@generata/core@1.2.0

Minor Changes

• b149d58: Brand WorktreeConfig so isolation requires the worktree() helper. The previous structural type let raw object literals assign through; the brand forces consumers to use the helper or omit the field.

• 95fd6ca: Chain-builder workflow API + typed factory agents. defineWorkflow now returns a fluent builder (.step().step().build()) and defineAgent accepts a factory (inputs) => config that types the agent's input contract. Workflow→agent wiring typos and forward-step references error in the editor; the engine no longer auto-leaks prior step outputs to agents (only what a stepFn explicitly maps reaches the agent). Breaking for code on the previous defineWorkflow({ steps: [...] }) shape — pre-1.0 surface, no published consumers yet, so a minor bump.

• d16de2b: Add shorthand for running workflows: generata <name> is now equivalent to generata workflow <name>. The first positional argument is treated as a workflow name unless it matches a reserved subcommand (init, add, agent, workflow, run, validate, metrics, skills, help). The longer form keeps working unchanged.

• 56082cf: Short-circuit critic retry on STATUS: halt and narrow Tool enum. The engine now breaks the critic-rejection retry loop when the worker reports a structural halt, since retrying cannot resolve a spec/plan-level conflict. The Tool enum drops "read", "glob", "grep" (no-op tokens that were never wired to the runner); shipped templates are updated to declare only the tools they use.

• cc1687a: Remove agent summaries feature. The post-agent humanize step added latency and cost to every run for a recap most users skipped, and the agentSummaries config option is now gone.

• a7b38c1: Remove the supervisor agent type. No shipped template used it, and the dynamic-workflow-generation path it powered added complexity to the agent CLI without a corresponding consumer. defineAgent({ type: "supervisor", ... }) is no longer accepted.

• bc019d7: Rename agent promptTemplate field to prompt. The shorter name reads naturally and matches the field's role - it's the prompt string the LLM sees, not a "template" in any generative sense (the factory's closure handles interpolation). Breaking for any code that references agent.promptTemplate or sets it on a defineAgent literal; pre-1.0, no published consumers, so a minor bump.

• c4d49df: Redesign the run header: rainbow generata banner with an italic tagline, a 7d · … weekly metrics line (new showWeeklyMetrics config, default on) that compares against the prior 7 days, and the prompt log path shown in-header when logPrompts is on. Prompt log files now live at <logsDir>/<kind>/<basename>-<runId>.log with collision-aware fallback. Agent type colours moved to cyan/magenta/orange so they no longer clash with status colours.

• 372554b: Show isolation mode (local or worktree: <path>) in the workflow start header so the run environment is visible at a glance and the worktree location is discoverable. Adds an optional isolation parameter to logWorkflowStart and exports a new WorkflowIsolation type.

• 45d9593: Add showPricing config option (default false). When off, runtime logs and notifications hide USD costs and show token counts instead. The generata metrics subcommand still surfaces cost as before. Set showPricing: true in defineConfig to restore the previous behaviour.

• d24a3b1: Typed outputs, first-class halts, and surgical bin permissions.

  • New outputs: Record<string, string> field on agents (key → LLM-facing description). Engine wires a per-agent emit bin with surgical Bash(<bin>:*) permission, parses the captured values, and merges them into the runtime params bag. Chain builder threads the literal output keys through TBaseParams so downstream stepFns destructure them with full type-safety.
  • First-class halts: agents call --halt "<reason>" via the emit bin to stop the workflow cleanly (no metric failure, downstream steps skipped, haltReason set). Replaces text-sentinel patterns (STATUS: halt, NO_ITEMS, etc.).
  • Factory-form onReject: StepOptions.onReject accepts a typed stepFn (params) => StepInvocation with the same contextual typing as .step(). Wrap factories in a stepFn to use them as rejection handlers.
  • Internal: verdict/params bin permissions tightened to surgical Bash(<bin>:*) patterns (no more broad Bash injection for critics/planners).

• f308457: Add verboseOutput config flag. Workflow runs now show a per-step spinner with tagline by default; set verboseOutput: true to restore the inline tool-event stream. Configs scaffolded by generata init opt in.

• a869be2: Add isolation: "worktree" opt-in to defineWorkflow. When enabled, the workflow runs against a fresh git worktree created from origin/main, while logs, metrics, and a configurable list of sharedPaths symlink back to the main checkout. The worktree is pruned at run end regardless of outcome. New CLI: --worktree / --local runtime overrides on generata workflow <name>, and a generata worktree prune recovery subcommand for orphaned worktrees.

• 0898ef6: Add baseRef to WorktreeConfig so workflows can override the default origin/main base for git-worktree isolation. A <remote>/<branch> value triggers a fetch first; a bare branch like "main" is treated as a local ref and used without fetching.

Patch Changes

• 1407041: Print help for bare generata invocation instead of Unknown command: undefined. New users running the CLI with no arguments now discover the available subcommands.
• 31a7774: Add test guidance to the coding template's markdown-slide-deck starter idea in NOTES.md. The seed now spells out the Ink testing approach (use ink-testing-library with lastFrame() and stdin.write rather than driving the built binary via expect/script) and requires a sample deck at examples/intro.md that doubles as the test fixture and README demo, exercising every splitter rule (--- separators, top-level # headings, fenced code blocks for cli-highlight, lists, and inline emphasis).
• 6a7fde7: Refresh the coding template's starter ideas and fix recency-biased idea selection. NOTES.md now ships with three modern, immediately-runnable TypeScript seeds (a terminal weather card via Open-Meteo, a markdown slide deck built on Ink, and a Carbon-style code screenshot generator using shiki + sharp) instead of the previous five utilitarian CLIs. The build-project workflow now seeds a random integer into spec-creator, which picks the unbuilt idea at random_pick mod N rather than the LLM's subjective notion of "most compelling" - removing the bug where the agent reliably chose the last item in NOTES.md. The post-install message in the manifest is updated to match.
• 72661af: Enforce tools[] restrictions for full-permission agents. Previously the runner ignored the tools[] array under permissions: 'full', silently disabling any declared restriction; it now emits --allowedTools alongside --dangerously-skip-permissions.
• 08feeae: Throw EnvProfileError from runWorkflow instead of calling process.exit(1). Lets CLI, tests, and library consumers decide how to handle the error.
• 3923f4c: Mark failed workflow steps with ✗ in the per-step CLI output. Previously a step whose agent exited non-zero still rendered with a green tick, even though the workflow summary correctly reported FAILED.
• dbf8a37: Fix workflow step output printing ↳ undefined <model> instead of the agent name. The registry's TypeScript loader was creating a fresh module graph per call, so an agent imported transitively by a workflow file was a different object than the one loaded directly into the registry - the name mutation only landed on one copy. Switched to Node's regular import() so both references share the ESM cache.
• 8f3b029: Run oxfmt across the codebase to baseline formatting against the formatter's defaults. No functional change; pure source reflow.
• 0a7d248: Show raw token counts under 1000 in logs instead of rounding to "0k tok". Small steps now report e.g. "543 tok"; counts ≥ 1000 keep the existing "Nk tok" format.
• c5daf72: Render engine bin invocations (emit, verdict, params) in plain English in verbose stream output instead of raw Bash: /abs/path/bin/... --flag "..." lines. Now shows e.g. Halted with reason: "...", Verdict: approve, Outputs emitted: spec_filepath="...", so it's clearer what the agent actually did.
• 049eb77: Colour only the ✓/✗ glyph in step-done logs. Id, duration, tokens, cost, and model now render in the default colour to reduce visual noise.
• a4530f9: Replace hardcoded /tmp paths with os.tmpdir() for cross-platform portability. Engine and agent-runner now work on Windows and systems with non-POSIX temp directories.
• ee7a67d: Harden subprocess handling for hung agents and missing critic verdicts. Adds a SIGKILL backstop 10s after a timed-out agent's SIGTERM so a non-responsive Claude CLI can no longer block the parent run, and retries the critic step (up to maxRetries) when it returns no verdict instead of halting on the first transient miss.
• 9055c29: Fix [object Object] rendering in generata help templates. The catalog format moved to { url, subdir } objects but the help command was still typing entries as plain strings; now destructures the fields and prints <url> (<subdir>).
• 8b99c85: Fix: flatten nested workflow names in slash command generation. Workflows under nested paths now generate commands with just the basename instead of namespaced paths, and collision detection prevents duplicate command names.
• e6543ad: Prevent sub-agents from recursively invoking their own workflow. Generated slash commands matching an agent's task description (via generata skills sync) combined with global "always invoke skills" directives caused agents t...

@generata/core@1.1.1

Patch Changes

• 07f2c36: Fix generata init failing with ERR_MODULE_NOT_FOUND for tsx when run via pnpm dlx from a directory that doesn't have tsx in scope. The bin now resolves the tsx loader via an absolute file URL.

@generata/core@1.1.0

Minor Changes

• 28ae9d2: - Removed @generata/sws and @generata/stock from the catalog (templates.json); only @generata/coding exists today.

  • Coding template manifest: dropped WORKDIR from requiredEnv (the working directory is now set in generata.config.ts, no longer prompted as an env var). Tidied bin hints and rewrote postInstall to match the current init flow.
  • Coding README: updated the env table and added a note about configuring workdir in generata.config.ts.

• d0792d8: @generata/coding template overhaul: replaced the 13-agent / 4-workflow pipeline with a single spec-driven build-project workflow built from 8 flat agents.

  • New flow: dream (spec-creator) -> plan (plan-creator) -> audit (plan-reviewer, retries plan up to 2x with feedback) -> execute (code-writer) -> verify (code-reviewer, archives the project on reject) -> readme -> tidy (plucks the used idea from NOTES.md).
  • Each project is self-contained under projects/<plan_name>/ with SPEC.md, PLAN.md, README.md, and code as siblings. The legacy code/ subdir convention is gone.
  • Reject path archives the failed project to projects/_archive/<plan_name>/ with a generated REASON.md.
  • NOTES.md ships pre-populated with five starter ideas so pnpm generata workflow build-project works on a fresh init.
  • Dropped: Cloudflare deploy, git committer, plan interview, ref enrichment, and the execute-plan / daily-plan / dream-and-build / deploy-project workflows.
  • Manifest stripped: no git / wrangler / Telegram requirements; only claude is needed.

  Breaking for existing users of the coding template - re-init to pick up the new pipeline.

• ac65068: Drop the name field from defineAgent() and defineWorkflow() - names are derived from each file's path relative to agentsDir. Workflows can now live anywhere under agentsDir; the workflowsDir config option has been removed. The CLI accepts either the canonical name (core/plan-dreamer) or just the basename (plan-dreamer) when the basename is unambiguous.

• e2dd9b7: generata init now detects which package manager invoked it (via npm_config_user_agent) and uses that PM for the install step in fresh projects. So npx @generata/core init ... runs npm install, pnpm dlx @generata/core init ... runs pnpm install, and so on. Existing projects with a lockfile still get matched against that lockfile.

• 9f42508: generata init now runs in non-empty directories. Existing files are preserved; template-file conflicts error unless --force is passed. Removes the previous "use 'generata add'" loop where users with existing files couldn't init or add.

• 184d89f: generata init now writes a default generata.config.ts if one doesn't already exist in the destination. Previously, init scaffolded agents/, package.json, .env, and slash commands but no anchor file, so subsequent commands like generata help workflows would fail with "No generata.config.ts found". The default config sets sensible Claude model tiers and points workdir at the destination directory; users can edit it freely. Existing config files are preserved.

• 0bd93ab: Make workDir optional in defineConfig. loadConfig now back-fills it from the directory containing generata.config.ts, so user configs no longer need to repeat the path.

• 70d6533: Added two new catalog templates alongside @generata/coding:

  • @generata/starter - bare-minimum scaffold (one worker agent, one workflow). Designed to be edited or thrown away as users build their own pipeline. Good for learning the model without inheriting a use case.
  • @generata/standup - daily standup generator. Reads yesterday's git activity and drafts a 3-section update (yesterday / today / blockers). Two agents demonstrating two-step composition.

  Both appear in generata help templates and generata init (no args) listings.

• c47b493: Template README.md now lands at README-<alias>.md in the destination instead of overwriting the user's own README.md. Alias is derived from the manifest name (@generata/coding → README-coding.md). Multiple templates can now coexist in one project (e.g. via add) without README conflicts. Templates can still override the path explicitly via installPaths in their manifest.

• 3a0bbe8: Workflow runs now print the final agent's text output below the step-done line, mirroring how single-agent runs (generata agent ...) display their result. Skips empty output, the interactive-session placeholder, and critic last-steps (whose verdict summary already prints). Applies to both static workflows and supervisor-generated workflows.

Patch Changes

• 7d6904a: fixed release action

• 1a99693: update github workflows

• 7d1d6ee: Rename the workdir field in generata.config.ts to workDir to match the camelCase convention of the other directory fields (agentsDir, workflowsDir, metricsDir, logsDir). The internal work_dir Jinja template builtin is unchanged.

  Breaking: existing config files must rename workdir: to workDir:. The init scaffolder now generates the new spelling.

• 611c893: generata init no longer crashes when the template scan can't import an agent or workflow file (typically because a fresh template clone has no node_modules). Files that fail to load are skipped with a single summary line; the workflow precheck still catches missing env vars at run time.

• f657c93: generata init no longer writes a 0-byte .env.example when the template declares no requiredEnv and no scanned agent declares envKeys. Prints a one-line note in its place so the step output stays accurate.

• cc3f90f: init and add no longer treat identical files as conflicts. If the destination file already has byte-identical content to what the template would write, the copy is skipped silently and not counted as a conflict. Only files whose content differs trigger the existing "re-run with --force" error. Re-running init/add against an unchanged template is now idempotent.