v0.13.0

Changed

• weld prompt show --raw now displays actual prompt templates instead of stubs
• weld prompt export --raw exports real templates used by each task type
• Research prompt overhauled with code-first philosophy
  • Core principles: "Read code, not docs", "Identify authoritative files", "Eliminate assumptions"
  • New sections: Authoritative Files, Existing Patterns, Integration Points, Constraints & Risks
  • Output requirements: Short artifact (1-2 pages), file:line references required, [VERIFY] markers for uncertain items
  • Includes Memento warning about fabrication without verified context
• Plan prompt enhanced with "Why Plans Matter" section
  • Emphasizes planning as highest-leverage activity
  • Good plan requirements: exact steps, concrete files, validation after each change, obvious failure modes
  • Warning: "Bad plans produce dozens of bad lines of code. Bad research produces hundreds."

Fixed

• weld prompt show --raw now displays actual full prompt templates instead of stub placeholders
  • Previously showed ~15 line stubs with "(Full template in ...)" references
  • Now shows complete templates (47-158 lines depending on task type)
  • Templates imported from actual source modules for accuracy

v0.12.0

Added

• Automatic shell completion installation during make bin-install for bash, zsh, and fish
• Auto-install shell completions on first CLI run for seamless onboarding
• Shell completion documentation with installation instructions for bash, zsh, fish, and PowerShell
• Shell completion for CLI commands:
  • weld interview apply questionnaire argument
  • weld research, weld plan, weld interview generate markdown file arguments
  • weld prompt export --format option
  • weld implement --step option and plan file argument
  • Phase and step number arguments
  • Task type arguments
• run_claude_interactive service function for fully interactive Claude sessions

v0.11.0

Added

• Documentation for weld prompt command to manage prompt customizations
• prompt personalization system for customizing weld command prompts
  • Configure via [prompts] section in .weld/config.toml
  • Global prefix/suffix applied to all prompts
  • Per-task prefix, suffix, and default_focus settings
  • Layered application: global_prefix → task_prefix → prompt → task_suffix → global_suffix
  • Includes example customizations in generated config template
• prompt viewer command weld prompt for managing prompt customizations
  • weld prompt list: Show all task types with customization status
  • weld prompt show <type>: Display customization details for a task type
  • weld prompt show --raw: Output raw template (pipe-friendly)
  • weld prompt show --focus: Preview with specific focus value
  • weld prompt export <dir> --raw: Export all templates as markdown files
  • weld prompt export --format toml/json: Export customizations as config

v0.10.0

Added

• Telegram bot support is now included by default (no longer requires telegram extra)
• Telegram bot extended prompt detection for y/n confirmations, default prompts ([Y/n], [y/N]), Continue?/Proceed?/Apply? questions, and arrow-key menus
• Universal /weld <command> for running any weld subcommand via Telegram
• Document upload handling with automatic save to .weld/telegram/uploads/
• Reply-to-document auto-injection of uploaded file path in commands
• Output file detection with Download button for created files
• Project sync method for Telegram bot to synchronize config projects with database on startup
• Telegram bot: runs table pruning to limit stored runs per user
• Telegram user management commands: weld telegram user add/remove/list
• Auto-prompt to install weld globally during weld telegram init
• --skip-hooks flag for weld commit to bypass pre-commit hooks (useful for Telegram bot or CI)
• Interactive prompt support in Telegram bot with inline keyboard buttons for command options
• weld telegram whoami command to show bot identity and authentication status
• weld telegram doctor command to validate Telegram bot setup and environment
• weld telegram projects add command to register projects for Telegram bot access
• weld telegram projects remove command to unregister projects
• weld telegram projects list command to show registered projects
• weld telegram command group for remote bot interaction
• Telegram bot commands: /doctor, /plan, /interview, /implement, /commit for remote weld execution
• Telegram bot /status command to view current run, queue, and recent history
• Telegram bot /cancel command to abort active runs and clear pending queue
• /use command for Telegram bot to switch between configured projects
• Rate-limited message editor for Telegram with exponential backoff retry
• Telegram message formatting utilities with proper Unicode chunking support
• Per-chat FIFO queue system for Telegram bot to ensure ordered run processing
• Async command runner for Telegram bot with streaming output and timeout handling
• TelegramFileError and TelegramRunError exception classes for more specific error handling in Telegram bot
• File path validation for Telegram bot /fetch and /push commands with path traversal protection
• SQLite state persistence for Telegram bot with user contexts, projects, and command run tracking
• Telegram bot configuration with user authentication and project registration
• Initial Telegram bot integration module structure

Changed

• weld plan default output now saves to same directory as input file with _PLAN.md suffix
  • Example: weld plan /path/to/SPEC.md → /path/to/SPEC_PLAN.md
  • Previously saved to .weld/plan/{filename}-{timestamp}.md
• weld implement --autopilot now respects --no-review flag (previously ignored)
• weld implement --autopilot now displays all active options (autopilot, auto-commit, no-review)
• Telegram support is now included in standard installation (no longer requires weld[telegram] extra)

Fixed

• Telegram bot status output now shows error messages instead of truncating them
• Telegram bot now shows clear feedback after selecting prompt options ("Command continuing...")
• Telegram bot prompt message is updated with final result when command completes
• Telegram bot now persists started_at timestamp when run transitions from pending to running

v0.9.1

Added

• weld plan now accepts multiple specification files
  • Combines all specs into a single implementation plan
  • Usage: weld plan spec1.md spec2.md spec3.md

v0.9.0

Added

• Early input validation with helpful error hints for all commands
  • Validates file paths before starting expensive operations (prevents wasted API tokens)
  • Clear error messages indicate what's wrong (e.g., "is a directory, expected a file")
  • Actionable hints suggest corrections (e.g., "Provide a file path: --output /path/output.md")
  • Applies to: plan, research, discover, review, interview, implement

Changed

• weld init now adds simple .weld/ entry to .gitignore instead of complex pattern
  • Previously used .weld/* with !.weld/config.toml exception to track config
  • Now ignores entire .weld/ directory (config is local-only)

v0.8.0

Added

• --autopilot flag for weld implement command
  • Executes all plan steps automatically without user intervention
  • After each step: runs code review with auto-fix, then commits
  • Stops on first Claude failure
  • Designed for CI/automation and unattended execution

Changed

• --dangerously-skip-permissions now defaults to True for weld plan command
  • Plans are generated without permission prompts by default
  • Use --no-dangerously-skip-permissions to restore previous behavior
• All Claude CLI invocations now use skip_permissions=True for smoother automated workflows

Fixed

• Plan prompt template restructured with explicit WRONG/CORRECT format examples
  • Shows concrete anti-pattern to avoid (bullet-point summaries)
  • Provides complete correct example with all required sections
  • Adds output format checklist as final reminder before generation
  • Explicitly forbids questions, follow-up options, and conversational closing
  • Clarifies CLI context: output goes directly to file, not interactive chat

v0.7.1

Fixed

• weld plan now produces more consistent output following docs/reference/plan-format.md
  • Clearer template structure reduces Claude deviations from required format
  • Complete Phase 2 example ensures all phases get proper Goal/Files/Validation/Failure modes sections

v0.7.0

Added

• --dangerously-skip-permissions flag for weld plan command
  • Allows Claude to explore codebase (read files, search patterns) during plan generation
  • Required when Claude needs file access to create comprehensive plans
  • Matches behavior of weld implement command
• --no-review flag for weld implement command
  • Skips post-step review prompt to avoid Claude CLI date serialization bugs
  • Workaround for "RangeError: Invalid time value" errors during review step
  • Useful when Claude CLI stats cache has issues
• Multiple transcript gist support in weld commit fallback mode
  • Now uploads separate gists for each session that contributed to staged files
  • Example: implement session + review session = 2 gists attached to commit
  • Each gist labeled with command type (implement, review, etc.)
  • Provides complete context for understanding changes
• Smart error recovery in weld implement command
  • When Claude crashes after making file changes, detects modifications and prompts user
  • Allows marking step as complete despite Claude error if work was actually done
  • Prevents losing progress when Claude has internal failures
  • Only prompts if file changes detected; genuine failures still return error

Changed

• Enhanced weld plan prompt with 10 implementation plan rules
  • Monotonic phases, discrete steps, artifact-driven output
  • Explicit dependencies, vertical slices, invariants first
  • Test parallelism, rollback safety, bounded scope, execution ready
  • Rules condensed to minimize context usage while enforcing plan quality
• Optimized weld implement prompt to skip redundant test runs
  • When Claude identifies a step as already complete (but not marked), checks git status first
  • If worktree is clean: marks step complete without running tests
  • If worktree is dirty: reviews uncommitted changes and proceeds without re-running tests
  • Significant time savings when resuming after crashes or re-running completed steps

Fixed

• Critical: Plan generation now strictly enforces required Phase → Step format
  • Added explicit format requirements to prompt to prevent conversational output
  • Added validation to reject plans that don't start with "## Phase" or lack steps
  • Fixed issue where Claude would output summaries/questions instead of structured plans
  • Plans now correctly follow docs/reference/plan-format.md specification
• Critical: weld commit now attaches transcripts from ALL relevant sessions
  • Fallback flow finds all sessions from registry that match staged files
  • Uploads one gist per matching session (e.g., implement + review)
  • Fixes issue where only one transcript was attached when multiple sessions contributed
  • Each transcript gets its own trailer line in commit message

v0.6.0

Added

• Interactive review prompt after step completion in weld implement command
  • Prompts user to review changes with optional auto-fixing via weld review --diff [--apply]
  • Non-blocking: review failures don't stop implement flow
  • Always available (independent of --auto-commit flag)

Fixed

• Interactive menu cursor position in weld implement - now automatically positions on first incomplete step
• Makefile bin-install target now forces package rebuild with --force flag to pick up source changes
• Critical: EOFError handling in auto-commit prompt (now handles non-interactive environments)
• Critical: File I/O error handling in review prompt (disk full/permission errors no longer crash)
• Critical: Exception handling for review prompt generation
• Directory naming for review artifacts (sanitize step numbers with dots, e.g., "1.2" → "1-2")
• Added model parameter to review Claude invocations (respects configured model)
• Removed redundant git status check in review prompt (performance improvement)
• Config validation with safe defaults for review feature (graceful handling of malformed configs)
• Result validation for empty Claude output in reviews
• Consistent error messages across all non-blocking failures
• Security documentation for skip_permissions behavior in review auto-fix mode