v0.4.0

Added

• Published /evals entry point for device evaluation with runtime context/token settings,
  fresh generation factories, cancellation, metadata, and per-attempt reports. The example
  includes an explicit Apple evaluation function using disposable fixture tools.

• Native CI compiles the Foundation Models metrics for iOS 26 device and simulator targets,
  then builds the complete Expo simulator app without signing.

• Apple compatibility guidance covering runtime budgets and provider boundaries for iOS 27.

• Standalone Expo starter export and tarball commands sourced from the example, with pinned SDK
  dependency, local assets, native context module, and customization instructions.

• CI and release-preflight checks that scaffold a fresh app outside the repository, install the
  packed SDK, and run typechecking, tests, Expo Doctor, and an iOS bundle.

• Public npm starter template packaging and a release job publishing the template after the
  matching SDK version, with a documented one-time trusted-publisher setup.

Fixed

• Install the packed SDK directly in starter smoke checks without resolving its unpublished
  release version from npm first.
• Classify wrapped provider codes for unavailable models and context overflow separately;
  preserve on-device guardrail handling, including refusal codes without English messages.
• Preserve earlier evaluation failures when the last run passes. Aggregate device/cloud counts
  and mean steps across every attempt; retain execution errors and reject invalid repeat counts.
• Clarify runtime-dependent context capacity and the limits of scripted evaluation results.

v0.3.0

Fixed

• Classify native unsafe-content refusals, including wrapped errors, as guardrails without
  retrying or sending them to fallback; recognize native Generable decode failures in the
  existing bounded structured-output validation paths.
• Retain nonempty best-effort replies and tool evidence in conversation memory, marked partial
  in subsequent prompts and memory compaction.
• Show cause-specific unfinished outcomes in the example, including refusal, repeated-step,
  model, and validation failures; avoid success icons and misleading smaller-request advice.

Added

• Standalone interactive docs/harness-walkthrough.html explaining the base harness, model/code
  boundaries, lifecycle hooks, token budgets, storage, and screenshot failure paths.

• Scripted provider errors via fakeModel([{ error: new Error(...) }]) for recovery regression tests.

• Optional @hellohelen-ai/goliath/filesystem module with in-memory, static, SQLite, and composite
  backends; scoped backend factories; bounded glob/grep/read tools; and revision-checked writes.

• Token-budgeted literal tool output via outputMode: "content", configurable per-result and
  active retrieval limits, and explicit approval requirements for opted-in writing tools.

• Example file tools with read-only sample documents, persistent shared notes, conversation-local
  scratch space, approval cards, and file suggestions; filesystem guide and API reference.

• Conversation-aware agent.run(text, { conversationId }), with isolated memory, request queues,
  context-window tracking, and fallback counts; existing calls use a separate default conversation.

• Memory factories for persistent history per conversation, per-run approval callbacks, and
  isSessionFallback(conversationId); hooks and tools receive conversation IDs outside prompts.

• Local SQLite persistence in the example: conversations, drafts, transcripts, approval decisions,
  and compacted agent memory survive restarts, with interrupted requests marked for explicit retry.

Changed

• Example task completion selects an ID from listTasks instead of matching titles; reject
  unlisted IDs before approval and expose IDs alongside titles in model-facing lookup results.
• The example exports one shared agent from agent.ts; Goliath manages conversation sessions
  while Zustand holds UI state and the runtime adapter handles approval buttons and cancellation.

v0.2.0

Added

• Optional outputSchema on createAgent or individual run calls for guided structured
  device answers with inferred RunResult<T>.output, schema-aware input budgeting, and one
  validation retry. Cloud fallback and extension stops remain text-only.
• A dark conversation UI for the example, with an inbox, search, suggested requests, inline
  task confirmations, and per-conversation agent memory.
• Console lifecycle logging in the example app, with Fast Refresh for the local harness source
  and development client support for iOS simulators and devices.
• Example UI split into hooks and focused screen components, with mock-tool suggestions and
  the shared docs stone logo.
• Zustand state for example conversations, drafts, navigation, and search, with coverage for
  background replies, concurrent sends, and conversation filtering.

Changed

• Separated example agent execution from conversation state, with ask/approve/cancel in
  an independent runtime and a small conversation hook connecting results and approvals to Zustand.
• Organized the example agent under src/agents/goliath/, with separate configuration, React
  integration, lifecycle logging, and task tools split into individual modules with shared mock data.

v0.1.0

This release contains a breaking rename; see Changed.

Added

• Typed lifecycle extensions with ordered async hooks for run start, recall, planning, tools,
  fallback, answers, memory, errors, and finalization. Includes per-run application context,
  private extension state, tool denial, explicit stops, and extension error diagnostics.
• Optional countTokens and async window callbacks, plus model factories for fresh generation
  contexts. Budget events identify native/provider counts versus estimates.
• A local Expo module for Apple context capacity and iOS 26.4+ token counting. The example and
  its CI checks now use the local library build.
• resolveInput for deterministic reference-to-ID handoffs, validated before confirmation.
  Full JSON-serializable tool outputs are retained in recent exchange records for application code.
• Per-call context checks and budget trace events for planning, argument generation, answers,
  and memory. Requests that still cannot fit escalate as context-budget before generation,
  or reject with GoliathBudgetError when lifecycle extensions are configured.
• An example Expo app in example/, typechecked in CI so it cannot drift from the public API. It
  is not part of the published package.

Changed

• Breaking. createGoliath is renamed to createAgent, and the returned Goliath type is
  renamed to Agent. The configuration object and its options are unchanged.
• Cap output tokens at every stage and include structured-output schemas and provider headroom
  in input budgeting. Disable automatic SDK retries; the harness still retries invalid plans
  and empty answers once.
• Apply the 600-character tool-result cap to custom toModelOutput formatters too. Applications
  that previously returned longer strings should filter or paginate their results.
• Estimate non-ASCII text conservatively and clip memory with the same estimator.

Fixed

• Generated tool arguments reuse the AI SDK's validated output so schema transformations run
  once; no-argument calls and extension-provided replacements still receive validation.
• Application and extension errors no longer trigger model-error fallback; fallback failures
  reject without a second handoff attempt.
• Session fallback now emits a complete trace and saves memory through the shared lifecycle.
  Model-error fallback retains the previous summary and the latest three exchanges without
  calling the failed device again; older exchanges are dropped on that route.
• Cloud answers emit answer events, and saved summary limits include the token estimator's margin.
• Recent exchanges now reach device prompts, including their completed/skipped action records.
• Duplicate writes are blocked before confirmation and execution. Object key order and argument
  mutation cannot evade the check. Read caches are invalidated after writes, and declared tool
  prerequisites must have succeeded before execution.
• Serialize turns on each instance. Failed memory saves and result formatters cannot erase a
  completed action, and failed fallback calls are no longer invoked twice.
• Planner compaction can no longer send a prompt that remains over budget. Final answers,
  including best-effort answers, also compact older step results when needed.
• Failed memory generation no longer replaces a completed answer with fallback. Preserve the
  previous brief and latest exchanges, emit memory-error, and avoid reusing a failed device
  session to summarize a cloud answer.

Deprecated

• The unused compressors option. Use afterTool and beforePlan extensions instead.

v0.0.3

Changed

• Build with TypeScript 7. Emitted JavaScript is byte-identical to the 5.9.3 build; only
  declaration property order changes.

v0.0.2

No changes to the library. This is the first release published from CI, and it exists to prove
that path works end to end.

Added

• CHANGELOG.md now ships inside the package.

Changed

• Releases are published from GitHub Actions over OIDC trusted publishing and carry a provenance
  attestation. No npm token exists for this package. Verify with npm audit signatures.