v1.1.0

Added

• ollama adapter — local-model backend (internal/ai/ollamacli.go, closes #30).
  teststop now ships a first-class ollama adapter that calls the ollama HTTP API at
  localhost:11434. No subprocess, no API key, no account quota.

  • Default model: qwen3.6:latest (36B Q4_K_M). Override with TESTSTOP_MODEL.
  • stream: false, think: false, num_ctx: 32768 per request.
  • Automatic <think>…</think> block stripping for qwen3 models on older ollama builds.
  • JSON-only output constraint appended to the mandate for local models (local models are
    less instruction-following than cloud models; cloud adapters are unaffected).
  • IsOllamaAvailable() pings localhost:11434 with a 2s timeout for fast detection.

• Auto-detection precedence changed: ollama → claude → copilot.
  TESTSTOP_CLI=auto (the default) now prefers ollama when the local server is reachable.
  Cloud CLIs (claude, copilot) are still fully supported but opt-in:
  TESTSTOP_CLI=claude or TESTSTOP_CLI=copilot. TESTSTOP_CLI=ollama forces ollama only.

  Quality tradeoff (measured, against waymark API project):

  Model	Scenarios	Time	Notes
  qwen3.6:latest	29–31	~3–4 min	High specificity; IDOR, race conditions, token expiry
  gemma4:latest	52	~8 min	Most thorough; covers auth, concurrency, edge inputs
  qwen3:4b	—	—	Not viable; outputs reasoning prose, never generates JSON

  Scenario depth and edge-case creativity from local models are moderately lower than
  claude's output. For production runs where quota is not a concern,
  TESTSTOP_CLI=claude remains the highest-quality choice.

• Resilient JSON parsing for local models (ParseScenariosFromJSON). Local models
  produce sloppier output than cloud CLIs. The parser now uses a three-pass strategy:
  (1) direct unmarshal; (2) extract […] from prose-wrapped output (qwen3:4b pattern —
  reasoning before JSON); (3) sanitize invalid JSON escape sequences then retry (gemma4
  pattern — emits \xNN hex notation which is not valid JSON). Cloud adapter output
  is unaffected (pass 1 always succeeds for claude/copilot).

• Design document (docs/design/ollama-adapter.md). Records the HTTP-vs-CLI decision,
  num_ctx rationale, think: false approach, and JSON-suffix strategy.

---

v1.0.1

Fixed

• claude CLI 2.1.x streaming format (internal/ai). claude 2.1.172 changed
  its --output-format json output from a single JSON object to a JSON array
  of streaming events. parseClaudeEnvelope failed to unmarshal the array into
  a struct, triggering the legacy fallback which fed the raw event array to
  ParseScenariosFromJSON. The parser then silently unmarshaled each event object
  (with no matching field names) into a zero-value Scenario — yielding batches
  of hollow structs with empty title, steps, priority, and confidence_area. Exit
  code was 1 ("review needed") rather than an error, so the defect was silent.

  parseClaudeEnvelope now handles both formats:

  • Single JSON object {…} — legacy path, unchanged.
  • JSON array […] — finds the last event where "type":"result" and returns
    its is_error / result fields. Error detection (rate-limit, auth, refusal)
    works correctly for both formats.

  ParseScenariosFromJSON also gains a hollow-batch guard: if every parsed
  scenario has an empty scenario_id AND title, an explicit error is returned
  instead of the silent zero-value batch. This is defense-in-depth that fires even
  if the envelope parsing falls back to raw output.

---

v1.0.0

Added

• .teststop/config.yaml support. The optional per-project config file is
  now real. Every key maps one-to-one onto an existing teststop run flag.
  Settings resolve with precedence config file < TESTSTOP_RUN_* env var <
  explicit CLI flag. A missing file is not an error; malformed YAML or an
  unknown key fails loudly. See .teststop/config.example.yaml for the keys.

• --ai-concurrency flag (default 1). Caps the number of concurrently
  running AI-mode scenario executions to prevent rate-limit exhaustion. Config
  key ai_concurrency; env TESTSTOP_RUN_AI_CONCURRENCY.

• Structured AI error detection. The Claude adapter now calls
  claude --output-format json and parses the outer envelope, surfacing
  structured errors (rate-limit, auth failure, refusal) with context instead
  of raw stderr. Non-zero exits and envelope is_error: true both map to
  informative error messages.

• E2E pipeline test (test/e2e/). A full reader → mandate → adapter →
  memory → reporter → exit-code integration test using a fake-claude fixture
  script. Runs without real tokens; skippable via -short.

• v1.0 contracts frozen (CONTRACTS.md). Exit codes, scenario JSON schema,
  run output envelope, memory file format, and environment variables are
  declared stable. Breaking changes require a major version bump.

• examples/waymark-demo/. Replayable demo artifact: 52-scenario run
  against the waymark MCP middleware repo, with the captured JSON report,
  Markdown summary, the mandate used, and replay instructions.

Changed

• pkg/scenario/types.go package comment declares the schema FROZEN at v1.0.
• internal/cli test coverage raised from 28.9% → 51.6%.

---

v0.3.1

Changelog

• 30af62a Merge: AI infra errors skipped + neutral cwd (#44 findings 1 & 3)
• 6cb391f fix: AI infra errors are skipped, not failures (#44 finding 1) + neutral cwd (finding 3)

v0.3.0

Changelog

• 8db34df release: v0.3.0

v0.3.0-rc.1

Changelog

• 009f867 Merge: predicted/executed reporting (#42) + concurrency exec mode (#43)
• c24e915 feat: concurrency exec mode (#43) + predicted-vs-executed reporting (#42)
• f2950f1 fix(executor): race winners are strictly 2xx; at-most-one passes

v0.2.1

Changelog

• 8f6df35 ci(docs): promote release tags to stable on manual docs deploy
• bcc77d8 feat(cli): add version command and --version/-v flags

v0.2.0

Changelog

• 483b89e Merge feat/executor-v0.2: dynamic scenario execution (#29)
• 72353de feat(executor): dynamic scenario execution against live systems (#29)
• 065587e fix(executor): emit ExecutionResult duration in milliseconds

v0.1.0

Changelog

• 1f0a3b4 feat(TASK-20260521-102217): bootstrap teststop build to compilable all-tests-green state
• b969e3b feat(phase/0): project infrastructure — go module, scaffold, README, MANDATE
• 23f2f83 feat(phase/1): foundation — scenario types + Cobra CLI scaffold
• f348af7 feat: add Apple Container isolated dev environment for coding agents
• 279aa78 feat: add Claude Code agentic infrastructure for autonomous development
• 64440da feat: implement phases 2-7 — full teststop v0.1 pipeline
• c800121 fix: replace API/SDK design with CLI-based AI adapter throughout
• 039b1ab refactor: remove reader and reporter packages