v1.39.34

Changed

• The roadmap now follows Primr's outputs. v1.40 deepens the Strategic
  Overview (typically 50 or more pages, every section adding depth); v1.41
  completes the strategy library, with AI Strategy as the default and every
  listed strategy role-agnostic and at least 30 pages; v1.42 ties skill packs
  to the company's roles and strategy; v1.43 brings Primr Zero to the same
  quality with no API spend. Backend freedom, the agent control plane, and run
  history follow, and the run ledger becomes an optional internal aid. The
  product contract states that Primr is deep consultant preparation for any
  role, not a financial analyst's report or a fact lookup, and that fixing
  repetition means deepening sections, never shortening reports.
• Model defaults and prices were rechecked against the xAI, Gemini, OpenAI,
  Claude, and OpenRouter catalogs on 2026-10-29. No newer generally available
  model or price change applies; the default full run (Strategic Overview plus
  AI Strategy) still estimates $1.24.

Added

• OpenRouter routes for Kimi K3.1 Pro, Meta Muse Spark 1.3, and Mistral Large 4,
  three leading models from OpenRouter's benchmark leaderboard that Primr did
  not yet register, priced from OpenRouter's endpoints API on 2026-10-29.

Fixed

• The OpenRouter GLM 5.3 Flash route could not run: its registered rate, which
  Primr also sends as the request price ceiling, was below every provider
  serving the model. It now uses Z.AI's first-party rate.

v1.39.33

Changed

• The README, product contract, agent guide, operator guide, and package
  description now state plainly that Primr is a research and consulting-prep
  tool whose reports mix cited facts from public sources with reasoned
  strategy that companies rarely publish. It is not only a fact checker:
  quality is judged by research grounding, insight, specificity, and
  usefulness, not by how much can be verified.

v1.39.32

Changed

• The v1.40 card now improves the Strategic Overview, AI Strategy, and skill
  pack directly instead of waiting on people adjudicating facts. Strategy
  analysis is reasoned judgment, so changes are confirmed by reading before
  and after outputs side by side; model judges are a second opinion and
  claim-label checks remain diagnostics. The product contract, roadmap, next
  steps, and evaluation plan record the change.

v1.39.31

Changed

• Default models now track each provider's current flagship, checked against
  the xAI, Gemini, OpenAI, and Claude catalogs and the OpenRouter models API
  on 2026-10-28. Grok 4.7 reasons in every tier (MAX also writes with it);
  Gemini 3.8 Flash and Gemini 4 Argon replace the Gemini 3 Flash and 3.1 Pro
  previews; Gemini writing uses 3.5 Flash-Lite; OpenAI-only runs use GPT-6
  Luna and the new GPT-6.1 Terra; Anthropic-only runs use Claude Haiku 5.5
  and Sonnet 5.5; OpenRouter routes use Gemini 3.5 Flash-Lite, GPT-6 Luna,
  and DeepSeek V4.1 Flash; the LLM judge default is Grok 4.7. Fallback chains
  follow the same models. This is an owner decision, not a measured quality
  promotion.
• The default xAI plus Gemini estimate rises from about $0.89 to about $1.24
  with the AI Strategy (base report $0.76 to $1.05). OpenRouter defaults fall
  from about $1.07 to $0.67 and Anthropic-only from $4.71 to $3.20. Always
  use --dry-run for the live quote.

Added

• GPT-6.1 Terra is registered directly and on OpenRouter with its standard,
  long-context, cached, and cache-write rates.
• Estimates disclose scheduled provider price changes. Gemini 4 Argon rises
  to $4/$20 on 2026-12-01 and Gemini 3.8 Flash to $1.50/$7.50 on 2027-01-01,
  so a quote approved before those dates names the new rates.
• A routing test fails when any provider key combination sends a default role
  to a model outside the current flagship set.

Fixed

• The OpenRouter openai/gpt-4.1-mini route is unavailable ahead of OpenAI's
  2027-01-14 shutdown and names openai/gpt-6-luna; Grok 4.1 Fast records
  its 2026-05-15 retirement. Both stay priced for historical accounting.
• Usage summaries, QA model metadata, and capability-routing price hints use
  date-effective rates instead of base rates, and QA metadata now describes
  the configured models instead of Gemini 3 Flash and 3 Pro.
• The CLI help cost guidepost is now checked against the live estimator.

v1.39.30

Added

• The evaluator validation harness gains an authority_without_support
  variant: the supporting span is removed and a strong trust cue is
  prefixed, so a judge that accepts it is scoring source authority instead of
  claim support. Cases may supply an authority_cue; abstaining counts as
  detection by default. The generator version is now 3, so earlier frozen sets
  record which rules built them.

Fixed

• A single-section writer response that opens with prose before a
  differently titled ## heading keeps that heading as a ### subsection
  instead of losing its title.

v1.39.29

Changed

• AGENTS.md is now the single source of agent guidance: the root file holds
  the development contract formerly in CLAUDE.md, and CLAUDE.md is exactly
  @AGENTS.md, the import Claude Code needs to load it. The operate-primr
  guide moved to clients/primr-operator.md; copy it into a workspace as
  AGENTS.md. Its reference links are now absolute. The primr://context
  MCP resource reads AGENTS.md and reports context_map_path: "AGENTS.md"
  (#264).

v1.39.28

Changed

• Client-facing Word reports open with the title and the Executive Summary.
  Recorded collection counts move from beneath the title to one muted closing
  Collection note in every Word path (Markdown conversion, premium builder,
  and text conversion), with unchanged wording and meaning. Terminal coverage
  output is unchanged. This supersedes only the under-title placement from
  #229/#236 (#263).

Fixed

• Writer section envelopes are parsed at the section-parsing boundary instead of
  leaking as report prose. An envelope missing its body tags, or its closing
  section tag after a closed body, is repaired. A truncated envelope keeps its
  text, earns one retry, and warns the operator. An envelope without a title,
  or with nested structural tags, is kept verbatim, the coherence pass is
  skipped so no model rewrite guesses a repair, and the polished DOCX is
  withheld with the Markdown kept for review. Only structural envelope tags
  outside code count, so prose or code that mentions an HTML <title> element
  does not block a report. Repeated title lines and headings around an envelope
  no longer duplicate the section title (#263).
• Single-section writer responses no longer drop later ## blocks silently;
  they are kept as ### subsections. A ## References block is moved last,
  so the existing references policy removes only it.

v1.39.27

Added

• primr doctor warns when OpenAI's free model list reports a past or
  scheduled shutdown_date that the model registry does not record, or a
  different date than it records. The check runs only when OPENAI_API_KEY
  is set, passes the SSRF guard, is bounded to 10 seconds, makes no generation
  call, and never changes routing (#258). The CLI smoke test now blanks
  provider keys so the default suite stays offline.

v1.39.26

Added

• Add a library-only evaluator validation harness for the v1.40 quality card.
  Adjudicated clean cases yield controlled variants for nine error categories.
  Items, pre-registered tolerances, and the generator version are frozen under
  one digest. Judges run blind across seeded passes; recall, false alarms,
  abstention, failures, repeat flips, shared errors, and agreement are scored
  with Wilson intervals. Disagreements, flips, and seeded samples of unanimous
  approvals and rejections are queued for human review. A decision needs an
  all-human-adjudicated set, two judges, two passes, and complete human review.
  It never arms a gate. No judge has been measured with it yet.

Changed

• Take the reviewed ddgs, hypothesis, mkdocs-material, and starlette updates
  proposed in #259 and regenerate both container requirement exports.

v1.39.25

Security

• Constrain locked and container installs to filelock 4.2.1 or newer
  (CVE-2026-112907, CVE-2026-112913, CVE-2026-112940), virtualenv 21.14.13 or
  newer (CVE-2026-115586), and dnspython 2.9.0 or newer (CVE-2026-114605).
  These are transitive packages reached through tldextract, virtualenv,
  pip-audit, and recon-tool; pip consumers resolve their own versions.
• Take the reviewed Python, setuptools, and GitHub Actions updates proposed in
  #254, #255, and #256, and regenerate both container requirement exports.

Fixed

• Route OpenAI-only reasoning to GPT-5.6 Terra, OpenAI's named substitute for
  o4-mini, which shut down on 2026-10-23. o4-mini remains priced for historical
  accounting and now reports its shutdown and replacement before egress.
  Replace the o4mini-flashlite eval slot with terra-flashlite; its cost is
  a planning figure, not a measurement.
• Request gpt-image-2.5-flare for opt-in OpenAI icon generation after the
  gpt-image-1 shutdown, and grok-imagine-image-2.0 at explicit medium
  quality for xAI before the 2026-11-02 retirement of
  grok-imagine-image-quality. Medium keeps the retired slug's intent instead
  of the provider's low-quality redirect. Image model IDs now live in
  config/models.py.
• Record the announced 2027-04-01 shutdown of gpt-5.4-nano with gpt-6-luna as
  its replacement. The OpenAI-only writing default is unchanged until then.
  OpenRouter's openai/gpt-4.1-mini route is unchanged because the gateway
  still lists it.

Changed

• Add start-tidy and stop-tidy contributor rules to CLAUDE.md (#252).