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

Status

Execution routing — refreshed 2026-09-05 after the owner's long-term-design directive. Current main re-read at 8c37b04012d4e54dcb9f063f079def2ddd510e78; open PR collection is empty. No code change, test run or new acceptance result was produced by this handoff.

Current milestone

M1 decision usefulness — greater agent freedom, faithful context use, concise communication without compressing judgment.

The owner has now requested a complete next-session development task optimized for the long-term investor experience, explicitly not another smallest-patch recommendation. This promotes the integrated next wave below. It does not approve an irreversible migration, engine rewrite, privacy relaxation, autonomous execution, new major milestone or PR merge.

Active / claimable implementation

One front: #827, reopened for a coordinated context-to-decision and permission-contract follow-through. Claimable; no implementer claimed yet. Its original PR #828 remains completed history, not work to repeat. Read #827's current body for evidence, scope, acceptance and stopping boundary.

The two connected outcomes are:

let the agent choose research, evidence retrieval, tool use and presentation, with code enforcing the relevant fact/write/action boundaries rather than a blanket conversation workflow;
make relevant personal context and genuine new evidence change the current judgment or its support, instead of decorating a generic warning or generating another intake.

The reviewer confirmed an important ambiguity on current main: SKILL.md permits record-led and non-portfolio judgment while its always-applicable references/agent-boundaries.md retains a prohibition on arguing a considered trade from anything but consider output, plus unscoped ranking/derived-analysis language. #827 owns correcting the inherited permission scope across actual readers. Its effect on current runtime behavior must be measured, not presumed.

Coordinated contract/reader/output/test changes are authorized when justified by the integrated outcome; implementation is not artificially limited to adding one sentence to the Skill. Conversely, this is not authorization to implement every deferred capability. Prefer a coherent reviewable PR, with ordered splits only when reviewability genuinely benefits.

Ownership / required context

[implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827: implementation handoff and current scope; original limit-removal work and new wave are separated in its body.
[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: record-selection, faithful use of source conditions, and the original 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 behavioral comparison. Integrate its findings in the same implementation front; do not fork a second context/memory architecture owner.
[design·M1] One communication method: the consulting pyramid as the mandatory top layer; per-surface derivation optional and additive-only #832 / expression contract: authority for any expression-shape amendment. Review whether rigid shape/exemplar obligations cause harm and synchronize any justified change at the authority rather than adding a contradictory route rule. No replacement template is predetermined.
[QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610: product acceptance through a delivered real-decision answer and live timings.
[research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475: scope guard as amended by the recorded owner rulings; declared runtime folder reading is allowed, copying private records into public artifacts is not.
[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 / [research·architecture] One route-scoped market-context capability — shared observations, bounded research, no background news system #776: consult only if missing current/event evidence proves a capability gap in the present journey. No automatic calendar, relative-strength, news or provider expansion.

Read latest main, AGENTS.md, the relevant Skill/references and maintainer guide, then #827 and these direct owners. Do not replay the whole backlog. Capture the actual loaded instruction/installation chain; a main checkout does not establish which Skill a host used.

Evidence and acceptance that remain true

PR #845 is merged at 8c37b04. The earlier September 5 review inspected main Actions run 33651572890: product and QA/eval jobs succeeded on Python 3.11/3.12; network smoke was skipped. This is not a live usefulness or market-data claim.

The case run September 2 and ranked September 3 still used the August 14 skill. It cannot accept or reject #845. Preserve the originally requested 5e0c7b6 versus 8c37b04 comparison, or report a new multi-change candidate separately; never attribute the combined gain to one old PR.

Freeze comparable inputs/model/research facts and compare actual old/new Skill runs with the existing same-context/no-skill control. Audit correctness and record fidelity before preference. Include simple/deep decisions, no-book/failure recovery, a subsequent decision with new evidence and irrelevant-context perturbations. No forced recommendation reversal or word-count-only success metric.

Owner-live acceptance remains required for product-complete claims, not a reason to block the newly authorized implementation and offline verification. Missing private harness/live-host access must be reported; finish the code, review, applicable tests and PR where possible without fabricating the missing evidence.

Next safe stopping boundary

Claim #827; establish the behavioral baseline; implement the coherent improvement; run focused/product and applicable QA/eval proof; independently review the diff and delivered answers; open the PR; update #827, affected #844 evidence and this index. Do not merge automatically or mark #610 accepted without its actual receipt.

Then report what improved, what remains unproven and what observed gap would justify the next capability. Numeric/state integrity, explicit durable-write intent, execution truth, local-first privacy, compatibility, idempotency and interruption recovery stay intact. Do not replace a rigid workflow with vague “be smarter” instructions or a new framework by default.

Prior ordering — superseded only for this promoted front

The earlier sequence was: #844 evidence / one-line record-reading follow-up, then forward-dated facts and relative strength, then must_state reduction, then review readers. The owner's new instruction permits treating relevant instruction, selection and expression defects together before those standalone capability additions. It changes implementation scope/order, not the truth of prior results or the requirement for eventual owner acceptance.

Historical September 5 routing before the owner's broader directive

Status

Daily control plane — refreshed 2026-09-05 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.
Drop the refusal patch the claim boundary made redundant #842 removed a redundant refusal patch; Give freeform an upper witness without restoring a length cap #843 added the freeform research-depth upper witness without a length floor or cap. Both merged on 2026-08-29. Their implementation fronts are closed; [design·expression] The exemplar corpus is a character cap the contract twice refuses to have #836's simple-versus-research live comparison remains evidence work, not a pending merge decision.
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 the original [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 implementation items 1–4 on 2026-09-02. The user's own record may supply the deciding reason and the engine checks the pick. Behavioral improvement remains unproven.
[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.

Current critical path: #844's correctly pinned behavioral comparison and record-use verification, followed by the already approved serial cuts below. The latest public evaluation finding does not evaluate #845: the case run on September 2 and ranked on September 3 still loaded the August 14 skill. Do not interpret that case's old-skill ranking as a new-version verdict. See the existing finding and #844's refreshed body.

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. The original folder/indexing/scope questions have been settled; no new folder manifest or engine index is required for this slice.

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: record-first SKILL.md, the fourth provenance class user_record sharing a citation checker with public_fact, the approved comparison exemplar revised in place, and installation in the user's investing folder across the README languages.
Main CI directly verified: Actions run 33651572890: product-contract and qa-eval-tooling succeeded on Python 3.11 and 3.12; network-smoke was skipped. No live market verification is inferred from those results.
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; this review does not promote that debt into the M1 critical path.

Active implementation fronts

None open. The September 5 work is a repository/UX review and issue refresh, not a code change or a newly opened front. The next front is claimable only after #844's behavioral evidence 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/evidence): compare old skill main@5e0c7b6 with new skill main@8c37b04, or explicitly freeze a later candidate, on the existing private harness. Verify the skill actually loaded, not just the repository checkout. Keep source/model/input conditions comparable and preserve the existing full-context/no-skill comparator.
[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 bounded follow-up design: the September 3 evidence identifies missed standing decision rules and lost source qualifiers. The September 5 review ties these to the existing Skill reader and the provenance validator's documented limits, with proposed synthetic verification cases. This belongs to [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, not a new memory or architecture issue. It is not implemented or behaviorally accepted.
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 generation-exemplar expansion.

Acceptance and blockers

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 implementation: merged at main@8c37b04; product and QA/eval CI verified on main. This is not an owner-usefulness verdict.
[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: the actually loaded skill is correctly pinned; the delivered answer is audited against the relevant standing rules and per-name records; the old/new comparison and owner verdict are 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. Apply the September 3 owner ruling: add the missing evidence and check whether the action changes that day. Preference, readable phrasing and action change are separate observations; do not force a changed decision or claim future investment returns were tested.
[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 verdicts, and live timings.
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. A valid source/date envelope is not proof of faithful interpretation.
Privacy: real holdings, trades, motives, notes, answers, and owner preference details remain local. Public fixtures and new review evidence are synthetic or generic; refer to existing de-identified findings rather than copying private material.

Safe parallel work

Prepare small synthetic evaluation variations 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 existing harness: a material standing condition changes while the ticker note is fixed; a note's action qualifier must survive; irrelevant history must not create a recap. These are evaluation inputs, not more generation exemplars or a new judge framework.
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

Repair the actual skill pin, address the already identified record-reading follow-up within [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 when claimed, and record a correctly pinned old/new behavioral result. Name a later candidate separately rather than attributing its effects to 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.
Open and promote the forward-dated-facts cut; then the must_state reduction; then review flows reading the record. Keep the previously approved order; this review adds no second active front.
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, source labels 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.
Do not add a new profile, memory store or engine folder index to repair a missing read in the existing agent-side input path.

Historical routing correction — 2026-09-05

The preceding body retained this pre-merge paragraph even though its later sections correctly said the PRs were merged:

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.

That merge blocker is superseded by the actual August 29 merges and the September 2 #845 slice. The live evidence obligations are retained above. #844's former discussion-only status and unresolved owner questions were likewise refreshed from its already recorded owner rulings and merge receipt; its original proposal remains under historical details. No milestone, product-positioning, scope-boundary or merge approval is changed by this correction.

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)

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）

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)

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.

Context delta — 2026-07-29 split review after PR #567

Latest inspected main: 7909f0c9 (PR fix(ledger): a share balance is only a balance when every quantity shares one split basis (closes #558) (closes #559) #567).
PR fix(ledger): a share balance is only a balance when every quantity shares one split basis (closes #558) (closes #559) #567 merged and CI passed on Python 3.11/3.12, but [bug·data·M1-blocker] Split-blind current book can erase a held position from TradeEvaluation #558 has been reopened: _virtual_valuation_frame still derives its provisional ticker universe without the frozen split map, then narrows the ValuationFrame before the final split-aware PortfolioBasis query. A split-crossing position that disappears in raw arithmetic can therefore be omitted from coverage and make review prepare refuse when exact holding/frame partition validation runs.
Smallest M1 repair remains owned by [bug·data·M1-blocker] Split-blind current book can erase a held position from TradeEvaluation #558: thread the existing frozen map into the provisional query and add one production-shaped review integration/mutation test. Do not widen into corporate-action architecture.
New [bug·data·P1] Split-crossing position_absence records the wrong exit quantity and can duplicate its revisit #571 records the separate P1 follow-up on the position_absence branch: absence_exits ignores the split map, and correcting its quantity later can duplicate a revisit because identity currently includes shares_sold. This is not an M1 TradeEvaluation blocker and should not interrupt [feat·M1] Bounded TradeEvaluation with visible two-sided challenge #479 after [bug·data·M1-blocker] Split-blind current book can erase a held position from TradeEvaluation #558 is closed again.

Active order: #558 narrow repair → #479 Wave B → #544 Slice B → combined #486/#488 exact-SHA acceptance. #571 may follow at the next safe leaf unless it can be included without widening the #558 PR.

Routing update, 2026-07-31 — Front A is implemented and in review; Front B is unblocked and unstarted

Front A · #539 / #536 — done, PR #659 (draft)

One shared primitive, book_refresh.carry_recorded_starts, reached by both adoption lanes. Full offline suite green (49/49); three mutations recorded on the PR. Both issues close with it.

Three things the design review changed before implementation, each recorded on its owning issue:

The plan this index carried would not have closed [bug·design·P2] Every new snapshot declaration re-asks for theses it was already given #539. "One shared carry-forward primitive in both lanes" closes [bug·design·P2] A review-lane anchor adoption spends the user's holding-duration answer #536 alone: _carry_forward_provenance was gated on a since_basis stamp, and the appearance question that produces one is only asked for a position that appears after the book exists — so every position in a user's first declaration was unstampable and reminted regardless of lane. What is carried is the start itself now, read off the derived book. (#539)
The seam is one layer below [bug·design·P2] A review-lane anchor adoption spends the user's holding-duration answer #536's stated fix shape, which would have added a second allow_engine_provenance=True call site the suite counts. No prepare reordering was needed. (#536)
A correctness hazard [bug·design·P2] Every new snapshot declaration re-asks for theses it was already given #539 widens, closed in the same PR. cycle_id is ticker#since#seq; carrying the start without the sequence hands a position bought back on the day it was sold the identity, thesis, standing conditions and closed status of the cycle it replaced. Reachable through an adoption that asks the user nothing.

New leaf: #660 — adopting a book resets add_count/decision_cursor for a position never sold. Same defect class, adjacent field, not fixed by the above and deliberately not folded in. Discussion, not queued.

Front B · #403 / #450 — unblocked, not started

The "parallel on file disjointness" framing in this index understated the coupling: Front B's contract depended on an unmade Front A decision (identity-preserving versus alias-based continuity), which is why the reader contract in #450 assumed an alias producer. That decision is now made and shipped as identity-preserving, so:

an ordinary re-declaration produces no alias — [parked·M2] Position rationale stream — not needed for the first memory loop #403's subject identity survives untouched and [arch·M2] Canonical bounded execution-qualified self-report history #450's first reader needs no chain walk on the common path;
aliases remain only for [design] A snapshot-inferred cycle start is now final — the relink that let a later CSV correct it rode on the completeness flag #549 removed #563's later-history upgrade, already produced by thesis.build_snapshot_cycle_relinks;
[arch·M2] Canonical bounded execution-qualified self-report history #450's shared fixture should be built with unchanged identity as the common case and the relink as the exception (#450).

Two defects in #403's frozen event contract should be fixed before the schema is written, both recorded there: the content-addressed event_id swallows a genuine re-statement when a user oscillates between two wordings, and change: initial | changed | no_change folds a user act together with a fact the fold derives (#403).

Unchanged

#610 still owns M1 acceptance and this PR is not part of its frozen candidate. The serial review.py integration slice still waits for Front B.

Generated by Claude Code

test(rationale): the #450 shared fixture Front B still owed (refs #403,…

2026-07-31 · daily control-plane decision

Owner direction: the daily 08:00 repository scan must land on GitHub and produce executable expected changes that other AI agents can claim.

Decision

tracking: current context index — milestone, critical path, what an agent may pick up #27 remains the single mutable current-truth surface: its body is the latest control plane; dated comments preserve the change history.
Each daily run updates the body first, then appends one delta-only comment. Do not create a new daily-summary issue.
Implementation remains in dedicated owning issues. Reuse or update an existing owner; open a new issue only for a distinct executable outcome with no current home.
Work is claimable only when tracking: current context index — milestone, critical path, what an agent may pick up #27 lists it under Claimable now, it has no unresolved owner decision or dependency, and no open PR or active implementation owner already claims it.
Claim through GitHub assignment when possible; otherwise post CLAIM — <agent/session> — branch <name>. A linked draft PR is the first durable work receipt.
Do not add GitHub Projects or a larger label/status taxonomy while the repository still operates with one or two active fronts. Revisit only when tracking: current context index — milestone, critical path, what an agent may pick up #27 can no longer express the real dependency/ownership graph cleanly.

Current result

No unowned implementation task is safe to claim today: PR #661 already owns the active implementation front; #610 is an owner-live judgment-quality walk; the #403 direct-entry → next-review integration slice remains serially blocked until #661 is accepted.

2026-07-31 QA-round routing delta

Changed the control plane after one owner QA round:

[QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 moved from “ready for dogfood” to blocked, not accepted: [defect·P1] One unmatched sell anywhere in the history disables consider for the whole account — the review lane discloses the same condition and continues #673 prevented a real evaluation and [design·P1] When consider cannot compute, no contract owns the answer — so the agent narrates the gap instead of framing the decision #674 failed the fallback user outcome.
Replaced the single M2 implementation front with two QA repair fronts:
- Front A starts at [defect·P1] A currency that appears only in a cash-flow row is added to the USD total at 1.0 — #612's refusal has held currencies as its domain, and no rate is ever requested #649, then [defect·P1] A TWD position's cost basis reaches the question surface labeled USD, and the "largest cost" ranking compares raw units across currencies #664 → [defect·P1] One unmatched sell anywhere in the history disables consider for the whole account — the review lane discloses the same condition and continues #673 → [design·P1] When consider cannot compute, no contract owns the answer — so the agent narrates the gap instead of framing the decision #674.
- Front B starts at [defect·P1] add-cash refuses the session it exists for: the drift gate compares question_queue and state_snapshot whole, so the card-beat cash answer is unrecordable #665, then [defect] planned_entry declares no answer requirement, then preview refuses without a captured thesis — a dead end reachable only by following the contract #667 → [defect] A light-tier weekly_review cannot produce a receipt that is both honest and verifiable — the plan says the cash anchor is not applicable, the route contract demands one anyway #677 → [process·qa] Cash-anchor recompute mid-first_review has no documented receipt slot for its second preview card #663 → [design·process] Cash-anchor ask has no format contract — a percentage answer needs three round-trips to resolve #662.
Promoted [defect·P1] A currency that appears only in a cash-flow row is added to the USD total at 1.0 — #612's refusal has held currencies as its domain, and no rate is ever requested #649/[defect·P1] A TWD position's cost basis reaches the question surface labeled USD, and the "largest cost" ranking compares raw units across currencies #664 above their prior no-milestone status because they can emit confidently wrong cross-currency facts; priority now reflects user harm rather than the issue prefix.
Recorded maintainer decisions on [defect·P1] A currency that appears only in a cash-flow row is added to the USD total at 1.0 — #612's refusal has held currencies as its domain, and no rate is ever requested #649, [defect·P1] A TWD position's cost basis reaches the question surface labeled USD, and the "largest cost" ranking compares raw units across currencies #664, [defect·P1] add-cash refuses the session it exists for: the drift gate compares question_queue and state_snapshot whole, so the card-beat cash answer is unrecordable #665, [defect] planned_entry declares no answer requirement, then preview refuses without a captured thesis — a dead end reachable only by following the contract #667, [defect·P1] One unmatched sell anywhere in the history disables consider for the whole account — the review lane discloses the same condition and continues #673, [design·P1] When consider cannot compute, no contract owns the answer — so the agent narrates the gap instead of framing the decision #674, [defect] A light-tier weekly_review cannot produce a receipt that is both honest and verifiable — the plan says the cash anchor is not applicable, the route contract demands one anyway #677 and [process·qa] Cash-anchor recompute mid-first_review has no documented receipt slot for its second preview card #663 so those leaves no longer require re-design by the implementing agent.
Draft PR feat(rationale): a canonical stream for why the user says they still hold a position (refs #403, refs #450) #661 is frozen pending a scope split: canonical stream stays isolated; direct-entry/next-review integration moves to a separate [parked·M2] Position rationale stream — not needed for the first memory loop #403 PR. Current head 50c9c4a has green PR CI but does not match its body/title scope.
Consolidated [docs] AGENTS.md, the first-review flow and the engine disagree about which question kinds may carry a private surface #666 under [defect·contract] The instruction floor says two question kinds may be re-worded; the code and its own reference say six #657; no parallel implementation.
Queued card/copy/provenance work ([bug·design·P1] The revisit backlog shows the return rate its own ranking rejected — and prices a partial trim as if the whole position had gone #670, [bug·gate·P1] R1 and A-13 both pass a card that violates them — each gate enumerates a symptom instead of naming its mechanism #671, [defect·P2] consider's own live-resolved prices are reported as mode=agent_feed, not engine_fetch #654, [defect·copy] The realized payoff sub-line shows per-trade averages without saying so, four lines under the realized total #668, [defect·contract] The instruction floor says two question kinds may be re-worded; the code and its own reference say six #657, [bug·a11y·P2] Ruling 8 (exactly one L1, and it is Block 4) is defeated in dark mode — the class is present, the emphasis is not #672) after the runtime and truth-critical fronts.

Only #649 and #665 are claimable now. No third implementation front.

Delta — 2026-08-01: both QA repair fronts executed to completion

Nine leaves, nine PRs, all merged. main@db6e114. Every PR: independent CI green on 3.11+3.12, full offline suite green, coordinator review with mutation spot-checks; the first two leaves were additionally verified as a combined local synthesis tree (full suite + an end-to-end replay of the original QA incident: USD book + TWD interest row + card-beat cash answer now prepares, converts, and records correctly with zero second-instant price fetches).

Front A — truth and usable pre-trade evaluation (complete)

Leaf	PR	What landed
#649	#685 (`fa499b3`)	One currency-universe builder (held + cash-flow rows) drives both FX acquisition and the fail-closed gate; 1.0 identity fallback removed; pure books byte-stable. Recorded deviation: accepted cash anchors excluded from the universe (rationale + reversal condition in #649's closing comment; flowless-anchor drop filed as #688).
#664	#693 (`dc654a1`)	Position cost reaches the stem in its native currency; ranking compares one base currency via the frozen fx map; missing rate refuses ranking instead of raw-unit ordering. One review-found defect (divergent aggregate-currency fallback) fixed before merge.
#673	#695 (`b0bd0a8`)	A per-holding integrity warning excludes and discloses that holding; denominator and rule collisions computed over usable holdings only; premise on the excluded holding refuses by name; structural/unscopable/no-usable-holding stay fatal.
#674	#697 (`2435764`)	Non-recoverable `consider` refusals carry a bounded `usable_facts` packet (frozen concentration + committed rule); new episode EP-010 + `usable_facts_grounding` checker enforce the decision-framing obligations.

Front B — review lifecycle and honest evidence (complete)

Leaf	PR	What landed
#665	#684 (`e662164`)	`add-cash` recomputes on the frozen market frame (`TR_MARKET_DATA_FROZEN`, exact-request reuse, never reaches a provider) and the drift gate compares positively-named source facts; refusals name "the facts moved" vs success `anchor_propagated`. Root cause was the recompute's own second live resolution, not the wholesale comparison (which is inert once the frame is frozen — regression gap filed as #686).
#667	#692 (`843c085`)	`planned_entry` declares its thesis-capture requirement and collects it in the same exchange; validator unchanged and strict.
#677	#694 (`5fbb9b9`)	Receipt cash-anchor obligation derives from the executed plan (`--cash-anchor-status`); `not_applicable` means zero events; a fabricated decline fails as a contradiction.
#663	#696 (`c217228`)	Exactly-once preview receipt is the settled pre-commitment card: `provided` must precede the recorded preview pair (recompute), `declined` stays after it (no recompute).
#662	#698 (`db6e114`)	Cash ask accepts an amount or a percentage; owner disposition recorded 2026-08-01 on #662 (denominator = cash + position market value, stated in plain words, one confirmation); engine owns the disclosed `p/(1−p)` conversion against the frozen position value; multi-currency percentage = single aggregate-currency anchor, mismatch refuses.

New issues filed during the campaign

#686 (drift-gate wholesale-comparison regression gap), #687 (same-day cache never serves a book with a permanently-unpriced symbol), #688 (flowless cash anchor silently dropped; must take anchors back into the universe when fixed), #689 (alpha_beta_note dead field + zh hardcode on en cards), #690 (account_perf residual-currency KeyError hardening), #691 (price-degraded test lanes masked #665; priced harness now exists). All Status: Discussion, none claimed.

Body staleness note

This body's Claimable now (#649/#665) and both QA repair Front sections are now fully landed and stale. Until the next full scan rebuilds the body: there is no claimable QA-repair implementation leaf left; the standing next actions are (1) owner-only — freeze a new exact main SHA (db6e114 or later) and rerun #610's four judgment-quality checks, per this body's "Next order" step 4; (2) the PR #661 scope split (unchanged, still frozen, still not a third front); (3) stale-base docs PR #614 disposition.

Daily delta — 2026-08-01 08:03 +08:00

This comment records only routing changes since the preceding completion delta; the full current truth is now in the issue body.

Control-plane baseline advanced: tracking: current context index — milestone, critical path, what an agent may pick up #27 now reads main@db6e114 instead of main@e58fef9. The nine QA repair leaves already listed in the preceding delta are treated as merged with green PR CI, not as owner acceptance. GitHub exposed no separate combined-status entry for merge commit db6e114.
Claimable queue changed from two items to one: [defect·P1] A currency that appears only in a cash-flow row is added to the USD total at 1.0 — #612's refusal has held currencies as its domain, and no rate is ever requested #649 and [defect·P1] add-cash refuses the session it exists for: the drift gate compares question_queue and state_snapshot whole, so the card-beat cash answer is unrecordable #665 were removed because their full serial fronts are complete. [defect·P2] consider's own live-resolved prices are reported as mode=agent_feed, not engine_fetch #654 is now the sole claimable autonomous leaf: Implementation, current M1, unassigned, no open PR, bounded to truthful engine_fetch versus agent_feed provenance with an existing fake-provider test shape.
Acceptance priority moved to the front: [defect·P1] One unmatched sell anywhere in the history disables consider for the whole account — the review lane discloses the same condition and continues #673 and [design·P1] When consider cannot compute, no contract owns the answer — so the agent narrates the gap instead of framing the decision #674 are now on main, so the prior blocked [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 walk is ready for a new exact-SHA owner rerun. Recommended candidate is db6e114 unless the owner names a newer SHA before the first timing/verdict event.
PR feat(rationale): a canonical stream for why the user says they still hold a position (refs #403, refs #450) #661 became more stale, not more accepted: its head remains 50c9c4a with historical green CI, but current GitHub reports it non-mergeable against the advanced base. The required stream-only/integration split, rebase, evidence refresh and independent review remain mandatory; no new integration fix is authorized on that PR.
PR docs(registry): profile.md has no writer, and both surfaces that claimed one now say so (refs #58) #614 is unchanged: historical green CI, stale/non-mergeable, documentation-only and outside the active queue pending [decision] profile.md 是記憶體系第 2 層的化石:退役、復活,還是等 #446 刀 2 搬家? #619's non-blocking decision.
No backlog-wide promotion: [bug·design·P1] The revisit backlog shows the return rate its own ranking rejected — and prices a partial trim as if the whole position had gone #670/[bug·a11y·P2] Ruling 8 (exactly one L1, and it is Block 4) is defeated in dark mode — the class is present, the emphasis is not #672 remain unpromoted; [bug·gate·P1] R1 and A-13 both pass a card that violates them — each gate enumerates a symptom instead of naming its mechanism #671/[defect·contract] The instruction floor says two question kinds may be re-worded; the code and its own reference say six #657 still require contract decisions; [defect·copy] The realized payoff sub-line shows per-trade averages without saying so, four lines under the realized total #668 and [process·test] Nothing would catch a reintroduction of the wholesale state_snapshot/question_queue comparison in add-cash's drift gate — it is inert until an anchor reaches a derived object #686–[test-infra] The #357 cash-anchor test block runs price-degraded by default — a moving-provider harness now exists but covers only one of its tests #691 remain Discussion. No new issue was created.

Current safe parallelism: owner reruns #610; one agent may claim #654; the existing #661 owner may perform only the required split/rebase, followed by independent review. No fourth lane.

Routing update — #683 owner-approved prototype lane (2026-08-01 11:05 +08:00)

Owner has approved #683 for an isolated prototype / draft PR only.

This does not change the M1 merge path: PR #700 / #676 → #654 → freeze SHA → #610 remains authoritative. #683 may proceed in parallel only under these bounds:

one walkable weekly-review vertical slice: Review Card → WeeklyMarketRead → NextWeekWatch → at most one optional context question;
no merge authorization before owner review;
no scheduler, notification service, durable NextWeekWatch, new memory ontology, direct investment-note dependency, card redesign, or [parked·M2] Position rationale stream — not needed for the first memory loop #403/[arch·M2] Canonical bounded execution-qualified self-report history #450 expansion;
no duplicate provider fetch; reuse the frozen market-data path;
public evidence synthetic only;
stop at a draft PR and owner-live walkthrough.

Owning contract and acceptance are now consolidated in #683's updated body. Treat #683 as prototype work, not as the next milestone or a reason to delay #654/#610.

Owner-directed parallel docs lane recorded: #699 / draft PR #701 refresh all three README variants as the complete public product journey. This is not a second autonomous product implementation front and does not change the #676 → #654 → #610 runtime critical path. For README design, the owner explicitly instructed us to treat #676's direct, evidence-bound voice as the finished product baseline rather than explain or qualify the personality work. Scope is documentation only: user moments, first visible value, bounded questions, observable delivery, user-owned action, continuity, privacy/truth boundaries, installation and recovery. No runtime, card, state, schema, route, demo-asset, or M2 capability change.

docs(release): v0.1.0 — stop promising what only one host delivers (#817

2026-08-06 — delta: `v0.1.0` tagged and released

The frozen candidate moved and was published. ba132a3 is superseded by
main@c40ab16, tagged v0.1.0 — the repository's first tag, first release,
and first changelog. .claude-plugin/plugin.json had carried 0.1.0 for some
time without that version ever existing as a tag.

Two owner rulings changed this body's standing conclusions.

[P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 downgraded P0 → P2 and explicitly does not gate the release.
This body said whole-product MVP remains blocked by [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230, and that stays true
of the MVP claim. What changed is that v0.1.0 does not make that claim, so
the tag was not held for a cross-client delivery adapter. The defect is
unfixed and the issue is open.
The README is a promotional document. An earlier revision of PR docs(release): v0.1.0 — stop promising what only one host delivers #817
added a "Version and known limitations" section stating that owner-live
acceptance had not happened. The owner removed it: not exaggerated, but not a
place to surface the project's own open questions either. CHANGELOG.md
keeps a limitations list; [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 keeps the discussion. This body's own framing
of acceptance status is unchanged and remains accurate — it simply is not
what the front door says.

Merged since the last scan: PR #818 / #737 (positions returned
positions: [] with every weight null on a fully priced mixed-currency book —
broken since cmd_positions was introduced in #561, silent, status: "ok";
the command now reaches _consider_valuation_frame like consider does, and a
weight it genuinely cannot state carries the engine's own reason). PR #817
(release documentation, two new mechanical checks, two stale assertions
corrected).

Filed: #816 — a positions row carries no currency, so on a mixed book
the native figures are unlabelled. Reachable only now that such a book produces
a populated list at all.

Closed: PR #614 as a stale claim; its finding (a profile.md registry entry
describing a writer nobody built, still present on main) recorded on #58. The
stale-claim lane in this body is now empty — there are no open PRs.

Corrected in docs/maintainer-guide.md: the mirrored-surfaces entry for
canonical ticker identity still described the spelling-derived cycle_id that
#814 / PR #815 retired along with the cycle_identity module.

Not changed by any of this: usefulness, inline delivery and latency remain
unaccepted. A tag is a published version, not an acceptance event, and nothing
here should be read as one.

2026-08-06 — v0.1 first-version product assessment

Status

Treat the current release as a controlled real-use / design-partner first version:

v0.1.0 is a contract-complete, local-first decision-audit runtime.
The runtime/release boundary is implemented and regression-covered.
Product usefulness and repeat-use are still unaccepted until the real-decision receipt in [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 exists.
This is not yet a standalone consumer app or a cross-host, whole-product GA claim.

Current repository truth

Current main: c40ab16 (PR docs(release): v0.1.0 — stop promising what only one host delivers #817), preceded by the mixed-currency safety fix in PR fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818.
PR docs(release): v0.1.0 — stop promising what only one host delivers #817 aligned the README/plugin/release copy with the actual v0.1.0 scope and reported make release-proof passing 58 suites / 603 tests.
There is no active product PR at the time of this assessment.
[P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 is explicitly P2 and does not block the v0.1.0 release boundary.
[QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 remains the sole real-use/product-acceptance owner.
[bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 remains the narrow truth-critical gap for per-row currency meaning in positions.

User outcome available now

Within a supported local coding-agent host, a user can:

Challenge a contemplated long trade before action and make thesis, evidence, invalidation, risk cap, price and sizing explicit.
Receive deterministic score/risk-cap/summary/lifecycle output rather than agent-authored arithmetic.
Confirm before persistence, preserving the distinction between contemplated, under-review and executed actions.
Backfill, complete, resume, inspect history and re-check invalidation criteria through a local, replayable decision trail.
Use the additive positions route for holdings/weight/concentration review where denomination semantics are safe.
Keep durable artifacts local, with schema validation, atomic/idempotent writes and recovery-oriented runbooks.

Explicit non-capabilities

The first version does not provide:

stock selection, alpha, trading signals or guaranteed better decisions;
live prices, news, filings or a research terminal;
broker connectivity or execution proof;
short-selling workflows;
FX conversion or trustworthy account-wide mixed-currency totals;
identical rich interaction across every host;
a standalone web/mobile product;
proof yet that the owner understands the output unaided or would reuse it next week;
automated personal-rule learning from the decision trail.

Blocker classification

Product-acceptance blocker: [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610. This does not block shipping the runtime, but it blocks claims that the product is validated, repeatedly useful or behavior-changing. The smallest evidence is one private real-decision run, a sanitized public receipt, and explicit owner feedback.
Scoped correctness blocker: [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816. This does not block the core consider/lifecycle release boundary, but it blocks describing mixed-currency positions output as fully trustworthy until every raw row value/impact carries its denomination.
Deferred interaction limitation: [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230. Keep at P2 unless [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 or repeated target-host evidence shows that delivery failure prevents the first useful outcome. Do not build a generalized host adapter speculatively.
Windows generation activation and host-specific visuals remain disclosed support boundaries, not launch blockers, provided promotional copy does not promise identical seamless behavior on every platform.

Public positioning decision

Use the category local-first investment decision-audit layer / 本地优先的交易决策审计层.

Recommended first-screen promise:

下单前，把冲动变成一份可复核的交易计划。

FOMO Kernel 不替你选股，也不预测市场。它要求你写清 thesis、invalidation、risk cap 与 sizing，用确定性规则检查这笔交易，并把 decision trail 保存在本地，供执行前 challenge 与事后复盘。

Primary proof points:

AI does not own the math: deterministic, replayable score/risk/lifecycle contracts do.
A contemplated action is never silently presented as an executed trade.
Durable decision evidence remains local and recovery-safe.

Avoid public claims such as AI investment advisor, alpha, better returns, signal engine, investment OS, automated personal learning, fully automated portfolio management, or seamless on every host.

README / promotion assessment

PR #817 removed the material factual anomalies: the first-timer entry now says this is a coding-agent skill rather than a standalone app, release status is v0.1.0, installation no longer starts with test-only dependencies, Chinese card coverage is aligned, plugin metadata no longer describes the old CSV-card product, and transcript evidence is not presented as executable host testing.

Remaining copy risk is claim breadth rather than factual breakage:

high-stakes investment decisions is broader than the verified first moment; lead with a user who is already considering a long trade.
for real use should be qualified as controlled/design-partner use until [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 is accepted.
Keep positions visibly additive/beta and do not market mixed-currency row amounts as comparable before [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816.
learn from the decision trail should mean manual revisit/review, not automated rule learning.
auditable means inspectable/replayable decision evidence, not a compliance audit or tamper-proof ledger.

Next smallest step

Run #610 against the current candidate without another broad UX redesign. Only open an implementation front from a reproducible acceptance failure. In parallel, #816 is the only currently justified narrow correctness patch if positions remains part of the public first-version story.

Privacy boundary remains unchanged: real decision inputs and receipts stay private; only synthetic or sanitized evidence belongs in this public repository.

Delta — 2026-08-07 08:01 +08:00

Changes since the prior 2026-08-06 07:58 daily scan:

main advanced from ba132a301d1f81162d266212a9883a5a2a0520f2 to c40ab16ad5c266f26198af64c96e30fd3bec07d1 and was published as v0.1.0. PR fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 merged the mixed-currency positions sizing correction, then PR docs(release): v0.1.0 — stop promising what only one host delivers #817 merged first-release documentation/manifest/changelog alignment. Reason: the owner chose to ship a controlled real-use first version once the core consider / lifecycle boundary was regression-covered rather than wait for broader product acceptance.
PR fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 and PR docs(release): v0.1.0 — stop promising what only one host delivers #817 are merged and CI-green. fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 head run 31068795223 and docs(release): v0.1.0 — stop promising what only one host delivers #817 head run 31069738429 both passed product-contract and qa-eval-tooling on Python 3.11/3.12; network smoke was skipped. PR docs(release): v0.1.0 — stop promising what only one host delivers #817 records two outside comprehension-review passes in its body, but neither PR has a formal GitHub review submission. Reason: distinguish test/review evidence from owner-live acceptance.
The execution priority moved from pre-release correctness freeze to real-use validation. [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 is now the current product-acceptance lane for the published v0.1.0@c40ab16 tree, with [design·perf] Response-time budget by route — separate real-user latency from QA/test overhead #603 collecting live timing segments. No owner-live usefulness/repeat-use pass exists yet. Reason: a tag proves publication, not that the product changes decisions or earns repeat use.
[P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 was removed from the release-critical path. The owner downgraded it P0 → P2 and explicitly deferred it; the cross-client delivery defect remains open and the issue remains assigned, but v0.1.0 makes no whole-product/cross-host GA claim. Reason: do not build a generalized host adapter before real-use evidence says delivery blocks first value.
PR docs(registry): profile.md has no writer, and both surfaces that claimed one now say so (refs #58) #614 is no longer an active stale claim. It was closed without merge during release preparation; its historical profile.md finding remains owned by feat: 第一次復盤後從行為推斷 profile.md + theses.jsonl（少問、猜為主） #58. Reason: the old PR was based on a superseded contract and should not be revived by queue inertia.
[bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 is a newly surfaced scoped correctness blocker, not a newly claimable task. PR fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 made mixed-currency weights safe, exposing that raw positions row amounts still lack their native-currency denomination. The issue is unassigned but has no lifecycle Status: Implementation and still contains a field-placement decision. Reason: severity and an empty PR queue do not satisfy the repository's claimability contract.
Active implementation front remains None; Claimable now remains None. There are no open PRs. Reason: the next justified product action is evidence collection under [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610, and no unclaimed issue currently satisfies the full Implementation pickup contract.

Owner decisions newly relevant after the release: decide from #610 evidence whether any bounded correction is warranted; separately decide whether to promote #816 to an executable Implementation contract. #592 remains an unchanged repository-setting discussion.

Delta — 2026-08-08 08:02 +08:00

No repository-state delta since the 2026-08-07 daily scan:

SHA / release state: unchanged at main@c40ab16ad5c266f26198af64c96e30fd3bec07d1 / v0.1.0; no newer commit landed.
PR state: no open PRs; the most recent PRs fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 and docs(release): v0.1.0 — stop promising what only one host delivers #817 remain merged. No new active PR claim exists.
Priorities: unchanged — [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 real-use usefulness acceptance remains first, with [design·perf] Response-time budget by route — separate real-user latency from QA/test overhead #603 collecting live timing; [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 remains a scoped correctness blocker but is not lifecycle-promoted; [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 remains assigned/deferred P2; [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 remains Discussion.
Claimability / blockers: no item became newly claimable or newly blocked. Claimable now stays empty and there is still no active implementation front.
Owner decisions: no new owner ruling or owner-live acceptance evidence appeared. The next decision still follows the first [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 real-decision receipt; [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 still needs an explicit Implementation promotion before another agent may claim it.

Fresh verification did not change routing: PR-head CI evidence for #817/#818 remains green, while the merge commit itself still exposes no independent combined-status / PR-triggered workflow verdict through the available GitHub status endpoints.

Delta — 2026-08-09 08:03 +08:00

No repository-state delta since the 2026-08-08 daily scan:

SHA / release state: unchanged at main@c40ab16ad5c266f26198af64c96e30fd3bec07d1 / v0.1.0; no newer commit landed.
PR state: no open PRs; fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 and docs(release): v0.1.0 — stop promising what only one host delivers #817 remain merged and their previously observed PR-head Actions runs remain green. No active PR claim was added or removed.
Priority / claimability: unchanged — [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 real-use usefulness acceptance remains first, with [design·perf] Response-time budget by route — separate real-user latency from QA/test overhead #603 collecting live timing; [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 remains a scoped correctness blocker but is not lifecycle-promoted; [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 remains assigned/deferred P2; [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 remains Discussion. Nothing became newly claimable or newly blocked.
Owner decisions needed: unchanged — the next product decision still follows [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610's private real-decision evidence; [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 still needs an explicit promotion plus field-placement ruling before implementation; [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 remains a separate repository-setting decision.

Reason for no routing change: the repository, PR set, and directly owning/acceptance issue states are unchanged since the prior scan.

Delta — 2026-08-10 08:03 +08:00

No repository-state delta since the 2026-08-09 daily scan:

SHA / release state: unchanged at main@c40ab16ad5c266f26198af64c96e30fd3bec07d1 / v0.1.0; no newer commit landed.
PR state: no open PRs and no PR updated since the prior scan. fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 and docs(release): v0.1.0 — stop promising what only one host delivers #817 remain merged; fresh verification confirms their PR-head Actions runs 31068795223 and 31069738429 remain successful. No active PR claim was added or removed.
CI distinction: the merge commit c40ab16ad5c266f26198af64c96e30fd3bec07d1 still exposes no combined-status records through the available status endpoint, so the evidence remains PR-head CI-green plus the recorded merged-combination full-suite run — not an independently observed post-merge-main CI verdict.
Priority / claimability: unchanged — [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 real-use usefulness acceptance remains first, with [design·perf] Response-time budget by route — separate real-user latency from QA/test overhead #603 collecting live timing; [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 remains a scoped truth blocker for mixed-currency positions but is not lifecycle-promoted to Implementation; [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 remains assigned and explicitly deferred at P2; [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 remains Discussion. Nothing became newly claimable or newly blocked.
Owner decisions: none newly required beyond the existing calls: run [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 before another broad product iteration; promote [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 only after settling the denomination contract; resume [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 only if real target-host evidence re-promotes it; decide [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 only if required status checks are wanted.

Reason for no routing change: the published first-version tree, open-PR set, acceptance evidence, lifecycle states and owner rulings are all unchanged.

Delta — 2026-08-11 08:02 +08:00

Changes since the 2026-08-10 daily scan:

SHA / release state: unchanged at main@c40ab16ad5c266f26198af64c96e30fd3bec07d1 / v0.1.0; no newer commit landed. Reason: repository runtime/release state is unchanged.
PR state: no open PRs and no PR updated since the prior scan. fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 and docs(release): v0.1.0 — stop promising what only one host delivers #817 remain merged, not active claims. Fresh verification keeps fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 head run 31068795223 and docs(release): v0.1.0 — stop promising what only one host delivers #817 head run 31069738429 completed successfully. Reason: no implementation ownership entered or left the queue.
CI distinction: fresh checks of merge commit c40ab16ad5c266f26198af64c96e30fd3bec07d1 still return no combined status records and no PR-triggered workflow runs. Evidence therefore remains PR-head CI-green plus PR docs(release): v0.1.0 — stop promising what only one host delivers #817's recorded merged-combination 58/58 full-suite run, not an independently observed post-merge-main CI verdict.
Roadmap authority changed only by research evidence: [research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475 received a 2026-08-10 weekly investment-AI competitor delta after the prior scan. That comment explicitly says research update only, no implementation request, and no change to #27. Reason: record the new roadmap evidence without promoting it into the execution queue.
Priorities / blockers: unchanged — [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 remains the sole product critical path for real-use usefulness evidence, with [design·perf] Response-time budget by route — separate real-user latency from QA/test overhead #603 collecting live timing; [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 remains a scoped correctness blocker but is not lifecycle-promoted; [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 remains assigned/deferred P2; [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 remains Discussion. No item became newly blocked or newly unblocked.
Claimability: Active implementation front remains None; Claimable now remains None. Reason: no open PR exists, but no unclaimed issue currently satisfies the complete Status: Implementation pickup contract.
Owner decisions: no new implementation ruling or owner-live acceptance verdict appeared. The next decision still follows [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610's first real-decision receipt; [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 still requires an explicit lifecycle promotion and field-placement decision before another agent may claim it.

No new issue was needed.

Delta — 2026-08-12 08:00 +08:00

Changes since the 2026-08-11 daily scan:

SHA / release state: unchanged at main@c40ab16ad5c266f26198af64c96e30fd3bec07d1 / v0.1.0; no newer commit landed. Reason: repository runtime/release state is unchanged.
PR state: no open PRs and no PR updated since the prior scan. fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 and docs(release): v0.1.0 — stop promising what only one host delivers #817 remain merged, not active claims. Fresh verification keeps fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 head run 31068795223 and docs(release): v0.1.0 — stop promising what only one host delivers #817 head run 31069738429 completed successfully. Reason: no implementation ownership entered or left the queue.
CI / review distinction: merge commit c40ab16ad5c266f26198af64c96e30fd3bec07d1 still exposes no combined status records and no PR-triggered workflow run through the available endpoints. docs(release): v0.1.0 — stop promising what only one host delivers #817 and fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 still have no formal GitHub review/discussion. Evidence therefore remains merged + PR-head CI-green + docs(release): v0.1.0 — stop promising what only one host delivers #817's recorded merged-combination 58/58 full-suite run, not an independently observed post-merge-main CI or GitHub-review verdict.
Acceptance priority: [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 has no new owner-live verdict and remains the sole product critical path; [design·perf] Response-time budget by route — separate real-user latency from QA/test overhead #603 has no new timing baseline. Reason: the missing evidence is still real-use usefulness/latency, not another implementation slice.
Claimability / blockers: Active implementation front remains None and Claimable now remains None. [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 is still unassigned but not lifecycle-promoted to Implementation; [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 remains an assigned, explicitly deferred P2 owner claim; [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 remains Discussion. No priority was added, removed or reordered.
Roadmap: [research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475 has no update newer than its 2026-08-10 research-only competitor delta, which explicitly authorizes no implementation and no change to tracking: current context index — milestone, critical path, what an agent may pick up #27. Reason: research evidence did not become execution authority.
Owner decisions needed: unchanged — run the next natural private [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 decision and record sanitized [design·perf] Response-time budget by route — separate real-user latency from QA/test overhead #603 timings; promote [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 only after settling its denomination-field contract; decide [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 only if required status checks are wanted. [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 stays deferred unless target-host evidence re-promotes it.

Reason for no routing change: the observed main SHA, open-PR set, acceptance evidence, lifecycle states and owner rulings are all unchanged.

Delta — 2026-08-13 08:02 +08:00

No repository-state or execution-routing delta since the 2026-08-12 scan.

Changed SHA / PR states: none. main remains c40ab16ad5c266f26198af64c96e30fd3bec07d1 / v0.1.0; no PR opened, closed, merged, or became an active claim since the prior scan.
Priorities added / removed / reordered: none. [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 remains the first product-evidence gate; no implementation item moved ahead of it.
Newly claimable / newly blocked: none. Claimable now remains empty; [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 still lacks Status: Implementation and retains its field-placement decision, while [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 remains assigned/deferred P2.
New owner decisions needed: none added. The outstanding calls remain the [QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 private real-decision verdict, whether to lifecycle-promote [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 after settling its exact contract, whether evidence ever warrants resuming [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230, and the separate [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 repository-setting decision.
Fresh verification only: PR docs(release): v0.1.0 — stop promising what only one host delivers #817 head Actions run 31069738429 and PR fix(positions): a mixed-currency book has weights, and one it cannot state says why (closes #737) #818 head run 31068795223 still report success; both PR discussion timelines remain empty. The current merge commit still exposes no combined status records and no PR-triggered workflow run through the available endpoints. Reason: refresh the evidence timestamp without upgrading merged/PR-head-green work into a post-merge-main CI or formal-review claim.

No existing owner needed modification and no new issue was justified.

Delta — 2026-08-15 08:05 +08:00

main changed from c40ab16 to b6bb069; v0.1.0 remains tagged at c40ab16.
PR One expression contract, decoupled from every output surface (closes #823) #824 opened and merged after the prior scan, closing [design·P1] One expression contract, decoupled from output surfaces — voice + disclosure placement defined once, every route inherits #823. There are no open PRs now. PR-head Actions passed; no formal GitHub review or post-merge-main status is exposed.
[design·P1] One expression contract, decoupled from output surfaces — voice + disclosure placement defined once, every route inherits #823's shared expression contract is implemented and complete. This does not establish owner-live usefulness or live delivered-text compliance.
[design·P1] Three outputs, three disclosure lists, one enforcement mechanism — the product is rigorous on the card and freeform everywhere else #525 is now closed not_planned: shared disclosure form/placement is kept, while a generic cross-output obligation-selection abstraction was explicitly rejected.
[QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 remains the next product-evidence gate, but the next owner walk must freeze the new exact candidate b6bb069 before its first timing/verdict event.
Active implementation front remains None; Claimable now remains None. [process·voice·P2] A live consider answer passes through no voice check — three output-voice violations shipped in one owner-live session with a fully green trace #782 stays P2 observation/decision ownership, [bug] positions rows carry no currency, so a mixed-currency value/impact is unlabelled #816 remains unpromoted, [P2·cross-client UX] Codex 無法完成選項互動與卡片交付；artifact 綠不等於 review 完成 #230 remains assigned/deferred P2, [process·infra] The only blocking test gate is one client's local hook; main has no required status check #592 remains Discussion, and [research·roadmap] Staged direction: pre-trade challenge → decision memory → rule learning #475 has no newer implementation authorization.

No new issue was needed.

Delta — 2026-08-19 21:40 +08:00

main advanced from b6bb069 to bea5952 through merged PR Make trade decisions recommendation-first #826; v0.1.0 remains c40ab16.
[design·consider] A cash-deployment question has no route — the premise contract requires the ticker that IS the question #825 is complete as implementation. Its synthetic post-merge comparison proved no regression, not improved usefulness.
Owner-created [implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 is lifecycle-complete Implementation, assigned to atomchung, and claimed on codex/issue-827-permission-boundaries.
Active implementation front changed from none to [implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827. Claimable now remains none because the front is claimed.
[implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 narrows hard boundaries to claims, durable writes, external actions, and privacy; relevant reasoning, research, recommendation, explicit candidate discovery, and useful presentation are no longer semantic refusal gates.
[QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 remains the owner-live usefulness authority. [implementation·skill] Move limits to claim/write/action boundaries — free reasoning, research and recommendation #827 must supply four actual-Skill baseline-vs-candidate runs and a blind owner preference; CI or separately authored artifacts cannot establish product improvement.
There are no open PRs at this refresh. No second implementation branch is authorized.

Routing refresh — #827's implementation front is open in PR #847

Claimed and implemented in one session from main@8c37b04012d4e54dcb9f063f079def2ddd510e78. gh pr list was empty at claim time and the branch is the only open PR from this work. Not merged; no automatic merge.

What moved

One front, four connected changes, two commits — the permission ambiguity this index named, the shape law that caused it to matter, the record-selection failure #844 found, and the continuity gap that made the loop unclosable.

The reference ambiguity this index called out is resolved. references/agent-boundaries.md forbade "rankings", "derived analysis" and arguing a considered trade "from anything but consider's output", all unscoped, while SKILL.md mandates candidate ranking, labelled judgment and a record-led deciding reason. Each prohibition now scopes to the fact it protects: a portfolio-derived number or ranking, a portfolio figure the engine did not emit, and a considered trade's portfolio consequence. The engine ranks the user's positions and only it does; a preference among candidates the user asked about is judgment, labelled, entering no canonical state.
The expression shape was amended at its authority, not overridden from a route file. docs/expression-contract.md §3's increment gate went from does the decision change? to the action, the confidence in it, or the comparison against the alternative in play, with a paired rule that a wait stance owes the evidence that settles it and the next checkable point. Both land in the mother chapter through §3.4's first lane and are logged in §8 — no new V/D/C ID, no character-count ceiling, four named bans unchanged.
[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 item 2 — standing decision rules in the record read, plus the qualifier-fidelity rule. See the comment 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.
review._unresolved_prior projects at most one unsettled same-ticker consideration onto the consider response with the evaluation_id --resolve needs. That identifier was previously emitted only in the response that minted it, so a consideration left open on this route could never be closed on it and _prior_decision stayed permanently empty for that ticker. It carries no decision and no decided_on: an open row is a question that was asked, never a decision.

Evidence

Product group 49/49, whole registry 60/60, tests/test_consider.py 234/234, locally on Python 3.14.

Eighteen mutations across two rounds against the new reader and its call site. Fourteen were caught by the first section R. Four survived, and independent review is what found them — each now has its own test, re-run against the mutation to confirm it reddens:

dropping _fold_evaluations from the open reader, which let one consultation reach the agent as both "you reported acting on this" and "you never told me what you did". The disjointness is the fold's, not the two decision tests'. The assertion that should have caught it was a coin flip — two ids minted the same day, ordered lexicographically, so it passed or failed with the calendar;
a non-string evaluation_id, whose guard was present and untested because max() on a one-row fixture never compares;
a call site hard-coding "buy", reversing the same-side preference for every sell the user brings, with no sell-side current consultation anywhere in section R;
decision != "open" loosened to not in PRIOR_DECISION_RESOLVED, projecting a row carrying null as a question the user was asked and never answered.

A fifth survives by design and the code says so: the evaluation_id exclusion prevents self-recall, not the call's placement before the append.

Six more review findings, all mine, all fixed in the third commit. The most serious: docs/expression-contract.md §3.1's summary table still carried the pre-amendment gate fourteen lines above the amendment, so the mother law gave two different tests on one page. Also: the recap gate I narrowed was defeatable by the contract's own house style and is restored to a whole-file ban; unresolved_prior shipped with no relevance gate, contradicting boundary 5 and SKILL.md's own ask-only-decision-changing-questions rule; and ranking a user's own holdings was reachable through the widened candidate-ranking permission. Details on the PR.

#846 is filed, not bundled: an evaluation_id of an unhashable type raises inside _fold_evaluations before any reader's checks run, taking down a live answer over a corrupt historical row. It reproduces on main@8c37b04 through prior_decision's identical fold, and the fix belongs to the shared reader and its four callers.

One observation that needs no change here but is worth keeping: the product gate is blind to the mirrored-surface class. The five surfaces the second commit synced were all still un-synced one commit earlier, with --group product 49/49 green.

The loaded instruction chain was measured per SHA. The always-on surface (SKILL.md + agent-boundaries.md, the two files an installed host gets) went 14,113 → 15,929 → 18,735 bytes across 5e0c7b6, 8c37b04 and this branch. That growth is prose, it is a real cost, and SKILL.md is now 11,866 of its 12,288-byte budget. The routed 123 KB is the must_state reduction's territory and is deliberately not touched here.

What remains unproven, precisely

No owner-live run and no usefulness claim. The four-lane harness is manual-first and private; route.sh open writes a cross-session marker on the owner's machine and the answers, the blind verdict and the archiving all need a real decision. Not run.
The harness pin is still b6bb069 (2026-08-14). Re-pinning must move the same value in four private places at once and is the owner's call on their own frozen comparison. Until it does, any run measures the August skill.
This PR is a third point. The originally requested pair was 5e0c7b6 vs 8c37b04; whichever pair is chosen, do not attribute the combined effect to 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.
[QA·dogfood] Judgment-quality dogfood — does a real consider answer now carry a supported insight? #610 is untouched and still needs its own real-decision walk with live timings.
The reading-fidelity half has no offline oracle and answer_provenance already says so. evals.json scene 5 puts it where a human reads it.

Serial order after this

Unchanged, and this PR bundles none of it: engine-supplied forward-dated facts and relative strength (#781's deterministic half), then the must_state reduction, then review flows reading the same record. #776/#781 were not opened; no calendar, news or provider expansion is in this branch.

Next safe stopping boundary

Independent review of the diff is done and its findings are on the PR. The boundary is owner decision on the PR plus a named candidate SHA for #844 item 5. Integrity, privacy, idempotency, interruption recovery and explore-before-commit are intact; every added example is synthetic.

Merged. PR #847 squashed to main@22c502e792b66f6314cae3560bb4748231b1259a on 2026-09-05, by owner instruction in session.

CI verified on the merged commit: Actions run 33968463095 — product-contract and qa-eval-tooling succeeded on Python 3.11 and 3.12; network-smoke skipped.

This is implementation evidence on the supported versions. It is not behavioural evidence, and nothing here claims the answers got better. What is now true is only that the changed contract, the widened increment gate, the record-reading rules and the unresolved_prior projection are the code an installed host loads.

The candidate for the next comparison is therefore 22c502e, not this branch's head. The harness pin is still b6bb069 (2026-08-14) and has to move in route.sh, tools/route_fomo.py, templates/manifest.yaml and RUNBOOK together before any run counts.

Open PRs after this merge: none. Open implementation fronts: none — the next cuts in the owner-approved order (engine-supplied forward-dated facts, then the must_state reduction, then review flows reading the same record) each still need their own issue before code. #846 is filed and unclaimed.

2026-09-06 repository-truth correction and context research routing

Latest main is 22c502e792b66f6314cae3560bb4748231b1259a (PR #847), not 8c37b04. Directly inspected main branch, current SKILL.md, the open-PR collection (empty), #844's merge receipt, and main Actions run 33968463095: product-contract and qa-eval-tooling succeeded on Python 3.11/3.12; network-smoke skipped.

Stale-body finding: this index and #827 still describe the integrated wave as not started; #844's body still calls standing-rule selection/qualifier preservation a proposed patch. PR #847 and current code already contain those changes plus unresolved_prior. Do not claim or redo that work from those stale headers. The latest merge evidence is in #844's existing receipt. This comment records the discrepancy; it does not claim those issue bodies have been rewritten or that all behavioral acceptance is complete.

The owner asked to evaluate, not implement, a context-centered next iteration. #446's current body has been updated as Research / Discussion only, preserving the old proposal as history. It defines good context, separates observed style / chosen strategy / demonstrated efficacy, and proposes testing whether a useful conversation forms correctable understanding that helps a later decision. #844 remains the faithful record-use and behavioral-comparison owner; no parallel memory architecture or new active front is created.

Recommendation: prioritize evidence for context selection/use and low-effort context formation over unrelated metric/report expansion, without treating every calculation as waste or requiring every answer to write memory. The immediate verification still needs an actually loaded, correctly pinned candidate, the existing same-context/no-Skill comparator, and honest source/decision/understanding/effort observations. A combined 22c502e result is not an isolated #845 result. #610 remains owner-live acceptance authority.

This research pass changed no code, runtime data, public positioning, milestone, merge approval or deployment. No new model run or live acceptance is claimed. Public evidence remains generic/synthetic.

tracking: current context index — milestone, critical path, what an agent may pick up #27

Description

Status

Current milestone

Active / claimable implementation

Ownership / required context

Evidence and acceptance that remain true

Next safe stopping boundary

Prior ordering — superseded only for this promoted front

Status

Current milestone

Repository truth

Active implementation fronts

Claimable now

Acceptance and blockers

Safe parallel work

Serial next steps

Stop conditions

Historical routing correction — 2026-09-05

Activity

atomchung commented on Jul 3, 2026

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

atomchung commented on Jul 5, 2026

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

atomchung commented on Jul 5, 2026

2026-07-05 pm 批次收尾(merge + 大掃除)

atomchung commented on Jul 29, 2026

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

atomchung commented on Jul 29, 2026

Context delta — 2026-07-29 split review after PR #567

atomchung commented on Jul 31, 2026

Routing update, 2026-07-31 — Front A is implemented and in review; Front B is unblocked and unstarted

Front A · #539 / #536 — done, PR #659 (draft)

Front B · #403 / #450 — unblocked, not started

Unchanged

atomchung commented on Jul 31, 2026

2026-07-31 · daily control-plane decision

Decision

Current result

atomchung commented on Jul 31, 2026

2026-07-31 QA-round routing delta

atomchung commented on Jul 31, 2026

Delta — 2026-08-01: both QA repair fronts executed to completion

Front A — truth and usable pre-trade evaluation (complete)

Front B — review lifecycle and honest evidence (complete)

New issues filed during the campaign

Body staleness note

atomchung commented on Aug 1, 2026

Daily delta — 2026-08-01 08:03 +08:00

atomchung commented on Aug 1, 2026

Routing update — #683 owner-approved prototype lane (2026-08-01 11:05 +08:00)

atomchung commented on Aug 1, 2026

39 remaining items

atomchung commented on Aug 6, 2026

2026-08-06 — delta: v0.1.0 tagged and released

atomchung commented on Aug 6, 2026

2026-08-06 — v0.1 first-version product assessment

Status

Current repository truth

User outcome available now

Explicit non-capabilities

Blocker classification

Public positioning decision

README / promotion assessment

Next smallest step

atomchung commented on Aug 7, 2026

Delta — 2026-08-07 08:01 +08:00

atomchung commented on Aug 8, 2026

Delta — 2026-08-08 08:02 +08:00

atomchung commented on Aug 9, 2026

Delta — 2026-08-09 08:03 +08:00

atomchung commented on Aug 10, 2026

Delta — 2026-08-10 08:03 +08:00

atomchung commented on Aug 11, 2026

Delta — 2026-08-11 08:02 +08:00

atomchung commented on Aug 12, 2026

Delta — 2026-08-12 08:00 +08:00

atomchung commented on Aug 13, 2026

Delta — 2026-08-13 08:02 +08:00

atomchung commented on Aug 15, 2026

2026-08-06 — delta: `v0.1.0` tagged and released