, one Test run, unpause;
    • "set up the structuring desk" (9.4, optional): the structuring interview. "SETUP step 9" still names the step that holds the structuring criteria.

    9.1 comes first: whichever phrase the owner says first, its two safety settings come before any owner-authority command and any routine (SETUP section 9, the bootstrap skill, agents/desk-lead.md, agents/structuring-desk.md, docs/STRUCTURING.md, the README, the FAQ and SECURITY.md). 9.2 needs 9.1's files, and 9.3 an adopted rulebook.

  • The first run's edge cases: when git refuses the checkout, the routine.py demo is skipped until the safe.directory fix of 9.3 (never --allow-dirty); the demo's CONFIG CHANGED line is named as the sample desk's; the live Opening Bell takes --token-file with the file fallback; a token that fails with auth: (often the production token pasted by mistake) is deleted before anything else; and the receipt's last line says a sandbox token was added only when the live bell succeeded.

  • Aegis Position Control may save rows it drafted from the owner's plain description, once the owner approves the exact text shown, byte for byte, with only the owner's numbers (agents/position-control.md, the role skill, aegis-positions-csv-book-of-record, docs/POSITIONS_CSV.md).

  • SETUP.md is rewritten for that shape (516 lines to 269). The bootstrap skill aegis-desk-bootstrap-and-onboarding (metadata version 0.2.0) now covers only the first run and points to SETUP section 9 for the rest.

  • scripts/opening_bell.py --demo: the Opening Bell on the checkout's own sample desk at its fixture instant (--offline --desk-root <checkout>/tests/fixtures/desk --now 2026-11-03T19:50:00Z, SPY by default). It writes nothing, reads no credential, and its first line says the data is a synthetic sample, not market prices. With --desk-root, --now, --fixture-set or --token-file it is a usage error (exit 2). The JSON payload gains demo (true or false). A new journey runs it exactly as published (the Journey record gains checkout_only, so it never runs against another tree). 8 more tests.

  • README.md is rewritten for a first-time reader: see it, try it, start it. A hero image, three badges, a "Try it now" with three commands ending in --demo, one Start prompt and its phrases, "Meet the desk", "A day on the desk", "Controls", the runtimes and the repository map. The anchors other pages link to (#start, #try-it-now, #meet-the-desk, #controls) are kept.

  • assets/: five original SVGs (the hero and the desk flow, each with a phone-width variant, and a terminal card of the real --demo output), built by the standard-library script assets/src/build_graphics.py, whose --check compares them with the committed files. tests/gates/test_readme_assets.py runs that check, holds every text colour at 4.5:1 contrast or more, and requires a markdown link to each SVG (in assets/src/README.md) so the link gate sees them. 4 more tests, 3052 in all (floor 3052).

  • Docs follow the staged path:

    • docs/FAQ.md answers "What does setup ask me for?" and "How do I get a Tradier sandbox token?", and marks unconfirmed facts as such;
    • SECURITY.md says when each Grok Bot control arrives, and is honest about the window before "add my positions": the sandbox token could place sandbox paper orders, while the repository code still has no write path;
    • docs/aegis-grok-bot.md describes the first run and the phrases;
    • the runtime pages drop "the tag is created at release";
    • docs/GROK_BOT_TEMPLATE.md, docs/STRUCTURING.md, docs/EXIT_RULES.md, skills/README.md and the anchors in docs/RELEASE.md follow the new SETUP numbering.
  • Skills aegis-routines-and-quiet-rules, aegis-expiry-and-assignment-protocol and aegis-rules-dsl-and-change-control name the phrase that starts their part of setup; template/grok-bot.json records the new sha256 of each, of the bootstrap skill, of aegis-positions-csv-book-of-record and of the three role skills regenerated from the edited agents.

  • Version bump to 0.1.1 (release_bump.py 0.1.1): every manifest, marketplace ref, tag pin, template hash, role skill and tool version moves together.

0.1.0 (2026-11-12)

  • Release workflow: .github/workflows/release.yml tags main and publishes the GitHub Release. It runs only when someone with write access starts it by hand (Actions, Release, Run workflow, on main), with the version (as in plugin.json) and a pre-release switch. A read-only gate job, which keeps no credential in its checkout, repeats the gate of docs/RELEASE.md step 5: the version matches plugin.json and has a section here, and bash scripts/check.sh and release_bump.py <version> --check pass. Only then does the release job, the one job with write access (contents: write), create the annotated tag on the same commit, push it, run check_grok_template.py and publish the Release with this section as its notes. A re-run after a failure reuses a tag already on that commit and an existing Release, and refuses a tag on any other commit. The inputs reach the shell only through env:. tests/gates/test_release_workflow.py pins these properties (12 more tests, 3040 in all; floor 3040).
  • A short option's act-by (FAI-689): the expiry checklist (the FINAL CALL, the T-1 preview and the PIN lines), the assignment watch and the morning sheet print a short option leg's act-by as the session close of its last trading day (16:00 ET; 13:00 ET on an early close), or broker_exercise_cutoff_et when that is earlier, because a short option is closed only while the market trades; its text ends ; the session close: a short option is closed only while the market trades and carries no early-close mark. A long leg keeps the broker cutoff, with the early-close rule and mark (it governs exercise and do-not-exercise instructions):
    • one helper, expiry.leg_act_by, serves the expiry sheet and the LET_ASSIGN ticket, whose printed text is unchanged;
    • ActBy.basis and the leg JSON's act_by_basis (session_close or broker_cutoff) name the instant that governed, described by the --schema of assignment_watch.py and expiry_checklist.py;
    • the FINAL CALL header names the earliest act-by of the listed legs when they differ, each leg line keeping its own;
    • each docket line of the Morning Desk Sheet and the EOD brief names its contract, side and size first (<trade_id> <OCC symbol> <long|short> <contracts> | ...), since the legs of one group can act by different times; a docket record whose legs lack them is not the record the desk writes;
    • 15 more tests, 3028 in all (floor 3028).
  • The Aegis Structuring Desk joins Stage 1 as a fourth Bot on the Floor. It answers only the owner's own request, never runs in a routine and never fires an alert:
    • scripts/exit_ticket.py (--action CLOSE, UNWIND_STOCK, LET_ASSIGN, MONETIZE_PUT, or ROLL with --new-expiration and --new-strike), scripts/roll_analysis.py (the replacements for a covered call's short call) and scripts/propose.py (a covered call, protective put or collar on a group that holds shares only), on the library scripts/aegis_desk/structuring.py, structuring_text.py and tickets.py: candidates from the held underlying's own monthly chains, priced with the European model and labelled as such, labelled against the owner's criteria and ranked by the objective the owner declared; one best candidate when one passes every criterion; tickets with the legs in a fixed order, their quotes, the cash at the conservative side and at the mid, and no order fields; no positions.csv lines (Position Control proposes them after a fill);
    • the owner's criteria structuring.toml and its strict reader (scripts/aegis_desk/structuring_config.py), rendered by desk_init.py from the structuring interview with its one-time disclosure; it joins the CONFIG CHANGED set acknowledged by ack.py config, and the structuring tools answer UNAVAILABLE while desk.toml, positions.csv or structuring.toml are unacknowledged, or while any group of positions.csv is quarantined;
    • the approval-level check (scripts/aegis_desk/approval.py), the first reader of desk.toml's self-attested approval_level, in the desk's convention;
    • a third footer and the alert lint's structuring mode (alert_lint.py --footer structuring), with a fixed leg grammar; the rules and market-snapshot footers refuse a leg reference, a ranked candidate or a structuring answer;
    • quote_cache.get_expirations, cached per ET date (an empty list is never cached), and synthetic structuring chains in the fixture desk;
    • routine.py refuses the three tools as routine steps; the owner-authority Auto-review text names them (333 of 1000 characters), so each run asks first;
    • tickets/, reports/rolls/, reports/proposals/ and the audit journal journal/structuring.jsonl;
    • the agent agents/structuring-desk.md and its generated role skill; the skills aegis-overlay-proposals, aegis-roll-decision-framework and aegis-ticket-and-unwind-order; docs/STRUCTURING.md, the owner and counsel reference; the provenance of the criteria defaults and the three objectives (docs/PROVENANCE.md); SETUP step 9's structuring criteria; the README, FAQ, monitoring, security and runtime pages for four Bots;
    • 429 more tests, 3013 in all (floor 3013).
  • Release tooling:
    • scripts/release_bump.py <next> [--check] moves the version in plugin.json, every mirrored manifest, the marketplace refs, the template's release and skill hashes, every pin check.sh step 4 checks, the role skills and the tool versions in docs/TOOLS.md, in one pass, writing every file or none; it refuses a version that is not plain or not greater, a tag that already exists (read from .git as files, worktrees included; a .git it cannot read refuses the run) and a tree that is not consistent at its version, never edits CHANGELOG.md or docs/RELEASE.md, and lists plain mentions of the old version for review;
    • the stub MCP server of the release deny check, tests/fixtures/release/blackhole_mcp.py (one tool that does nothing), and its --mcp-config file;
    • docs/RELEASE.md, the 0.1.0 checklist and receipt, with who does each step and empty evidence rows;
    • 82 more tests, 2584 in all (floor 2584).
  • Stage-1 content:
    • four role profiles (agents/desk-lead.md, agents/exit-sentinel.md, agents/position-control.md and, with the Structuring Desk, agents/structuring-desk.md) and their four generated role skills; eighteen Stage-1 skills, each with its LICENSE and ATTRIBUTION.md (one adapted: the bootstrap skill, which carries its notice line in its body);
    • the broker-write gate hooks/broker_write_gate.py, a backstop that denies a shell command or MCP tool call that looks like a broker write and gives no opinion on reads, registered by hooks/hooks.json (Claude Code and Grok Build) and hooks/cursor-hooks.json (Cursor, fail-closed, answering ask when it has no opinion); check_manifests.py asserts the wiring exactly, with no other key at any level;
    • check.sh step 2 requires an adapted skill's in-body notice line; ack.py connectors --owner-tapped records the owner's readback of connected apps and MCP servers, which the doctor re-asks after 30 days; routine.py's --desk-root help names its /workspace/aegis-desk default;
    • manifest descriptions of the Stage-1 desk without counts; the Cursor manifest's rules names rules/aegis-desk.mdc;
    • template/grok-bot.json claims every skill and role skill by its full-file sha256 and exports the bootstrap skill; its status stays unpublished until the owner publishes it after the tag;
    • SETUP.md (the Stage-1 runbook), the README, SECURITY.md, the architecture, FAQ, positions, monitoring and provenance pages, the page on your exit rules, the runtime pages for Grok Bot, Grok Build (with copy-paste [permission] deny rules for ~/.grok/config.toml), Claude Code, Cursor, Codex and Gemini CLI, the template page, the skill index and the attribution table;
    • 2502 tests (floor 2502).
  • Counsel review before the public release (an owner action; nothing here is signed off):
    • the alert texts (EXIT SIGNAL, WATCH, UNAVAILABLE, CONFIG CHANGED, the pre-close and watch-alive lines) and both fixed footers;
    • the threshold-level sheet and the exit-plan card, which carry no order shape ([levels] order_shape_enabled stays false);
    • quoting diagnostic output: a Bot quotes it only when the owner asks, unchanged, in a fenced block headed by the exact command, and the proposed positions.csv lines of csv_recon.py --diff appear only in Aegis Position Control's chat;
    • the Structuring Desk (rows 1 to 9 of docs/RELEASE.md step 4): the structuring footer and the disclosure, the texts of the tickets, the roll analyses and the proposals, the approval convention and the REFUSED line, the criteria defaults and the interview, the FAQ answer and the README description, whether publishing software that names a best candidate raises investment-adviser questions, and whether the release may advertise the Structuring Desk.
  • Levels, routines and the doctor:
    • the threshold-level sheet (levels.py, trigger_levels.py): each adopted rule as an option or spread price, an underlying level, a band or a date, with its arithmetic shown, rounded on the class's tick grid in the direction that keeps a resting take-profit inside the rule, take-profit levels marked resting-eligible and loss levels alert-only, and exit-plan cards;
    • the routine manifest (routines/manifest.toml, routines.py) with the Stage-1 routines, their posting policies, dependencies and weekly budget (lean 41, active 61 runs), the Grok Bot Auto-review texts (routines/autoreview.toml) and routine_audit.py;
    • routine.py, the single reviewed routine entry point: it refuses to run a step on a checkout that is not the pinned release, runs the chain in-process, lints every part and the whole message, and writes the heartbeat, journal and run record;
    • the desk doctor (doctor.py, desk_doctor.py): twelve checks, including sandbox-token verification and broker-looking environment names;
    • the Morning Desk Sheet and EOD Desk Brief (summaries.py, eod_summary.py), including P/L per open trade;
    • desk_init.py, which renders desk.toml byte for byte from the interview;
    • opening_bell.py, with a market-snapshot footer that the lint accepts only on rule-free text;
    • gen_tools_doc.py (docs/TOOLS.md) and gen_role_skills.py (the role skills), both drift-checked by check.sh;
    • csv_recon and expiry_checklist now write their records (data/positions/<date>.recon.json, reports/expiry/<date>.<mode>.json);
    • the no-write gate also rejects loader and reflection routes, laundering through the subprocess-allowed modules, git configuration through arguments or the environment, and environment writes;
    • two placeholder agents; twenty-four more journeys (81 in all).
  • Rules engine: the strict exit-rulebook grammar with parameters bounded in their safe direction, the metric catalogue on the book's fields, Kleene evaluation and the verdict precedence EXIT_SIGNAL > UNAVAILABLE > WATCH > HOLD, the adoption lint and per-rule shas (scripts/aegis_desk/exit_rules.py); the per-(group, rule) state machine with time-based confirmations on quote stamps, bounded re-arm, a cooldown that never holds a re-arm past the next session's open, first observation and carry-forward (rule_state.py); the alert texts with their fixed footer and emitted-number documents (alerts.py), the lint that proves every number and identifier was emitted and that no trade language appears (alert_lint.py), the signal and alert journals (signals.py); the tools exit_check.py, rules_diff.py, rules_adopt.py (owner-tapped, immutable versions, refused in regular trading hours unless --allow-rth), book_check.py (posting rules, re-post ladder, CONFIG CHANGED first, dry run without --post), ack.py and alert_lint.py; the example rulebook rules/examples/exit_rules-1.0.0.toml (each rule illustrative until adopted), adopted rulebooks in both fixture roots, red-team lint fixtures; twenty-eight more tool journeys.
  • Book follow-up: every book leg carries fresh_bid and fresh_ask (the liquidation sides of a grade A-E quote; a zero ask stays 0, no offer), and extrinsic_bid now uses the assignment watch's geometry (any grade A-E fresh bid, the usable underlying mid stamped within 60 s, the strike as a plain decimal), so the book and the expiry sheet agree on every quote state; mark_source is mid only with both sides fresh. quote_cache.py and quote_qc.py classify the session with the research calendar's full closures like the book tools (quote_qc.py also takes its prior close from it; an unparsable calendar is exit 1 calendar_invalid). The command-journey step test lists its temporary directories under a private root, so concurrent local runs cannot collide.
  • Book of record: the hand-edited positions.csv parser with the desk's money conventions (scripts/aegis_desk/positions_csv.py), strategy grouping, the Decimal payoff and P/L kernels, the book (book.py: one quotes request, quality grades, IV and Greeks, group metrics, portfolio totals that name every excluded group, the end-of-day snapshot and report; the session classified with the calendar's full closures), the reconciliation engine with every settlement branch (recon.py), the owner-sourced research calendar reader with its freshness lint (calendar_events.py), the expiry sheet with dividend and pin tiers (expiry.py: a dividend tier only while the contract is alive on the last cum-dividend day; the act-by instant is the earlier of the broker cutoff and an early close), daily-bar history in the quote cache; the tools positions_check.py, csv_recon.py, book.py, calendar_check.py, assignment_watch.py and expiry_checklist.py; a second fixture root tests/fixtures/desk_recon/ covering every settlement branch and three more fixture sets (T-1, expiry day, post-expiry); seventeen more tool journeys, including multi-step chains on one fixture copy.
  • Data plane: the GET-only Tradier market-data client (scripts/aegis_desk/tradier_client.py: pinned hosts, allow-listed paths, literal GET and no body on every request, no redirects, no proxy, TLS 1.2 or later, one receipt per request, credential redaction, refusal of a 200 that echoes the credential, loopback fake-server tests asserting the method on every request) with an offline transport over the recorded fixture desk tests/fixtures/desk/ (five sets: regular, stale, after the close, before the open, holiday); the quote cache with the desk-wide 60-per-minute ledger, per-run caps and a run-wide stop after a quota or credential error; quote quality grades with session-aware freshness; heartbeat records; shared vendor types, desk paths and the strict desk.toml reader; the tools quote_cache.py, quote_qc.py, heartbeat.py and market_clock.py; --now on the shared command-line contract; nine tool journeys. The no-write gate also rejects subclasses of urllib.request.Request, assignment of a request's _data or target fields, set_proxy, and reflection into third-party HTTP clients.
  • Repository gates: scripts/check.sh (structure, Python 3.9 floor, version pins, test-count floor, command journeys), check_manifests.py, check_no_write_paths.py (AST no-write gate and prose scanner), check_grok_template.py, cli_journeys.py; CI on Python 3.9, 3.10, 3.12 and 3.13 plus the QuantLib oracle workflow; negative fixture trees for every gate.
  • Engine library scripts/aegis_desk/: Black-Scholes-Merton pricing, implied volatility, desk-unit Greeks and units ported from the Aegis quant engine with their golden values; NYSE market clock with embedded 2026-2027 calendar and DST rules; OCC symbol parser; strict vendored TOML reader; atomic state and file locking; the shared command-line contract and exit-code table.
  • Repository scaffold: manifests, licences, placeholder role and skill, documentation skeletons.