feat: 第一次復盤後從行為推斷 profile.md + theses.jsonl（少問、猜為主）

[atomchung]

問題

第一次復盤結束後，~/.trade-coach/profile.md 和 theses.jsonl 應該由 Claude 從交易行為推斷建立，但現在完全沒有執行——兩個檔案都是空的。

這導致：

• 下次給規矩時不知道用戶的風格，給出的規矩可能不適配
• 持倉 thesis 為空，無法做 trigger 對帳
• 用戶冷啟動沒有任何「你的交易 DNA」可以 review

設計原則：少問、猜為主

可以猜（行為算得出來）	必須問（why 算不出來）
持有風格（中位持倉天數）	加碼是定投還是凹單
集中傾向（AI 暴險 %）	sizing 有沒有算過上限
紀律缺口（出場 / 加碼哪個弱）	賣出是 thesis 到位還是怕回吐
標的 thesis why（從加碼行為推）	—

猜的一律標 inferred，用戶看到不對再改，不強逼填寫。

預期行為

第一次復盤收尾時（Step 4 之後）：

1. 建 profile.md：從引擎行為數據猜 3 條個人原則（持有風格 / 集中傾向 / 紀律缺口），以人話寫入，每條標「(AI 推斷，可修改)」
2. 建 theses.jsonl：對每個當前持倉，從 ticker_diagnosis + 加碼行為推斷 why / 三個 trigger，標 maturity=inferred，直接 append，零提問；只對「行為矛盾、金額最大的 1 檔」問一句確認
3. 規矩適配性：給候選規矩前先對照 profile（定投型 vs 短打型 → 給不同的規矩模板）

關聯

• SKILL.md 收尾 part 2/3 已有規範但未執行
• 產出供 feat: 用戶隨時可讀自己的策略庫（profile + theses + 承諾歷史） #57（策略庫查看介面）讀取使用

[atomchung]

Status check, 2026-07-27: the direction is ratified, but this issue's target file is a fossil

Scoping #446 (the distillation layer) revisited this issue. Two findings.

The direction here — infer rather than ask — is now the ratified default. Owner, 2026-07-27:

一切以行為為主 … 現階段不主動問，這是後續討論點 但還不關鍵

and, on whether the product should open with an explicit "tell me your goals and principles" step: 「綁在交易行為記錄中 有多少紀錄多少」. There is no dedicated profile-building interview. What this issue proposed in 2026-07 is what the product does.

But profile.md does not exist and never did. coach.py's directory listing still describes it as 「交易目標 + 個人原則（第一次復盤時建立，Claude 直接寫檔）」. Verified against current main: no writer, no reader, and no instruction anywhere in SKILL.md, references/, or any flow that would create it. git log -S"profile.md" shows the line has been in the directory description since ee686b2 (2026-07-13, PR #183) without a corresponding implementation.

The structured model that actually carries this responsibility today is rules.jsonl + conditions.jsonl + profile.json (the last holding only max_position_pct and muted_rules). The free-text profile.md is a fossil of the earlier design, left in a user-facing directory description that says it exists.

Two things follow, neither of them this issue's original scope:

1. The directory description should stop advertising a file the product never writes.
2. theses.jsonl — the other half of this issue's ask — is real and populated; the inference half of this issue is partly shipped via the per-position inferred thesis in flows/first-review.md.

Retitling or closing this issue in favour of #446 is worth considering once the profile layer's shape settles. Not closing it now, since the "infer, don't ask" principle it states is still the live default and is cited from #446.

Refs #446, #403.

[atomchung]

Current state of profile.md, 2026-07-30 — registered, exportable, and written by nothing

Found while auditing #597's evidence section against main@c767018. This issue asked for profile.md to be inferred at the end of the first review; it still is not, and the surrounding surfaces have since drifted in a way that hides that.

What exists today:

• profile.md is registered in coach.py's DATA_FILES (("profile.md", "text", "交易目標 + 個人原則(第一次復盤時建立,Claude 直接寫檔)")), so data-status / data-export / data-reset all see it and it has its own dedicated test in tests/test_coach_data_cli.py — the registry-completeness scan cannot see it, because the writer is supposed to be the agent rather than an engine/ module.
• README.md and both localized READMEs still tell the user to cat ~/.trade-coach/profile.md for "your trading goals + 3 personal principles (the baseline each review compares against)".
• No instruction surface tells any agent to write it. skills/fomo-kernel/SKILL.md, skills/fomo-kernel/flows/*.md and skills/fomo-kernel/references/*.md contain no occurrence of profile.md. The only profile text in references/ is data-contract.md, and both mentions are profile.json (muted rules, standing single-position cap) — a different file that holds only max_position_pct and muted_rules.
• docs/maintainer-guide.md nevertheless asserts the opposite as settled fact: profile.md is listed as one of two already-registered entries "the scan cannot see" because "the agent writes it directly per SKILL.md". SKILL.md does not say that.

So the current shape is: a registered, user-visible, exportable and resettable file that nothing produces, documented in the maintainer authority as agent-written and advertised in the README as a review baseline.

Two separate things follow, and they are not the same size:

1. The maintainer-guide.md claim is wrong and should be corrected, independently of whether this feature is ever built. It is a mirrored-surface drift in the document that other surfaces are told to trust, and it is the kind of statement a future audit would use to conclude the writer exists.
2. Whether profile.md should be written at all is this issue's call. [research·product] Philosophy bootstrap: decision guidance before transaction history #597 (book-free decision framing) explicitly ruled that profile.md must not become a second canonical truth for provisional principles, because it has no append-only provenance, confidence, supersession or replay contract. That ruling is about [research·product] Philosophy bootstrap: decision guidance before transaction history #597's own scope and does not decide this issue — but it does mean an implementation here should choose deliberately between reviving the agent-written file, re-homing the content into an append-only projection, or retiring the registry entry and the README line.

No file changed and no scope claimed by this comment; recording the measured state so the decision is not made from the README.

[atomchung]

Claiming the doc-only half: the false-writer claim has two carriers, not one

Following the 2026-07-30 comment above, which separated finding 1 (docs/maintainer-guide.md asserts profile.md is agent-written) from finding 2 (whether the file should exist at all, this issue's call). Taking finding 1 only. Verified against main@94b7e5f.

The un-recorded half: the same false sentence lives in a second surface, and that one is the original.

• docs/maintainer-guide.md:137 — `profile.md` (the agent writes it directly per `SKILL.md`)
• tests/test_coach_data_cli.py:303-304 — profile.md (written directly by the agent per SKILL.md, never by engine code)

git log -S puts the test comment first (6a82ec8, #471, when the registry scanner was written) and the guide's copy second (78aedc4, #587, when the shared contract left CLAUDE.md). So the guide did not invent the claim; it inherited it. Correcting only the guide would leave the assertion alive in the file the guide's own row points at as the coverage evidence — the mirrored-surface defect docs/maintainer-guide.md names as this repository's most frequent.

The test comment also carries a second, smaller error the guide does not: it says both unseen entries "are written from outside that scope", when profile.md is not written from anywhere.

Re-verified, not taken from the earlier comment. grep -rn "profile" skills/fomo-kernel/SKILL.md skills/fomo-kernel/flows/ skills/fomo-kernel/references/ returns exactly two hits, both profile.json (data-contract.md:68,70 — muted rules and the standing single-position cap). No profile.md occurrence on any agent instruction surface.

A third mention checked and deliberately left alone: tests/test_sizing_literal_gate.py:58 cites "the two paths it names as outside its own reach" as a scope-naming precedent. That statement stays true after the fix — the scanner does still name them — so it carries no drift.

Scope of what I am doing now: correct both locations together, and have each name the other as its mirror so the pair cannot drift again. Doc/comment text only.

Not touching, because they are this issue's call and not mine:

• coach.py's DATA_FILES entry and its user-facing description 「第一次復盤時建立,Claude 直接寫檔」, which is the same false claim in the one place a user reads it (data-status). Recorded above already; deregistering it would change data-export/data-reset coverage.
• README.md:111 / README.zh-TW.md:111 / README.zh-CN.md:108, which still tell the user to cat it as "the baseline each review compares against".

Both are downstream of the revive / re-home / retire choice this issue owns. The correction I am landing makes the maintainer authority stop testifying that a writer exists, so that choice gets made from measurement rather than from the guide.