dsh-persona-forge 0.3.1 → 0.8.3

package/README.en.md

@@ -9,7 +9,7 @@ character's voice by the harness LLM before it is sent — following the model
 the session is already using. Choose whether the rewrite goes straight out or
 lands back in the composer for review.
 
-> **Status:** early (`0.3.x`). The card format and the HTTP contract are stable
+> **Status:** early (`0.8.x`). The card format and the HTTP contract are stable
 > enough to build on; expect additive changes and read the changelog before
 > upgrading.
 
@@ -30,12 +30,36 @@ an explicit action.
 
 ### A word on what personas actually do
 
-Persona instructions change **style, register, length and compliance** — how
-readily the model agrees and how willing it is to push back. They do **not**
-raise the model's capability ceiling: there is no dormant potential that
-naming an identity unlocks.
+First, a fact: a persona does **not** unlock hidden capability. There is no
+dormant potential that naming an identity wakes up — the ceiling is the ceiling.
 
-Two consequences worth knowing before you use this:
+But the other half is true too: **a persona really does change how the model
+thinks.** It changes style, register, length and **compliance** — how readily
+the model agrees, how willing it is to push back, and how seriously it will
+commit to an absurd metaphor.
+
+This project was born in that second, slightly unreasonable place. The
+inspiration is a 30-second video on Bilibili:
+
+> [**“ds：好的，我现在是万机之神了（”**](https://www.bilibili.com/video/BV1Ata26pE92) — by 身居幻想乡の长乐安生
+
+The comments put it better than any spec sheet:
+
+> Another AI is you pretending to be Guan Yu — however well you act, you are not
+> Guan Yu. But when **Big Fat Fish** pretends to be Guan Yu, she genuinely
+> believes she is Guan Yu, and really does ride out and cut down Hua Xiong.
+
+"Okay, I am the Machine God now" — the model is not *humouring* the bit. It has
+**accepted the identity**. What this plugin does is make that moment reusable:
+put a character card in front of your draft, and let DeepSeek speak as that
+character.
+
+**It is tuned for DeepSeek Flash.** Flash is fast, cheap and plentiful — the
+tier you actually use all day — and letting it answer you once in the voice of
+the Machine God is the entire point. It runs on other models too, but that
+"one line and it is in character" feel was tuned against Flash.
+
+**So do not treat it as a performance switch.** Two things worth knowing:
 
 1. A persona that frames you as a supplicant and the model as an authority
    makes the model **less likely to question your plan**. For coding work that
@@ -43,8 +67,12 @@ Two consequences worth knowing before you use this:
    deference.
 2. "Longer and more ornate" is easy to mistake for "better". It usually is not.
 
-So: use it for tone and framing, keep **review mode** on, and judge the output
-by whether your requirements survived — not by how impressive it sounds.
+Use it for tone and framing, keep **review mode** on, and judge the output by
+whether your requirements survived — not by how impressive it sounds.
+
+> The two bundled cards are the two tempers of that video: **万机之神 · 欧姆弥赛亚**
+> is the solemn Mechanicum litany, and **肌肉集团 · 硬邦邦** is its antidote — a
+> machine-gun-tempo crude client who says it all in fewer words than the litany.
 
 ---
 
@@ -128,33 +156,71 @@ you want them gone.
 
 1. Click 🎭 in the composer tool row and pick a character. **No persona**, at the
    end of the same list, turns rewriting off — that is the state the control
-   starts in, and your choice is remembered per session.
-2. Set **intensity** and **send mode**:
+   starts in. **Which character you are using is remembered per session.**
+2. The selected character grows its own **"Settings for <name>"** row right
+   under it. Open it for **intensity**, **send mode**, **rewrite mode** and
+   **fidelity**:
    - Intensity is a four-rung ladder: **Light · Medium · Strong · Zealot**
      (轻 · 中 · 重 · 狂热). Light adds a touch of flavour; Medium is clear but not
      dominant; Strong gives the voice real weight while keeping the request the
      first thing the reader sees; Zealot is ritual throughout. A card carries
-     its own default, and this choice overrides it for your session.
+     its own default, and this choice overrides it.
    - **Review first** — the rewrite appears in the comparison panel below the
      composer; you read it and then press Send there. Your draft is untouched
      until you confirm.
-   - **Send directly** — the rewrite is sent as your message immediately, and
-     the panel stays open afterwards showing exactly what went out, because a
-     sent message cannot be recalled.
+   - **Send directly** — the rewrite is sent as your message immediately, and the
+     panel then **closes itself**. There is nothing left to confirm, and a panel
+     left up would block the next message.
 3. With a character selected, the composer's normal **Send** button — or
    **Enter** — *is* the rewrite trigger.
 
-**More settings**, at the bottom of the menu, is folded shut by default. It holds
-**rewrite mode** and **fidelity** — per-session OVERRIDES of a card's own
-declaration that most people set once and never touch, so they do not take space
-in a menu whose job is choosing a character. When something *is* overridden, the
-collapsed row names it, so a choice you forgot about cannot hide behind the fold.
+<img src="assets/menu.png" alt="The character menu: three cards, No persona, and the selected character's expanded settings (intensity, send mode, rewrite mode, fidelity)" width="400">
+
+**Everything that configures a character lives in its own "Settings for <name>"
+fold, and follows the CHARACTER rather than the session.** "硬邦邦 always goes out
+Zealot, un-reviewed" is a fact about that character rather than a temporary
+choice in one conversation — so configure it once and picking it in another
+session gives you the same settings. Switching character **re-reads that
+character's own settings** rather than carrying the previous one's intensity
+across.
+
+That fold holds four things:
+
+| | |
+|---|---|
+| **Intensity** | Light · Medium · Strong · Zealot |
+| **Send mode** | Review first / Send directly |
+| **Rewrite mode** | Card default / Model rewrite / Template |
+| **Fidelity** | Card default / Voice only / Constraints too |
+
+Every row explains itself on **hover** rather than in permanent prose. The folded
+row itself **summarises the current settings**, so you can see what a character is
+set to without opening it.
+
+**All four follow the character card** — pick it in another session and you get
+the same set.
+
+**The card directory has exactly one entry point: the settings page.** It is not
+in the rewrite menu — it is occasional housekeeping, and it should not sit in the
+path of choosing a character.
+
+**The rewrite column is editable.** What the model wrote is a starting point,
+not a verdict: fixing a dropped identifier or a tone that is slightly off by hand
+is faster than another round trip. Once you edit it the panel says so, and
+**Send** and **Fill** both carry *your* text.
+
+<img src="assets/review-panel.png" alt="The review panel: the original draft on the left, the editable rewrite on the right, and beneath them the local check verdict, model and elapsed time, plus Send / Restore original / Rewrite" width="820">
+
+**While the panel is open, the composer's send arrow is deliberately inert.** The
+panel is the decision point; pressing send now would either push the un-reviewed
+draft out or start a second rewrite of the same draft. Pick an action in the
+panel, or close it, and the composer comes back.
 
 The rewrite runs **once per draft**: after a rewrite lands, sending goes through
 as-is instead of rewriting the output again. Editing the draft re-arms it.
 
-Your draft is preserved: *Restore original* puts it back, and *Fill only* puts
-the rewrite in the composer without sending.
+Your draft is preserved: *Restore original* puts it back, and also rewinds the
+"already rewritten" memory, so a rewrite you cancelled can be run again.
 
 ### How send interception works, and its one risk
 
@@ -186,40 +252,28 @@ message.
 
 ### The preservation check
 
-After each rewrite, a second call on the same model compares the rewrite
-against your original and reports whether technical content survived:
-
 The biggest risk is not a voice that sounds wrong — it is a rewrite that
-**quietly loses your request**. Two layers cover it, and they are complementary
-rather than redundant.
-
-**Layer 1: the local check — always on, free.** It compares the **technical
-tokens** of the draft against the rewrite, in this process: identifiers, file
-paths, versions, numbers, `--flag`s, acronyms like `API`. Anything lost is
-**named** ("Lost: a.js"). It calls no model, so it costs no tokens and adds no
-latency — and it cannot be turned off.
+**quietly loses your request**. The plugin compares the **technical tokens** of
+the draft against the rewrite, in this process: identifiers, file paths,
+versions, numbers, `--flag`s, acronyms like `API`. Anything lost is **named**
+("Lost: a.js").
 
-**Layer 2: the model check — off by default, switchable in the menu.** It makes
-one **extra model call** per rewrite, to see what a token comparison cannot see
-at all:
+It calls **no model**, so it costs no tokens and adds no latency — and it cannot
+be turned off. It always runs.
 
-- a requirement reworded into something **weaker** (every word survives, the
-  demand does not);
-- a constraint **added in prose**.
+**What it cannot see** (and the panel says so, so you do not assume more): a
+requirement reworded into something **weaker** while every word survives, a
+constraint **added in prose**, a quantity written in Chinese words (「两次」), or
+a bare lowercase English word that is not a path, flag or identifier.
 
-The cost is doubled latency and doubled tokens, which is why it ships off: the
-most common and most damaging failure — a dropped filename, number or version —
-is already proven for free by layer 1, so paying a model call on every send to
-re-check it is a waste. Turn it on when the drafts are technical enough to
-warrant it.
+> A second layer once existed: one **extra model call** per rewrite aimed
+> exactly at those cases. It has been **removed** — it doubled the latency and
+> the token cost of every send, and the most common and most damaging failure (a
+> dropped filename, number or version) was already proven for free. For the
+> residue, **the rewrite column is editable**, which is more direct than another
+> round trip.
 
-The switch lives in **Settings → Persona** ("Enable the model check") and is a
-global preference: it answers "do I want to pay for the second call", which is
-decided once rather than re-asked on the way to every send. With it off the panel
-states plainly that the local check ran and spent no tokens, and says what that
-check does **not** cover — a narrow verdict has to declare its own limits.
-
-Both layers are **advisory** and never block a rewrite: they annotate, not stop.
+The check is **advisory** and never blocks a rewrite: it annotates, it does not stop.
 
 ---
 
@@ -232,9 +286,9 @@ in makes the character appear without a restart**.
 ~/.dsh/persona-cards/<id>.yml
 ```
 
-Use the 🎭 menu's *Open card directory*, or the settings page, to get there.
-You can also create and edit cards entirely in **Settings → Persona**; that
-writes the same files.
+Use *Open directory* in **Settings → Persona** to get there (the rewrite menu no
+longer carries an entry). You can also create and edit cards entirely in
+**Settings → Persona**; that writes the same files.
 
 On first run the bundled cards are **copied into your card directory** so the
 folder is a usable starting point. A copy shadows the bundled card of the same
@@ -267,25 +321,66 @@ draft box sits at the **top of the form that opens**, not as a second competing
 entry on the page. A hint line on the settings page tells you it exists, so you
 do not have to press a button to discover it.
 
-Describe the character in a sentence:
+<img src="assets/settings.png" alt="The 角色扮演 settings page: New character card and the expanded AI draft's Generate draft are boxed, and the footer names the harness default model this run will use" width="420">
+
+**You** write one sentence: **who the AI plays, who you are, and how they talk.**
 
-> a grumpy but utterly reliable old lighthouse keeper who speaks in short
-> sentences and nautical metaphors
+> the AI plays a grumpy but utterly reliable old lighthouse keeper, and I am his
+> apprentice; he speaks in short sentences and nautical metaphors
 
-The model fills in the whole card — the style text, a name, an icon, a
-description, and the **conversion examples for all four intensity levels**.
+**The AI** then fills the form by the character-card fields: a name, an icon, a
+description, the style text, and the **conversion examples for all four
+intensity levels**.
+
+> **The rewrite is always YOUR message, never the character's line.** If you say
+> "the AI is my mother", the example must be **the child speaking to her mother**
+> (`妈，帮我把标题改成首页好不好？`), **not** the mother speaking
+> (~~`乖，妈帮你把标题改成首页`~~). Get this backwards and the card is useless —
+> the rewrite comes out as the character answering for you.
+
+The result **only fills the form — it is never saved automatically**. Two things
+are deliberately fixed: **fidelity** is always voice-only, and **rewrite mode**
+is always model rewrite. A generated card should not quietly acquire the power to
+add constraints to your request; raise it by hand if you want it.
+
+> **All four intensity sets must come back, or the generation counts as failed.**
+> The examples are the longest part of a card, and a model occasionally runs out
+> of output budget part-way through them. That used to be accepted as-is, which
+> left the missing levels with **nothing to demonstrate** — it looked fine and was
+> quietly worse. It now reports an error; retrying once is usually enough.
 
 **Which model it uses has to be stated precisely:**
 
 | Case | Which model |
 |---|---|
-| The **rewrite** from the 🎭 menu | **Follows the current session** (the provider/model of its last request header) |
+| The **rewrite** from the 🎭 menu | **Follows the model the current session is using** — preferring the one you have just **picked in the model dropdown** (even before you send with it), then the provider/model of the session's last request |
 | **AI draft** in the settings page | The **harness default model** — because `settings.section` receives only `{ close }` from the shell and therefore **has no session id to follow** |
 
 These are usually the same model (the default is usually what the session is
 using), but **not guaranteed**: if you switch models inside a session, the
 settings page's draft still uses the default.
 
+> The rewrite reads the host's own precedence: **pending pick → last request
+> header → harness default**. So changing the model in the dropdown takes effect
+> immediately, with no message to send first. Reading the request header alone
+> sent a rewrite to the **previous** model whenever you had switched and not yet
+> spoken — that was a bug, and it is fixed.
+
+> What is followed is the **provider / model**; the **reasoning depth is never
+> specified explicitly**. The rewrite sends no `reasoningEffort` field and lets
+> the harness decide per route.
+>
+> That is not laziness, it is the only portable spelling. The harness validates
+> the field *before* handing the request to any provider: when a route's model
+> declares **no** reasoning metadata, **any present value throws** — including
+> the empty string (the host's own "Provider default" sentinel) and including
+> `off`. `hy4-preview` on `workbuddy-xdpool` is exactly such a route, and
+> pinning `low` there broke every rewrite outright.
+>
+> The cost: on a reasoning model the rewrite may still spend thinking time
+> before emitting text. The panel's elapsed seconds and the 120s deadline are
+> the current answer to that.
+
 So you do not have to guess, the settings page **shows the model it will use**
 (`provider / model`) and says that it is the harness default; after generating it
 shows the model it **actually** used. If neither can be resolved (no message sent
@@ -319,6 +414,14 @@ mode: llm                  # llm = model rewrite (recommended) | template
 fidelity: style            # style = voice only | strategy = may add constraints
 intensity: medium          # light | medium | strong | zealot (the card's default)
 
+# The fixed wrapper for template mode: no model call, {{input}} is replaced
+# verbatim with what you typed. Every card should carry one, or "Template" is a
+# switch that does nothing.
+template: |
+  万机之座在上，吾等是您卑微的信徒，谨向欧姆弥赛亚祷告：
+  {{input}}
+  恳请您垂听吾等的祈求。愿机魂安宁，愿数据各安其位。
+
 # What the character IS. This is an instruction to the rewriting model, not
 # finished copy: the output is written fresh for each of your messages.
 style: |
@@ -326,32 +429,63 @@ style: |
   【发言者方向】用「信徒」的口吻说话，向万机之神恳求……
   语气：庄严、古奥、仪式化。
 
-# Conversion demonstrations: input draft → persona voice.
-examples:
-  - from: 帮我写个快速排序
-    to: >
-      万机之座在上，吾等恳请您赐予枢机之序的奥义：
-      请为我们编写一个快速排序算法，使重复之数各归其位。
-
-# Optional: different demonstrations per intensity. If omitted for a level,
-# the flat `examples` list is used instead.
+# Conversion demonstrations: input draft → persona voice, one set per level.
+#
+# Four sets is all you need. An earlier version also had a flat `examples:` list,
+# but it was only the fallback for a level with no set of its own — on a card
+# with all four populated it was never reached, yet still had to be maintained.
+# Write only this now. (A flat `examples:` in an older file is still READ and
+# still falls back; it is just no longer something you need to write.)
 examplesByIntensity:
   light:
     - from: 帮我写个快速排序
       to: 万机之座在上，恳请您为我们编写一个快速排序算法。
+  medium:
+    - from: 帮我写个快速排序
+      to: >
+        万机之座在上，吾等恳请您赐予枢机之序的奥义：
+        请为我们编写一个快速排序算法，使重复之数各归其位。
+  strong:
+    - from: 帮我写个快速排序
+      to: 伟大而不朽的万机之神，恳请您赐予枢机之序的奥义……
   zealot:
     - from: 帮我写个快速排序
       to: 伟大而不朽的万机之神……
+
+# The SECOND set, shown when fidelity is `strategy`. All four levels again.
+#
+# It demonstrates "voice + behavioural framing" — framing being an instruction
+# about HOW to do the work (do not over-plan, do not stop at the first plausible
+# cause), never a restatement of the request. Without this set a user who raises
+# fidelity gives the model nothing to imitate, and it pads instead.
+examplesByIntensityStrategy:
+  light:
+    - from: 帮我写个快速排序
+      to: 万机之座在上，请为我们编写一个快速排序算法，径直给出可用之器。
+  medium:
+    - from: 帮我写个快速排序
+      to: >
+        万机之座在上，恳请您赐予枢机之序的奥义，使重复之数各归其位。
+        另有一请：请直取可行之解，不必先为吾等铺设长篇筹谋。
+  strong:
+    - from: 帮我写个快速排序
+      to: 伟大而不朽的万机之神，恳请您赐予枢机之序的奥义……勿止步于初见可行之解。
+  zealot:
+    - from: 帮我写个快速排序
+      to: 伟大而不朽的万机之神……其一，勿以理论充作交付；其二，勿满足于最初可行之解。
 ```
 
 ### The two rules that make a card work
 
-**1. State the speaker direction.** The character is the *assistant's*
-identity, but the text being rewritten is *yours*. So the rewrite must speak in
-your voice — a petitioner addressing the character, never the character issuing
-orders. This is the single most common way a persona card goes wrong: a card
-that has the god commanding the assistant inverts the relationship and the
-model follows the wrong half of the framing. Write it out explicitly.
+**1. State the speaker direction.** The text being rewritten is always *yours*,
+but **which side the character is on is your decision**: the AI can play the
+character (and you address it as a petitioner, a client or a subordinate), or
+**you can be** the character (speaking in its voice to the AI). Both are valid.
+The only thing that goes wrong is **leaving it unsaid** — an implicit direction
+leaves the model guessing, and guessing wrong inverts the whole relationship (a
+petitioner's prayer comes out as the god issuing orders). Write it out
+explicitly. The two bundled cards take one side each: 欧姆弥赛亚 is "the user is
+a believer, the AI is the Machine God"; 硬邦邦 is "**the user is** the client".
 
 **2. Make your examples preserve technical facts.** Examples are the strongest
 style anchor, and the model imitates them — including their mistakes. If an
@@ -367,29 +501,50 @@ substance. Run it after editing a card.
 
 | Value | Meaning | Use when |
 |---|---|---|
-| `style` | Voice only. The request and its constraints are untouched. | Default. Almost always this. |
-| `strategy` | The rewrite may **add behavioural constraints** that fit the character. | Deliberate cases — e.g. the bundled *肌肉集团* card, which adds "don't over-plan, don't search, don't ship beginner-tier solutions". |
-
-`strategy` genuinely changes your request, not just its wording. Keep review
-mode on for those cards.
-
-**Fidelity is a property of the card. It can be overridden per session, but only
-ever narrowed, never widened:**
-
-- **Narrowing (allowed)** — a `strategy` card can be asked to behave as
-  voice-only for one message. That only ever **takes away** its power to add
-  constraints, so it is strictly safer than the card's own default.
-- **Widening (refused)** — a `style` card cannot be told to become `strategy`.
-  Widening lets the rewrite add constraints you never typed, which is the card
-  author's decision, not something to grant with a click on the way to Send. To
-  widen, edit the card.
-
-The control appears under *More settings*, and **only when the card is
-`strategy`**. On a `style` card it is not rendered at all: both options would
-resolve identically there, and showing a control that cannot do anything — plus a
-paragraph explaining why the other one is refused — is worse than showing
-nothing. The server enforces the same rule, so the UI never offers a control it
-would ignore.
+| `style` | Voice only. | Default. Almost always this. |
+| `strategy` | The rewrite may **add behavioural framing**. | Deliberate cases — e.g. the bundled *肌肉集团* card, which adds "don't over-plan, don't search, don't ship beginner-tier solutions". |
+
+**"Constraint" means two different things in this document — do not conflate
+them:**
+
+| | What it is | May it change? |
+|---|---|---|
+| **Technical content** | What you are asking for, which file, which number, which limit | **Never, under any setting** |
+| **Behavioural framing** | How the assistant should approach the work — how direct, whether to plan first, what counts as good enough | `strategy` may add it |
+
+What `strategy` adds is the **second** kind, and it **genuinely changes the
+request you send** — it puts something in your message that you did not type. So
+keep **Review first** on for those cards.
+
+> The prompt used to conflate the two: the hard rules said "never add, remove,
+> widen or reinterpret a requirement" while declaring themselves an override of
+> everything *including the character description*, and the `strategy` permission
+> was written before them. The model obeyed the absolute prohibition, so
+> `strategy` was a **no-op** — flipping "Constraints too" in the UI changed
+> nothing about the output. The hard rules now cover technical content only, and
+> say explicitly that they do not forbid behavioural framing.
+
+**Fidelity is a property of the card, but you can override it per character, in
+both directions:**
+
+| Option | Meaning |
+|---|---|
+| **Card default** | Use whatever the card declares on the settings page |
+| **Voice only** | Wording changes; your request does not |
+| **Constraints too** | The rewrite may also add the character's working style (e.g. "no long planning") |
+
+**"Constraints too" genuinely changes your request**, rather than only rewording
+it — it adds things you never typed. So pair it with **Review first**.
+
+> This rule used to be one-way: a `strategy` card could be narrowed to `style`,
+> but a `style` card could never be widened. The reasoning was that a click on
+> the way to Send should not grant the power to add constraints to your request.
+>
+> That reasoning does not hold up: **the card is your own file**, and the
+> settings page has always let you set `fidelity:` directly. Refusing in the menu
+> did not remove the capability — it only made the choice **disappear from the
+> place where you actually work**. All three options are now available on every
+> card.
 
 ### `mode: template`
 
@@ -409,16 +564,12 @@ Set in the plugin's loader row config:
   name: dsh-persona-forge
   config:
     sendMode: review   # review | direct — the initial composer default
-    factCheck: false   # default for the MODEL check (off by default; see
-                       # "The preservation check"). The local token check does
-                       # not depend on this: it always runs and is free.
     watchCards: true   # reload the catalog when the card directory changes
 ```
 
-`factCheck` controls only the default for **layer 2**; the menu switch overrides
-it per session. Setting it to `true` adds **one extra model call to every
-rewrite** (doubled latency and tokens) and is only worth it when drafts are
-technical enough.
+The technical-token check depends on no configuration: it always runs and is
+free. The retired `factCheck` key is ignored, so a profile still carrying it
+changes nothing.
 
 ---
 
@@ -430,8 +581,8 @@ refused). Mounted under `/persona-forge`.
 
 | Route | Body | Returns |
 |---|---|---|
-| `/persona-forge/cards` | `{}` | `{ cards, diagnostics, directory, sendMode, factCheck, draftModel }` |
-| `/persona-forge/rewrite` | `{ cardId, text, sessionId?, intensity?, mode?, fidelity?, modelCheck? }` | `{ text, provider, model, mode, cardId, fidelity, intensity, elapsedMs, facts, factCheck }` |
+| `/persona-forge/cards` | `{}` | `{ cards, diagnostics, directory, sendMode, draftModel }` |
+| `/persona-forge/rewrite` | `{ cardId, text, sessionId?, intensity?, mode?, fidelity? }` | `{ text, provider, model, mode, cardId, fidelity, intensity, elapsedMs, facts }` |
 | `/persona-forge/cards/save` | `{ card }` | `{ id, path }` |
 | `/persona-forge/cards/delete` | `{ id }` | `{ removed }` |
 | `/persona-forge/cards/reset` | `{ id }` | `{ id }` — replaces the user card with the bundled one |
@@ -475,7 +626,8 @@ npm run check                 # both
 `check.mjs` needs no harness and no model. It covers the built-in YAML reader
 (including its refusals and dump/parse round trips), card normalization, the
 bundled cards' fact preservation, prompt construction, output normalization,
-the fact-check parser, and client-bundle syntax.
+the local fact comparison, route resolution order, the send-gesture decision,
+and client-bundle syntax.
 
 `host-check.mjs` mounts the real plugin entry on a stub Cordis context and
 drives the real HTTP routes over a real socket, with a stubbed `llm` service.
@@ -506,7 +658,7 @@ naming the line.
 | `lib/index.js` | Host entry: card store, directory watcher, route registration |
 | `lib/store.js` | Card loading, normalization, save/delete |
 | `lib/yaml.js` | The built-in YAML subset reader and writer |
-| `lib/prompts.js` | Rewrite system prompt, draft framing, fact-check prompt |
+| `lib/prompts.js` | Rewrite system prompt, draft framing, AI-draft prompt |
 | `lib/rewrite.js` | The model calls and output normalization |
 | `lib/routes.js` | HTTP routes and session-model resolution |
 | `lib/loopback.js` | The loopback trust fence |