v1.3.0 — OpenRouter, live pricing, response cache, Claude 5 + GPT-5.6 defaults

Added

• Per-query cost in API mode (-g / -c): each provider block, the header
  panel and the Completed in footer show the dollar figure derived from the
  OpenRouter catalog and the response's token metrics, and --json gains
  responses.<provider>.cost_usd plus meta.total_cost_usd (providers +
  judge). An unpriceable model is omitted rather than shown as $0.00, and a
  total ending in + means at least one response could not be priced. CLI mode
  shows nothing about dollars because it is subscription-billed. --raw and
  --brief are unchanged.
• --budget <usd> for batch mode (also CONCLAVE_BATCH_BUDGET): stops
  dispatching new items once cumulative estimated spend reaches the cap,
  lets in-flight items finish, and exits non-zero with a summary naming the
  cap, the completed count and the skipped count. Undispatched items stay out
  of the checkpoint so --resume continues the run. Because cost is measured
  post-hoc, overshoot by up to --workers items is expected.
• Opt-in response cache: --cache[=TTL] (24h by default) or
  CONCLAVE_CACHE_TTL=<hours> reuses an identical provider response instead of
  paying for it twice; --no-cache overrides an env-enabled cache. The key is
  a sha256 of the mode, provider, model and the full prompt including file and
  stdin context, so any context change is a miss. A hit is tagged (cached) in
  the progress line and provider block, carries cached: true in --json, and
  costs nothing. Judge synthesis is never cached. Works in CLI and API mode;
  batch mode honours it per item. conclave cache stats and
  conclave cache clear manage the store. ADR-011.
• make check: one gate running go vet, gofmt -l, go test and
  conclave models --check, plus a .github/workflows/check.yml running the
  same on ubuntu-latest and windows-latest.
• Tests for internal/batch, which had none: item parsing (malformed line
  skipped, missing id assigned, duplicate id dropped), worker fan-out,
  rate-limit retry and retry exhaustion, checkpoint resume, cost estimation
  precedence, and the budget stop, plus checkpoint_test.go for
  load/append/corrupt-line handling.
• Runtime pricing catalog (internal/pricing): conclave caches OpenRouter's
  public models feed under the user cache directory, refreshes it in the
  background once per CONCLAVE_PRICING_TTL hours (default 24), and never
  blocks a query on the network once a cache exists. CONCLAVE_NO_PRICING=1
  disables it. ADR-009.
• Model drift warning: when a configured model id is not in the catalog,
  conclave prints one stderr line naming the newest listed alternative and
  proceeds. Suppressed under --json, --raw, -q.
• conclave models [provider] [--check|--refresh|--json|--all] to inspect
  current ids, context sizes and API prices, and to gate releases
  (--check exits 2 when a compiled default is missing, 3 when the catalog is unreachable).
• OpenRouter as an API-mode backend: in -g mode any provider token written
  as an OpenRouter slug (deepseek/deepseek-v4-pro, anthropic/claude-opus-5)
  routes through openrouter.ai/api/v1/chat/completions, with the slug as
  both provider name and model id. Works in the provider list and --judge;
  --all never auto-includes OpenRouter models. Key OPENROUTER_API_KEY
  (env, .env, or OS keyring); conclave init and keyring list know it.
  Preflight checks /auth/key and reports an exhausted spend limit as "no
  credit". Drift warnings, display names and batch cost estimates resolve
  the slug directly in the pricing catalog. CLI mode rejects slash tokens
  (API-only, pay-as-you-go). ADR-010.
• The judge is now resolved (and preflighted) before the panel runs, so a
  judge that cannot be built fails before any provider is paid for. A
  slash-routed judge that is not in the OpenRouter catalog is refused
  outright (--skip-preflight sends it anyway); panel members only warn.
• Provider lists are trimmed ("a, b" works); malformed slugs (/model,
  model/) and a bare openrouter token get specific errors; an
  OpenRouter error envelope inside an HTTP 200 surfaces as an error.

Fixed

• Batch mode no longer discards a result that has already been paid for. A
  worker whose send raced a cancelled context threw the result away, so an item
  that had been queried and billed left no output line, no checkpoint entry and
  no trace it had run. Results are now always handed to the writer.

• An interrupted batch exits non-zero and says how many items were not
  dispatched. It previously exited 0, letting a pipeline read a partial JSONL
  as the complete answer.

• Provider decorators no longer hide a provider's auth check. Embedding the
  Provider interface does not promote the optional Preflighter, so wrapping
  a provider silently skipped its preflight; every decorator now implements
  Unwrap and RunPreflight follows the chain.

• The response cache no longer deletes an entry before replacing it, which
  opened a window where a concurrent reader saw nothing, and it no longer leaks
  a temp file when the replacement is refused.

• --budget warns when it cannot bind: on a non-batch query, where it does
  nothing, and when the pricing catalog is unavailable, where estimates cover
  only the built-in providers.

• Batch mode no longer checkpoints items that a cancellation stopped from ever
  running, which made --resume skip them permanently.

• A batch run now keeps a checkpoint even without --resume, so the resume
  hint printed by the budget-stop and interrupt summaries is actually true.

• Error paths in batch mode now carry the spend they incurred, so a failing
  judge model no longer makes --budget unenforceable.

• --json gains meta.total_cost_partial, marking a total that understates
  the real spend because something could not be priced. The styled output
  already showed this as a trailing +.

• Runtime failures no longer print the whole usage block after the error,
  which buried the batch summaries. Argument misuse still shows usage.

• conclave models --check exits 2 on real drift and 3 when the catalog is
  unreachable, so make check and CI branch on the code instead of matching
  message text.

• --json now reports the timeout actually in force as
  execution.timeout_seconds. The field existed but was never populated, so it
  always read 0 regardless of -t.

• gemini CLI mode: pass -p (gemini-cli 0.58 treats a positional prompt as
  interactive mode and never returns headless) and --skip-trust plus
  GEMINI_CLI_TRUST_WORKSPACE=true (exit 55 in any un-trusted directory).
  When gemini-cli still fails on auth (Google retired the free Code Assist
  OAuth tier it defaults to) and a GEMINI_API_KEY is present, the query
  falls back to the direct Gemini API with the same model. Set
  security.auth.selectedType to gemini-api-key in ~/.gemini/settings.json
  to keep the CLI route.

• Preflight budget raised 2s → 5s; claude/codex cold starts on Windows were
  tripping it. codex login status prints to stderr, which the old check
  never read.

• Auto-init no longer runs when stdin is not a terminal. A subprocess with
  no provider keys visible used to block forever on an invisible prompt and
  look like a 110s+ hang.

• openai CLI-mode preflight asks codex login status before demanding
  OPENAI_API_KEY, so ChatGPT-subscription users are no longer rejected.
  Remediation text for gemini/openai now says which mode needs what.

Changed

• Default models bumped (all verified live 2026-09-08): openai
  gpt-5.5 → gpt-5.6-sol, claude claude-opus-4-8 → claude-opus-5,
  grok grok-4-1-fast-reasoning → grok-4.6 (the only id the grok CLI
  offers), glm glm-5.2 → glm-5.3. Cheap models: grok → grok-build-0.1,
  glm → glm-5.3-flash. conclave models --check now passes clean.
• Batch-mode cost estimates now use live per-model prices from the catalog;
  the hardcoded table in internal/batch/processor.go is demoted to an
  offline fallback (and its gpt-5-nano input price corrected 0.10 → 0.05).
• docs/MODEL_REGISTRY.md refreshed against the 2026-09-08 feed: GPT-5.6
  Sol/Terra/Luna, Claude Fable 5 / 5.1, Opus 5, Sonnet 5, Gemini 3.5–3.8
  Flash, Grok 4.20–4.6 and Build 0.1, GLM 5.3 / 5.3 Flash. Notes that all
  prices are API-mode only. Adds a Drift Watch section: grok-4-1-fast-*
  and glm-4.6v-flashx are no longer listed on OpenRouter.
• Line endings pinned to LF via .gitattributes for .go, .md, .yml and
  the Makefile, so gofmt agrees on Windows and Linux. A fresh worktree on an
  autocrlf machine can still show CRLF until re-checked out; see
  docs/CHECK_GATE.md landmines.
• make check enumerates Go files via go list instead of gofmt -l ., which
  recursed into nested .claude/worktrees/* checkouts and failed on other
  sessions' files.
• Makefile VERSION now tracks the release (it reported 1.1.0 for 1.2.0 code).