ProseForge Agent v0.4.1

Agent runtime closure hotfix for ordinary users.

Fixed

• pf-agent setup --quick now writes the actual provider profile key names
  used at runtime, including DASHSCOPE_API_KEY, ARK_API_KEY,
  ZHIPU_API_KEY, and XAI_API_KEY.
• Provider resolution now honors .pf-agent/config.yaml model and
  api_key_ref overrides instead of always using bundled profile defaults.
• pf-agent service check now reports missing real provider keys instead of
  claiming every requested provider is configured.
• Agent tool calls such as provider.configure no longer report fake success;
  unavailable tool actions return clear next steps.
• The product TUI no longer prompts configured users to run setup again, and
  placeholder workflow packs are marked as coming soon.
• POSIX installer PATH persistence and PowerShell installer command execution
  are safer and less likely to mislead new terminal sessions.

ProseForge Agent v0.4.0

General Agent Runtime release.

Added

• pf-agent is now positioned as a general personal Agent CLI Runtime with
  ProseForge Novel available as the first built-in workflow pack.
• The product TUI shows setup, provider/model, workspace, active workflow, and
  next actions instead of a novel-only or blank demo surface.
• pf-agent workflow list and pf-agent workflow use <name> expose the
  workflow pack model.
• pf-agent chat / pf-agent run now build a shared Agent Kernel with default
  tools, intent routing, session storage, injection safety, and project
  retrieval where a project is bound.
• ProjectChatRetriever provides deterministic project evidence retrieval from
  manifests, bible files, timelines, relations, foreshadowing notes, chapters,
  scene artifacts, and existing RAG chunks.

Changed

• pf-agent run now resolves the requested or configured provider instead of
  always constructing the fake provider. Real providers without keys return
  setup guidance instead of silently falling back.
• Offline runs explicitly require fake/local provider behavior and reject
  explicitly requested network providers.
• chat --show-citations now reports real project evidence refs when project
  files exist.
• Legacy TUI and REPL fallback paths now receive the same default tools,
  retrieval, session, and safety wiring as the main chat surface.

Fixed

• Intent router tool targets are now aligned with the default tool registry,
  including provider.configure.
• Missing tool registries now return a traceable error instead of crashing with
  a NoneType tool execution failure.
• Tool permission errors now show both required and current permission levels.

ProseForge Agent v0.3.2

First-run UX and installer repair release.

Added

• pf-agent setup --quick is now an interactive guided setup path for
  normal users. It prompts for provider, API key, model, language/persona,
  workspace, and optional ProseForge root, then writes keys to
  .pf-agent/.env while keeping config.yaml free of raw secrets.
• Startup dotenv loading now includes .pf-agent/.env, so keys written by
  setup are available to providers on the next command.
• The full-screen TUI now renders a welcome panel on an empty transcript
  with provider status, setup guidance, and common commands instead of a
  mostly blank terminal.

Changed

• The one-line installers now install or upgrade the PyPI tool by default,
  put the manager bin directory first on PATH for the current and future
  shell sessions, run pf-agent doctor by absolute path, then launch
  pf-agent setup --quick.
• Installer skip switches were added for automation:
  -NoSetup, -NoLaunch, -NonInteractive on PowerShell and
  --no-setup, --no-launch, --non-interactive on POSIX.
• TUI message prefixes now use You:, Agent:, and Error: for a clearer
  modern agent CLI transcript.
• README and README.zh-CN now document the guided first-run installer path
  and the v0.3.2 release.

ProseForge Agent v0.3.1

Terminal UI default hotfix.

Fixed

• Bare pf-agent now launches the terminal UI by default after first-run
  bootstrap. In an interactive terminal this opens the full-screen TUI; in
  non-interactive pipes and tests it falls back to the line-based TUI. The
  explicit pf-agent chat command remains available for scripted chat turns.
• Help text and both READMEs now describe pf-agent as the TUI entry point
  instead of the old chat REPL, matching the v0.3.x user-facing promise.

ProseForge Agent v0.3.0

Onboarding + hygiene batch. Two-user pain points from v0.2.x feedback:
"is this command real or a scaffold?" and "how do I taste value in 60
seconds?"

Added

• Readiness badges on every operator report (Report.readiness,
  values ready | contract | planned). Non-ready reports render with a
  visible [contract] / [planned] badge in the title on markdown and
  terminal output, and expose readiness on JSON. The 106 dispatch
  sites that go through _planned_report all inherit "planned" for
  free; gateway slack/discord/signal/email/whatsapp check, remote
  execution backends (docker/ssh/singularity/modal/daytona), and the
  fake skills-hub search now emit "contract" explicitly.
• --strict global flag — refuses to run any command whose report
  is not readiness="ready", exiting 3 with a one-line pointer at the
  README readiness table. Placed next to --offline.
• pf-agent quickstart — first-60-seconds offline loop. Delegates
  to the existing DemoRunner, walks intake → phase plan → daily
  workbook → chapter draft → review → export dry-run → memory candidates
  → closeout → report pack, all with the fake provider, no API keys.
  Prints a numbered checklist and every artifact's path.
• Doctor onboarding check — warns on empty workspace with
  Run pf-agent quickstart for the 60-second offline tour.
• flake-hunt scheduled workflow — weekly pytest --count 10 -q
  on ubuntu × py3.11; opens an issue on failure. Uses pytest-repeat.
• integration pytest marker — labels the ~35 tests that touch a
  real subprocess, real HTTP over urllib, or a real sqlite file, so
  they can be run in isolation via pytest -m integration. New
  CI job runs the integration subset separately on every PR after the
  main matrix passes.
• Readiness table in both READMEs — replaces the earlier bullet
  list with a Command | Status | Notes table so a future PR can update
  one row instead of editing prose.

Changed

• CI now uses [ci] extras — pytest-rerunfailures + pytest-repeat.
  The main test job passes --reruns 2 --reruns-delay 1 to mask the
  handful of remaining time-sensitive tests while their causal
  rewrites land. Local pytest -q semantics unchanged.
• Two flaky tests rewritten causally —
  test_plugin_sandbox_timeout_returns_before_callback_finishes and
  test_run_returns_timeout_when_func_blocks now assert on
  Event.is_set() / a callback_finished flag rather than wall-clock
  ceilings.
• Doctor wiring wording — remote gateway and remote execution
  bindings now say "contract-only — see README readiness table",
  matching the new badge terminology.

ProseForge Agent v0.2.0

First public release.

Added

• Real stdio MCP transport: MCP servers now run as actual subprocesses
  speaking JSON-RPC over stdin/stdout (previously placeholder-only).
• Real local execution environment: the local backend executes commands
  via subprocess under the existing permission ceiling (previously dry-run
  plans only).
• Real Telegram gateway transport: long-polling getUpdates and
  sendMessage over HTTPS using the shared stdlib HTTP transport.
• Local HTTP service on port 8765 (stdlib http.server) exposing the
  agent service API (/health, /chat), wired into the Docker
  entrypoint so EXPOSE 8765 is meaningful.
• Release automation: tag-triggered GitHub Actions workflow that runs the
  test gate, builds sdist/wheel, publishes to PyPI (trusted publishing),
  pushes a Docker image to GHCR, and creates a GitHub Release with
  install scripts attached.

Changed

• Version bumped to 0.2.0.
• License declaration unified to Apache-2.0 across README, packaging
  metadata, and the LICENSE file (previously inconsistent).