dsh-persona-forge 0.3.1 → 0.8.3

package/CHANGELOG.md

@@ -6,6 +6,654 @@ adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+## [0.8.3] - 2026-10-11
+
+### Fixed
+
+- **The review panel is no longer cut off at the top of a conversation.** It
+  renders in the composer stack, whose seat is `flex: none` — so a tall panel does
+  not compress the composer, it pushes the whole stack **up**. With no transcript
+  above to scroll, the overflow simply left the viewport: starting a session and
+  rewriting immediately showed only the panel's title bar, and during 改写中 that
+  was all you got.
+
+  The panel is now bounded and scrolls **inside** itself, with the header pinned
+  so 关闭 stays reachable. It can be any height without ever pushing itself
+  off-screen.
+
+### Changed
+
+- **The composer menu names what the card actually declares.** "跟随卡片" was a
+  black box: the rewrite-mode and fidelity rows showed the same words whether the
+  card declared `llm` or `template`, `style` or `strategy`, so what would actually
+  run was invisible. They now read **跟随卡片（模型改写）** and
+  **跟随卡片（可加约束）**, and the folded summary line lists fidelity too — it is
+  the setting with real consequences.
+
+- **The generated example set no longer repeats the generic rule.** The card
+  generator had been told, literally, to write `方块系统、合成表、生存模式` for
+  "写一个我的世界" — which is the *same* example the rewrite rule already carries,
+  so the set it produced was guaranteed to be redundant. The prompt now separates
+  the two explicitly: generic detail is already covered (do not write it),
+  character-specific demands are what this field is for. The bundled cards were
+  rewritten the same way — 欧姆弥赛亚 asks for a record left for those who come
+  after and refuses shortcuts; 硬邦邦 refuses a proof-of-concept handed over as a
+  deliverable.
+
+## [0.8.2] - 2026-10-10
+
+### Fixed
+
+- **The card form now shows the second example set and the template.** 0.8.0
+  added both to the card format but left the settings form on the old shape, and
+  `save` writes the payload as the **whole card** rather than merging — so the
+  moment you edited a card's name, its implementation-detail example set was
+  deleted. The form now renders a textarea per intensity level for both sets, and
+  a template field for every mode.
+
+- **A template is no longer lost when saving an llm card.** It was sent only for
+  a template-mode card, so an llm card that had one lost it on the next save —
+  and template mode is offered on every card, which is how that switch ended up
+  silently falling back to a model call.
+
+- **The form's copy matches what the fields now mean:** the second set is labelled
+  "may add implementation detail", and the fidelity hint says the same instead of
+  the retired "behavioural constraints" wording.
+
+### Added
+
+- **A check against this whole class of bug.** For every card field added since,
+  the form must both read it and write it back into the payload — checking only
+  that the name appears somewhere would pass on the read alone, which is exactly
+  how the write-back went missing. Confirmed red when the write-back is removed.
+
+## [0.8.1] - 2026-10-10
+
+### Changed
+
+- **"Constraints" now means what it should: concrete IMPLEMENTATION requirements.**
+  The strategy fidelity was read as permission to add behavioural framing ("do
+  not over-plan"), which is not what a constraint is. It now permits adding the
+  concrete features and limits a competent take on the request implies but the
+  user did not spell out — for "写一个我的世界": 方块系统、合成表、生存模式、存档.
+
+  The hard rules said the opposite: "never add, remove, widen or reinterpret a
+  requirement", while declaring themselves an override of everything. That
+  cancelled the permission and made the setting a no-op. They are now about
+  PROTECTING what the draft states — nothing stated may be removed, weakened or
+  contradicted — and explicitly leave ADDING to the fidelity block.
+
+  Both bundled cards' strategy sets now demonstrate real additions (边界用例、
+  三数取中、复现之证、入参校验) rather than tone advice.
+
+### Known limitation
+
+- **The local fact check cannot see an added requirement written in Chinese.**
+  compareFacts reports an addition only for a token it can prove is structural
+  — a path, number, flag or code span. 方块系统 and 生存模式 are bare words, so they
+  pass unreported, and the panel's "新增" line stays empty even when the rewrite
+  added them. The added detail is still visible in the editable result column,
+  which is where a reviewer would catch it.
+
+
+## [0.8.0] - 2026-10-10
+
+### Added
+
+- **Every card now carries TWO example sets**, and which one the model is shown
+  depends on the fidelity you chose. `examplesByIntensity` demonstrates the voice
+  alone; `examplesByIntensityStrategy` demonstrates the voice **plus the
+  behavioural framing** the rewrite may add ("do not over-plan", "do not stop at
+  the first plausible cause").
+
+  This fixes a real failure. Raising fidelity used to show the model the
+  *voice-only* examples — every one of which merely added a greeting — so it had
+  nothing to imitate and invented padding: "写一个我的世界" came back as
+  "就想要一个我的世界", a restatement rather than a framing.
+
+- **Every card now carries a `template`.** Template mode fell back to a model call
+  on every card, because none of them had one: the switch appeared on and the
+  model ran anyway. The template substitutes `{{input}}` verbatim, so the wrapper
+  supplies the mouth and the user supplies the request.
+
+- **The card generator produces both sets and a template**, and refuses a card
+  missing any of them, naming which. `DRAFT_MAX_TOKENS` goes to 8192, because a
+  complete card is now roughly twice the size of one example set.
+
+### Changed
+
+- **`fidelity` is no longer a coin flip on a card with nothing to add.** The
+  bundled cards' strategy sets demonstrate concrete framing, so the setting
+  produces visibly different output instead of a louder restatement.
+
+  A card without `examplesByIntensityStrategy` still falls back to the voice-only
+  set, so cards written before this release keep working; only the generator
+  requires the new fields.
+
+## [0.7.3] - 2026-10-10
+
+### Fixed
+
+- **"Constraints too" (fidelity: strategy) actually does something now.** 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 to add behavioural framing was
+  written *before* them. The model obeyed the absolute prohibition, so flipping
+  the control changed nothing in the output: a setting that appeared to work and
+  did nothing.
+
+  The root cause is that **"constraint" was doing two jobs**. A message carries
+  (a) technical content — what is asked, under which files, numbers and limits —
+  and (b) behavioural framing — how the assistant should approach the work. The
+  hard rules now lock (a) absolutely and say explicitly that they do **not**
+  forbid (b); the fidelity block owns (b) and names what it may add.
+
+- **Direct send closes its panel.** The message was sent correctly, but the panel
+  stayed up — and a settled panel sets `panelOpen`, which **swallows the
+  composer's send arrow**. The next thing you typed could not be sent until the
+  panel was dismissed by hand. Reported as "direct send works, but the review
+  card still shows up". Review mode still keeps its panel: that one *is* the
+  decision point.
+
+  A refused submit leaves the panel up, so the text is still reachable by hand.
+
+## [0.7.2] - 2026-10-10
+
+### Fixed
+
+- **AI card drafting no longer writes the character's own line as the rewrite.**
+  Given "the AI is my mother, I am her child", the generator produced examples
+  where the **mother** spoke:
+
+  > `from`: 帮我把标题改成首页
+  > `to`: 乖，妈帮你把标题改成首页。 ← the MOTHER talking
+
+  That is the character answering, not the user asking — the rewrite is always
+  the **user's** message, so the card was unusable and the role was inverted.
+
+  The prompt was the cause: it said the later step rewrites the user's message
+  "into that voice", and that `to` is the request "rewritten in the voice". Both
+  leave *whose* voice open, and the generator resolved it to the character's.
+
+  The generator is now told plainly that **the output is always the user's
+  message and the character's own speech never appears**, with the direction
+  restated for both sides and a **worked wrong/right pair** using exactly this
+  case. The bundled cards already obeyed the rule — in both, `to` is the user
+  speaking (the client, the believer), never the character — so they are cited as
+  the standard.
+
+## [0.7.1] - 2026-10-10
+
+### Fixed
+
+- **AI card drafting no longer runs out of output budget part-way through the
+  examples.** The draft call asked for **2048** output tokens, but a complete card
+  is bigger than that: each bundled card is ~1700 characters of CJK body, which
+  is well over 2000 tokens once the JSON wrapper and four intensity sets are
+  counted. `examplesByIntensity` is the longest field and the one written last, so
+  it was exactly what got cut. The budget is now 4096, matching the rewrite call.
+
+- **A generated card must carry all four intensity sets, or the generation is
+  refused.** The parser accepted whatever rungs arrived, so a truncated answer
+  produced a card that looked complete in the form while two of its four settings
+  had **nothing to demonstrate** — the voice quietly vanished at those levels.
+  It now reports which levels are missing and asks the user to retry.
+
+  This was the second half of the same bug: raising the budget makes truncation
+  rare, and refusing the partial card means it can no longer pass unnoticed when
+  it does happen.
+
+## [0.7.0] - 2026-10-10
+
+### Fixed
+
+- **AI card drafting no longer fails when the model appends a sentence.** The
+  recovery step that slices the outermost `{…}` out of the answer ran **only when
+  the text did not start with `{`**. A model that emitted clean JSON and then one
+  line of commentary — a very common habit — therefore skipped recovery and was
+  parsed whole, failing with:
+
+  > the model returned invalid JSON: Unexpected non-whitespace character after
+  > JSON at position 1051
+
+  Recovery now runs whenever the raw text does not parse on its own, so a leading
+  sentence, a trailing sentence, a code fence, or any combination of them all
+  survive. Text with no object in it is still rejected, and a legitimate object is
+  never rewritten on the way in.
+
+### Removed
+
+- **The flat `examples:` list, from the form and from the bundled cards.** It was
+  the fallback for a card whose per-intensity sets were missing, but
+  `examplesFor` prefers `examplesByIntensity` — so on the bundled cards, which
+  populate all four rungs, the flat list of four was **dead data** that still had
+  to be read and kept in sync. The two bundled cards carried 4 flat examples plus
+  8 per-intensity ones, and only the 8 ever ran.
+
+  A card file that still carries `examples:` is **still read** and still falls
+  back to it; the form simply no longer offers a second place to write the same
+  thing. Saving a card preserves any flat list it already had, because `save`
+  writes the payload as the whole card rather than merging — without that, editing
+  a legacy card's name would silently delete its examples.
+
+## [0.6.4] - 2026-10-10
+
+### Changed
+
+- **The AI-draft box stopped saying the same thing three times.** The direction
+  rule was carried by the placeholder, restated in the description ("who the AI
+  plays, who you are, how they talk"), and then explained again in a paragraph
+  above the button. Three tellings of one instruction is not emphasis, it is
+  noise.
+
+  The paragraph is gone. The **placeholder** keeps the rule — it is the strongest
+  teaching signal in the box, and it demonstrates the direction in the shape of a
+  real example rather than describing it. The description is back to what the
+  form does: which card fields the AI fills.
+
+  The prompt layer still enforces the direction, and the card-authoring section
+  of the README still explains both sides in full; that is where an author
+  writing a card by hand needs it.
+
+### Added
+
+- **A check against restating the rule.** The direction must not reappear as a
+  separate paragraph, and the description must not repeat the placeholder. The
+  regression was confirmed to fail the suite.
+
+## [0.6.3] - 2026-10-10
+
+### Fixed
+
+- **The character may be on EITHER side of the conversation, and the prompts no
+  longer pretend otherwise.** Both the rewrite prompt and the card generator
+  asserted flatly that *the character is the ASSISTANT's identity*. That is false
+  for half the cards, and the bundled set proves it:
+
+  | Card | Who is the character |
+  |---|---|
+  | 万机之神 · 欧姆弥赛亚 | the **AI** — the user addresses it as a petitioner |
+  | 肌肉集团 · 硬邦邦 | the **user** — "**the user is** the client", addressing their own team |
+
+  The rewrite prompt now states both cases, tells the model to take the direction
+  from the card's style text, and defines the assistant-side default for a card
+  that is genuinely silent. The generator is told the same, and takes the
+  direction from the idea it is handed.
+
+- **The AI-draft direction note no longer states the wrong rule.** It said "你是
+  对角色说话的那个人，**不是角色本身**" — an absolute that contradicts 硬邦邦,
+  where you *are* the character. It now says the character can be on either side
+  and that picking one and saying so is all that is required.
+
+  The brief example and description keep asking for the same three things (who
+  the AI plays, who you are, how they talk) — the ambiguity was the bug, not the
+  two-sidedness.
+
+## [0.6.2] - 2026-10-10
+
+### Fixed
+
+- **The AI-draft example now states the speaker direction.** The placeholder was
+  "一个脾气暴躁但极其可靠的老灯塔守夜人，说话简短、爱用航海比喻" — a character
+  and nothing else. It never said whether **you** play that character or the
+  **AI** does, which is precisely the ambiguity this plugin exists to avoid: the
+  character is the *assistant's* identity, and the message being rewritten is
+  always the *user's*.
+
+  The mismatch was sharpest because the generator's own system prompt **requires**
+  the style field to state the direction explicitly, while the example it was
+  handed omitted it. The example is the strongest teaching signal in that box, so
+  it now demonstrates what it demands:
+
+  > 例如：AI 扮演一位脾气暴躁但极其可靠的老灯塔守夜人，我是他的学徒；他说话简短、爱用航海比喻
+
+  The description asks for the same three things (who the AI plays, who you are,
+  how they talk), and the box carries a standing line stating the rule. Both
+  bundled cards already named it — 欧姆弥赛亚 is "a supplicant's prayer to the
+  Machine God", 硬邦邦 is "**the user** is the client".
+
+- **The AI-draft copy said who does what, and what "the spec" actually is.**
+  It read "用一句话描述角色的思路，AI 会按同一套规范把表单填好" — with no
+  subject, so it was unclear whether *you* describe the character or the AI
+  does, and "同一套规范" pointed at an internal quality standard the user cannot
+  see or act on.
+
+  Both languages now name the actor and the fields: **you** write one sentence
+  about the character, and the AI fills the form by the character-card fields —
+  name, icon, description, style text, and a conversion example set per
+  intensity level. The hint also states the two things a generated card
+  deliberately pins: **fidelity is always voice-only** and **rewrite mode is
+  always model rewrite**, so a generated card never quietly acquires the power
+  to add constraints to your request.
+
+### Added
+
+- **Checks pin the direction into the brief.** The placeholder must demonstrate
+  who plays whom and must name the user's own role; the description must ask for
+  it; the rule must be stated in the box; and the copy must carry no markdown,
+  since the shell renders it as plain text. Confirmed red against the old
+  placeholder.
+
+- **Checks pin that copy to the behaviour.** The description must name the four
+  fields the generator fills and must not fall back to "规范", and the generated
+  card must stay pinned to `fidelity: 'style'` / `mode: 'llm'` with the copy
+  saying so. Confirmed red against the old wording.
+
+## [0.6.0] - 2026-10-10
+
+### Changed
+
+- **Fidelity is now free in both directions, on every card.** It used to be
+  one-way — a `strategy` card could be narrowed to `style`, but a `style` card
+  could never be widened — and the control was rendered only on `strategy` cards.
+  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 the user's own file**, and the
+  settings page has always let them set `fidelity:` directly. Refusing in the
+  menu did not remove the capability, it only made the choice disappear from the
+  place where the work happens. The server rule is now symmetric, and all three
+  options — card default, voice only, constraints too — appear on every card.
+
+  The honest framing is kept: "constraints too" genuinely changes the request, so
+  the tooltip says so and points at review mode.
+
+- **The settings fold no longer explains itself in permanent prose.** Every row
+  and every option now carries its explanation as a **tooltip** on the control.
+  The fold had grown to four rows plus six lines of standing explanation, which
+  is a wall of text for a menu whose job is picking a character.
+
+- **Rewrite mode and fidelity now follow the character too**, not the session.
+  They were already displayed under the character while being stored per session,
+  which meant the setting you saw and the setting you got could disagree after
+  switching character. All four settings are now read from, and written to, the
+  card's own keys.
+
+### Added
+
+- **The check suite guards the READMEs.** The status line and the inspiration
+  link had both already rotted unnoticed, so three checks per README now pin the
+  current minor version, the video link and its author, and the named target
+  model.
+
+## [0.5.2] - 2026-10-10
+
+### Changed
+
+- **The README's "what a persona changes" section now says where this came
+  from.** The old text was a flat disclaimer — personas do not raise the
+  capability ceiling, do not expect a performance switch. That is still true and
+  still stated, but it read like a warning label on a project whose whole point
+  is delight.
+
+  It now credits the source: [**“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 — a 30-second video whose comment section describes the
+  effect better than any spec sheet ("another AI is you pretending to be Guan
+  Yu; when Big Fat Fish pretends, she really rides out and cuts down Hua
+  Xiong"). It also names **DeepSeek Flash** as the model this is tuned for,
+  which was never said anywhere.
+
+  Both READMEs also carried a stale `0.3.x` status line; they now state `0.5.x`.
+
+### Added
+
+- **The check suite now guards the READMEs.** The status line and the
+  inspiration link had both already rotted unnoticed, so three checks per README
+  now pin the current minor version, the video link and its author, and the
+  named target model.
+
+## [0.5.1] - 2026-10-10
+
+### Changed
+
+- **"More settings" is gone; everything it held moved into the character's own
+  fold.** Having two folds split one idea across two places — intensity and send
+  mode under the character, rewrite mode and fidelity somewhere else — so
+  **rewrite mode** and **fidelity** now sit in the same **"Settings for
+  <name>"** fold as the rest. The folded row's summary names the intensity, the
+  send mode and the rewrite mode, so the whole configuration is legible without
+  opening it.
+
+- **Every rewrite mode now says what it does.** The segmented control named
+  three options and explained none of them, so the difference between a model
+  rewrite and a template was something you had to guess:
+
+  - *Card default* — use whatever this card declares on the settings page.
+  - *Model rewrite* — calls the model to rewrite your draft in the character's
+    voice. Most flexible; costs one model call.
+  - *Template* — fills your draft into the card's preset template with no model
+    call. No latency, no tokens, but the voice is fixed.
+
+### Removed
+
+- **"Open card directory" left the rewrite menu.** The settings page already has
+  an *Open directory* button, so the menu row was a second door to the same room
+  — and it put occasional housekeeping in the path of choosing a character.
+
+## [0.5.0] - 2026-10-10
+
+### Changed
+
+- **Intensity and send mode now belong to the CHARACTER, not to the session.**
+  "硬邦邦 always goes out Zealot, un-reviewed" is a property of that character
+  rather than a choice about one conversation, so both are stored per card:
+  configure it once and picking that card in another session gives you the same
+  settings.
+
+  They also moved out of the shared block above the list and under the card they
+  apply to, as a folded **"Settings for <name>"** row. Sitting above every card,
+  they read as global switches for the whole menu — and switching character
+  silently carried the previous card's intensity over. The fold's summary line
+  shows the current rung and send mode, so the state is visible without opening
+  it.
+
+  Switching character now **re-reads that card's settings**; without that the
+  previous card's values stayed on screen and were then written to the new
+  card's key.
+
+- **Picking a character no longer closes the menu.** Choosing a character and
+  configuring it are one act, and closing on pick forced a reopen to touch
+  intensity at all.
+
+- **The card directory moved into "More settings".** It is housekeeping you do
+  occasionally, not something to scroll past on the way to choosing a character.
+
+## [0.4.1] - 2026-10-10
+
+### Removed
+
+- **The model-based preservation check.** It made a second model call on every
+  rewrite to catch what the free token comparison cannot see — a requirement
+  reworded weaker, a constraint added in prose. It is gone, along with its
+  prompt, its verdict parser and every wire field (`factCheck`, `modelCheck`).
+
+  It doubled the latency and the token cost of every send to re-check what the
+  local comparison answers for nothing, and the settings page never offered a
+  working switch for it — a documented feature with no reachable control. For
+  the residue it was meant to cover, the rewrite column is now editable, which
+  is more direct than another round trip.
+
+  The retired `factCheck` config key is **ignored rather than rejected**, so a
+  profile still carrying it loads unchanged and simply makes no second call.
+
+### Fixed
+
+- **Rewriting in one session no longer makes every other session look busy.**
+  The composer button read the panel's phase **globally**, so while any single
+  session was rewriting, every open composer rendered "改写中…". The panel is one
+  shared surface, so its phase is global — but what the button shows must be
+  scoped to its own session. It now reads the session-filtered panel, the same
+  value the send guards already used.
+
+## [0.4.0] - 2026-10-10
+
+### Added
+
+- **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. **Send** and **Fill** carry the
+  edited text, and the panel says when the text differs from what the model
+  produced, so nothing goes out silently different.
+
+- **While the review panel is open, the composer's send arrow is inert.** The
+  panel is the decision point. Pressing send with it up either pushed the
+  un-reviewed draft out or started a **second** rewrite of the same draft —
+  the duplicate that appeared once the panel was on screen. Closing the panel,
+  or choosing an action in it, re-arms the composer. Direct-send mode is
+  unaffected: it submits programmatically rather than through a gesture.
+
+### Removed
+
+- **The "Fill only" button, from the normal flow.** With an editable result,
+  writing the text into the composer without sending is the same act as
+  reviewing it, so the button was a second control for one idea. It survives
+  only for a shell that exposes no submit action, where it is the sole way to
+  get the text out.
+
+### Fixed
+
+- **An edited result is no longer rewritten again.** `apply` recorded only the
+  untouched original as "already rewritten", so sending a hand-edited result
+  from the composer came back as a fresh rewrite of the user's own edit. Any
+  settled text is now recorded. *Restore original* still deletes the memory
+  outright, because that is a rewind rather than a settled result.
+
+## [0.3.5] - 2026-10-10
+
+### Fixed
+
+- **The same sentence typed twice is rewritten both times.** The send guard
+  skipped a rewrite when the draft matched the last rewrite's result **or** its
+  original text. Restoring the original already clears that memory, so matching
+  `original` bought nothing and silently disabled the persona for any repeated
+  draft — type "讲个笑话", send it, type it again, and it went out
+  un-rewritten.
+
+- **Interception can no longer disarm itself.** The send listeners were bound
+  to the composer card, found through the plugin's own button. A card that was
+  not found meant the whole feature silently did not arm — the plugin looked
+  dead, with no error anywhere. The listeners now sit on `window` and each
+  guard scopes its event to the composer card it belongs to, so a miss costs
+  one event rather than the feature.
+
+### Changed
+
+- **The send-gesture decision is now a pure function** (`decideSend`) returning
+  `{ action, reason }`, and the client exposes `__internals` so the check suite
+  can boot the real bundle and drive it. This decision fails **silently** in
+  both directions — swallow a gesture that should have been sent and the
+  message vanishes; let one through that should have been rewritten and the
+  persona quietly does nothing — and neither throws, so the only symptom is "it
+  stopped working". Thirteen behavioural cases now cover it, replacing
+  source-shape assertions. The repeated-draft case was confirmed to fail
+  against the old predicate.
+
+  The pure-guard-plus-seam approach follows
+  [`dsh-prompt-hardener`](https://github.com/ffyfox/dsh-prompt-hardener), which
+  reached the same conclusion about this class of silent failure.
+
+## [0.3.4] - 2026-10-10
+
+### Fixed
+
+- **One wedged rewrite no longer freezes the plugin for good.** The review
+  panel is a single shared surface, but the guards around it were global while
+  the send interceptor guarded per session. A single rewrite left stuck at
+  `loading` therefore poisoned everything after it: `start()` hit its global
+  `busy` guard and **silently did nothing** — no rewrite, no error — while in
+  another session the interceptor let the gesture through to the native send.
+  Reported as "after it hangs, it just sends my text and the plugin never
+  intercepts again".
+
+  - `runRewrite` now refuses only a concurrent run for the **same** session, so
+    one wedged session cannot block another.
+  - `start()` guards on the **per-session** busy, matching `runRewrite` and the
+    interceptor. A send gesture is now either rewritten or sent; it is never
+    swallowed into nothing.
+  - A client-side watchdog settles the panel after 135s even if the host never
+    answers. The host's own deadline is not enough: this half must stay usable
+    without depending on the other half being alive.
+
+## [0.3.3] - 2026-10-09
+
+### Fixed
+
+- **A rewrite now runs on the model you picked, even before you send with
+  it.** Switching models in the composer and then rewriting used the
+  **previous** model: the route came only from the session's last request
+  header, which does not advance until a message is actually sent. Observed
+  live as the composer showing `hy4-preview` while the rewrite ran on a stale
+  `workbuddy-xdpool-global / deepseek-v4.1-flash` route.
+
+  The host's own precedence, in `api-session`'s `selectionFor`, is
+  **pending → request header → harness default**; the plugin implemented only
+  the last two. It now reads the `modelSelection` projection's `pending` first,
+  so a dropdown pick takes effect immediately. Every step is still optional and
+  falls through to the next, so a harness without that projection behaves
+  exactly as before.
+
+- **A rewrite whose model never answers now fails instead of hanging
+  forever.** `for await (const chunk of llm.stream(...))` only exits when the
+  adapter stops yielding. A provider that accepts the request and then sends
+  nothing — a rate-limited model — yields no chunk, so the 120s deadline's
+  `abort()` was invisible to the loop and the call never returned: the panel
+  spun with no result, no error, and no timeout.
+
+  Each pull is now raced against the deadline, so the loop leaves on its own
+  and reports `TIMEOUT` instead of waiting on a stream that will never speak.
+  The stream is abandoned rather than drained, because draining a wedged
+  provider is the hang this guards against.
+
+  Two regression tests cover both, and both were confirmed to fail against the
+  old code: the route test goes red, and the deadline test hangs the suite
+  outright (`Detected unsettled top-level await`) — the reported symptom.
+
+## [0.3.2] - 2026-10-09
+
+### Changed
+
+- **The rewrite stops naming a reasoning effort, so it works again on routes
+  that declare none.** An earlier attempt to speed slow rewrites up pinned
+  `reasoningEffort: 'low'`; on `workbuddy-xdpool` / `hy4-preview` every rewrite
+  then failed outright with
+  `does not support reasoning effort "low"`.
+
+  The harness validates the field before it reaches any provider: when a
+  route's model declares **no** reasoning metadata, **any present value**
+  throws — including `''`, which is the host's own "Provider default" sentinel
+  in the ACP session-config layer, and including `'off'`. Absence is therefore
+  the only portable way to ask for the route's default, and the call now leaves
+  the field out.
+
+  The latency this was meant to fix is **not** fixed by it. On a reasoning
+  model the rewrite may still spend thinking time before it emits anything; the
+  120s deadline and the elapsed-seconds readout from 0.3.1 remain the honest
+  answer there.
+
+## [0.3.1] - 2026-10-09
+
+### Fixed
+
+- **A rewrite waiting on a slow model no longer reads as frozen.** The panel
+  showed 改写中 with no sense of time, so a slow rewrite was indistinguishable
+  from a hang. The host has had a 120s deadline all along, so the panel was
+  never stuck forever — but two minutes of a silent spinner looks exactly like a
+  freeze. It now ticks elapsed seconds, and past 45s says the model may just be
+  slow and how to give up.
+
+### Removed
+
+- **The settings page's model-check checkbox.** It duplicated the plugin row's
+  own `factCheck` config — a second setting for the same thing, sitting on a
+  page that is about cards. The host config is the single control and already
+  defaults to off. The checkbox, its dictionary keys and the client-side
+  preference behind it are gone.
+- **The permanent drift-check instruction line.** The button's tooltip already
+  carried the same sentence. The page now renders only the check RESULT, which
+  is real feedback rather than the same copy twice.
+- Three dictionary keys no route or component referenced:
+  `settings.ai.briefLabel`, `settings.changed`, `error.busy`.
+
 ## [0.3.0] - 2026-10-09
 
 ### Added
@@ -325,7 +973,29 @@ First public release.
   tools, and never enters the session log. What the session records is the text
   the user actually sent.
 
-[Unreleased]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.3.0...HEAD
+[Unreleased]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.8.3...HEAD
+[0.8.3]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.8.2...v0.8.3
+[0.8.2]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.8.1...v0.8.2
+[0.8.1]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.8.0...v0.8.1
+[0.8.0]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.7.3...v0.8.0
+[0.7.3]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.7.2...v0.7.3
+[0.7.2]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.7.1...v0.7.2
+[0.7.1]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.7.0...v0.7.1
+[0.7.0]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.6.4...v0.7.0
+[0.6.4]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.6.3...v0.6.4
+[0.6.3]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.6.2...v0.6.3
+[0.6.2]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.6.0...v0.6.2
+[0.6.0]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.5.2...v0.6.0
+[0.5.2]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.5.1...v0.5.2
+[0.5.1]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.5.0...v0.5.1
+[0.5.0]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.4.1...v0.5.0
+[0.4.1]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.4.0...v0.4.1
+[0.4.0]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.3.5...v0.4.0
+[0.3.5]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.3.4...v0.3.5
+[0.3.4]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.3.3...v0.3.4
+[0.3.3]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.3.2...v0.3.3
+[0.3.2]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.3.1...v0.3.2
+[0.3.1]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.3.0...v0.3.1
 [0.3.0]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.2.1...v0.3.0
 [0.2.1]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.2.0...v0.2.1
 [0.2.0]: https://github.com/sugarmaster666/dsh-persona-forge/compare/v0.1.2...v0.2.0