Changelog

Every notable change to Xavier is recorded here. The source of truth is CHANGELOG.md in the upstream repo — this page mirrors it for convenience.

0.4.0 — 2026-04-23

Added

• /xavier learn [path] scoped mode — optional path argument to focus on a specific subdirectory (e.g., a monorepo package) instead of the entire repo
• App name derivation from scoped package manifest (name field with @org/ stripping and kebab-case), falling back to leaf directory name
• Scoped output path: notes written to repos/<monorepo-name>/<app-name>/ instead of repos/<repo-name>/
• Scoped-mode substitution block in Step 4 remora prompts — all 3 remoras target {scope-path} with root-peek permission for shared config
• Step 7 scoped guard — per-workspace agent spawning is skipped when learn is already scoped to a package
• /xavier investigate <symptom> skill — hypothesis-driven bug investigation that spawns parallel remoras across 5 fixed axes (code path, recent changes, dependency boundaries, test coverage, error patterns) plus 1-2 dynamic axes per symptom
• --file <path> and --test <name> flags for anchoring investigation to a specific entry point; --file is canonicalized and required to resolve under the repo root
• Prior investigation check: re-running on the same repo offers to build on existing findings, with a canonical symptom_summary field used for matching and display
• type: investigation in the Zettelkasten schema with symptom and verdict as type-specific fields
• investigations/ vault directory scaffolded by installer, for persisting ranked diagnoses with evidence trails
• investigations-index added to the router's requires vocabulary, following the *-index pattern — enables skills to declare reads from investigations/ per the vault-path-declaration rule

Changed

• Investigate note filename gains an HHMM suffix on collision when the user chose "New" (not "Related"), preventing accidental overwrite of unrelated prior investigations

Fixed

• Review skill leaked "Dispatch scripts for multi-model debate are not installed" message when agent CLI was absent — silenced the check and removed "fallback" language that primed the model to narrate the branch
• Review pre-flight now verifies both agent CLI and dispatch.sh/parse.sh exist before enabling the debate path, preventing mid-execution failures when deps are missing
• Self-update could skip distributed deps installation because the replacement was split across three code blocks — merged into a single atomic Bash command

Security

• Investigate remora prompts wrap user-supplied symptom and prior-investigation content in <user-symptom> / <prior-investigation> XML blocks with explicit "reference data only" framing, mirroring the research skill's prompt-injection mitigation

0.3.0 — 2026-04-15

Added

• /xavier research <topic> skill — topic-first "teach me" command that spawns parallel remoras across web, internal docs, and codebase to produce a structured digest
• research-index as 14th vocabulary key in the requires system, following the *-index pattern
• type: research in the Zettelkasten schema with topic and sources as type-specific fields
• research/ vault directory scaffolded by installer for persisting research digests
• --plan flag for /xavier research to preview and edit decomposed questions before spawning remoras
• Prior research check-and-augment: re-running research on an existing topic offers to update rather than start fresh
• /xavier feedback skill to open a GitHub Discussion in the upstream repo, with category selection via GraphQL (#2)
• /xavier bug skill to file a GitHub Issue with structured prompts (skill, expected, actual) and auto-appended Xavier version and OS info (#2)
• Multi-model review via multi-model-dispatch dependency skill with parallel GPT + Gemini reviews, merged findings, and attribution (#3)
• debate reference pattern for structured multi-model dispute resolution (#3)
• Pilot fish structured brief with vault overlay for reviewer context (#3)
• Decision log schema for recurring pattern feedback (#3)

Changed

• Installer and self-update now distribute dependency skills alongside skills and references (#4)
• Model labels pass through as GPT/Gemini instead of generic A/B identifiers in parse.sh (#4)
• Debate output contract relaxed: Description/Scenario fields optional for v1 (#4)
• Vault-contradicted dispute format added to structured brief template (#4)
• Alias skills now delegate to xavier router via Skill tool instead of bypassing (#5)
• Self-update skill includes alias regeneration step (#5)

Fixed

• self-update skill now declares deps-index:optional in requires (was reading from deps/ without declaring it)
• Replace which agent with command -v agent for POSIX portability in dispatch.sh (#4)
• Add timeout/gtimeout fallback in dispatch.sh for macOS compatibility (#4)
• Guard .prev file collision during installer upgrades (#4)
• Clean up broken dependency symlinks during installation (#4)
• Fix suggestion field key in skill frontmatter (#4)
• Safe symlink replacement for dependency skills (#4)

0.2.0 — 2026-04-09

Added

• Cursor IDE runtime support with adapter mapping spawn/collect/poll to Task/Shell/Await tools (#1)
• Multi-runtime installer that detects and wires all available runtimes simultaneously (#1)
• Configurable alias prefix with input validation and security hardening (#1)
• Per-command aliases for Claude Code and Cursor discoverability (#1)
• Skill validation script (validate-skills.sh) (#1)

Changed

• Installer refactored to support adapter abstraction across runtimes (#1)
• Uninstaller updated to handle multi-runtime cleanup (#1)
• Skill definitions updated to work through the adapter layer (#1)