@operatingtag/dsh-unclesam 0.0.0-stage → 1.0.1

package/CHANGELOG.md

@@ -0,0 +1,928 @@
+# Changelog
+
+All notable changes to Uncle Sam, newest first.
+
+Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Versions use
+[semantic versioning](https://semver.org/): breaking behaviour changes bump the major, features the minor,
+fixes the patch.
+
+---
+
+## 1.0.1
+
+**The tool line no longer tells the website what is installed on your machine.**
+
+MEASURED, on screen, by the person using it: every chat turn was typing this into the box, above their own
+sentence -
+
+    Tools: ask_user_question, cordis_inspect_list, cordis_inspect_query, create_goal,
+    demergo_brave, demergo_busy, demergo_click, demergo_clipboard, demergo_cursor,
+    demergo_drag, demergo_elements, demergo_glide, +282 more
+
+That list was real. It was the harness's tool inventory for the session - 283 tools, including a separate
+screen-control plugin installed on the same machine - and because a web chat has no system channel, every
+word of it was typed into `chat.deepseek.com` and sent to a third party. An inventory of a private computer
+leaving the computer, on every turn, for no benefit.
+
+- The preamble now states a **count** and the call format, and names nothing:
+
+      [283 tools available · to call one reply only with {"tool":"name","arguments":{}}]
+
+- **Proved by test, not by intention.** `dev/test-preamble.mjs` asserts that no tool name reaches the prompt
+  on either a greeting or a real task, using real names from this machine (`demergo_`, `cordis_`,
+  `ask_user_question`) plus a synthetic set. A name leaking back in fails the suite.
+- **The call format survives verbatim**, because without it the model answers in prose instead of emitting a
+  call. Live tool use re-tested after the change - a call still happens.
+- Things that had to change with it: the panel's "last thing sent" view recognised the old `Tools: ` label,
+  and `test-answer-language.mjs` asserted it. Both now recognise the new wording. The suite caught the second
+  one; it was not spotted by reading.
+
+Honest loss: a model can no longer browse the catalogue and pick a tool nobody told it about. On a turn where
+the harness attaches tools, their names must still be sent - a model cannot call a tool it has never heard of.
+This only affects the turn where no particular tool is being offered.
+
+---
+
+## 3.9.17
+
+**A page that is thinking is not a page that has died.**
+
+MEASURED, by the person reading Brave: a request for about a thousand words was answered, and the plugin
+abandoned it after five silent seconds with "no text arrived". The page had not been silent - it was writing
+its reasoning into a Thinking block, which is a different element from the answer the reader watches. From
+inside the plugin that looked exactly like a dead page. Three earlier theories blamed the website; the
+screenshot showed the answer.
+
+- The reader now also samples the page's total text length. The answer element still decides what is emitted,
+  but any growth anywhere in the document counts as a sign of life and resets the pause clock. This is
+  deliberately blunt: it needs no knowledge of the reasoning element's class name, which the site can rename
+  at will.
+- The change is strictly permissive. The sign of life can only ever delay a stall from being declared; it can
+  never cause one. Nothing about what reaches the transcript changed - the reasoning is still excluded, and the
+  live test confirms no reasoning text appears in the answer.
+
+New: `dev/test-paced-think.mjs` - one case on `deepseek-reasoner`, the model whose state is DeepThink on, so
+the page thinks before it answers. Two runs, both correct, both clean of reasoning text.
+
+Honest limitation: the exact thousand-word case was not re-run, because a reply that long cannot finish inside
+the one-minute rule. The mechanism is covered by the two shorter thinking runs and by the property above.
+
+Tests: 14 suites green, npm run check clean.
+
+---
+
+## 3.9.16
+
+**The savings figure can now show a small number.**
+
+It printed two decimals, so anything under a cent read `$0.00` - which is exactly what a first test produces.
+One turn of ordinary chat costs about a hundredth of a cent at the API's list rate, so a person could use this
+for an evening and watch a counter that never left zero. A number that cannot move is worse than no number: it
+reads as broken rather than as small.
+
+    one small turn   $0.0001  ->  0.01c
+    a few turns      $0.0042  ->  0.42c
+    one cent         $0.01    ->  $0.01
+    a session        $0.42    ->  $0.42
+    too small        $0.000001 -> <0.01c
+    zero             $0.00    ->  0.00c
+
+Whole cents and above keep the familiar `$0.42`. Anything smaller is stated in cents, and anything smaller
+still says `<0.01c` rather than pretending to be zero. Currencies other than USD have no cents convention
+here, so they keep four decimals. Zero reports as zero rather than as a small positive, which can happen
+legitimately when the configured rates are zero.
+
+Tests: 14 suites green, npm run check clean.
+
+---
+
+## 3.9.15
+
+**The coffee link moved out of the banner and into the Brave recommendation.**
+
+It was pinned above the title, where it read as an advertisement and pushed the actual settings down the page.
+It now sits immediately above the Brave recommendation, so the ask belongs to the same thought as the text
+around it - this page is about which browser to use and why the work is worth supporting.
+
+Still two links, still one destination: this one and the support block at the foot of the panel. A test
+asserts the placement rather than merely its existence, so it cannot drift back to being a banner.
+
+---
+
+## 3.9.14
+
+**Search tested against the real page for the first time - and the two faults that exposed.**
+
+The search branch was the last surface with no live coverage. One paced case, model `deepseek-chat-search`:
+
+    🔴 Uncle Sam Web search - "what is the latest stable version of Node.js?"
+    Sources (2): nodejs.org · ...
+    The latest stable version of Node.js is Node.js 26.13.0, codenamed 'Lithium'.
+
+Two blocks, the person's own words as the query, real sources, the label, no tool envelope, 9.6 s.
+
+- **The five-second pause limit was aborting every search.** MEASURED: the first run died at 5.1 s with
+  nothing on screen. A web search fetches before it writes, so silence there is the page working. A search
+  call now gets thirty seconds of silence before it is treated as dead; an ordinary message keeps the five
+  second rule unchanged. Both are overridable (`DSH_STALL_MS`, `DSH_SEARCH_STALL_MS`).
+- **Search was left switched on after a call.** The plugin only ever turned the site's toggles on, never off,
+  so the next ordinary message would search the web without anybody asking - slower, and a silent change of
+  behaviour. Every switch the plugin turns on is now restored at the end of the call, and again at the start
+  of the next one if the call threw. A switch the person turned on themselves was never in that list, so it is
+  left exactly as they set it.
+
+New: `dev/test-paced-search.mjs` - one case, sixty-second run limit, browser killed at the end, written
+report.
+
+NOT YET PUBLISHED to npm.
+
+---
+
+## 3.9.13
+
+**A tool call on its own carried no signature, and the tool path is now tested against the real page.**
+
+Two things, both found by running the surface that had never been run.
+
+The tool-call path had no live coverage until now. Three cases that need a tool produced correct calls -
+`read`, `glob`, `pwsh`, each with the right arguments - but every one of them came back with **no
+`🔴 Uncle Sam` label**, because a call-only reply has no prose and the label was only written when there was
+prose to write it beside. That breaks the one rule the label exists for. The label now goes out whenever a
+token is a web one, with or without prose. A call is not an empty reply: something happened, so naming the
+route above it is information rather than noise.
+
+The pro se-beside-a-call shape - the one 3.9.8 fixed - was then exercised deliberately, by asking for a
+sentence before the call. All three replies kept their prose intact to the final full stop:
+
+    "I'll read package.json now. "          + read({"path":"package.json"})
+    "I'll list the files now. "             + glob({"pattern":"*"})
+    "I'll run a command to print "hello". " + pwsh({"command":"Write-Output 'hello'"})
+
+Before 3.9.8 those sentences lost their last characters to an offset error.
+
+New: `dev/test-paced-tools.mjs` - four cases with real tool schemas attached, five seconds apart, a
+five-second stall limit, a sixty-second run limit, the browser killed at the end, and a written report.
+
+NOT YET PUBLISHED to npm.
+
+---
+
+## 3.9.12
+
+**Back to MIT.**
+
+The whole point of this plugin is that it costs nothing and asks nothing. A licence that turns a free tool into
+a negotiation works against that, so the stricter licence is withdrawn and the coffee request takes its place.
+
+- LICENSE is MIT, and it is a long one on purpose: the grant and the disclaimer first as the law needs them,
+  then a plain-English summary of the same thing, then an explicit list of what the licence does NOT ask for
+  (no royalties, no copyleft, no attribution in your UI, no patent grab, no revocation).
+- **The coffee ask is in the licence file, and it is labelled as a request, not a condition.** MIT obliges
+  nobody to support the project, so the ask is set apart, says it carries no legal weight, and says that
+  ignoring it changes nothing. A test asserts both the link and that wording.
+- The README leads with the one-line deal and keeps the same ask.
+- All 204 source headers now declare MIT.
+- **3.9.7 to 3.9.11 were published under the Business Source License 1.1 and those grants remain in force.**
+  The LICENSE and README both say so rather than implying the change reaches backwards.
+- `dev/relicence.mjs` is deleted: it existed only to perform the opposite change, and leaving it would leave a
+  script whose job is to undo this decision.
+
+Tests: 14 suites green, npm run check clean.
+
+NOT YET PUBLISHED.
+
+---
+
+## 3.9.11
+
+**The fourth reply in a chat was arriving and being thrown away.**
+
+MEASURED, by the person watching the browser: the site answered the fourth message - they read the answer on
+screen - while the plugin reported no text at all and abandoned the call as a stall. Three runs in a row
+stalled on the fourth message and only the fourth, with different questions each time. A page that answered
+cannot be a page that stalled, so the fault was in the reader.
+
+Cause: "is this a new reply?" was decided by comparing the NUMBER of assistant blocks with the count taken
+before sending. The site adds a `Thinking` block as the answer begins, and recycles earlier blocks, so the
+count stops being a reliable signal after a few turns. The reader then either saw the reasoning block instead
+of the answer, or concluded nothing new had appeared and returned empty.
+
+- The baseline now records the shape of the last reply as well as how many there are, and a reply counts as
+  new when a block has been added OR the last block's text has changed. Length and tail are compared rather
+  than the whole text, so the injected expression stays small on a 70 ms poll.
+- `readReply` reports whether it matched by growth or by replacement, which is what made this visible.
+- The blank-chat check used by the fresh-chat path keeps a plain count, where it is correct.
+
+Verified live, paced at five seconds, with the window on screen and watched while it ran:
+
+    1/4  4417ms  "Hello! How can I help you today?"
+    2/4  1606ms  "The capital of Australia is Canberra."
+    3/4  2654ms  "The three traditional primary colours…"
+    4/4  2133ms  "Goodbye, and take care!"       <- the case that failed three runs running
+
+Four of four, no stalls, browser killed, report written.
+
+NOT YET PUBLISHED.
+
+---
+
+## 3.9.10
+
+**Debugging, and a five-second pause limit.**
+
+Driven by three runs that ended in a fault with nothing but terminal scrollback to explain them.
+
+- **A pause longer than five seconds now abandons the call.** Measured from the last sign of life - a
+  character of the answer, or the page becoming ready - not from the start of the call, because launching a
+  browser is legitimately slow and must not look like a hang. The error carries code `STALLED` so a test can
+  tell it apart from a slow answer and stop cleanly. `DSH_STALL_MS` tunes it, floored at one second.
+- **A stalled case stops the run** instead of moving to the next one, and the browser is killed at once.
+- **`dev/report.mjs`** writes every case to `dev/reports/<run>-<stamp>.jsonl` and `.txt` AS it finishes,
+  so a run killed by the one-minute rule still explains what happened up to that point. Each row carries the
+  message, the reply, the elapsed time, the chunk count and the failure code.
+- The paced English test now uses the reporter and the stall limit, and exits non-zero when anything failed.
+
+NOT YET PUBLISHED.
+
+---
+
+## 3.9.9
+
+**An empty assistant block could hold a call for ten minutes.**
+
+MEASURED. A test call hung for over eight minutes. The page had created the assistant block, so the poll's
+"is there a new block" test was true, but it never wrote a character into it. The no-reply guard lived INSIDE
+the "no new block" branch, so that shape skipped the guard entirely and the loop ran to the ten-minute cap.
+Anyone watching saw a spinner with no explanation.
+
+- The guard is now evaluated on "has any text arrived", independently of whether a block exists.
+- The stall counter counts an empty new block as no progress, so the page-active nudge can fire for it too.
+- The error names the timeout that actually applied instead of a hard-coded 90 seconds.
+- `DSH_FIRST_TOKEN_TIMEOUT_MS` overrides the grace period, so a test fails in twenty seconds instead of
+  ninety. A value below one second or a non-number is ignored, so a typo cannot disable the guard.
+
+**New operating rule, from the person: no test may run longer than 60 seconds**, and the browsers it opened
+must be killed when it stops. `dev/guard.mjs` enforces it - a 60-second deadline, browser cleanup on expiry,
+on normal exit, on Ctrl-C and on a crash. Proven: a 3-second deadline in a 10-second script exited 124 having
+killed the browser.
+
+Also added: `dev/test-no-reply-guard.mjs` asserts the guard cannot be re-nested inside the branch that
+bypassed it, and the paced English test that reproduced the stall.
+
+NOT YET PUBLISHED.
+
+---
+
+## 3.9.8
+
+**Words were being deleted from answers that contained a tool call.**
+
+The tool branch worked out how much prose had not yet streamed by taking the scrubbed whole reply and slicing
+it at an offset measured against the unscrubbed one. Scrubbing trims leading whitespace, so every newline the
+model put before its sentence shifted the offset and the remainder started mid-word:
+
+    "\n\nHi. {…} more text here"    came out as  "Hi. ore text here"
+    "  Hello there. {…} and then"    came out as  "Hello there. nd then"
+    "\n\n{…} trailing words"       came out as  "ailing words"
+
+Models prefix prose with a newline constantly, so this was not rare - it was every tool-calling turn whose
+reply was not flush left. The held slice is now scrubbed on its own, so the offset and the text always belong
+to the same string. Four shapes are asserted in dev/test-tool-tail.mjs.
+
+Also: a browser launch now records the call stack that asked for it, to browser-launches.log and to stderr
+when DEBUG_WEB_CHAT is set. Four times tonight a browser appeared with the route on API and every explanation
+was a theory; the next one will name its caller.
+
+NOT YET PUBLISHED.
+
+---
+
+## 3.9.7
+
+**Licence change: MIT to Business Source License 1.1.**
+
+Free for individuals and for any organisation under GBP 250,000 annual revenue. Above that a commercial
+licence is required, arranged through the issue tracker. Four years after each release it converts to the
+Apache License 2.0 and becomes ordinary open source for everyone.
+
+- LICENSE rewritten with the parameters; package.json declares BUSL-1.1.
+- Every source header now declares BUSL-1.1. A new test fails the suite if any file still names the old
+  licence, so the LICENSE file and the code cannot drift apart.
+- README explains the terms in plain English before the legal text.
+- **Versions 3.9.6 and earlier remain MIT.** That grant is irrevocable and is not being withdrawn; this
+  licence starts at 3.9.7.
+
+NOT YET PUBLISHED to npm.
+
+---
+
+## 3.9.6
+
+**The ask is at the top, and the story now matches the work.**
+
+- "BUY ME A COFFEE" is a link at the head of the panel as well as in the support block at the bottom. Same
+  destination, same words, drawn twice - a person who opens the page should not have to scroll to the end
+  to find out how to help.
+- The support copy was rewritten in all 51 languages. It began as a provider that spoke English; it now
+  answers in fifty-one languages, and the text says so. The ask is stated as what the work actually takes:
+  an evening to add a language, several more to test it properly, and the part nobody sees.
+- The copy implies the cost and never states it. No dollar amount, no day count - a test asserts that, so the
+  rule cannot drift when the text is edited later.
+
+NOT YET PUBLISHED. Installed into the local harness for testing only.
+
+---
+
+## 3.9.5
+
+**Harness notices are no longer mistaken for the person's message.**
+
+MEASURED, not guessed: in one real transcript, 1112 records carried `role: 'user'` but only 770 were typed
+by the person. The other 342 were harness machinery filed under the same role - time context (163), runtime
+context (58), model-selection notices (42), tool-job notices (41), the skill catalog (10), compaction
+checkpoints (8), goal notes (7) and more.
+
+- The recorded failure: the person switched from Web Chat to Web Reasoner+Search mid-conversation, and the
+  search query became "[model changed: assistant turns above this point were generated by …]". The search
+  found nothing, so the answer was written from memory while wearing a "Web search" heading.
+- Only a message whose `source.kind` is `user` counts as the person speaking, in both the prompt builder
+  and the search-query picker. Messages with no source at all are still kept, so the API path is unaffected.
+- A greeting no longer gets the tool list. "hola mi gran amigo" prompted "What would you like me to do? I
+  have a broad toolset available…" because the preamble invited it. A greeting carries no task, so no tool
+  can be the answer; a real request still gets the full tool line.
+- When search is on and the page reports no sources, the reply now says the answer came from the model's
+  own memory rather than presenting it as researched.
+
+Two of the plugin's own tests were asserting the old behaviour - a greeting with tools, and a tool line
+above a "hola" - and were corrected rather than left to lock the bug in place.
+
+NOT YET PUBLISHED. Installed into the local harness for testing only.
+
+---
+
+## 3.9.4
+
+**The language choice now survives a restart.**
+
+- Reported: "I started harness and it was auto-set to english." The choice had been made and working; a
+  restart erased it, the panel silently reverted, and because the page is only told the language when a
+  choice exists, answers came back in English too.
+- The choice is written to the harness profile's storage folder as it is made, and restored on load.
+- Cleared choices persist as cleared; a missing, corrupt or read-only file falls back to defaults instead
+  of throwing; a write repairs a corrupt file.
+- The dropdown shows what the host actually has. If no choice is stored it reads "auto", which is the
+  honest state - previously it could show Español while the page was told nothing.
+
+MEASURED BEFORE FIXING: three prompt shapes were sent to the real page - the current one, an inline
+instruction and a leading sentence. All three came back in Spanish, which proved the page obeys the
+instruction and the fault was the lost setting, not the wording.
+
+NOT YET PUBLISHED. Installed into the local harness for testing only.
+
+---
+
+## 3.9.3
+
+**The answer no longer waits for you to open the window.**
+
+- Reported: a message was sent, the elapsed timer climbed past a minute, and the reply only appeared once
+  the browser window was raised by hand. The plugin minimises the window on purpose, so this was the normal
+  case, and it had never been tested that way.
+- Cause: Chromium throttles timers and rendering in a minimised or occluded window. The reply was finished
+  inside the page while this client - which reads it by polling every 70ms - was not allowed to see it.
+- Fix: the browser is launched with background timer throttling, occluded-window backgrounding, renderer
+  backgrounding and intensive wake-up throttling disabled, and the window-occlusion calculation removed.
+  The page is also told it is active at every connection and again immediately after minimising.
+- If a read still stalls, the page is nudged once and then the call fails with a sentence that says to open
+  the window, instead of a timer that climbs forever.
+
+NOT YET PUBLISHED. Installed into the local harness for testing only.
+
+---
+
+## 3.9.2
+
+**Full-file audit; sixteen bugs fixed.**
+
+- A tool reply that began with prose used to have its prose emitted TWICE: the comparison that decided what
+  to append included the route label, so it never matched. Emitting is now anchored on the streamed
+  position, and a brace-first reply gets its label too.
+- Search answers no longer leak tool envelopes: the answer block is scrubbed and a tool call found in a
+  search reply is still executed.
+- Opening the settings panel no longer launches a browser: inspect() reads or reports "nothing to listen
+  to", and only an explicit action starts one.
+- A repeated answer only counts as a fault when it arrives within a minute; asking the same question an
+  hour later is a correct answer, not a stuck page.
+- Six strings that were rendered in English regardless of language now follow the dictionary: the trust
+  panel, the session/provider status details, the "switching…" chip, "Last thing sent", the savings line,
+  the tools caption and the "auto" language label. All six translated into all 51 languages.
+- The settings tab title follows the harness locale (Uncle Sam / 山姆大叔).
+- "Start fresh" and "Show browser" are greyed again when no browser is running - the helper that did this
+  had silently lost it.
+- Dead config wired up, broken indentation fixed, and a prompt-preview edge case guarded.
+
+NOT YET PUBLISHED. Installed into the local harness for testing only.
+
+---
+
+## 3.9.1
+
+**Answers now come back in the language the panel is in.**
+
+- A Spanish panel used to answer "hola" in English, because the page only ever saw the single word. One
+  short line now goes with the message: [Answer in Spanish].
+- The language name is the English one, so the instruction is unambiguous - "Spanish" rather than "Español".
+- English readers get nothing extra: the instruction is omitted when the panel is English.
+- Tested against all 51 languages, including that the tool line, the language line and the person's own
+  words appear in that order.
+
+NOT YET PUBLISHED. Installed into the local harness for testing only.
+
+---
+
+## 3.9.0
+
+**Fifty-one languages.**
+
+- Ten more: Nepali, Sinhala, Khmer, Burmese, Georgian, Azerbaijani, Kazakh, Uzbek, Lithuanian and
+  Slovenian.
+- The review page and the plain-text sheet now cover all 51: 4,437 strings.
+
+NOT YET PUBLISHED. This build is installed into the local harness for testing only.
+
+---
+
+## 3.8.0
+
+**Forty-one languages.**
+
+- English, Chinese, Spanish, German, French, Portuguese, Russian, Japanese, Korean, Italian, Arabic,
+  Hindi, Dutch, Polish, Swedish, Danish, Norwegian, Finnish, Czech, Slovak, Hungarian, Romanian, Greek,
+  Bulgarian, Croatian, Serbian, Ukrainian, Thai, Vietnamese, Indonesian, Malay, Filipino, Bengali,
+  Tamil, Telugu, Marathi, Persian, Urdu, Turkish, Swahili and Hebrew — each covering the whole panel,
+  the model picker and the error messages.
+- Every language is written colloquially rather than literally: "buy me a coffee" uses the idiom for
+  treating someone in each language, not a word-for-word translation.
+- Six of them (Bengali, Tamil, Telugu, Marathi, Swahili, Hebrew) are flagged in dev/TRANSLATION-NOTES.md
+  as worth a native review.
+
+---
+
+## 3.7.0
+
+**Forty-one languages, and the tool line cut down to one.**
+
+- Each panel gets a language selector; the words follow it instead of the harness detection.
+- The tool line is still sent, because the chat box has no system channel, but the call format is now a
+  clause on that same line rather than a second block above the person's own message. It is also shown in
+  the panel's "last thing sent" view, where it belongs.
+
+---
+
+## 3.6.2
+
+**The rest of the page, translated.**
+
+- Every remaining English string now has a key: the model-route card, the page introduction, the whole Brave
+  recommendation with its three reasons, the headless explanation, the browser-in-use line, the Edge
+  suggestion and the composer switch with its tooltips.
+- The title is translated too: El Tío Sam, Onkel Sam, Oncle Sam, 山姆大叔.
+- The panel key list, the locale build and the tests now agree on 86 strings.
+
+Fixed while doing this: the build checked for missing keys before merging the translations, so it reported
+keys as missing that it had just supplied.
+
+---
+
+## 3.6.1
+
+- The language dropdown is painted with the harness's own theme tokens instead of `transparent`, so its
+  field and its open list follow the theme rather than falling back to the system's white popup. That was
+  the bug: a native select draws its list with system colours, which on a light system sat white over a
+  dark panel.
+
+---
+
+## 3.6.0
+
+**Five languages, chosen by the harness, overridable in the panel.**
+
+- English, Spanish, Chinese, German and French, each covering the whole panel, the model picker and the
+  error messages.
+- The language follows the harness's own setting. No IP lookup, no cookies, no network request: the harness
+  already knows, so the plugin asks rather than guesses.
+- A small selector under the title lets the reader override it - someone reading English happily on a system
+  set to another language should not be forced into that language.
+- A test fails the build if any language loses a key or a %s placeholder, so a translation cannot silently
+  fall back to English.
+
+---
+
+## 3.5.5
+
+- The author section opens with the coffee link itself, in blue, instead of a heading. Two short sentences
+  follow, then the machine and its specification.
+- The name is gone from the panel: there is one person behind this and the reader already knows who.
+- The duplicate link above the title is removed, so the page asks once rather than twice.
+
+---
+
+## 3.5.4
+
+- The reason for the machine is now the real one: a faster build-test-fix cycle. The previous wording claimed
+  the plugins needed to run models locally, which was not true - the harness does its work over the network
+  and the browser, not a local GPU.
+
+---
+
+## 3.5.3
+
+- The goal machine is now the ASUS ProArt P16 (H7606WX): AMD Ryzen AI 9 HX 370, AMD XDNA NPU at 50 TOPS,
+  and the same 24 GB RTX 5090. Replaces the Intel machine, which was the only 18-inch option but not the
+  better fit for running models locally.
+
+---
+
+## 3.5.2
+
+- Coffee link is now a blue uppercase link with no icon and no explanation.
+- The author section states the goal: the exact machine being saved for, its specification, and that every
+  coffee moves it closer.
+
+---
+
+## 3.5.1
+
+- The coffee link moves to the top of the settings page, above the title, and is a plain link rather than a
+  button. The bottom button is gone; one ask on a screen is enough.
+- The trust list no longer marks reassurances with a red cross, which read as warnings.
+
+---
+
+## 3.5.0
+
+**Customer-facing additions.**
+
+- **Savings counter** in the panel: calls, estimated tokens, and the money those tokens would have cost
+  through the API, at the published list rate. Labelled as an estimate because the page reports no usage.
+- **Start fresh** button: resets the browser to a blank conversation, which was previously impossible from
+  inside the harness.
+- **Show browser** button: raises the Uncle Sam window without hunting the taskbar.
+- **Last thing sent**: the exact text typed into the chat box, so what was sent is visible without opening
+  the tab.
+- **Trust panel**: a plain list of what the plugin does and does not do.
+
+---
+
+## 3.4.0
+
+**Queue watchdog, so a dropped stream cannot freeze the plugin.**
+
+- A call that holds the serialisation slot is released by force after 20 minutes. If the harness abandons
+  a stream, the release used to be skipped and every later request waited on a promise that would never
+  settle - the plugin appeared frozen until restart. Nothing legitimate holds a slot that long.
+- Deliberately still no minimum interval between requests. A competitor inserts a random 2-4 second gap
+  because it re-sends the whole transcript every turn; Uncle Sam continues the chat in the page and sends
+  only the newest message, so that gap would cost latency and buy nothing.
+
+---
+
+## 3.3.1
+
+- Attribution completed: Operating Tag everywhere a person looks, with the copyright holder named in
+  LICENSE only.
+- README install command, badge and installer updated to the `@operatingtag` scope.
+
+---
+
+## 3.3.0
+
+**Published under the Operating Tag scope.**
+
+- Package published under the `@operatingtag/dsh-unclesam` name.
+- Author field now reads **Operating Tag (@operatingtag)**.
+- Repository links updated. The display name is unchanged: this is still Uncle Sam.
+
+---
+
+## 3.2.0
+
+**Windows installer.**
+
+- `install.cmd` and `install.ps1` install or remove the plugin without any clicks in the GUI: they find
+  the harness profile, back up both files it touches, link the package and add the profile row.
+- `-DryRun` prints every action and writes nothing.
+- `-Uninstall` restores the profile to exactly its previous state, verified by a round-trip test on a
+  throwaway profile.
+- Execution policy is bypassed for that one process only; no machine setting is changed.
+
+---
+
+## 3.1.0
+
+**Support link added.**
+
+- The author section in the plugin panel now carries a **Buy me a coffee** button, pointing at a real
+  payment link rather than inviting the reader to go and find one.
+- README gains a Support section with the same link.
+
+---
+
+## 3.0.2
+
+**Installation documented and verified end to end.**
+
+- README gains an Installing section: requirements, both install routes, what a successful install looks
+  like, the first-run sign-in, and how to remove it again.
+- Repository, bugs and homepage fields added to the manifest.
+- `check` and `pack:check` scripts added.
+- Verified by packing the tarball and installing it into a clean directory: all twelve files land, the
+  entry point exports `apply()`, and the prompt module runs from the installed copy.
+
+---
+
+## 3.0.1
+
+**Renamed the package to `@operatingtag/dsh-unclesam`.** The display name stays *Uncle Sam* — the
+package identifier is what npm resolves, the title is what a person reads.
+
+- Package name, bundle row id and client module id all updated together, so the client half still loads.
+- Provider route `deepseek-web` deliberately unchanged: it is recorded inside saved model selections, and
+  renaming it would orphan existing conversations.
+
+---
+
+## 3.0.0
+
+**First publishable build.**
+
+- Every source file carries a header with its name, version, author and licence.
+- `LICENSE` added (MIT), `author` and `license` fields filled in the manifest.
+- Manifest gains keywords and a Node engine requirement; `README` and `LICENSE` ship with the package.
+- Development scratch files moved out of the package so the distributed files are only the real ones.
+
+---
+
+## 2.9.2
+
+**Three fixes found by screen recording a real session.**
+
+- **Tool-call JSON no longer reaches the visible reply.** The model was writing
+  `{"tool":"…","arguments":{}}` immediately followed by prose on the same line; the previous pattern both
+  missed it and could not balance the nested braces. Replaced with a brace-counting scan.
+- **Tool list capped at twelve names, with the remainder counted.** A session carrying thirty tools turned
+  the preamble into a wall of text in front of the user's own message.
+- **DeepThink and Search are no longer reset.** The plugin treated `false` as an instruction to switch them
+  off, so a user who enabled them by hand had them silently cleared on the next message. Only an explicit
+  `true` acts now.
+
+---
+
+## 2.9.1
+
+- Tool preamble reduced to names only, with descriptions dropped. The previous "compact" form was compact
+  for three tools and enormous for thirty.
+
+---
+
+## 2.9.0
+
+**Plain conversation sends only the user's words.** The harness system prompt is no longer typed into the
+chat box, so the browser shows what the person typed and nothing else. Tool turns still describe the tools,
+because a model cannot call a tool it was never told about.
+
+---
+
+## 2.8.6
+
+- Tool-call leakage scrubbed from the visible text as well as discouraged in the instruction, covering
+  calls written before prose, fenced calls, truncated calls and stray narrator lines. Ordinary JSON in
+  prose is left untouched.
+
+---
+
+## 2.8.5
+
+**Latency reduced by roughly 2.8 seconds per message.**
+
+- The fixed 1,400 ms wait after sending replaced with polling, which exits as soon as the composer clears.
+- Reply settle window reduced from 2,000 ms to 600 ms.
+- Poll interval tightened from 110 ms to 70 ms so the settle window is measured accurately.
+
+---
+
+## 2.8.4
+
+- Plugin display name changed to **Uncle Sam** across the settings tab, panel heading and description.
+  Package identifier unchanged at this release.
+
+---
+
+## 2.8.3
+
+- Switch labels changed to `DeepSeek Harness` and `Uncle Sam`, matching the signature on replies so the
+  control and the answer use the same name.
+
+---
+
+## 2.8.2
+
+- The status dot no longer greys out on startup. Availability now only affects the colour when the active
+  route is the web one — a grey light on a working API route invited people to fix something that was not
+  broken.
+
+---
+
+## 2.8.1
+
+- Labels corrected to `Api` and `Uncle Sam`, with the dot red on the web route and blue on the API route.
+
+---
+
+## 2.8.0
+
+- Switch redesigned from a bordered two-segment pill to a single quiet chip: a small status dot and one
+  lowercase word, matching the weight of the controls beside it.
+- Hover surface and pointer cursor make it read as clickable without shouting.
+
+---
+
+## 2.7.0
+
+- Switch moved from below the composer card into the composer tool row, on the left, where there is room.
+  The right side of that row is already full, so adding there squeezed the model name or pushed the send
+  button.
+
+---
+
+## 2.6.2
+
+- Composer row tidied: model label enlarged and vertically centred, switch right-aligned.
+
+---
+
+## 2.6.1
+
+- Tool preamble wrapped in a labelled code block so it reads as harness scaffolding rather than raw JSON
+  appearing in the conversation.
+
+---
+
+## 2.6.0
+
+**The switch now changes the current conversation.**
+
+- The session id is read from the component prop the slot provides, instead of being searched for in the
+  client service, sidebar links and address bar. Those searches failed, so the host saved the choice as the
+  default for new conversations only, and the badge could show WEB while replies came from the API.
+- Settings keeps no conversation in view, so it changes the route for new conversations and says so.
+
+---
+
+## 2.5.0
+
+- One browser chat per harness conversation, so two sessions no longer share a single transcript and answer
+  each other's questions.
+- Session-title and history-compaction calls get their own slot and never touch the user's chat.
+- Internal calls carry no route signature.
+
+---
+
+## 2.4.0
+
+- **Sending made reliable.** Enter is retried with escalation — a real key press, an in-page key event, a
+  click on the send button — and a send only counts as successful when the composer clears.
+- Reported cause: a 16-character message was typed in full and Enter was ignored, with no error anywhere.
+
+---
+
+## 2.3.0
+
+- A message that is typed but never submitted now fails immediately with the character count still in the
+  box, instead of waiting for a first token that will never arrive.
+
+---
+
+## 2.2.1
+
+- Reply signature changed to `🔴 Uncle Sam`, documented in the plugin description and configuration as a
+  joke rather than a claim of authority.
+
+---
+
+## 2.2.0
+
+- Web replies are signed so the route that produced an answer is visible in the transcript itself, not only
+  in a badge beside the composer.
+
+---
+
+## 2.1.1
+
+- Panel rows renamed from two ambiguous "Browser" labels to `Running` (is it up) and `Using` (which engine).
+
+---
+
+## 2.1.0
+
+- **First-run flow:** the browser window stays visible until the user has signed in, then minimises itself.
+  Previously the window minimised immediately, leaving a signed-out browser hidden with no way to log in.
+
+---
+
+## 2.0.0
+
+**Collapsed to one browser model.**
+
+- Browser selection simplified: Brave when installed, Edge otherwise. Brave is recommended but never
+  required, and nothing is installed on the user's behalf.
+- Headless mode removed. It was built and measured: the site answers headless clients with a 403 from
+  CloudFront before the application loads, so no flag or setting could make it work.
+- Window is minimised to the taskbar after launch, verified rather than assumed.
+- Mode switches removed; there is one behaviour.
+
+---
+
+## 1.20.0
+
+- Returning to the API no longer refuses. When no API model was remembered the plugin now falls back to the
+  deployment default, then to any routable model, and only declines if the harness offers none.
+
+---
+
+## 1.19.0
+
+- Probe port follows the display mode, so a headless browser launched correctly on its own port is not then
+  looked for on the wrong one.
+- A launch that never settles can no longer leave the badge reporting "starting the browser…" forever.
+
+---
+
+## 1.18.0
+
+- Web calls continue the browser chat instead of restarting it, so the harness conversation and the browser
+  conversation are the same conversation.
+- Prompt size capped: tool descriptions budgeted, and a hard ceiling on the whole prompt.
+- Setup panel waits for the single-page app to hydrate before reporting "not signed in".
+
+---
+
+## 1.17.0
+
+- Author attribution added, and the owner name shown in the plugin panel.
+
+---
+
+## 1.16.0
+
+- Setup panel also registered as a top-level settings page. The plugin tab lives inside the Plugins section,
+  which is not where someone lands after clicking a plugin in the Plugins list.
+
+---
+
+## 1.15.0
+
+- Window minimisation performed as an action rather than a launch flag. `--start-minimized` was measured to
+  have no effect; the result is now read back and reported.
+
+---
+
+## 1.14.0
+
+- Brave recommended, with the reasoning stated in the panel, and Edge used automatically when Brave is
+  absent.
+
+---
+
+## 1.13.0
+
+- Display modes introduced: headless, minimised, visible.
+
+---
+
+## 1.12.0
+
+- Setup panel added, showing browser, session, provider and readiness, with sign-in and check actions.
+
+---
+
+## 1.11.0
+
+- Choosing the web route starts the browser itself, so the switch no longer sits on "unavailable" until a
+  message happens to be sent.
+
+---
+
+## Earlier builds (1.0.0 – 1.10.x)
+
+Pre-release iteration on the provider itself: registering `deepseek-web` as a model provider, driving the
+composer through the DevTools Protocol, streaming deltas into harness chunks, and the two-way switch. These
+builds were development snapshots and were not distributed.
+
+---
+
+## Origins
+
+The web route exists because the site requires a proof-of-work challenge per message and is fronted by
+Cloudflare Turnstile. Driving a real signed-in browser satisfies both without impersonating the site's
+private API.