@prplkane/roundtable 0.0.0-stage → 0.6.3

package/CHANGELOG.md

@@ -0,0 +1,513 @@
+# Changelog
+
+## 0.6.3
+
+### Assistant integration
+
+- Add `roundtable setup` to connect user-level MCP tools and a workflow skill for Codex,
+  Claude Code and Antigravity across projects. Include checks and removal, preserve unrelated
+  settings and refuse to overwrite edited entries. No automatic npm install hook is added.
+- Add a dependency-free stdio MCP adapter using the existing JSON client. Start a hidden
+  local board on demand, expose questions and bounded extensions, and keep reviews running
+  when the host disconnects. Report completion only after a ready chairman verdict.
+- Recommend one critique round for normal agent reviews and reserve zero for quick
+  comparisons. Report actual completed rounds so agents cannot mistake blind positions
+  for a discussion. Bundle the skill in the npm payload and test installation outside a checkout.
+- Count attempted rounds separately from rounds with usable replies, so failed critiques do
+  not inflate review depth. Share the HTTP client and validation bounds between MCP and CLI;
+  remove the subprocess watchdog that could interrupt a valid long wait.
+- Check server health without waiting for provider catalogues; retry transient readiness
+  errors and refuse incompatible versions, modes or explicit storage directories. Add idle-only
+  server status/start/stop/restart commands for installed updates.
+- Preserve all unmanaged TOML declaration forms, literal dollar strings, selected modes and
+  ports. Honor Claude profile overrides, check the Node executable and repair missing managed
+  files. Test a real npm archive installation after moving its source checkout.
+
+### Review workflow
+
+- Make attended checkpoints finite in both the browser and agent client: 15 seconds for a
+  correction window and two minutes for a question by default, configurable separately.
+  An unanswered window continues with unknowns intact; it never supplies an answer or accepts
+  the recommendation. The hard wall clock still applies.
+- Replace permanent holds on incidental browser clicks with bounded extensions while writing
+  or requesting More time. Show question countdowns, draw attention in the tab title, and
+  retain automatic-continuation notices after reload.
+- Return an explicit question handoff, answer template, deadline and commands to coordinating
+  agents. Add `agent hold` for a bounded extension and `wait --after-round` to await a response
+  without asking the same question again. Empty checkpoints need no user confirmation.
+
+### Documentation
+
+- Add Linux first-connection steps, including a direct GitHub archive install, assistant
+  registration, discovery checks and a provider-free demo.
+- Show the app's pixel logo tile in interactive help and foreground startup, drawn with half
+  blocks in the orchid and aubergine of the board header, with the wordmark and installed
+  version beside it. Respect terminal width, UTF-8 support and `NO_COLOR`; keep redirected
+  output and machine protocols undecorated.
+- Add a Known issues entry to the guide, linked from the README: Codex calls can stall, what
+  Roundtable does about it and what it cannot do.
+
+## 0.6.2
+
+- Simplify model names to default and lightest annotations, removing inferred plan restrictions.
+- Let Claude's next quota refresh recover after a free `/usage` answer without limits.
+  Keep the safeguards for answers that may have called a model.
+- Show the latest quota on every provider card and replace the first-call promise with
+  "quota unavailable" when no figure is known.
+- Make Clear include interrupted sessions as well as finished ones. Running reviews and
+  reviews waiting for input remain protected, with the full selection checked before deletion.
+- Remove deletion success banners from the history list.
+- Clarify that usage statistics describe your own agents' Roundtable calls, stay on your
+  machine, and are not sent to Roundtable's developers. Document restarting after updates.
+- Take the version shown in the board footer and sent in the Codex app-server handshake from
+  `package.json`; both still said 0.6.1.
+
+### Review fixes
+
+- Stream Codex replies over WebSocket, so a response that never starts is caught. A request
+  can be held for minutes before its response starts (one traced call: 173 seconds). Over SSE,
+  Codex starts its idle timer only once response headers arrive, so that wait had no limit short
+  of Roundtable's call limit; over WebSocket the timer starts when the request is sent. Against a
+  local server that never answers, SSE waited until it was killed, while WebSocket reconnected at
+  each idle limit. On the real backend, 18 WebSocket calls at low, medium and high effort all
+  answered, with no reconnect on a healthy call; the three stalls were resent at their idle
+  limit. After its reconnects Codex falls back to HTTPS for the rest of the call, where the wait
+  for response headers is again bounded only by the call limit, and a WebSocket that does not
+  open within 10 seconds does the same. Reconnects are now counted from Codex's retry log too,
+  since its events leave out the first WebSocket retry.
+- Reconnect a silent Codex request after 15 seconds instead of up to 95. Each request a Codex
+  call makes either streams within a few seconds or stays silent, on every model and search mode
+  tried, through `codex exec` and `codex app-server` alike, with or without Roundtable's
+  settings. A Codex seat, chairman or searcher at low effort (or none or minimal) now reconnects
+  after 15 seconds of silence, up to five times, and a reconnect resends only the silent request.
+  This does not make stalls rarer (in testing, 15, 35 and 60 second limits each saw one or two
+  slow calls in eight); it makes a silent one cost about 15 seconds instead of 95 for a seat at
+  the default call limit, where stalled low-effort seats took 101 to 220 seconds. Other efforts
+  keep the limit derived from the call limit, since a request that reasons longer may stay silent
+  while healthy (GPT-6.1 Sol at high effort sent no stream events for 15 to 17 seconds).
+- Count a Codex searcher's reconnect notices as progress, and start once more a searcher call
+  whose reconnects all failed. The reconnects are bounded, and a reconnect keeps a finished
+  search, so stopping the call at a reconnect threw that search away. The fresh attempt follows
+  the 0.6.1 rules: at least 30 seconds must remain, and both attempts share the time limit.
+- Keep a search whose sources include a date the searcher could not give as a real `YYYY-MM-DD`
+  date. One such date (a month only, a year only, "unknown") used to reject the whole source
+  bundle and stop the review before discussion; that source's date is now saved as not reported
+  and the other sources keep theirs.
+
+### Security and usage
+
+- Ask a Codex searcher for one search call carrying several queries. Each further call is another
+  request that can stall, and it resends everything gathered so far: input grew from about 22k
+  tokens for one search to 103k for four. In testing, one call returned as many sources (3 on
+  average) at 23k input tokens on average instead of 35k, with a median of 13 seconds instead of
+  41. The search limit is unchanged: a second call is still allowed when the first returned
+  nothing usable. Claude and Antigravity search one query per call and get no such hint.
+
+## 0.6.1
+
+### Security and usage
+
+- Require acknowledgment of Antigravity's account-risk notice before provider calls, model
+  lists or quota reads. Google's [FAQ](https://antigravity.google/docs/faq) says third-party
+  wrappers are unsupported under consumer plans and third-party access can lead to account
+  suspension or termination. The board records the acknowledgment locally; it can be
+  withdrawn and is requested again when the notice is updated.
+- **Upgrading Antigravity workflows:** agent requests now need
+  `options.acknowledgeTermsRisk: true`, after the person agrees. `doctor` needs
+  `--acknowledge-terms-risk` for anything beyond `agy --version`. These are separate from
+  limited-isolation consent. See the [setup guide](docs/workflow.md#antigravity-google).
+
+### Review fixes
+
+- Restore GPT-6.1 Terra in the Codex picker and order model families from stronger to lighter,
+  keeping plan restrictions visible. Custom model IDs remain editable instead of reverting to Default.
+- Align card controls and effort indicators; put provider warnings in an expandable row.
+  Refresh the status after an account notice is acknowledged or withdrawn.
+- Start the searcher query from the question and options, with a preview and an editable custom query.
+  Other context and discussion answers stay out of research.
+- Put the chairman's recommended option above its answer in bold uppercase. Rename history's
+  cleanup action to Clear; its confirmation still identifies the finished sessions to remove.
+
+- Stop a Codex searcher call after 40 seconds without progress and allow one fresh attempt
+  when at least 30 seconds remain. Both attempts share the searcher's time limit and are
+  recorded. Reconnect notices do not count as progress. Missing usage from the first attempt
+  remains unknown and can block the second under a token cap. Each attempt has its own
+  search-operation limit. This bounds recovery; it does not repair a failing network route.
+
+### Documentation
+
+- Explain provider regional availability, the limits of login/status checks, and how to
+  investigate stalled connections. Clarify that Roundtable leaves VPN, proxy and DNS settings
+  unchanged; network comparisons are optional diagnostics.
+- Clarify that Codex's stream-idle timer does not cover every connection stage. Roundtable's
+  overall call limit still applies while the CLI waits for a response to begin.
+
+## 0.6.0
+
+### Review fixes
+
+- Keep successful research when the searcher repeats a source URL. Repeated URLs become one
+  citation, with distinct excerpts combined when they fit the existing limit. Every record
+  still passes URL, date and length checks; no extra model call is made to repair duplicates.
+- Offer each Codex model only the effort levels it takes. The Luna models stop at `max`
+  (`codex debug models`), but every Codex model was offered `ultra`. The pips, the chairman's
+  effort list and the server's check now follow the chosen model, and picking a model that
+  does not take the chosen level returns the seat to the CLI's default.
+
+### Security and usage
+
+- Read Claude's remaining quota with `claude -p /usage`, which the CLI answers itself with no
+  model call, at start, on Refresh status, after each Claude reply and when a debate ends. Claude
+  no longer shows "not reported" until its first reply after every server start, and its
+  per-model weekly limit (for example "7d Fable") now appears beside the session and weekly
+  ones. Only the limit lines are kept; the local activity summary in the same answer is
+  discarded. A reply's own figure stands in when the read gives nothing. An answer that shows a
+  turn, tokens or a cost, one with no plan limits (API-key billing), or a read that times out
+  stops further reads until restart. Calibration now treats reset times a few minutes apart as
+  the same window, since `/usage` prints them to the minute.
+- Read Antigravity's remaining quota with `agy -p /usage`, which the CLI answers itself with no
+  model call. Its limits come per model group (Gemini; Claude and GPT; Gemini 4 Argon on some
+  plans), and the seat lamps, the quota line and the quota picture show the windows of the
+  group each seat's model draws on, with the used share and the reset time. An answer that
+  shows a turn, tokens or a conversation, or a read that times out, stops further reads until
+  the server restarts, so a CLI change can cost one call, never one per reply. Measured quota
+  rates for Antigravity are kept per model and only against that model's group.
+- Bench only the used-up Antigravity model group, not the whole login: when an Argon seat
+  reports a used-up quota, the Gemini and Claude seats on the same login keep answering. A
+  seat on the default model, whose group cannot be told, and a logged-out CLI still take the
+  whole login out, as for every other vendor.
+- Keep Roundtable's calls out of each CLI's own history. Claude seats now pass
+  `--no-session-persistence` and Codex seats `--ephemeral`, so a stateless seat, chairman or
+  searcher call no longer leaves a session, prompts included, under `~/.claude/projects` or
+  `~/.codex/sessions`. Only seats in resume memory, whose next round continues the session,
+  still save it. Antigravity has no such flag.
+- Read Antigravity's AI credits with `agy -p /credits`, which the CLI answers itself with no
+  model call, beside the `/usage` read and at most every ten minutes. The usage table shows
+  "AI credits left", which you previously had to type in from the credits page.
+
+### Local operation and interface
+
+- Keep Windows history and login profiles in `%USERPROFILE%\.roundtable` instead of
+  `%LOCALAPPDATA%\Roundtable`. Windows redirects the AppData writes of anything started from a
+  packaged app, such as the Claude or Codex desktop app and their terminals, into a private copy
+  per app, so each app saw its own history and its own one-server lock. At startup Roundtable
+  now lists every older folder that still holds a review the data folder lacks: a checkout's
+  `debates/`, the earlier AppData folder and each packaged app's copy of it. It never moves them
+  itself. A login profile made under the earlier AppData default keeps working until it is moved.
+- Stream Codex replies over SSE with an idle limit that fits inside the call limit and two
+  reconnects, through a provider entry named per call that points at the same ChatGPT backend
+  with the same login. The built-in provider offers no idle or reconnect setting, and the CLI's
+  documented stream-idle default equals the default call limit, so a stalled reply ran the full
+  five minutes before failing; it now reconnects after 95 seconds at that limit and fails sooner
+  once the reconnects are used up. That a custom provider may use the ChatGPT login is not documented;
+  `SECURITY.md` names both ways a future CLI could react. Every reconnect is counted into the
+  attempt's saved progress and shown as a "reconnected ×N" mark on the reply card, in failure
+  details, error logs and the agent state, so a recovered stall stays visible.
+- Wait out an API endpoint's short rate limit once instead of benching its host for the rest of
+  the debate. A `429` whose `Retry-After` asks for 30 seconds or less, the delay `agy` itself
+  waits out, is retried once after that wait when it ends before the wall clock does. Without
+  `Retry-After`, with a longer one, or at a second `429` the host is benched as before; `402`
+  and every other status are unchanged. A `429` now counts as zero tokens, since the server
+  refused the request before the model ran, so a seat's token cap no longer blocks the retry or
+  later rounds as unknown spending. Stop, finish and the wall clock end a retry's wait at once,
+  for this wait and a network failure's, and a retry is skipped when another seat benched the
+  same login during the wait. The limited attempt is recorded with the kind `throttled`.
+
+## 0.5.0
+
+### Security and usage
+
+- Fix the Antigravity tool guard for CLI 1.4, which reports each tool as a `step_update` with
+  the tool inside it. The guard read only the older shapes, so a tool that no Antigravity
+  permission rule can deny, its web search, went unnoticed. Any reported tool step now stops
+  the seat at once, as it does for Claude and Codex.
+
+### Local operation and interface
+
+- Move a reply card as it opens and folds: the body under its seat and position unrolls over
+  the board's base motion time and rolls up a little faster before the card folds. Reduced
+  motion opens and folds it at once. The card head is also a button, so Enter or Space opens
+  and folds it, and focus stays on it while other replies land.
+- Codex calls disable analytics/telemetry uploads and unlimited connection retries. A valid
+  completed answer with measured usage survives a CLI that hangs during shutdown; incomplete
+  answers still fail without a timeout retry. Search and reconnect stages appear in diagnostics.
+- Treat Codex warning items as diagnostics instead of tool use, and surface a failed turn's
+  provider error promptly. Unknown tools remain blocked. Resolve npm and native CLIs in PATH
+  order so a later desktop binary cannot replace the terminal's selected installation.
+- Stop presenting Codex defaults from user configuration that isolated calls ignore. Keep
+  model and effort choices explicit when reproducibility matters.
+- Keep parent Codex app routing and session variables out of child CLIs and quota reads.
+  Use a shorter decision-review base prompt for Codex instead of its coding-agent prompt;
+  the tool restrictions remain unchanged. The temporary instruction file is removed after
+  each call, including failures.
+- Charge Codex resume calls by the increase in session token counters. Previous rounds are no
+  longer counted again against usage and budget caps. Missing or decreasing baselines remain
+  unknown; fresh sessions start their accounting from zero.
+- Show the chairman at work: while the verdict call runs, the Conclusion panel draws a pixel
+  sheet being written under a quill instead of staying hidden. Reduced motion shows a still page.
+- Keep a round or detail the person folded or opened across redraws; before, every server
+  event reopened the latest round and closed everything else.
+- Give the option columns their own grid, so two options fill the row with no blank third
+  column beside them when a note or the thinking strip is present.
+- Move reported quota to one quiet line above the board; the question now heads the review
+  state, the conclusion and the rounds directly.
+- Show a plain boot screen when setup takes longer than a moment: the logo in pixels, its
+  ring breathing slowly in and out while a light runs round the seats, and one line of text.
+- Box only the main controls. Row Delete appears on hover or focus (always on touch), Show all
+  and Clear finished are quiet text at the foot of the debates list, the review state and CLI
+  seats are unboxed, recommended states take their colour, and partial token sums show "+".
+- Fewer controls in the header: Compact is gone; each reply card starts folded to its seat and
+  position, even in the round that is still coming in, and opens by a click on its head;
+  theme and sound are pixel icons; Save error log moved to the status line. The debates list
+  folds and the browser remembers it.
+- Lead the board with the question as its heading and the review state as a small tag beside
+  it. Usage details is a toggle that never moves; its table groups seats under one line per
+  account, right-aligns numbers and marks partial sums with "+" as the debates list does.
+- Start every fold closed: debates, More context, the Quota picture, login profiles, and rounds
+  once a review is finished. More context spans the panel; run settings take the width they
+  need. Quota cards share one layout, with a shade per seat and a hatched chairman in the bar.
+- Rework the paper theme on night's structure: a faint plum page, white fields with visible
+  borders, seat colours closer to night's hues at text contrast, softer position tints.
+- Make research a visible role, the **searcher**: a card added with **+ Add searcher** beside
+  the chairman, with an open pixel eye as its mark. It has its own vendor, model,
+  effort and login profile instead of borrowing the first Codex seat's. Pick Claude, Codex or
+  Antigravity on its card; it starts on that vendor's lightest model at low effort, and agent
+  requests may set all four in `options.research` (Codex when no vendor is named). Claude
+  searches with WebSearch as its only tool, and its own search count and search-model tokens
+  are included; on Antigravity any tool but its web search stops it. It
+  replaces the research panel and its switch, and is named on the board, in Usage details and in
+  the quota picture. While it works the review state reads "Searcher at work" and the sources
+  panel shows its eye moving down a column of results, lighting the ones it keeps; with reduced
+  motion the picture stays still. Its card title and Effort carry a help tip, the README's
+  "How a review works" names it as an optional step, and the chairman's Effort has the tip too.
+- Show Antigravity's limited isolation on the card that runs on it, seat, chairman or searcher:
+  one warning line whose tip says in three sentences what cannot be sealed off, what Roundtable
+  denies and stops, and that choosing Antigravity accepts it. Choosing it there is the consent
+  the board sends, so there is no separate switch and no refusal when the table opens. Agent
+  requests still send `allowLimitedIsolation: true`; without it the server refuses Antigravity.
+- Fix a stray brace in the stylesheet that dropped the quota readout grid, and debate titles
+  squeezed to one letter at medium widths.
+
+## 0.4.0
+
+- Add optional Codex web research before discussion, with an explicit public search question,
+  time limit and reported search-operation limit. Save one cited evidence bundle for every
+  seat, recovery and export; show research spending separately. Failures stop before discussion,
+  interrupted research never automatically reruns, and debate tools remain disabled.
+  Project safe research fields in agent snapshots, including older saved failures. Include
+  shared source context in setup estimates and admission checks for discussion and verdicts.
+
+- Save safe provider progress on timeout, cancellation and failure, including first output,
+  last stage, idle time and byte counts. Show it in failure details and diagnostic downloads;
+  preserve complete usage received before a stalled exit without retrying unknown spending.
+
+- Enforce the provider tool guard on every discussion and verdict call, including recovered
+  reviews. Stop reported tool activity without a repair; retain spending and exclude rejected
+  text from later prompts. Inspect bounded CLI streams while running, strengthen Claude and
+  Codex isolation flags, send Antigravity prompts on stdin and require explicit consent to
+  Antigravity's remaining isolation limits with session-only tool denial rules.
+  Invalid, incomplete or oversized streams stop the seat without retry or session reuse.
+  Agent help documents Antigravity's explicit consent option and remaining isolation limits.
+
+- Clear the board's reconnecting state as soon as its event stream reconnects, including
+  while waiting for human input with no new events to replay.
+
+- Explain when an earlier timeout or interruption left a seat's spending unknown. Show that
+  cause on the blocked reply and distinguish unknown usage from a projected cap crossing.
+
+- Recover interrupted discussions in fresh provider sessions seeded with completed rounds
+  and saved notes, retaining earlier spending and deadlines. Preserve discussion completion
+  reasons when a later verdict is cancelled.
+
+- Resolve protected export directories through filesystem aliases, refuse linked record/index
+  files, and exclusively create temporary files. Public exports now require an empty staging
+  directory (an existing `.git` is preserved) and never delete destination contents.
+  Preserve existing export/index permission bits on POSIX; new files respect the process umask.
+
+- Redact known endpoint credentials from decoded replies and error excerpts before recording
+  them. Endpoint configuration continues to belong to the trusted local user.
+
+- Quote peer replies, repair text and discussion notes as attributed input records. Reject
+  duplicate reply and verdict headers while preserving ordinary Markdown in content.
+
+- Preserve Resume on server shutdown without recording a human Stop; close event streams and
+  retain cancelled-call spending. Accept emphasis around whole reply and verdict header lines
+  while preserving punctuation inside option labels.
+
+### Review fixes
+
+- Normalize replies, verdict headers, option punctuation and persona parentheticals in linear
+  time, including adversarial whitespace and Unicode input. Preserve exact option labels
+  ending in quotes, stars or punctuation before applying formatting tolerance.
+- Flush a dispatch intent before every provider call; unfinished attempts remain unknown
+  after restart and cannot bypass admission caps. Settled calls are counted once.
+- Stop engines before cancelling server calls and refuse dispatch during shutdown.
+- Rebuild failed reviews after cancelled turns settle so repaired storage can resume without
+  restarting the server. Keep unknown spend and transcript-recovery warnings.
+- Read history summaries asynchronously, cache unchanged records, bound inactive replays and
+  avoid adding global pool/quota events to idle finished records.
+- Remove invalid key-variable values before saving or restoring browser preferences.
+- Preserve unreadable pool/calibration files, report their failure and block writes until
+  repair and restart; replace valid settings atomically.
+
+### Security and usage
+
+- Require HTTPS for hosted endpoints, including names with private-address-like prefixes.
+  Limit decoded API responses, including errors, to 1 MiB.
+- Check budgets before every provider attempt, including retries and fallback chairmen.
+  Reserve for concurrent calls and the verdict; refused calls add no usage.
+- Recheck the wall-clock deadline after a retry delay, before dispatching another provider call.
+- Carry the saved cumulative cost into resumed CLI calls so a server restart does not charge it twice.
+- Keep missing token counts unknown and block further calls to that vendor under a token cap,
+  including after restart. Show unknown or partial totals while retaining reported costs.
+- Make preflight reads cached-only. Authenticate refreshes, share overlapping checks, and
+  bound concurrent work, queue size and cache size.
+- Send a Content-Security-Policy, `nosniff` and a no-referrer policy with every response, so
+  only the server's own scripts can run in the page that holds the launch token.
+- Pin CI actions to commit IDs and make the dependency-update cooldown explicit.
+
+### Records and recovery
+
+- Treat an expired input pause as final even when its timer fires before the overall deadline
+  timer. Do not start a chairman call after the pause expires.
+
+- Add Delete and Clear finished with confirmation and protection for active sessions. Keep
+  exported files and usage statistics after cleanup. Default history to the OS user-data folder,
+  outside the checkout, and allow one server per data folder.
+- Make Stop cancel the whole review, including automatic and manually requested verdicts.
+  Keep Stop visible during synthesis and distinguish cancellation from successful completion.
+  Cancel active calls even if a transcript write fails; fatal engine errors cancel them too.
+- Add Finish with verdict after the first usable round, so ending discussion and aborting the
+  whole review are separate choices in the board and agent client.
+- Keep agent-submitted notes distinct from human notes in the board, prompts, replay and exports.
+- Ignore explicit no-question placeholders when deciding whether to pause. Preserve the blind
+  round pause and questions that still need an answer, without another model call.
+- Recover an incomplete final transcript line with an exact backup before appending.
+  Show recovery notices; keep corrupt completed records visible and block further writes.
+- Preserve Resume notes before the first round and Stop notes entered during a round.
+  Honor Stop without waiting for another input gate and prevent further seat retries.
+- Retain the original chairman's model, effort and profile when requesting a missing verdict.
+  Restore the human decision and reason on reload.
+
+### Local operation and interface
+
+- Present the final answer and its reasons before supporting detail. Distinguish material
+  clarifications from optional questions, and collapse panel positions and run metadata.
+- Add Copy result with a manual clipboard fallback. Keep the agent verdict contract and full
+  exports unchanged; old reviews remain readable without another model call.
+- Ask for a direct, conditional conclusion when important facts are missing, and allow a
+  nuanced synthesis for reflective questions. Do not manufacture questions or thresholds.
+
+- Keep optional cat and alien personas in the public build, with their prompts and artwork in
+  a readable catalogue. The public-tree test checks the catalogue before publication.
+- Add preview status and clear feedback and support links to the README.
+- Lay out a round as one column per option, in option order, from the first reply on. A reply
+  lands at the foot of its column and nothing else moves; seats still thinking wait in a strip
+  under the columns. Before, cards regrouped as each reply arrived and the page jumped.
+- Tint a card against the majority as soon as replies arrive (provisional, "so far" in the
+  tooltip) and settle the tint at the end of the round, instead of colouring only then.
+- Show one banner with a **Reload** button when the board belongs to an earlier server
+  session, and stop retrying from that page. Before, a board left open across a restart
+  reported a page error and retried its diagnostics every 15 seconds, refused each time.
+- Tell the chairman of an unattended review that any seat question is still open, so it lands
+  under its open questions instead of reading as agreement.
+- Recommend the chairman one effort step above the seats, in the agent recipe and the guide:
+  the verdict is the call that decides, and a low-effort chairman dismisses the hardest
+  objection in a word.
+- Play a distinct rising cue when the table stops for you (after the blind round and for a
+  seat's question), so a review left in another tab can call you back. The gavel sounds stay
+  for opening and ending.
+- Add an **Unattended** switch to the round settings, off by default: the review skips every
+  pause, and the seats are told to state what they would have asked as assumptions instead of
+  asking, since nobody answers; the question line leaves their contract, so its tokens leave
+  every call too. For a review nobody will watch. The JSON client takes it as
+  `options.unattended`.
+- Explain each mood on hover: the mood list, the mood on a reply card and the mood of a seat
+  that asked a question all carry its one-line description.
+- Add a **Compact** switch in the header: every card folds to its seat and stance, so a round
+  reads as who chose what; a click on a card opens that reply. The switch is remembered.
+- Name every seat that asked a question above its answer field (name, lens and mood, with a
+  persona's avatar), not only the lens. Seats still see each other by lens only.
+- Show the effort a blank seat really runs at when the CLI's own default is known:
+  "default (max)" for Codex, read from its config.
+- Optional personas on Antigravity run on the lightest Gemini the account lists (Flash Lite
+  where offered, otherwise Flash at low), instead of a fixed id.
+
+- Seats and the chairman reply in the language of the decision; section headers stay English.
+- Add **Choose…** beside both export fields: it opens the system's own folder dialog, so the
+  folder no longer has to be typed.
+- Give a chairman whose first attempt held no recommendation the whole question again on its
+  retry. Before, a stateless retry carried only the format errors, so it could not know the
+  decision it was judging.
+- Use the same external data folder for live diagnostic workspaces and headless calibration
+  reads as for saved reviews. Respect `ROUNDTABLE_DEBATES_DIR` in both commands.
+- Reduce the demo quick start to clone, enter the folder and start the board.
+
+- Mark session opening and completion with the chairman seat's pixel gavel and a quiet generated sound, with
+  separate endings for a verdict, no verdict and a failure. Add a sound toggle. Keep the gavel out
+  of saved-review replay and the JSON agent client, and keep the card still under reduced motion.
+  Events a board missed while its connection was down now arrive as new, so a review that ends
+  during the drop still shows its ending.
+- Give the JSON agent client a `board` link in every result with a review id, so an agent can
+  hand a person the page to watch. Add a small-review recipe for lenses, moods, effort, model
+  and context fields to `agent --help` and the workflow guide.
+- Redraw the grill, steelman, builder, contrarian and chairman icons, and the session gavel with them.
+- Add per-seat token limits. Missing or partial usage blocks further calls for that seat,
+  including retries and resumed reviews; other seats can continue.
+- Select Haiku 4.5 by its explicit model id instead of the `haiku` alias.
+- Keep the folder picker available while its dialog is open, and distinguish cancellation from
+  a timeout or launch failure. Errors are visible beside the export field.
+- Trim a reply with too many DISAGREE or NEW bullets to the limit instead of re-prompting the
+  seat; the extra call cost more than the lost bullet. Show agents `gate.reopenCalls` when the
+  seats already agree at round 0, so answering only reopens the table when it could change a
+  position.
+- Put delayed hover explanations on text labels. Separate positions from arguments, emphasize
+  rejection words, and collapse usage into quota bars with expandable details.
+- Collect questions and answers in one area, combine identical questions without a model call,
+  and wait for final-round answers before the verdict. Reject stale-round submissions.
+- Close discussion controls when rounds end. Explain Stop, provider timeout and review expiry
+  separately, including expiry while waiting for input; retain those reasons on replay.
+- Add JSON agent commands for setup checks, estimates, start, bounded waits, answers and recovery.
+  Keep launch tokens in memory and require explicit actions for calls. Shorten response targets.
+- Give added seats distinct starter lenses and recover empty lenses in saved drafts. Highlight
+  and focus missing or duplicate lenses and empty custom moods before a review starts.
+- Explain lenses, moods and round settings with delayed, keyboard-accessible hints. Add a
+  gentle reveal for hints and new replies, respect reduced motion, and keep replays still.
+- Add local, rotating diagnostic logs and a **Save error log** download. Give failed actions
+  reference IDs and next steps; keep pending browser diagnostics through reloads. Exclude
+  review content, credentials, URLs and personal paths from the diagnostic data.
+- Report failed quota saves and distinguish provider failures from malformed verdicts.
+- Show Demo or Live mode, the active data folder, and backup guidance. Default to two seats
+  and one critique round; remember saved settings and prefer ready logins on first use.
+- Retain drafts after failed requests, explain errors, and prevent duplicate pending actions.
+  Add setup/history retries and explicit status refresh; confirm pauses after acknowledgement.
+- Improve keyboard focus, effort targets, narrow layouts and space for the action bar.
+  Keep saved dollar caps editable when no dollar estimate is available.
+- Serve the guide and fonts locally. Clarify setup, billing, tool restrictions and provider
+  retention in the documentation and doctor output.
+
+## 0.3.0
+
+First public release.
+
+- **Debates between the coding agents you already pay for.** Claude, Codex and Antigravity seats
+  on their own subscriptions, any number per vendor, each with its own lens, mood, model and
+  thinking level. A blind first round, critique rounds with a fixed reply format, stops for
+  consensus, no movement, round limit, wall clock, budget or you.
+- **You take part.** The board waits after the blind round, stops for any question a seat asks
+  you, and carries your notes into every later round.
+- **A sceptical chairman.** A fresh session writes the verdict: a recommendation, the dissent,
+  kill criteria and open questions. It recommends; you accept an option on the board.
+- **Usage tracking.** Per account: remaining quota where the vendor reports it, the projected
+  draw per seat, and token or dollar caps checked against available usage estimates.
+- **Failures you can read.** Each seat card has a lamp (ready, nearly used, out, not checked)
+  from a free pre-flight check. A used-up quota or a lost login benches that account for the
+  debate instead of failing every round.
+- **API endpoint seats.** An OpenAI-compatible endpoint can take a seat: a local server (Ollama,
+  LM Studio, vLLM, llama.cpp) or a pay-per-token API (DeepSeek, Moonshot, OpenAI). Every card and
+  record says whether a seat is a CLI on a subscription or an API.
+- **Decision records.** Export a debate as a markdown record and an index into any codebase, so
+  your coding agents can read what was decided.
+- **The `roundtable` command** starts the board and opens it in your browser (as
+  `npx @prplkane/roundtable` once the package is published); `doctor` shows what your machine
+  can seat.

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 prplkane
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.