v0.3.2

Fixed

• Filter injected runtime context that was previously bypassing the final reusable prompt boundary for plain string history content and prompt-history text values.
• Make CLI stdin guarding idempotent across repeated runCli calls and avoid double stdin unref on help.
• Bound history discovery and scanning with deterministic caps and warning diagnostics for truncated roots or oversized files.
• Align the agent-discovery contract documentation with the current same-window handoff model in MVP.
• Tighten existing-skill duplicate scoring to reduce false positives from short generic skills suppressing broader candidates.
• Add a decision spike for bounded history cache/index strategy, with privacy and opt-in guidance before any implementation.

v0.3.1

Fixed

• Guard history timestamp conversion in Claude and Codex parsers against out-of-range finite
  values so malformed records are skipped without crashing scanning.
• Surface skill validation warnings in interactive output and report agnix validation failures
  as structured validation errors without throwing.
• Add integration coverage for local fallback generation, ecosystem mirroring, single-target writes, and overwrite-denial.
• Add CLI help output for --help, -h, and help.
• Emit prompt-dump source summaries and non-info diagnostics to stderr, keeping prompt rows on stdout.
• Show interactive history warning diagnostics from discovery and scanning so malformed records and skipped files are visible before ranking.
• Ask for extra history sources even when defaults are already discovered, and merge extra scans without dropping defaults.
• Continue history discovery when one source directory cannot be traversed.
• Install SIGINT/SIGTERM handlers only once in the CLI runtime.
• Enforce selected skill name validation so generated SKILL.md names must match the chosen skill target.
• Restore invalid primary skill writes by removing newly created targets or restoring overwritten targets after validation failure.
• Add unit coverage for release note extraction CLI behavior, including missing sections and empty release entries.

v0.3.0

Added

• Add an agent-led discovery handoff that opens the selected local agent with session/history paths, asks it to present a Markdown candidate table with an opinionated recommendation, and lets the user choose and implement skills inside that same agent window.

Changed

• Preserve Ritual's local repeated-workflow ranking as a fallback when agent discovery is declined, unavailable, or exits unsuccessfully.
• Allow agent discovery to inspect readable session/history paths even when Ritual's local prompt extractor finds no user prompts.
• Keep discovery itself file-free until the user approves a skill implementation inside the agent window.
• Require agent discovery to account for existing project/global Claude and Codex/agents skills before proposing new candidates.
• Require the agent to ask whether selected skills should be installed project-local to the current command path or global under the user's home directory before writing files.

v0.2.0

Added

• Add ritual prompts and ritual --prompts to dump the latest extracted user prompts in descending date order, with --limit/-n support for changing the default 100-prompt count.

Changed

• Scan Claude project transcripts plus Codex active and archived session transcripts by default, while preserving prompt-history file discovery and honoring CLAUDE_CONFIG_DIR and CODEX_HOME.
• Prefer the most recent matching prompts in workflow previews so newly repeated prompts are visible when reviewing candidates.

Fixed

• Exclude newer Codex # AGENTS.md instructions context records from user-prompt extraction.
• Skip slash commands, low-signal acknowledgements, skill-call context, structured payloads, terminal transcripts, standalone attachments, injected runtime events, local page checks, rendered output dumps, app logs, risk reports, CI log dumps, generated handoffs, assistant completion summaries, and assistant review reports during prompt extraction, and deduplicate mirrored prompt-history/session records before ranking or dumping prompts.

v0.1.1

Fixed

• Run the CLI when package managers invoke the ritualai binary through a .bin symlink.

v0.1.0

Added

• Existing-skill duplicate detection that skips recurring prompt candidates already covered by project-local or global Claude/Codex skills.
• Initial product requirements for Ritual as an interactive CLI that turns repeated Claude and Codex prompts into reusable skills.
• Technical specification for the TypeScript, Biome, typecheck, Bun package-manager workflow, npm registry distribution, and tag-driven GitHub Actions release architecture.
• Production TypeScript CLI scaffold with strict ESM build output and Bun-managed ritual executable.
• Local Claude and Codex JSON/JSONL history discovery, resilient user-prompt extraction, and source diagnostics.
• Local prompt normalization, lexical clustering, recurrence ranking, candidate selection, and two-match fallback handling.
• Project/global scope selection, Claude and Codex/agents target resolution, safe skill-name sanitization, and overwrite protection.
• Explicitly approved claude or codex drafting with an embedded versioned skill-generation prompt.
• Direct skill writes to selected target paths with built-in SKILL.md validation and optional agnix.
• Unit and integration tests covering parsers, ranking, path safety, validation, and the full fixture-driven session flow.
• GitHub Actions CI and tag-driven Bun release workflow.
• Release-note extraction from CHANGELOG.md for GitHub Releases created by CI.

Changed

• Launch Claude or Codex as an inherited terminal session for skill drafting, passing the generated skill prompt as the final argv argument and validating the draft file the agent writes.
• Simplified the skill generation flow by removing the draft workspace, global-write confirmation, launch confirmation, final-write confirmation, editor prompt, validation warning prompt, and draft cleanup prompt.
• Refactored the Ritual CLI entrypoint into a reusable runtime lifecycle wrapper with signal handling, stdin cleanup, help normalization, and top-level error funneling while preserving the one-command/no-flags MVP contract.
• Switched the project package manager from npm to Bun, replacing package-lock.json with bun.lock and updating local, CI, and release commands to use bun.
• Made the interactive candidate-review flow friendlier by replacing raw internal slugs, lexical-coherence details, and near-miss terminology with plain workflow summaries and actions.
• Move directly from choosing a repeated workflow into local-agent draft setup instead of showing a second approval menu.
• Detect available local draft generators and ask the user to choose Claude Code or Codex before creating the SKILL.md draft, instead of silently preferring the first executable found.
• Summarize scanned history sources and repeated diagnostics so users see concise totals instead of a long line-by-line dump for every skipped session file.
• Cleanly handle terminal prompt cancellation and release stdin after the interactive CLI settles.
• Use Codex prompt history from ~/.codex/history.jsonl as the default Codex source instead of full session transcripts.
• Use Claude prompt history from ~/.claude/history.jsonl, honoring CLAUDE_CONFIG_DIR, as the default Claude source instead of full session transcripts.
• Configure Bun release publishing to use npm automation token authentication in CI.
• Rename the npm package to ritualai while keeping ritual as a binary alias.

Fixed

• Let the ranking progress spinner animate during repeated-workflow clustering by yielding through the async ranking path.
• Removed stale draft wording from the direct skill generation flow and made successful progress spinners clear without printing ok ... status lines.
• Avoid Node's unsettled top-level-await warning during interactive prompts by deferring stdin unref cleanup until the CLI session settles.
• Parse real Codex session response_item.payload records so user-authored prompts are extracted from ~/.codex/sessions.
• Ignore injected Codex context blocks such as environment metadata, app/plugin instructions, skill instructions, and AGENTS.md context when extracting user prompts.
• Avoid generating suggested skill names from attachment/path noise such as image filenames and local filesystem fragments.