v0.14.1 — benchmark quiz task + Ollama default bump

[0.14.1] - 2026-04-19

Added

• Benchmark harness now understands zrun providers. tools/run-benchmark.mjs
  forwards stack.provider as ZRUN_PROVIDER on spawn so a zrun-ollama stack
  and a zrun-zhipu stack can coexist on the leaderboard instead of collapsing
  into one row. buildRunMatrix also honours per-task scoredRepetitions /
  warmupRepetitions overrides so a diagnostic task can run at a different
  cadence from the 1+3 standard.
• New quiz task + benchmarks/fixtures/quiz/ fixture: a codebase
  comprehension diagnostic that asks the agent to read bundled snapshots of
  zrun/src/tools/index.ts, zrun/src/config.ts, and
  benchmarks/manifest.json and answer four identifier-lookup questions.
  Scored by regex hit-rate per question with an ANSI-stripping
  iteration-log fallback for models that emit answers to stdout instead of
  calling write_file.
• Four new Ollama stacks in benchmarks/manifest.json covering the models
  the quiz was run against: zrun-ollama-llama32-native,
  zrun-ollama-qwen35-9b-native, zrun-ollama-gpt-oss-20b-native, and
  zrun-ollama-qwen25-cline-native. Existing zrun-glm51-native gains an
  explicit provider: "zhipu" for schema consistency.
• tools/quiz-rank.mjs renders the quiz rows from leaderboard.json as a
  ranked markdown table with per-question hit breakdown pulled from
  runs.jsonl notes.

Changed

• Ollama provider default model changed from llama3.1:8b to
  gpt-oss:20b. In the 2026-04-19 codebase comprehension quiz, gpt-oss:20b
  was the only installed local model to score a clean sweep;
  llama3.2:latest, qwen3.5:9b, and qwen2.5-7b-cline all failed at the
  tool-call or response-writing layer. Users who want the older default can
  pin it via providers.ollama.model in zrun/config.json or pass
  --model llama3.1:8b per run. zrun/config.json.example and
  docs/ollama-integration.md updated to match.

Removed

• zrun/config.ollama.example.json. The combined zrun/config.json.example
  covers every shape this file did — an Ollama-only setup is just
  config.json.example with defaultProvider flipped to "ollama" and/or
  the zhipu block removed. Reduces maintenance surface and the chance of
  the two examples drifting out of sync on future changes.

v0.14.0 — zrun multi-provider config with runtime switching

[0.14.0] - 2026-04-19

Added

• ZRun config.json gains a multi-provider shape: both Zhipu AI and Ollama
  settings can live in the same file under providers.zhipu and
  providers.ollama, alongside an optional defaultProvider field. No more
  hand-editing JSON to swap backends.
• Runtime provider selection via four cascading signals (highest first): the
  --provider <name> CLI flag, the ZRUN_PROVIDER env var, auto-inference
  from --model <name> when the model is uniquely owned by one provider's
  registry (e.g. --model llama3.1:8b routes to ollama, --model glm-5.1
  routes to zhipu), and defaultProvider in the config.
• Ollama provider registry now carries a curated list of known model families
  (llama3.1, qwen2.5, codellama, deepseek-coder-v2, gpt-oss, etc.)
  used for --model X → provider inference. The actual installed model list
  is still discovered at runtime from /api/tags.
• Startup banner now shows both provider and model (e.g.
  [zrun] Provider: ollama Model: llama3.1:8b) so it's always obvious which
  backend fired for a given run.

Changed

• ZRunFileConfig (on-disk) is now distinct from ZRunConfig (resolved).
  The normalizer promotes a legacy flat config ({ apiKey, model, baseUrl }
  at top level) into providers.zhipu automatically, so existing setups
  continue to work without any migration.
• API-key-missing error now names the right config path
  (providers.zhipu.apiKey) instead of the deprecated flat apiKey.
• zrun/config.json.example rewritten to the multi-provider shape with
  both zhipu and ollama entries populated as a starting point.
• Root package.json and VERSION aligned at 0.14.0;
  zrun/package.json bumped 0.3.0 → 0.4.0.

v0.13.0 — zrun provider abstraction + Ollama

[0.13.0] - 2026-04-19

Added

• ZRun gains a provider abstraction layer. ZRunConfig now carries a provider
  field (defaults to "zhipu" so existing configs keep working unchanged) and both
  apiKey and baseUrl are optional — each provider declares its own defaults via
  src/providers/index.ts.
• Ollama support via provider: "ollama". Uses Ollama's OpenAI-compatible endpoint,
  health-probes /api/tags with a 5s timeout on both the probe and the model
  listing so a hung local instance can no longer stall CLI startup. When model is
  unspecified, auto-selection prefers the provider's declared defaultModel if
  installed, else sorts alphabetically so repeat runs are deterministic.
• zrun/test-ollama.ts — a read-only /api/tags probe that reports installed
  models and exits non-zero when the server is unreachable or has no models pulled.
• docs/ollama-integration.md — end-to-end Ollama setup, model recommendations,
  troubleshooting, migration from Zhipu AI, and security notes.
• .automaker/features/ records for zrun-provider-abstraction,
  zrun-ollama-integration, and zrun-ollama-docs-testing.

Changed

• src/config.ts loadConfig is now async to accommodate the Ollama health probe.
  The only call site is src/index.ts, which awaits it at startup.
• src/client.ts no longer reuses the provider name as a dummy API key; providers
  that don't require auth (Ollama) get a named NO_AUTH_DUMMY_KEY sentinel so the
  OpenAI SDK constructor accepts a non-empty value.
• Root package.json version aligned to VERSION (0.11.2 → 0.13.0); the two had
  drifted during the 0.12.0 release.

Fixed

• Zhipu default baseUrl in src/config.ts was missing the /api/ path prefix
  (https://api.z.ai/coding/paas/v4 → https://api.z.ai/api/coding/paas/v4).
  config.json.example already had the correct URL, so users who copied the
  example were unaffected; only users relying on the in-code default would have
  hit a 404 at first request.

v0.11.0

Added

• Benchmark harness with fixture manifests, disposable workspaces, cohort comparisons, and machine-readable result artifacts.
• Benchmark documentation covering fixtures, outputs, and report generation.
• New benchmark package scripts for full runs, dry runs, and report-only rebuilds.

Changed

• Codex CLI execution now uses the bash-oriented bypass path to avoid Windows sandbox shell mismatches during unattended runs.
• Feature validation now requires a dependencies array, with [] as the canonical empty value.
• ZRun defaults now allow longer sessions by raising maxTurns to 500.

Fixed

• Prompting and audit guidance now enforce a stronger verification gate before audit findings are converted into remediation work.