v2026.8.4

What's Changed

Features

• The streaming pane reflows when you resize the window. Wrap width was
  chosen once per answer, at the moment it landed, and baked into the
  scrollback as real newlines — so answers that arrived before and after a
  resize sat side by side at different widths, which reads as broken output.
  When a new width holds still for a moment, every answer so far is
  re-rendered to it, at the close prompt as well as mid-run. The redraw clears
  the pane's scrollback along with the screen, so no old-width copies survive
  in history; the cost is that it resets copy-mode position.
• Response banners name the model that answered. Every CLI-backed provider
  used to read default, because the council passes no model flag and so had
  nothing to name. It does have something: each CLI records its selection in
  its own config, and reading it is free and needs no extra call — codex from
  $CODEX_HOME/config.toml, grok from ~/.grok/config.toml, kimi from
  ~/.kimi-code/config.toml, and agy from its CLI settings. The value also
  reaches the cache key, which is the point: reconfiguring a CLI's model now
  stops the council serving back an answer the previous model gave.

Fixes

• A CLI provider no longer gets a deadline four times stricter than an API
  one. The API providers wrap curl in curl_with_retry, so
  COUNCIL_TIMEOUT bounds one attempt and their real ceiling is
  (COUNCIL_MAX_RETRIES + 1) times it. The CLI providers never retry, so the
  same number was a hard deadline. One value the user sets once meant two
  different things, and a council carrying a large --file dropped kimi on
  that difference while the API path survived far longer on the same run. CLI
  providers now default to COUNCIL_CLI_TIMEOUT (1200s), matching the API
  ceiling. Retries are deliberately not extended to the CLI path: a CLI that
  is slow is not failing transiently, and retrying a fifteen-minute run four
  times is worse than reporting it.
• A kimi timeout reports a timeout. When the watchdog killed the CLI, bash
  printed its own report of the signalled child — Alarm clock: 14, the PID
  and the entire command line — on the provider's stderr, which the council
  stores verbatim as the error text. The pane showed that instead of the
  timeout message the script already emitted, carrying an absolute plugin path
  into the display.
• The ask command can no longer be talked out of running the council. A
  question ending "in exactly this shape and nothing else, no extra sections"
  was read as governing the command's own reply — and the command's own reply
  is the provider display, so it answered directly and queried nobody. The
  question is now declared as data: formatting directives inside it constrain
  what each provider is asked to return, never the command's output, and never
  authorize skipping the run.
• codex's model is read from CODEX_HOME. codex scopes its configuration
  to that variable, so a relocated config left the banner naming a model codex
  itself ignores.
• A config the council can't parse no longer takes the run down with it.
  Reading a CLI's model left the parser's own stderr and exit status
  unhandled, so a corrupt config printed a parser complaint as that provider's
  error text and returned non-zero from a function called under
  set -euo pipefail. One unreadable file could end the whole council run
  instead of degrading one banner to default.
• Three ways a valid config was misread. A commented table header
  ([models] # mine) lost its entire table, so an annotated config silently
  reverted to default and orphaned every answer cached under a real model
  id. TOML literal strings (model = 'gpt-x') were skipped, so the feature
  did nothing with no error to explain it. And a bracketed line inside a
  multi-line value hijacked the parser's idea of which table it was in — with
  a body line of exactly [] re-entering the root table, so quoted lines
  inside a string were read as configuration.
• The close prompt no longer stacks. Resizing twice with nothing on screen
  printed a second and third [esc/ctrl-d] close; and when the pane's input
  was already gone, the same dead tty failed the width query, which read as a
  resize and wiped the reader's scrollback on the way out.

Changed

• Two blank lines between provider blocks instead of four. The response
  renderer's trailing margin and the banner's leading margin were written
  without each other in view, and both survived into the gap.
• The council answers to its own name. The ask command described the
  situations it should volunteer in but never the phrasings that ask for it
  outright, so "a full council review" matched nothing. Direct mentions now
  trigger it, and the plugin's keywords cover the full provider set.

Docs

• The README leads with a screenshot of five providers answering the same
  question with the synthesis alongside.
• Architecture and testing docs updated: COUNCIL_CLI_TIMEOUT and why the CLI
  bound differs, the file tree, and corrected per-suite test counts.

Other

• Extracted json_value alongside toml_value so the per-provider config
  table reads as a table.

Full Changelog: v2026.8.3...v2026.8.4