tracking: current context index — milestone, critical path, what an agent may pick up

[atomchung]

Status

Daily control plane — refreshed 2026-09-02 (post-merge) against main@8c37b04012d4e54dcb9f063f079def2ddd510e78.

This issue is the mutable routing surface for the current milestone. GitHub open state is not an execution queue; lifecycle and claim semantics remain in docs/issue-lifecycle.md.

Current milestone

M1 decision usefulness — concise expression without compressing judgment.

The supported vertical slice remains pre-trade review / challenge: engine-owned portfolio consequences, sourced current facts, contextual judgment, an actionable recommendation, honest delivery, and canonical persistence only after explicit intent. Missing inputs narrow claims, durable writes, or external actions; they do not disable research, reasoning, recommendation, or useful presentation.

Completed evidence:

• [implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 / PR Move limits to claim, write, and action boundaries #828 moved remaining whole-turn limits to claim, write, and action boundaries. Its frozen actual-Skill owner blind test preferred the candidate in 3 of 4 scenes, with no integrity finding.
• [design·M1] Owner-live: answers exceed the reading budget — verbosity is the next usefulness bottleneck after #827 #830 / PR feat(challenge): delete the obligation whitelist, keep the seven facts that earn their place (closes #830) #831 removed obligation recitals; [design·M1] One communication method: the consulting pyramid as the mandatory top layer; per-surface derivation optional and additive-only #832 / PR One communication method: the consulting pyramid as the mandatory top layer, per-surface derivation additive-only (closes #832) #833 established one answer pyramid; [implementation·skill] Put one canonical exemplar into each surface's generation path — examples-first, closing #832's residual force #834 / PR Put one canonical exemplar into each surface's generation path (closes #834) #835 put canonical exemplars on the generation path. The frozen four-scene rerun then materially shortened every target answer.
• [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 remains the owner-live product-acceptance authority for one real consider decision. It is unblocked but still not accepted; green CI and synthetic expression witnesses do not substitute for that walk or its live timings.

The present critical path remains #836. Its freeform upper-witness implementation is mechanically complete on PR #843; the remaining gates are the owner's merge decision and the frozen simple-versus-research owner-live comparison.

Owner ruling 2026-09-02 (#844): the skill is installed inside the user's investing folder and the user's own written record is the first evidence — the deciding reason may come from it and the engine's consequence checks the pick. "No private repo" was a development-side rule (public artifacts stay synthetic; the engine ingests nothing), not a runtime prohibition. #842 and #843 are merged; the expression fronts are closed.

Repository truth

• Latest main: 8c37b04012d4e54dcb9f063f079def2ddd510e78, squash merge of PR The user's own record comes first — install inside the investing folder, quote it as theirs, let the engine check the pick (refs #844) #845 (the [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844 slice): SKILL.md "The user's own record comes first", the fourth provenance class user_record sharing one citation checker with public_fact, the owner-approved consider_three_way_comparison revised in place so its deciding reason is the user's own recorded condition, install under <investing folder>/.claude/skills/ in three README languages, the repository description and lede carrying the positioning. Product group 49/49 and CI green on 3.11/3.12 before merge; a ten-angle review's fifteen findings were fixed in the same PR.
• Open PRs at this refresh: none.
• [research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475 remains the scope guard, amended by the owner ruling recorded on [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844 and in a [research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475 comment (2026-09-02): reading the user's own folder is declared, not silent, and nothing is imported into the engine.
• [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 remains the owner for unprotected main.

Active implementation fronts

None open. The next front is claimable only after #844's behavioural evidence (below) or an explicit owner promotion of the next cut.

Claimable now

• [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844 item 5 (owner): old skill (main@5e0c7b6) vs new skill (main@8c37b04) on the private four-lane harness, blind pairwise owner preference, on the next natural decision. Recorded on [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844; it decides whether this slice is product improvement.
• Next cuts in the order the owner approved on 2026-09-02, one at a time, each its own issue before code: engine-supplied forward-dated facts ([design·P1] What does this product mean by risk? Structure at one instant is the whole current answer, and the owner's definition includes what is about to happen to it #781's deterministic half — next earnings and ex-dividend dates) and relative-strength readings; then the must_state reduction the 2026-08-02 harness ruling asked for; then review flows reading the same record instead of asking for a thesis.
• No new expression rule, length heuristic, or exemplar expansion.

Acceptance and blockers

1. The user's own record comes first — install inside the investing folder, quote it as theirs, let the engine check the pick (refs #844) #845: merged at main@8c37b04 with CI green on Python 3.11 and 3.12 and the review's fifteen findings fixed.
2. [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844 behavioral acceptance: old skill vs new skill on the private four-lane harness, blind pairwise owner preference, on the next natural decision. Recorded on [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844, never inferred from green tests.
3. [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 product acceptance: still requires one private real-decision consider walk with visible delivery, honest four-check verdicts, and live timings.
4. Integrity: deterministic numbers and state remain engine-owned; model judgment never fills a missing engine fact; the user's own record is quoted as theirs and never relabelled; no contemplated action is described as executed.
5. Privacy: real holdings, trades, motives, notes, answers, and owner preference details remain local. GitHub contains only synthetic or de-identified evidence.

Safe parallel work

• Prepare the frozen, synthetic scene prompts for [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844's harness comparison (recorded book + the user's own notes for the names in play; no real note content).
• Read-only audit of [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610's current rerun instructions before the next real walk.
• Draft the issue for the forward-dated-facts cut ([design·P1] What does this product mean by risk? Structure at one instant is the whole current answer, and the owner's definition includes what is about to happen to it #781's deterministic half) — design only, no code before the owner promotes it.

Serial next steps

1. Run [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844's harness comparison on the next natural decision; record the verdict on [design·M1] The user's own investing folder is the evidence for why a name — install inside it, read it as the user's record, let the engine check the pick #844.
2. Open and promote the forward-dated-facts cut; then the must_state reduction; then review flows reading the record.
3. Freeze a new main candidate and run [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610's real consider walk with live timings before calling M1 product-accepted.

Stop conditions

• Do not infer intelligence or usefulness from shorter text alone.
• Do not infer product acceptance from green tests or a synthetic exemplar.
• Do not reinstate a character cap, fixed question count, mandatory counter-case, or research/presentation ceiling.
• Do not compensate for a weak answer by adding more standing rules; first classify whether the miss is content selection, decision condition, missing evidence, realization, or route reliability.
• Do not expand Give freeform an upper witness without restoring a length cap #843 past freeform without owner-live evidence.
• Do not expose private decision content in public issues, PRs, fixtures, or QA receipts.
• Do not compensate for a missing note by inventing one, and do not read an AI-maintained status field as the user's belief.

[atomchung]

2026-07-04 凌晨 loop session 盤點(多 session 對齊用)

本輪開了 7 個 PR(全 CI 綠,待 review/merge),issue 全量 41 → 已解 10 個:

PR	closes	一句話
#69	#44	MIT LICENSE + 英文 TL;DR;repo description/topics 已直接 gh 設定生效
#70	#42 #43	隱私措辭架構中立 + card-feedback.yml 表單(Stage 0 回收管道落地)
#71	#61	TR_JSON/state 契約測試(38 斷言)+ CYCLE_ID_RE 單一事實源 + SKILL 收尾片段煙霧測試
#72	#62	合成價格路徑測試(11 條:賣太早邊界/β=2/α=rf 恆等式/處方四分支)+ 每週非阻塞 network-smoke job。⚠️ stacked on #71,先 merge #71
#73	#55 #56	Step 2 強制 AskUserQuestion;Step 3 規矩用戶親選 + FINAL_RULE=SKIP 語意
#74	#47	onboarding:需要 Claude Code 明示 + PEP 668 venv 三行 + ModuleNotFoundError 補救
#75	#68	eval-design ↔ EVALS ↔ card-spec 三檔對齊 + C-5/C-6 新判準。C-2 判準依賴 #73 先/同 merge

merge 順序建議:#71 → #72;#73 → #75;其餘任意(hunks 不重疊)。

剩餘 code 隊列(下個 session 可直接接,無需討論):

• [P1·第一印象] mock_trades 頭號洞已漂移 + README demo 數字失真(NVDA 佔 92%、round-trip +1940%) #45 mock 漂移(建議走修法 b「demo 凍結價格快照」,可與 [測試·P1] 「全程離線、確定性」宣稱在本機是假的:實測連網、隨行情漂、同輪混用成本/市值基礎;dedup 語意只釘一半 #64 一起修)+ README 卡/PNG 重拍
• [測試·P1] headline 選卡是「測試自己測自己」+ severity 無錨:tier 權重翻轉、稀釋 10 倍都全綠存活 #63 [測試·P1] 「全程離線、確定性」宣稱在本機是假的:實測連網、隨行情漂、同輪混用成本/市值基礎;dedup 語意只釘一半 #64 [測試·P2] 測試工程衛生:CI 版本/依賴不鎖、_DRIVER_MAP 全域耦合、test_state_loop 混居 engine 目錄 #65(測試 P1/P2:headline 自測自/離線宣稱/工程衛生——[測試·P1] 「全程離線、確定性」宣稱在本機是假的:實測連網、隨行情漂、同輪混用成本/市值基礎;dedup 語意只釘一半 #64 已有先例:test: TR_JSON/state 契約測試 — SKILL 唯一消費介面上紅綠燈 (closes #61) #71 的 yfinance shim 可直接推廣)
• [P2·誠實] 輸入層靜默丟資料:load() 無「讀入/跳過」計數,與卡面的數字誠實不一致 #50 load() 讀入/跳過計數 · [P2·邊界] 非美股/多幣別:台股港股 CSV 會靜默算錯 — 至少明示邊界 + 偵測警告 #51 非美股偵測警告 · [P2·留存] weekly loop 缺用戶端操作指引:第二週該匯什麼 CSV?卡尾沒有回訪 hook #52 weekly loop 用戶指引 · [P2·上手] demo 模式的 Step 2 沒有特例:對假交易問用戶動機,答不出 #53 demo Step 2 特例
• [UX·P1] 卡上數字 4 處 edge case 沒守:雙 ROI 並列、∞、mock alpha、insufficient_data gate 漂移 #21 卡面數字 edge case · [UX·P3] The share-safe card is documented but undiscoverable to the user (narrowed from 5 discoverability holes) #22 discoverability · 支援同日同價的多筆獨立成交（目前會被去重誤殺） #14 同日同價去重誤殺
• [GTM·P2] Demo materials describe the step-② dialogue but never show an answer changing the card #46 差異化展示(README 加 Step 2 對話示例)· [P1·分發] 多 agent 可移植性:Codex / Antigravity 已原生支援 SKILL.md 標準,只差安裝指引 + 實測 #49 多 agent 安裝指引(要實測 Codex/Antigravity)

產品方向 bucket(留 owner 整體討論,本輪不動):
#3(Stage 0,唯一真命門)· #12(投資 OS Epic)· #16(風格偵測 research)· #31–33(已裁非 MVP)· #34–38(Phase 2 批次)· #54(舉證制,綁 eval harness B-1 落地)· #57/#58(策略庫/profile 推斷)· #60(e2e 驗收,綁 harness)· #15(做空)· #8(chat 版 PARKED)· #6 #5(P2 舊件)

—— loop session by Claude(worktree zen-hugle-eb5897)

[atomchung]

2026-07-05 開發過程盤點 → 5 個 process issue（多 session 對齊用）

對這個 repo 一個月開發史（108+ 番號）做了一輪「並行／需求不清晰／變更」三類問題盤點，缺口逐條拆成 issue（皆含證據＋解法選項＋建議，留 owner 拍板）：

issue	一句話	便宜的第一步
#110	修 bug 只修發現的實例，同根因缺全面掃描（#93→#98→#108 拆股 fixture 三連）	CLAUDE.md 加一條掃描慣例
#111	zoom-out review（跨 diff 疊加＋persona 壓測）驗證有效但只跑過一次，未固化成批次 merge 前節奏	CLAUDE.md 固化觸發點＋三 smell trigger
#112	關鍵產品假設無時效標記，過時結論被跨 session 複讀（Stage 0 話術案例，owner 2026-07-05 已糾偏）	#3＋BACKLOG.md:83 補更正註記
#113	跨檔同步靠人記（eval 三檔／README 雙語／demo 卡多處），缺機械 checker	CLAUDE.md 先列鏡像對照表
#114	worktree/branch 無清理節奏，大掃除後又堆回 11 worktree / 20+ branch	批次 merge 收尾順手清＋SOP 下沉 CLAUDE.md

盤點的第四類（agent 執行品質：工具結果須硬證據驗證、別在錯的 base 下結論）刻意沒開 issue——那是 session 工作紀律不是 repo 缺陷，已在 session 工作記錄層固化；若要下沉可併入 #110/#111 落 CLAUDE.md 時一起。

並行註記：本批開出期間，#109（knowledge/ai-errors.md 跨 repo AI 錯誤帳本 PR）由另一 session 同時段冒出——與本批無語意對撞，主題與上述第四類親緣，特此互認留痕。

—— 盤點 session by Claude（worktree infallible-feistel-973c96）

[atomchung]

2026-07-05 pm 批次收尾(merge + 大掃除)

本輪 merged(依序):

• docs(claude-md): 並行開發慣例 — 認領/掃同根因/zoom-out 節奏/鏡像對照表 (closes #111, closes #114) #118 CLAUDE.md 並行開發慣例(自動 close [process·P1] zoom-out review 固化：批次 merge 前跑跨 diff 疊加檢查 + mock persona 壓測 #111/[process·P2] worktree/branch 無清理節奏：大掃除後又堆回 11 worktree / 20+ branch #114)→ main 1016dcd
• chore(hooks): PreToolUse commit gate — block engine commits when tests red #119 PreToolUse commit gate hook → main 0d11da2
• docs(readme): 分檔雙語,英文版設為 GitHub 主入口 + 繁中移至 README.zh-TW.md #107 README 分檔雙語(英文主入口 + README.zh-TW.md)→ main 2c4161a。衝突解法=英文版補譯 fork 後 fix(demo): demo_weeks 對齊 SKILL 收尾契約(ai_pct 四鍵)+ 比例人話顯示 + README 曝光 #116/refactor(engine): drop is_demo — call is data-agnostic, demo = fake data through the real call #91/feat(skill): 試駕模式 — 沒資料也能走完四步,三防護(不落盤/標演練/卡標示範) (closes #53) #117 的 8 處變更,zh-TW 直接取 main 最新全文;CLAUDE.md 的雙語規則改寫進 docs(claude-md): 並行開發慣例 — 認領/掃同根因/zoom-out 節奏/鏡像對照表 (closes #111, closes #114) #118 的鏡像對照表行
• test: 5 個 engine 邊界情境 persona fixture(pyramid/insufficient/noisy-broker/rotator/panic-seller) #81 五組 engine 邊界情境 fixture → main 54a3abf。合併前修了 sample_rotator 的 NVDA 拆股前價格(同 [bug·P0] 拆股標的 fixture 用今日拆股後價填在拆股前交易 → adjust_for_splits 雙重縮放,污染出場紀律/未實現% 訊號 #93/fix(mock): mock_trades NVDA 拆股前建倉補名目價,消除 adjust_for_splits 雙重縮放 #108 型,qty÷10 price×10),其餘 4 fixture 掃過無拆股窗口;README 兩檔 sample 計數 7→12 同步

大掃除(#114 的 SOP 首次執行):worktree 10→3(留 main / 本 session / 新並行 session xenodochial-clarke-36abde);本地 branch 34→5;remote branch 35→4。刪除標準=PR MERGED 或 commit 可達 main,逐一驗證後才刪。

刻意保留:

• remote claude/beautiful-torvalds-44cb09:無 PR,2026-06-20 雙前端 OS PRD 存檔(docs/requirements.md),刪了會丟內容,owner 決定去留
• claude/naughty-chatelet-8853d0(test: 回歸測試套件 + 一鍵 runner + CI(防止迭代改壞引擎) #25 CLOSED)已刪,內容被現行測試套件取代;如需可從 PR test: 回歸測試套件 + 一鍵 runner + CI(防止迭代改壞引擎) #25 頁面 restore branch

當前 open PR(皆 owner 決策項):

• docs(knowledge): add ai-errors.md for cross-repo AI error ledger #109 ai-errors.md 進 public repo?(可合/可 pending)
• docs: 隱私措辭誠實化 + Step 4 反饋落地管道 (closes #42, closes #43) #70 隱私措辭誠實化(closes [P0·分發] Step 4 反饋迴路沒有落地管道 — Stage 0 收不到「戳中沒」 #42/[P0·信任] 隱私措辭過度承諾:「交易不離開你的電腦」在 LLM 架構下不成立 #43):注意 README/SKILL 在它 fork 後已被大改,合前需 rebase + 語意重驗(README 隱私段現行措辭已部分吸收其意圖)

—— 批次收尾 session by Claude(worktree sleepy-boyd-6cae2f)

[atomchung]

Context-index delta — 2026-07-29 21:xx Asia/Taipei

The body predates the latest merge sequence and should not be read as the exact current SHA/PR map.

• Latest main: b035d159 (fix(data): every source records the book at its own time, and completeness stops disqualifying one (closes #549) #562).
• Open PRs: none at this check.
• fix(data): every source records the book at its own time, and completeness stops disqualifying one (closes #549) #562 head CI: tests completed successfully.
• Current product milestone remains M0, unaccepted.
• New next agent-executable leaf before formal [QA·M0] Owner-live current-book and delivery acceptance on Claude and Codex #486: reopened [bug·contract·M1-prep] Host entry points still teach retired incomplete-snapshot semantics after #562 #549, limited to synchronizing AGENTS.md / SKILL.md with fix(data): every source records the book at its own time, and completeness stops disqualifying one (closes #549) #562's canonical recorded-book semantics and adding a drift test. [QA·M0] Owner-live current-book and delivery acceptance on Claude and Codex #486 remains the owner-live exit gate after that correction.
• M1 status changed from discovery-only: feat(safety): a two-sided answer is checked against the result it claims to describe (refs #414) #547/test(eval): the trigger matrix gets a corpus and a scorer, and nothing that can call a host (refs #458) #548/refactor: the stored artifact is a TradeEvaluation, in every place it is named (refs #479) #551/feat(consider): what the user said gets frozen beside what the engine computed (refs #479) #555 landed the independent provenance validator, trigger corpus/scorer, TradeEvaluation vocabulary and DecisionContext/persistence foundation. Overall M1 remains unintegrated, release-unverified and UX-unaccepted; [feat·M1] Bounded TradeEvaluation with visible two-sided challenge #479 Wave B and [QA·M1] Owner-live TradeEvaluation acceptance #488 still own those gates.
• Closed component states on [safety·M1] Production provenance gate for decision-surface answers #414/[eval·P1] Implicit trigger reliability after explicit-entry M1 #458 must not be read as proof that production validator wiring or host/locale release evidence has passed; [QA·M1] Owner-live TradeEvaluation acceptance #488's explicit preconditions still govern.

Detailed M1 delta is recorded on #479; the host-contract blocker is recorded on #549 and promoted on #486.