Skip to main content

Agent Layer publishes full release notes and artifacts on GitHub Releases. The complete history lives in CHANGELOG.md and is embedded below.

Where to read release notes:

note

If you pin versions in a repo, read the release notes before upgrading (see /docs/concepts#version-pinning and /docs/upgrades).

If you are looking for a specific change, search the changelog by version tag (for example v0.6.0).

Changelog

All notable changes to this project will be documented in this file.

v0.20.0 - 2026-09-09​

Added​

  • Live harness model discovery for Claude, Codex, Grok, Antigravity, and GitHub Copilot CLI.
  • Wizard prefetches model options across all supported harnesses at startup and falls back gracefully to defaults or custom input if discovery fails.
  • GitHub Copilot CLI headless SDK model discovery and doctor model health checks.

Changed​

  • Standardized Antigravity model parsing to native model IDs.
  • Configuration validation allows dynamically discovered models across all supported agent harnesses while retaining custom model entry.

v0.19.0 - 2026-09-07​

Added​

  • CLI commands al dispatch inspect and al dispatch output for observing running or terminal dispatches and reading final answers or captured events.
  • MCP tools dispatch_inspect and dispatch_output.
  • Dispatches support selector targeting by conversation handle or exact invocation_id, with bounded provider process-group termination confirmation.

Changed​

  • Provider cancellation requests now verify process-group termination before considering work stopped.

v0.18.6 - 2026-09-05​

Added​

  • HTTP MCP servers can declare client-managed OAuth authentication with auth = "oauth"; al doctor reports these servers without attempting unauthenticated tool discovery.

Changed​

  • Agent Dispatch now preserves provider conversation identity across failed handoffs and continuations, records provider activity timestamps, and requires bounded proof of provider process-group termination before releasing ownership.
  • Dispatched child processes suppress inherited update warnings so machine-oriented dispatch output stays focused on the requested work.

Fixed​

  • MCP validation now rejects unsupported authentication values and safely removes HTTP-only authentication fields from stdio server configuration.

v0.18.5 - 2026-09-02​

Changed​

  • Completed benchmark studies can regenerate reports from their immutable manifests without revalidating current project treatments or starting provider and verifier processes, and recovery-only runs print the regenerated report path.
  • Upgrade commands reject targets newer than the running CLI and direct users to run the target release binary.

Fixed​

  • Release artifact builds now require the target migration manifest and smoke-test upgrade planning and application with the built binary, preventing releases with incomplete embedded upgrade metadata.

v0.18.4 - 2026-09-01​

Added​

  • DeepSWE planner export schema v3 pins per-task published sample size and variance so one-repetition studies can report labeled published_proxy inference when observed Welch variance is unavailable.

Changed​

  • Study reports prefer observed Welch inference when every required task has at least two completed repetitions in both arms. Schema-v3 selections otherwise use shared published-proxy variance; schema-v2 selections without those pins remain unavailable at one repetition.
  • Reports record inference source and published-variance provenance and surface selector versus executed model or reasoning mismatches.

v0.18.3 - 2026-08-30​

Added​

  • al dispatch start --role and MCP dispatch_start role retain a caller-defined workflow role as dispatch evidence.
  • al benchmark run --recover-only finalizes retained evidence-backed verifier test timeouts without starting a provider or verifier.

Changed​

  • Official al benchmark init studies pin embedded benchmark-safe core rules and workflow skills, keep the project's instructions and skills as unreferenced audit snapshots, and require plan-reviewer, implementer, and code-reviewer dispatch roles.
  • Benchmark progress reports configured timeout budgets, attempt limits, concurrent active cells, and each experiment's effective workflow.
  • Paid Grok cells pause before inference when the repo-local OAuth credential has 30 minutes or less remaining.

Fixed​

  • Evidence-backed verifier test timeouts become explicit zero-score test_timeout outcomes instead of remaining eligible for verifier replay.
  • A cell failure stops assigning new paid work without cancelling sibling cells that are already running.

v0.18.2 - 2026-08-29​

Added​

  • Paid benchmark cells now persist provider-completion checkpoints and support verifier-only replay from the exact retained provider patch without another provider call.

Changed​

  • Benchmark progress now reports the active environment, provider, or verifier phase with elapsed time and its effective timeout.
  • Benchmark cancellation gives Pier a bounded graceful-cleanup window before process-group escalation.

Fixed​

  • Provider streams and patches survive verifier failures, cancellation, cleanup errors, and artifact-promotion failures; failed promotion retains and reports its staging path.
  • Linux cleanup repairs verifier-owned logs for running and stopped containers before removal, including root ownership repair through a temporary container when needed.
  • Failed verifier replays retain their checkpoint and staging evidence instead of becoming resumable paid failures, while sanitization preserves a byte-exact private replay patch.

v0.18.1 - 2026-08-28​

Changed​

  • Benchmark runtime preflights now persist content-addressed receipts and reuse them on later identical invocations; al benchmark run and al benchmark readiness warn when DeepSWE linux/amd64 task containers require host emulation.

Fixed​

  • Retained Grok and Antigravity streams are now validated before paid provider calls, and failed stream-preflight artifacts are cleaned up.

v0.18.0 - 2026-08-28​

Added​

  • Added al update to update the global CLI through its current installation method: Homebrew formula upgrades use Homebrew, while script installations preserve their existing prefix through the official installer.

Changed​

  • Upgrade guidance now uses al update for the global CLI and keeps al upgrade as the separate repository migration workflow.

Fixed​

  • MCP discovery now avoids duplicate built-in Agent Dispatch entries and does not pass version-dispatch state into built-in MCP child processes.
  • Global CLI update detection now fails explicitly for development, dispatched, ambiguous, or nonstandard installations instead of risking an unsafe overwrite.

v0.17.9 - 2026-08-28​

Changed​

  • Benchmark study preparation now preflights runtime per task and experiment and retains each task image across its serial cells before reclaiming it, with cleanup preserved after failed or cancelled runs.
  • Benchmark overlay certification now derives identity from the pinned image and exact overlay source while still rebuilding and rechecking overlay images; Grok benchmark execution uses Pier's disposable-container devbox sandbox without a host Bubblewrap dependency.

Fixed​

  • Benchmark run heartbeat output now resets on real progress and reports after 60 seconds of inactivity, avoiding misleading duplicate heartbeats.

v0.17.8 - 2026-08-27​

Added​

  • al benchmark init <selection.json> now scaffolds a self-contained bare-versus-Agent-Layer study with a benchmark-safe provider configuration and snapshots of the current instructions and skills. Added docs/BENCHMARK.md with the recommended workflow and advanced controls.

Changed​

  • Benchmark run and readiness commands now choose host-aware concurrency automatically, preflight Docker capacity before image pulls, reclaim task images by default, and report stage/task progress with long-running heartbeats. Readiness can target the tasks selected by a study.
  • CI coverage is now behavioral rather than quota-based, and shell syntax validation is part of the complete CI gate; coverage remains available as diagnostic reporting.

Fixed​

  • Antigravity structured terminal responses now retain the complete final answer even when it exceeds the structured metadata key limit.

v0.17.7 - 2026-08-27​

Changed​

  • Agent Dispatch now uses structured provider output internally for every supported agent, including Antigravity display-name models, while returning only the extracted plain-text answer to callers.

Fixed​

  • Antigravity exact-slug dispatches now allocate their structured event capture before provider execution instead of failing before launch.

v0.17.6 - 2026-08-27​

Changed​

  • Release catalog certification now runs before tagging in eight isolated, bounded shards. Release publication reuses only a successful certification for the exact tag commit instead of repeating the full Docker audit.
  • Benchmark readiness reports task-level progress and supports deterministic task filtering, sharding, and per-task timeouts.

Fixed​

  • Empty MCP preflight contracts now encode servers as an array, preserving the benchmark study contract when no MCP servers are enabled.
  • The quill-shared-toolbar-focus readiness image now installs pinned xauth, allowing the complete pinned DeepSWE catalog to certify successfully.

v0.17.5 - 2026-08-27​

Fixed​

  • Empty MCP preflight contracts now encode servers as an array, preserving the benchmark study contract when no MCP servers are enabled.

v0.17.4 - 2026-08-26​

Changed​

  • Release preflight now certifies the pinned DeepSWE benchmark catalog before building release artifacts, with bounded Docker image cleanup during the catalog audit.
  • Hosted benchmark readiness now bootstraps a minimal Agent Layer workspace in the release checkout before certification.
  • DeepSWE benchmark selections now translate published provider model identities to the exact canonical identities accepted by the benchmark CLI and report supported models when selectors are invalid.
  • The .agent-layer/.env file is optional when no Agent Layer secrets are needed; unreadable or malformed files still fail explicitly.
  • The implement skill now limits plan and code-review corrections to the requested input scope.

Fixed​

  • Docker cleanup uses the correct force flag for containers, volumes, and task images.

v0.17.3 - 2026-08-26​

Added​

  • al benchmark run supports pinned Antigravity 1.1.21 and Grok 1.0.5 coordinators, provider-native Agent Layer treatments, structured coordinator/dispatch evidence, and usage-based API-equivalent cost accounting without running a full benchmark during development.
  • Antigravity and Grok benchmark adapters use the same repo-local subscription OAuth boundaries as their Agent Layer clients. Antigravity prefers its repo-local OAuth fallback and otherwise exports only the native keyring OAuth profile into the contained CLI fallback file; Grok stages only .grok-config/auth.json.

Changed​

  • Agent Dispatch's tested Antigravity baseline is now 1.1.21 because benchmark child usage requires that release's structured headless output. Older Antigravity releases fail the existing provider-version preflight instead of running under an unevidenced stream contract.

v0.17.2 - 2026-08-25​

Added​

  • Added review/oversized for otherwise-cleared entries over 100 files or 250 MiB (including oversized immediate children and top-level files), and review/symlinks for top-level or move-breaking links.
  • Interactive upgrades can record intentional unknown files and directories from .agent-layer/ and docs/agent-layer/ in a gitignored .agent-layer/upgrade-keep-list. Kept paths are omitted from future upgrade plans and deletion flows.

Changed​

  • al wizard replaces the workflow-bundle yes/no with an instruction choice (None, Rules, or Rules and memory) and a single skills catalog. Agent Layer development skills (/implement, /ship-pr, and the rest of that set) are one catalog row; selecting it installs those skills, and deselecting it removes them. Instructions and skills are independent.
  • Consolidated the hidden al organize-scratch safety model around one complete metadata/hazard walk and outcome pipeline. Dry runs are now strictly read-only and print the full proposed review list; apply preserves unresolved prior review entries and records actual moved, collision, failed, and unattempted outcomes. Directory/file size limits, credential-bearing content, authored assets, nested checkout markers, unreadable paths, and move-breaking symlinks now conservatively force review.
  • al organize-scratch now accepts non-repository and untracked/non-ignored roots but always refuses repository roots, including empty or unborn repositories, as well as subtrees containing tracked content. Git fact failures fail closed under stable English diagnostics. Registered entry, nested, foreign main, and foreign linked worktrees are all protected and repaired from the correct repository context, including newline-containing paths and external linked registrations owned by a moved main checkout, with failed or stale repairs returning non-zero. Invalid --keep paths now fail explicitly instead of being ignored.
  • make dev is now a fast formatting and lint loop (make fmt then make lint) instead of chaining coverage and release tests. make ci remains the complete pre-PR verification gate.
  • Public website pages and best-practice guides were rewritten for clearer, more direct language.

Fixed​

  • Listing .agent-layer/tmp in .agent-layer/upgrade-keep-list now keeps that directory and skips the grouped tmp deletion prompt. Individual files under tmp remain ineligible for the keep list. The interactive keep-list checklist includes .agent-layer/tmp last, unchecked.
  • Codex project trust is now seeded using the physical repository path after symlink resolution, so a repo opened through a symlink matches the managed [projects."<root>"] trust key. Trust-root resolution fails explicitly when the path cannot be canonicalized.
  • Nested upgrade keep-list file entries no longer treat ancestor directories as fully kept, so sibling unknown paths remain eligible for deletion.
  • Upgrades now preserve the .agent-layer/ and docs/agent-layer/ git tracking choices stored in .agent-layer/gitignore.block while still applying new managed ignore rules from the release template. Match, al upgrade plan, and overwrite preview compare against that merged target, so customized tracking alone is not reported as an update. A # after a managed tracking pattern is rejected as unsupported, because Git treats that text as part of the pattern rather than a comment.
  • Locked source loading removes empty immediate child directories under .agent-layer/skills/ before validation, so sync, launch, dispatch, and skill-import share the same cleanup. Nonempty or malformed skill directories still fail strictly.

v0.17.1 - 2026-08-21​

Added​

  • al skills diff <name> compares live base, local, upstream, and destination trees as an ordinary Git unified diff.
  • Pull and push merge conflicts now leave a Git workspace under .agent-layer/tmp/skill-conflicts/<name>/. Finish the merge with ordinary git commands and al skills resolve <name>.
  • ship-pr includes a stateless read-pr-comments.sh command that prints every PR comment kind, including review-thread resolved/outdated state, as readable Markdown.

Changed​

  • The Claude instruction shim is now .claude/CLAUDE.md instead of root CLAUDE.md. When Grok is enabled, sync sets [compat.claude] agents = false in repo-local .grok-config/config.toml, and Grok launch/dispatch/VS Code set GROK_CLAUDE_AGENTS_ENABLED=false, so Grok does not load both generated instruction files. The 0.17.1 migration deletes Agent Layer-generated root CLAUDE.md files; hand-authored files at that path are left untouched.
  • Recreating a deleted contribution branch reuses its prior publication checkpoint when that commit remains in the destination's history, avoiding conflicts with changes already merged from the earlier branch.

v0.17.0 - 2026-08-20​

Added​

  • Grok Build CLI is a first-class Agent Layer client: al grok, [agents.grok] (enabled, model, reasoning_effort, disable_memory), always-on GROK_HOME=<repo>/.grok-config, native .grok/config.toml MCP and [permission] projection, seeded folder trust, --sandbox mapping, shared .agents/skills/ skills, doctor warning for a missing or older-than-1.0.5 grok, al probe grok, wizard coverage, Agent Dispatch target grok, and notifications.chime via .grok/hooks/agent-layer-chime.json. Upgrade 0.17.0 defaults agents.grok.enabled to false.

Fixed​

  • Grok sync now recognizes generated .grok/config.toml files by a stable marker across header revisions, requires the credential-bearing .grok-config home to be a real directory, and rejects symlinks before seeding repo trust outside the isolated home. Sync, al grok, Grok dispatch, and al vscode create that home at 0700 and tighten an existing real home the Grok CLI created at 0755, instead of failing with a chmod 700 instruction. Grok dispatch and capability probes also bound retained provider output with accurate truncation notices.

v0.16.3 - 2026-08-09​

Added​

  • Added one content-addressed DeepSWE study workflow: al benchmark run <study.toml> executes explicit experiments from a website selection, resumes immutable cell evidence, and produces JSON and HTML reports with fixed-selection Welch comparisons and Holm-adjusted p-values. al benchmark readiness remains the developer preflight for the pinned task catalog.

Changed​

  • Removed the obsolete benchmark baseline, benchmark treatment, benchmark report, matrix runner, and benchmark correct-scores implementations. Compatible historical selection evidence is read narrowly by the study reporter and corrected automatically from preserved verifier artifacts; private campaign state is never scanned or rewritten.

Fixed​

  • The built-in Agent Dispatch MCP server now starts from the repository root embedded during al sync, so installing a newer global al no longer causes MCP clients with an older repo pin to bypass that pin. The launcher keeps the legacy al dispatch mcp-server argument contract when it reaches the cached binary, preserving compatibility with older pinned releases, and retains a same-project caller working directory for dispatched agents. Existing generated MCP files need a one-time refresh with the fixed binary (AL_VERSION=v0.16.3 al sync); this does not change the repo pin.

v0.16.2 - 2026-08-08​

Added​

  • Added al benchmark correct-scores to regenerate versioned canonical results for affected stored DeepSWE runs from preserved verifier evidence. Benchmark reports consume the canonical result and fail rather than emit a known-incorrect score when neither it nor the required verifier artifacts remains available.

Changed​

  • The generated .agent-layer/.gitignore no longer ignores skills-imported/ or skills.lock.json. In a project that commits .agent-layer/, imported skills and their lockfile are now tracked, so a fresh clone or CI run has the exact imported skill content — including local edits — without a network al skills pull, and a pinned import reproduces without one. The two must be committed together: the imported tier is fully managed, and al sync rejects a skill directory that has no skills.lock.json entry. .agent-layer/skills-imported/.staging/, the import transaction's scratch space, stays ignored. .agent-layer/.gitignore is agent-owned and rewritten on every al init/al upgrade, so existing projects pick this up with no migration; run git add .agent-layer/skills-imported .agent-layer/skills.lock.json once after upgrading to start tracking them. Note that the root .gitignore block still ships /.agent-layer/ uncommented, which ignores the whole directory and overrides this; projects that want their Agent Layer configuration in version control comment that line out in .agent-layer/gitignore.block.

Fixed​

  • Repeated al skills push calls to a contribution branch now record the last successfully published tree separately from the source lock, so review-driven edits and explicit reversions add commits to the same pull-request branch without conflicting with or silently preserving the previous push. Compatible edits made directly on the destination branch are still retained. This release reads existing skill lock schema version 1 files and writes version 2 whenever it next changes a lock; every consumer of a committed version 2 lock must upgrade, while older Agent Layer binaries reject it with an explicit unsupported-version error instead of misreading new fields.

v0.16.1 - 2026-08-06​

Fixed​

  • al init and al upgrade no longer block when .agents/skills/ or .claude/skills/ contains content that cannot be matched to a canonical Agent Layer skill. Those client roots are exclusively owned disposable projections, so inspecting their contents before replacing them contradicted the documented sync contract and prevented some v0.15.0 projects from upgrading to v0.16.0.

v0.16.0 - 2026-08-06​

Git-backed Agent Skill imports and a consolidated skill and instruction bundle.

Added​

  • Git-backed Agent Skill imports. al skills add <repository> <selector>... records an import in a new [[skills.imports]] config block, al skills pull fetches every configured source and reconciles it with local content, al skills status reports local state without touching the network, al skills reset <name> discards one skill's local edits, al skills remove drops a selector, and al skills push publishes local changes back to a configured destination. Imported skills live in .agent-layer/skills-imported/<skill-name>/ and stay editable; pull merges upstream changes without discarding local edits. Each import block carries its own ref, tracking (tracked or pinned), and write_policy (none, branch, or direct); write_policy = "branch" requires an explicit non-primary push_branch, and al skills push refuses to write to the destination's actual default branch. add, remove, reset, and push prompt before persistent configuration changes, destructive reset, or remote publication; non-interactive callers confirm with --yes. The generated .agent-layer/.gitignore ignores both skills-imported/ and the skills.lock.json state file.
  • A bundled skill-sync catalog skill for managing those imports from an agent. Its pre-approved shell surface is limited to al skills and al sync.
  • al organize-scratch --root <dir> sorts a scratch directory into reports/, artifacts/, and review/ folders and writes an ORGANIZE-REVIEW.md listing everything that still needs a human decision. It only moves entries: nothing is deleted, overwritten, or merged, and an entry whose destination is already taken is left in place. Defaults to a dry run; pass --apply to move. Registered Git worktrees are left alone unless --move-worktrees is given, in which case their registrations are repaired after the move. The command is a maintenance aid and is hidden from al --help.

Changed​

  • Skill sources now use one strict, byte-exact tree contract. User-managed and Git-imported skills both require uppercase SKILL.md, reject symlinks and other non-regular nodes, preserve unknown/provider-specific frontmatter and every resource byte, and project from one locked immutable snapshot. Agent Layer now owns .agents/skills/ and .claude/skills/ completely: enabled roots are replaced wholesale and disabled roots are removed. The v0.16.0 migration blocks before mutation if a pre-existing client root contains an entry that is neither marker-bearing output from a released Agent Layer version nor a directory matching a source-tier skill directory with a canonical regular SKILL.md.
  • The workflow skill templates are consolidated. implement replaces the plan-work, review-plan, implement-plan, fully-implement-plan, and full-workflow chain, and ship-pr absorbs address-pr-comments and fix-ci as references. boost-coverage, clean-and-fix-code, improve-codebase, review-uncommitted-code, run-and-fix-all-checks, schedule-backlog, simplify-codebase, and verify-work are removed; their work is covered by the remaining skills. Existing installs keep their copies of the removed skills until they are deleted by hand; al upgrade reports them as orphans rather than removing them.
  • debug-and-fix-issue is removed. No remaining skill covers reproduce-then-diagnose debugging; run those investigations directly, or hand the diagnosis to implement once the cause is known.
  • docs/agent-layer/ROADMAP.md is no longer a memory file. The template is removed and the memory instructions, audit-memory, audit-documentation, and the implement-backlog loop mode no longer reference it; BACKLOG.md is now the only scheduling surface. An existing ROADMAP.md is left untouched on upgrade and simply becomes an ordinary project document. al upgrade still understands the memory_roadmap_v1 ownership policy recorded in manifests from earlier releases, so upgrading a repo that has one continues to work.
  • The agent-dispatch catalog skill is renamed to dispatch-agent, so agents now invoke it as /dispatch-agent and the implement, ship-pr, and auto-skill-loop skills instruct that name. The v0.16.0 migration renames an installed .agent-layer/skills/agent-dispatch/ directory to .agent-layer/skills/dispatch-agent/ and fails if a populated destination already exists. The Agent Dispatch feature, its al dispatch commands, and the built-in agent-layer MCP tool names are unchanged.
  • The managed instruction templates and the dispatch-agent, audit-documentation, audit-memory, audit-tests, auto-skill-loop, and interface-audit skills are rewritten for brevity. The Explicit-only. description marker is dropped in favor of per-client invocation metadata.
  • The bundled instruction set is consolidated from five files to two: 00_rules.md absorbs 01_base.md and 03_tools.md, and 02_memory.md becomes 01_memory.md. The user-managed 04_conventions.md template is removed along with the user-owned instruction ownership rules; a project tailors instructions by editing the managed files or adding its own. The v0.16.0 migration renames 02_memory.md so its content carries forward. Existing 01_base.md, 03_tools.md, and 04_conventions.md files are left on disk — instruction fragments are user-editable and may hold project-specific rules — and al upgrade reports them as orphans rather than removing them. Delete them by hand once you have moved anything you want to keep into 00_rules.md.

Fixed​

  • al dispatch now delivers approvals.mode to headless agents, which previously could not edit files in any mode except yolo. codex exec starts in a read-only sandbox and claude -p denies Edit/Write, so a dispatched agent approved to run commands failed on its first write. Dispatch now sends sandbox_mode to Codex (read-only for none/mcp, workspace-write for commands/all) and --permission-mode to Claude (dontAsk / acceptEdits). Editing follows command approval because Codex has no separate edit-approval rule: its sandbox is the edit gate. yolo is unchanged. Explicit agents.codex.agent_specific.sandbox_mode and agents.claude.agent_specific.permissions.defaultMode still win. Note that workspace-write keeps .git read-only and denies network access including loopback, so dispatched agents cannot commit or bind test servers without explicit agent_specific passthrough.
  • Claude dispatches now pass the allowlisted commands and MCP servers as --allowedTools. Claude applies a project's permissions.allow rules only after its workspace trust dialog is accepted, and that dialog never appears under -p, so the generated settings file left every approval inert unless the repository happened to have been trusted in an earlier interactive session.
  • A dispatched Claude agent that is denied a tool call now fails the dispatch instead of reporting success. Claude returns exit 0, is_error: false, and a fluent final answer for a run whose writes were all denied, so a denied dispatch previously looked complete.
  • Git subprocesses now resolve the repository from the path they are given instead of honoring an inherited GIT_DIR. Git exports its repository-discovery variables to every hook, and they take precedence over git -C <path>, so benchmark provenance and pinned-checkout validation could report on a different repository than the one being measured when a run started from a Git hook.

Internal​

  • Added the v0.16.0 migration and template ownership manifests. The migration claims the client skill projection roots, renames the managed 02_memory.md instruction file to 01_memory.md, and renames the agent-dispatch catalog skill directory to dispatch-agent.
  • The generated Homebrew formula no longer sets a redundant version field; Homebrew derives it from the release asset URLs.

v0.15.0 - 2026-07-31​

Agent Dispatch MCP and benchmark tooling release.

Added​

  • Agent Dispatch now exposes a built-in agent-layer MCP server with dispatch_options, dispatch_start, dispatch_wait, dispatch_continue, and dispatch_cancel, using the same asynchronous handles, states, and durable result files as the CLI. The server is projected into enabled Codex, Claude, Antigravity, VS Code, and Copilot CLI clients.
  • Added al benchmark baseline, al benchmark treatment, and al benchmark report for website-planned DeepSWE comparisons backed by immutable campaign evidence.

Changed​

  • The bundled Agent Dispatch skill now uses the built-in MCP tools for bounded leaf work, while the CLI remains the human and scripting surface.
  • Workflow skills were simplified around explicit root-to-leaf boundaries, review gates, and delegated delivery.
  • Renamed the bundled playwright-cli skill to playwright while preserving the Playwright CLI command surface.

Internal​

  • Added the v0.15.0 migration and template ownership manifests. The migration renames the managed playwright-cli skill directory.

v0.14.0 - 2026-07-22​

Breaking Agent Dispatch lifecycle release. Dispatch now exposes asynchronous, handle-based conversations with durable result files instead of synchronous named conversations and fanout operations.

Added​

  • al dispatch start, wait, continue, and cancel. start and continue return a handle immediately; wait blocks for the terminal state and returns the completed invocation's durable Markdown result_path.
  • A stable JSON contract for every successful dispatch command. Independent start calls provide parallel work without a public fanout resource.

Changed​

  • Breaking: al dispatch now requires the explicit lifecycle commands above. The synchronous fresh/resume syntax and fanout, inspect, history, list, and delete commands are removed.
  • Breaking: per-agent agents.<agent>.dispatch.default_agent tables are removed. start requires --agent; the release migration deletes retired tables before strict config decoding.
  • Completed results are persisted atomically before their invocation becomes completed; failed and cancelled conversations may be continued with the same handle.

Fixed​

  • Dispatch and client launchers preserve cancellation signals, and sync ignores its transient lock file.
  • Updated golang.org/x/text to v0.39.0 to remediate GO-2026-5970, which the release binary vulnerability scan detected in v0.38.0.

Internal​

  • Added the v0.14.0 migration and template ownership manifests. The migration supports the established 0.10.2+ upgrade range and deletes retired per-agent dispatch default tables.

v0.13.0 - 2026-07-17​

Breaking Agent Dispatch redesign and workflow-skill refresh. Dispatch now starts fresh, final-answer-only provider turns by default, and exposes explicit continuation, inspection, history, cancellation, and fanout operations rather than streaming provider traffic.

Added​

  • al dispatch resume <name>, fanout, inspect, history, cancel, list, and delete. Named conversations have atomic mappings and immutable run-UUID records; fanout runs one shared prompt across independently retained targets.
  • Separate fresh, resume, and inspection capability facts in al dispatch options --json, including installed provider versions and compatibility warnings. Claude Code 2.1.207, Codex CLI 0.144.1, and Antigravity 1.1.1 are the tested baselines; newer versions remain available with a warning, while older, unreadable, and malformed versions fail before launch.
  • Private bounded run evidence, 30-day retention for eligible evidence and mappings, process-group supervision, final-answer replay, a safe pre-start retry, and Antigravity’s isolated documented log-file ID extractor with fail-loud not resumable handling.
  • al hook chime, a project-shareable completion handler for Claude, Codex, and Antigravity. It filters non-terminal lifecycle events, fails open, uses the macOS system sound or Linux canberra-gtk-play when available, and never claims that a provider stop event proves task completion.

Changed​

  • Breaking: al dispatch no longer streams answer text, progress, raw events, or provider diagnostics. Standard output contains only a successful final answer; standard error begins with one compact identity line.
  • Breaking: the v1 dispatch_capable and streaming options fields were removed from the public JSON contract. Use separate capability facts instead.
  • The default dispatch.max_depth is now 3, allowing two nested dispatches after the initial call. Built-in workflows remain root-to-leaf and use external dispatch only for bounded leaf judgment.
  • Antigravity's generated settings file is now shared state: Agent Layer patches only its managed model, permissions, and passthrough paths while preserving native workspace, approval, and trust settings.
  • Codex VS Code launches now configure only extension-relevant runtime features; ordinary Codex CLI settings and status-line configuration are not projected into the VS Code path.
  • Model suggestions were refreshed for Claude, Codex, Copilot CLI, and Antigravity in the wizard and dispatch option catalog.
  • Version-binary handoff moved from internal/dispatch to internal/versiondispatch to distinguish it from Agent Dispatch.
  • auto-skill-loop now provides fix-issue-log, implement-backlog, improve-interfaces, and improve-codebase modes plus repository-added mode files. It selects adaptive fresh work, preserves local blockers, batches and ships centrally, keeps /ship-pr isolated in its shipper dispatch, independently gates exact-head merge authorization, and reconciles each result without preplanning the full source.
  • Instruction-only assets now live under their owning skills' references/ directories; output and machine-readable resources remain under assets/.
  • Bundled workflow skills are now more explicitly root-to-leaf: orchestration owns transitions and delivery, while Agent Dispatch is restricted to named, bounded leaf roles. The refresh adds clean-and-fix-code, debug-and-fix-issue, fully-implement-plan, review-uncommitted-code, run-and-fix-all-checks, and verify-work as the current workflow surface.

Removed​

  • Retired the standalone fix-issues skill after preserving its explicit filters, batching, dispositions, and one-delivery behavior in the fix-issue-log mode.
  • Retired superseded workflow skill names and wrappers, including audit-and-fix-uncommitted-changes, debug-issue, review-scope, verify-against-plan, prune-new-tests, simplify-new-code, complete-current-phase, finish-task, repair-checks, and multi-agent-plan-review. The migration manifest renames managed skill directories where a direct replacement exists; retirement remains ownership-aware and reviewable during upgrade.

Fixed​

  • Dispatch now waits for provider completion evidence and terminates failed process groups, preventing terminal failure while owned descendants continue.
  • Projection preparation is serialized before provider launch, avoiding concurrent generated-skill writes while allowing independent targets to run concurrently afterward.
  • Dispatch now preserves the caller's working directory across linked worktrees, records resolved targets for reliable resume, validates provider diagnostics, and closes cancellation and recovery races without releasing a live process claim prematurely.
  • al sync, launchers, and the root command preserve cancellation signals and logical repository roots more reliably. Upgrade rollback retries now reset scoped targets before restoring a failed snapshot, allowing a partial rollback to converge safely.

Internal​

  • Added the v0.13.0 migration and template ownership manifests. The migration supports upgrades from the 0.10.2 line and renames managed workflow skill directories and their instruction resources.
  • Updated the Go toolchain to 1.26.5, modernized terminal UI dependencies, and added release-binary vulnerability scanning.

v0.12.3 - 2026-07-10​

Patch release for the v0.12 line. Adds current Codex model and reasoning-effort suggestions, and fixes al sync so generated skill resources stay reconciled safely with their source trees.

Changed​

  • Codex model suggestions now include gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna in the wizard and Agent Dispatch options.
  • Codex reasoning-effort suggestions now include max and ultra in the wizard and Agent Dispatch options.

Fixed​

  • al sync now reconciles generated skill resources (scripts/, references/, and assets/) with their source skill directory: it removes stale resources and handles file/directory transitions without replacing the generated SKILL.md.
  • Skill-resource sync now detects source symlinks without following them and safely replaces generated destination symlinks rather than writing through them.

Internal​

  • Added v0.12.3 migration and template ownership manifests. The migration has no operations; existing configurations need no manual migration.

v0.12.2 - 2026-07-06​

Patch release for the v0.12 line. Tightens bundled skill and instruction contracts, fixes installer prompt routing edge cases, adds public skill-architecture documentation, and hardens the release workflow before signing.

Added​

  • Added the Agent Layer-specific Skills approach documentation page, including the target root-skill model, workflow-skill boundaries, and the mapping from target modules to current bundled skills.
  • Added template tests covering bundled skill and instruction contract expectations so future wording changes are checked automatically.

Changed​

  • Bundled skill templates now state scope, deferral, review, and orchestration rules more precisely across planning, review, issue fixing, PR handling, verification, and autonomous-loop workflows.
  • Multi-agent plan review instructions now make reviewer orchestration and synthesis responsibilities clearer, while participant terminology is standardized across skill and instruction docs.
  • The Agent Dispatch catalog skill now activates only for explicit external dispatch targets or skill-directed dispatch, leaving generic subagent work to built-in subagent behavior.
  • Universal skill, CLI-skill, and instruction design guides were generalized for portable authoring guidance and now separate client/specification requirements from Agent Layer-specific conventions.
  • Release preflight documentation now reflects that CI, release-script checks, and upgrade-doc validation run before tagging.
  • The release workflow now runs make ci before importing signing credentials.

Fixed​

  • Installer and upgrade prompt routing now flow through a shared prompt router, preserving required overwrite/delete prompt validation while centralizing optional prompt fallbacks.
  • Statusline source replacement prompting is now gated on the actual optional prompt capability, preventing zero-value prompt implementations from being treated as wired callbacks.
  • Zero-value prompt fallback behavior is covered for unified overwrite, grouped tmp deletion, statusline source, config defaults, and skills migration prompts.
  • Agent Dispatch now forwards SIGINT and SIGTERM to the full dispatched process group, so shell-launched child processes terminate promptly on interruption.
  • End-to-end assertions were updated for the retired generated Codex .codex/AGENTS.md shim.

Internal​

  • Added v0.12.2 migration and template ownership manifests. The migration has no operations; the release updates managed templates and docs only.

v0.12.1 - 2026-07-04​

Patch release for the v0.12 line. Fixes binary Homebrew publishing, preserves Codex hook trust state across chime refreshes, and retires the generated .codex/AGENTS.md shim that duplicated root AGENTS.md when Codex used repo-local config.

Changed​

  • al sync no longer writes .codex/AGENTS.md. Codex reads root AGENTS.md as project instructions, and repo-local CODEX_HOME caused .codex/AGENTS.md to be loaded as home-level instructions as well.
  • Workflow instruction and skill templates were tightened for tradeoff handling, full-workflow spec approval, and prune-new-tests reviewer output.
  • Release documentation now clarifies binary Homebrew tap PR handling.

Fixed​

  • Codex chime refresh now preserves existing hook trust state stored inside the managed chime markers and remains idempotent across repeated syncs.
  • Homebrew formula generation now includes both macOS ARM64 and macOS Intel release assets.
  • Homebrew binary formula installs now mark al executable before generating shell completions.
  • The release workflow now verifies all binary assets needed by the Homebrew tap, including al-darwin-amd64.

Internal​

  • Added v0.12.1 migration and template ownership manifests. The migration deletes only Agent Layer-generated .codex/AGENTS.md files; hand-authored files at that path are left untouched.

v0.12.0 - 2026-07-03​

Consolidates all unreleased work since v0.11.0 into one coherent release. Adds typed Antigravity model selection, configurable Agent Dispatch depth, repo-local Codex home opt-in, shared Codex TOML patching, provider turn-stop chimes, serialized sync writes, new workflow skills, signed/notarized macOS binaries, and binary Homebrew delivery.

Added​

  • Antigravity model selection is now a first-class Agent Layer setting. agents.antigravity.model is projected into generated Antigravity settings, al wizard can discover choices from live agy models output with catalog fallback, and al dispatch --model plus al dispatch options now report and accept Antigravity model overrides. Antigravity reasoning level remains encoded in the selected model display string; --reasoning-effort is still unsupported for Antigravity.
  • dispatch.max_depth allows nested al dispatch chains beyond the default depth of 1. AL_DISPATCH_ACTIVE now carries the current dispatch depth, invalid or empty values fail loudly, and config validation rejects non-positive depths.
  • agents.codex.local_config_dir controls whether Agent Layer sets CODEX_HOME=<repo>/.codex for al codex, Codex dispatch, and al vscode. The default is false, preserving Codex's normal global/project config layering; set it to true to keep repo-local Codex auth, sessions, logs, and runtime state.
  • notifications.chime is a global opt-in for provider turn-stop chimes. Sync projects provider-native stop hooks for Claude and Codex and an Agent Layer-owned Antigravity plugin, while preserving user-owned hooks and plugin content.
  • al sync now serializes concurrent writes for the same project with a repo-local lock, preventing overlapping generated-file updates during parallel launches or dispatches.
  • .codex/config.toml is now treated as shared state. Agent Layer patches only managed Codex keys, MCP projection, project trust, feature toggles, and statusline entries while preserving unrelated user or Codex runtime TOML, comments, multiline strings, and plugin settings.
  • al wizard adds a Git tracking step for .agent-layer/ and docs/agent-layer/, implemented by rewriting the managed .agent-layer/gitignore.block source before sync.
  • New built-in workflow skills: auto-skill-loop, full-workflow, interface-audit, and multi-agent-plan-review. ship-pr also gained a bundled monitor-pr.sh helper for polling pull request readiness and filtering review-bot noise.
  • Release builds now sign Darwin binaries with Developer ID, enable hardened runtime, notarize them with Apple, and write checksums after signing. The release workflow also documents the required signing and notarization secrets.
  • Homebrew delivery now uses prebuilt macOS and Linux release binaries with per-platform checksums instead of building from the source tarball.
  • Shared provider option catalogs now back wizard and dispatch option suggestions for supported agents, with corrected Codex model/reasoning suggestions and fable in the Claude model catalog.

Changed​

  • al claude, al codex, al copilot, and al agy now replace the al process with the target agent CLI instead of spawning a child process. Agent exit codes now pass through directly and the old Error: <agent> exited with error: ... wrapper line is gone.
  • macOS permission prompts are now attributed to the actual signed Agent Layer binary that launches the agent. Users may see one prompt per agent after upgrading; old al entries in Privacy & Security are cosmetic.
  • approvals.mode = "yolo" no longer emits the VS Code chat.tools.global.autoApprove setting. YOLO still sends full-auto controls to Claude, Codex, Copilot CLI, and Antigravity.
  • al wizard no longer offers a workflow-bundle refresh when Agent Layer workflow files already exist. The workflow-bundle prompt is install-only for missing bundle files and preserves existing files; use al upgrade for managed workflow updates.
  • Agent-specific passthrough config now uses the shared ProviderPassthrough type, and Agent Layer-owned Antigravity model config is rejected under agents.antigravity.agent_specific.model.
  • The Claude status line now rounds weekly-limit reset time up to the next whole day/hour so a partial remaining unit stays visible, drops the # prefix on the session id, and avoids per-untracked-file git diff --no-index processes when counting untracked line changes.
  • Agent Dispatch target metadata now uses the shared provider option catalog, reports Antigravity model override support, preserves inherited CODEX_HOME unless Codex local config is enabled, and wraps stdout write failures with target-specific dispatch exit errors.
  • Release tooling now verifies published binary assets before opening the Homebrew tap PR and renders the full binary formula from release checksums.
  • CI/local workflow documentation now notes that make ci includes make test-race, and release docs clarify that tagged migration and ownership manifests are immutable release artifacts.

Fixed​

  • Upgrades move agents.antigravity.agent_specific.model to agents.antigravity.model before strict runtime validation rejects the passthrough key.
  • Wizard Git tracking recognizes managed patterns with inline comments and avoids duplicating them.
  • Antigravity wizard model options are prefetched before prompting, so the wizard can show live model choices rather than a stale or empty catalog.
  • Codex reasoning suggestions and restored shared agent option suggestions now stay valid across CLI, wizard, and dispatch surfaces.
  • Codex TOML patching now preserves multiline strings, plugin defaults, inline comments on managed keys, and unrelated shared config while updating Agent Layer-owned entries.
  • Strict config validation now fails loudly on malformed nested AskUserQuestion overrides and invalid AL_DISPATCH_ACTIVE depth values.
  • Gitignore management was hardened to prevent managed block data loss and handle inline-comment tracking defaults correctly.
  • al doctor diagnostics and CI dead-code enforcement were corrected, including loader-error handling that previously weakened the dead-code gate.
  • The hidden deprecated al mcp-prompts path is now a no-op stub instead of a live prompt delivery surface.
  • Release and formula update tests now cover signed binary build outputs, checksum extraction, notarization hooks, and binary Homebrew formula rendering.

Security​

  • The Go toolchain/dependency floor was updated to address reachable Go standard-library vulnerabilities found by local vulnerability analysis.
  • Codex trust-block generation now rejects invalid UTF-8 repository roots before writing project trust entries.
  • The release workflow now checks out code with persisted credentials disabled in the release build job.

Documentation​

  • README, site docs, troubleshooting, concepts, reference, FAQ, and security pages were refreshed for current Antigravity, MCP, Codex, upgrade, and release behavior.
  • docs/RELEASE.md now documents Developer ID signing, notarization, binary Homebrew delivery, release asset verification, and the complete set of required release secrets.
  • Agent instruction templates were tightened: production code should validate inputs and returned errors defensively, repeated failed fixes should trigger research before another attempt, and memory files are described as living records of the current working tree.
  • Skill design and CLI skill design docs were tightened, and generated workflow skills now include clearer subagent/review constraints.

Internal​

  • Shared semver parsing/comparison moved into internal/version, and release message helpers were consolidated.
  • A reusable TOML patching engine (internal/tomlpatch) now powers safer shared-config edits.
  • Upgrade migration logic was split to keep skill migration code separate from the main migration coordinator.
  • Markdown heading slug generation now avoids quadratic suffix deduplication on large documents.
  • Test coverage was expanded across dispatch depth, Antigravity model discovery, Codex shared-config patching, chime cleanup, sync locking, release tooling, wizard option discovery, gitignore tracking, and upgrade migration contracts.

v0.11.0 - 2026-06-03​

Replaces Gemini CLI support with the agy-backed Antigravity integration, adds Agent Dispatch for focused second-agent work, ships explicit opt-in Claude/Codex status lines, expands the wizard into a workflow-bundle, CLI-skill, and deterministic answer-file setup flow, hardens upgrade/CI release workflows, and publishes the public best-practice guides from canonical repository docs.

Added​

  • Agentic status lines for Claude Code and Codex, explicit opt-in per provider. al wizard and interactive al upgrade can write agents.claude.statusline = true or agents.codex.statusline = true and seed missing editable sources once. al sync then projects .agent-layer/claude-statusline.sh to .claude/claude-statusline.sh and wires statusLine into .claude/settings.json; for Codex it injects a managed [tui].status_line block into .codex/config.toml from .agent-layer/codex-statusline.toml (skipped when you define agent_specific.tui.status_line yourself). The Claude line renders model, reasoning effort, context %, weekly usage limit, session, directory, git branch/dirty state, lines changed, and session cost; it requires jq on PATH and degrades to a one-line hint when absent. Absent or explicit false disables the statusline. Existing source files are user-owned: sync, wizard, and non-interactive upgrade never overwrite them; interactive upgrade can show a diff and ask before replacing them.
  • al dispatch and al dispatch options for focused headless second-agent work across Codex, Claude, and Antigravity. Dispatch accepts prompt text from arguments or stdin, can invoke a portable Agent Layer skill with --skill, supports target selection via --agent (including random), reports model/reasoning override support through text or JSON options output, streams target answer text to stdout, writes wrapper status/errors to stderr, and uses stable wrapper-owned exit categories. Agent Layer-launched clients receive AL_DISPATCH_CALLER_AGENT; dispatched targets receive AL_DISPATCH_ACTIVE=1 so nested dispatch is blocked.
  • al wizard now offers five per-agent feature disable toggles, folded into the existing model step (most default to No, keeping the client's native behavior; the Codex apps toggle is the exception and defaults to Yes, disabling apps): disable Codex browser/computer-use (features.browser_use/in_app_browser/computer_use) and built-in apps (features.apps); disable Claude's IDE open-file reading (env.CLAUDE_CODE_AUTO_CONNECT_IDE), auto-memory (autoMemoryEnabled), claude.ai connectors (env.ENABLE_CLAUDEAI_MCP_SERVERS), and the AskUserQuestion tool. The first three Claude toggles and the Codex toggles write their client-native agent_specific key only when you opt in; the AskUserQuestion toggle writes a typed agents.claude.disable_question_tool flag and al sync injects the matching permissions.deny entry plus a PreToolUse hook into .claude/settings.json, merged with (never replacing) any deny/hook entries you already have. All toggles read back from existing config so re-running the wizard preserves your choice. The Codex "apps" prompt is reworded to the shared "Disable …?" form (its enabled-state storage is unchanged).
  • al wizard now asks whether to install or refresh the Agent Layer workflow bundle (instruction files, memory templates/docs, and built-in workflow skills). Bare al init creates only operational scaffolding with empty instructions/ and skills/; answering "no" in the wizard leaves existing workflow files untouched. Answering "yes" refreshes managed bundled instruction files and workflow skills, creates missing 04_conventions.md and memory docs/templates, and preserves existing user-owned conventions and memory files.
  • al wizard now includes an opt-in CLI skill catalog for tavily-web, playwright-cli, find-docs, and agent-dispatch. Selected catalog skills are copied into .agent-layer/skills/<id>/; unselected catalog skills are removed. al doctor checks installed catalog skills for required binaries and reports missing tools without blocking agent launch.
  • al wizard now detects MCP servers in config.toml that are not part of the default catalog and asks about them in a dedicated step, separate from the catalog multiselect. Selected servers stay enabled; unselected servers are set to enabled = false with their definition preserved — disabling never deletes a custom server (it has no catalog template to restore from). The step is skipped when there are no custom servers, and the apply summary lists any custom servers being disabled.
  • al wizard --answers <file> for deterministic JSON-driven wizard runs. Answer files can script select, multi-select, confirm, input, and secret-input prompts; the runner bypasses terminal detection, rejects --profile/--yes conflicts, validates unknown fields, invalid options, multiple JSON values, missing prompts, and unused answers, and gives e2e coverage a stable way to exercise the real wizard flow without PTY automation.
  • al doctor now prints a context size summary after its checks: estimated instruction tokens, skill catalog-metadata tokens (against a ~4,000-token budget), MCP totals (enabled servers, total tools, total tool-schema tokens), and an estimated total of the always-loaded token costs (instructions + skill catalog + MCP tool schemas). Configurable metrics show their configured threshold; the skill catalog shows its fixed token budget. The summary always prints — even when values are under threshold and even with noise_mode = "quiet" or al --quiet doctor — because it is informational, not a warning. Thresholds left unset show (no limit set); components that can't be measured are named in an (excludes …) note on the total, and unreachable MCP servers are excluded from totals with a note. A one-off al --quiet doctor run suppresses warning-only doctor output while still surfacing failures.
  • Antigravity support via al agy, backed by agy --gemini_dir=<repo>/.agy and repo-local settings under .agy/antigravity-cli/.
  • al probe agy capability probe reports Antigravity permissions, MCP config migration, and runtime MCP discovery status as JSON.
  • Public best-practice website guides for skill design, CLI skill design, and instruction design. The release publisher now generates /skill-design, /cli-skill-design, and /instruction-design from the canonical repository docs with public page headers, and /best-practices links the guide set together.
  • config_delete_key and config_replace_string upgrade migration operations and a v0.10.2 migration that moves agents.gemini.enabled to agents.antigravity.enabled, rewrites MCP client lists from gemini to antigravity, deletes stale Gemini and retired Antigravity desktop config keys, and defaults Antigravity to disabled when no prior Gemini setting exists.
  • make al-agy developer convenience target.
  • make lint-ci-local developer target for no-Docker CI-parity golangci-lint runs. It uses disposable Go build, module, and golangci-lint caches with GOOS=linux GOARCH=amd64 CGO_ENABLED=0, and the command is documented in COMMANDS.md.

Changed​

  • The shipped al init template no longer disables Claude Code's AskUserQuestion tool by default. Fresh installs now allow the tool (matching Claude Code's native default) instead of seeding agent_specific.permissions.deny = ["AskUserQuestion"]. The wizard's new opt-in toggle instead sets agents.claude.disable_question_tool = true, and al sync injects the permissions.deny entry plus a PreToolUse hook into .claude/settings.json — merged with (never replacing) any deny/hook entries you already have. The hook is what enforces the block under YOLO/bypassPermissions, where permissions.deny is skipped entirely. Existing repos are unaffected (config.toml is never overwritten on upgrade); run al wizard to opt back in.
  • al wizard no longer deletes default MCP servers you unselect. Previously, unticking a catalog default in the wizard pruned its [[mcp.servers]] block (and any hand-customization) from config.toml. Now an unselected default that already exists is kept with enabled = false, matching how custom servers are handled — the wizard never deletes a server block. Selecting a default that is absent still adds it from the embedded catalog; leaving an absent default unselected leaves it absent. Fully removing a server is now a manual config.toml edit. Profile/--yes runs are unaffected.
  • The al upgrade migration report no longer prints [no_op] rows (migrations that ran but changed nothing because the target was already in the desired state). Only [applied] and [skipped_*] rows are shown. The report header, target/source versions, and any source-resolution notes still print so diagnostics are preserved. This removes the wall of noise that dominated upgrade output when most migrations were already satisfied.
  • The recurring agent-layer update available warning (shown on al sync and al <client> runs) now tells users how to turn it off: set version_update_on_sync = false under [warnings] in .agent-layer/config.toml. The default stays on, and al doctor's update check is unaffected.
  • Shipped agent instructions (01_base.md) now: direct agents to add real logging/instrumentation to the code to gather evidence on repeated failures instead of guessing; tell agents to zealously preserve context by delegating context-heavy work to subagents; and explicitly encourage scratch scripts and temporary files under .agent-layer/tmp for debugging.
  • Shipped instruction templates were also tightened: the hard-rule template now leads with grounding unknowns, several generic rules moved out of always-loaded rules, response-style guidance was made explicit, tool-routing now prefers local files/CLIs before MCP or web retrieval when local sources can answer, and relevant skills should be activated automatically from their descriptions.
  • The default MCP server catalog now contains the external-tool servers context7, tavily, fetch, and playwright. The previous ripgrep and filesystem catalog entries were removed from the embedded wizard catalog; existing hand-authored MCP server blocks are preserved, and custom servers are now handled by the dedicated custom-server wizard step.
  • The release website publisher now stages pages before copying, generates public guide pages from canonical Markdown sources, escapes plain-text angle brackets for MDX safety, builds guide tables of contents from real headings, and keeps versioned-doc publishing idempotent for the target tag.
  • CI and release workflows now pin third-party GitHub Actions to immutable commit SHAs with version comments across both workflows.
  • Shared skill projection now treats Antigravity as the supported shared-skill consumer in place of Gemini CLI.
  • Fresh al init now defaults [agents.antigravity] enabled = false (the prior true default was scoped to the retired Antigravity desktop launcher). Existing repos keep their migrated value from the v0.10.2 migration; users on the retired desktop launcher will have their pre-existing enable flag replaced by the rename from agents.gemini.enabled (or the new default if no Gemini config existed). The v0.10.2 row in site/docs/upgrades.mdx documents the replacement behavior.
  • Claude reasoning_effort no longer requires an Opus model. Agent Layer previously failed config validation for agents.claude.reasoning_effort unless agents.claude.model was an Opus variant — and rejected it outright when the model was unset. That hard error is removed: the value is passed through to Claude Code for any model (including when no model is set), and Claude Code is the authority on which model/effort combinations apply (e.g. Sonnet now supports effort levels). al wizard correspondingly offers the reasoning-effort prompt for any enabled Claude model instead of only Opus, and no longer clears the choice when you switch models. Copilot CLI reasoning_effort is still rejected, since that client exposes no effort control.
  • Wizard and installer status-line source handling now share the same exported source metadata, including canonical paths, legacy Claude source path, template path, and permissions, instead of maintaining separate mappings.

Fixed​

  • al doctor no longer warns merely because docs/agent-layer/ is absent in a bare-initialized repo. Bare al init intentionally does not create optional workflow memory docs; doctor now warns only when .agent-layer/instructions/*.md references docs/agent-layer and the directory is missing.
  • al doctor now reports the configured Copilot CLI enablement state. The default config enables [agents.copilot_cli], but doctor previously omitted that agent row even though sync generated Copilot CLI artifacts.
  • Copilot CLI's generated .copilot/mcp-config.json now keeps an explicit empty "mcpServers": {} object when no Copilot MCP servers are enabled, matching the documented shape and the other MCP client writers.
  • al wizard now redirects to al upgrade when config.toml contains a legacy key that only a migration can fix (e.g. a leftover [agents.gemini] table). Previously the wizard announced it would help fix the config, then ran sync, which re-validated strictly and hard-failed with a raw config validation failed error — a dead end, because the wizard's config patch preserves unknown sections verbatim and never runs the rename migration. The wizard now detects this class (a new ErrConfigNeedsUpgrade sentinel wrapped by config validation) and prints a clear "run al upgrade, then re-run al wizard" message, exiting cleanly instead of failing at sync.
  • al doctor's Antigravity version check now accepts the bare version string (e.g. 1.0.2) that agy --version prints in Antigravity 1.0.x. Previously the check required an agy-prefixed line (agy 1.0.0) and reported [FAIL] Could not parse Antigravity version against a working install. Multi-line build-timestamp noise is still rejected: the bare form is only accepted when the entire --version output is a single version triple.
  • Upgrade planning now treats wizard-managed CLI catalog skills as managed files instead of unknown files, and unpinned dev/legacy upgrades can trigger source-agnostic operations across the target-supported manifest chain when legacy Gemini config or missing source-agnostic defaults prove there is migration work to do. Pinned upgrades keep their normal source-to-target chain.
  • al wizard now reseeds a missing Claude or Codex statusline source when the effective post-wizard config has that provider's statusline enabled, even if you did not re-toggle the setting in the current run. Previously an enabled repo with a deleted source file could complete the wizard and then fail during sync.
  • Upgrade diff previews for a missing .agent-layer/claude-statusline.sh now show legacy .agent-layer/statusline.sh content when that legacy file will seed the new source, and the legacy source is classified as known so upgrade unknown-file scans do not prompt to delete it.
  • The Claude statusline now counts untracked-file line changes without spawning one git diff --no-index process per untracked file, avoiding prompt stalls in repos with many unignored files.
  • The upgrade-profile-overwrite-claude e2e scenario now scopes MCP sanitization assertions to the affected server blocks instead of scanning the whole config.toml, avoiding future false failures when another server legitimately uses similar fields.

Removed​

  • The al wizard "Default MCP server entries are missing from config.toml: … Restore them before continuing?" confirm prompt. Missing catalog defaults are now just unselected options in the MCP multiselect — select one to add it, leave it unselected to keep it absent. The prompt was effectively a no-op anyway (it defaulted to "yes" while the multiselect still showed missing defaults unselected, so the rendered config tracked the multiselect, not the answer).
  • Gemini CLI sync/client projection, including generated .gemini/settings.json, .gemini/policies/agent-layer.toml, the global ~/.gemini/trustedFolders.json write, and the root GEMINI.md instruction shim, has been replaced by Antigravity projection. The v0.10.2 migration cleans up any orphan GEMINI.md in existing repos. Historical release notes below remain unchanged.
  • al gemini subcommand removed entirely (no deprecation window). Existing scripts must switch to al agy; invoking al gemini now produces cobra's standard "unknown command" error.
  • Duplicate carry-forward Gemini-to-Antigravity migration operations from the v0.11.0 manifest. Unknown-source upgrades now discover those source-agnostic operations from their original v0.10.2 manifest through the supported-chain planner.
  • Internal buildCodexConfig test-only shim. Tests now call the System-aware Codex config builder directly.

v0.10.1 - 2026-05-17​

Adds diff-scoped cleanup skills (prune-new-tests, simplify-new-code) and wires fresh-context reviewer subagents into five existing skills to prevent narrative-driven rationalization. Introduces al init --here for in-place installs in subdirectories of existing repos, deep-merges Claude agent_specific configuration so permissions.deny is additive, auto-writes a Codex per-repo trust stanza, and adds a human-gated merge phase to ship-pr.

Added​

  • al init --here flag installs Agent Layer in the current directory without walking up to an ancestor .agent-layer/ or .git. Lets users add a separate .agent-layer/ inside a subfolder of an existing repo. When al init resolves to an already-initialized ancestor, the error now points at --here so the option is discoverable.
  • agent_specific.permissions.deny = ["AskUserQuestion"] shipped in the install seed (internal/templates/config.toml). Fresh al init now disables Claude Code's structured clarification-question tool by default; remove the line to keep it. Existing repos are unaffected.
  • prune-new-tests skill — burden-of-proof pruning of tests added in the current uncommitted diff. Each added test must defend its existence with a concrete production-code mutation that would flip its assertion or it is auto-deleted. Surviving coverage gaps are reported, never backfilled. Uses a fresh-context reviewer subagent so the implementer's narrative cannot rationalize speculative tests into surviving.
  • simplify-new-code skill — diff-scoped scope-creep removal. Scans the current uncommitted diff for agent-added speculative flexibility, premature abstractions, dead branches, impossible-case error handling, defensive scaffolding, clever patterns, and half-finished work, and auto-applies simplifications while preserving the user-requested behavior. Uses a fresh-context reviewer subagent that identifies scope creep by pattern, not by comparison to the request.

Changed​

  • Built-in skill frontmatter descriptions are shorter and more routing-focused, reducing Claude skill listing budget pressure while keeping key trigger language.
  • simplify-code skill renamed to simplify-codebase and scoped explicitly to the codebase (full repository or explicit paths). The implicit "if uncommitted changes exist, scope to the diff" branch is removed — simplify-new-code is the diff-scoped sibling. All in-repo callouts updated: diff-context skills (debug-issue, complete-current-phase, fix-issues, implement-plan, audit-and-fix-uncommitted-changes) now point at simplify-new-code; codebase-context skill (improve-codebase) points at simplify-codebase.
  • implement-plan skill inserts mandatory cleanup phases between implementation and verification: Phase 4 runs prune-new-tests whenever tests were added, Phase 5 runs simplify-new-code whenever production code was added or modified, and Phase 6 (renumbered from 4) handles plan-vs-implementation verification.
  • audit-and-fix-uncommitted-changes skill adds Phase 0.5 (Pre-pass Cleanup) that runs prune-new-tests and simplify-new-code before any review round, so reviewers don't spend budget on code about to be pruned.
  • verify-against-plan skill restructured around a plan-anchored, narrative-blind fresh-context reviewer subagent. Phase 2 now delegates plan-vs-implementation comparison to a subagent that sees only the plan and the post-implementation state — never the implementer's narrative, prior conversation, or deviation rationalizations.
  • address-pr-comments Phase 6 (reply audit) restructured around a fresh-context reviewer subagent. The original author's replies are audited by a subagent that sees only the comment, the reply, and (for Fixed verdicts) the named commit's diff — not the agent's prior reasoning when authoring the replies.
  • improve-codebase Phase 3 per-chunk re-audit restructured around a fresh-context reviewer subagent. The post-fix chunk and the originating findings are the only inputs; the fixer's narrative is excluded so sunk-cost reasoning ("we just fixed that") cannot rubber-stamp incomplete fixes.
  • Claude agent_specific is now deep-merged into .claude/settings.json for object values (arrays and scalars still replace at their key). Previously, top-level objects were replaced wholesale. permissions.deny is additive and does not trigger an override warning; permissions.allow continues to warn when present.
  • Codex sync now writes [projects."<repo root>"] trust_level = "trusted" to repo-local .codex/config.toml, preserving Codex's exact absolute-path trust semantics without requiring a per-repo agents.codex.agent_specific.projects passthrough. The agent_specific.projects override warning is path-aware: it only fires when the user's projects map contains the managed repo root (real collision), not when it lists unrelated paths that coexist with the managed entry.
  • ship-pr skill adds a human-gated Phase 9: the agent merges the PR only when the user replies with the exact phrase I approve merging PR #<N> matching the run's PR number, using an unambiguous GitHub merge method or pausing for a strategy choice when multiple methods are available; on a successful merge it deletes the source branch locally and remotely. The skill refuses to delete the repository's default branch.
  • ship-pr skill adds an upfront "Continuation rule" framing sub-skill returns as intermediate, not terminal. Addresses a recurring failure where the orchestrator stopped after audit-and-fix-uncommitted-changes returned, mistaking the sub-skill's closeout summary for ship-pr's completion.

v0.10.0 - 2026-05-07​

Consolidates skill projection into a shared .agents/skills/ directory for non-Claude clients, migrates Gemini sync to the Policy Engine, splits the MCP catalog from the install seed, adds xhigh reasoning effort for Claude, and improves the upgrade experience with automatic post-upgrade sync and opt-in diff preview.

Added​

  • Skills synced to a shared .agents/skills/<name>/SKILL.md tree for non-Claude clients (Codex, Gemini, Antigravity, VS Code/Copilot, Copilot CLI). Per-client directories (.codex/skills/, .gemini/skills/, .agent/skills/, .vscode/prompts/, .github/skills/) are retired and cleaned automatically by al sync. Projection rules and ownership contract documented in docs/SKILL-CLIENT-SPEC.md.
  • chat.agentSkillsLocations written to .vscode/settings.json pointing at .agents/skills/ so VS Code Copilot picks up the consolidated location.
  • xhigh as a valid reasoning_effort value for Claude. Custom (unknown) effort values now pass through with a warning instead of a hard validation error.
  • _generatedBy: agent-layer provenance field added to generated .mcp.json and .gemini/settings.json for stronger ownership detection during upgrade readiness checks; replaces the weak mcpServers-only signature.
  • Internal MCP server catalog (mcp-catalog.toml) embedded in the binary and consumed by al wizard for its MCP server multiselect; decoupled from the install seed so fresh config.toml files start with a minimal [mcp] section.

Changed​

  • Gemini sync migrates from the deprecated tools.allowed field to the Policy Engine. WriteGeminiSettings no longer emits tools.allowed in .gemini/settings.json and instead writes a policyPaths: [".gemini/policies"] pointer; WriteGeminiPolicies generates .gemini/policies/agent-layer.toml with one [[rule]] block per allowed command (toolName = "run_shell_command", commandPrefix, decision = "allow", priority = 100, allowRedirection = true). allowRedirection = true preserves the previous tools.allowed behavior for headless workflows that pipe output (e.g., git ... > file). Resolves the Gemini CLI deprecation warning.
  • al upgrade now runs al sync automatically on success so retired projection paths and freshly-introduced templates are reconciled without a manual follow-up. Sync warnings surface on stderr; sync failures are wrapped (upgrade applied; sync failed: <err> (run \al sync` to retry)) with errors.Is` preserved, and the "Upgrade successful." banner is suppressed when sync fails so the failure is unmissable.
  • al upgrade overwrite prompts now show a compact summary (file path with +N -M line stats, colorized when output is a terminal) and ask "View the full diff?" (default no) before printing unified diff bodies.
  • .agent-layer/tmp/ excluded from pre-upgrade snapshots; rollback does not restore tmp content. Interactive bulk-delete prompt is scoped to non-tmp content; a separate grouped prompt handles tmp unknowns. Non-interactive deletion requires --apply-tmp-deletions; --apply-deletions alone never touches tmp.
  • config.toml install seed no longer includes inline MCP server entries; the [mcp] section now points to al wizard for server configuration. Existing repos are unaffected.
  • Stale .claude/skills/ and .gemini/skills/ directories flagged for cleanup during upgrade readiness when those agents are disabled.

Fixed​

  • .mcp.json no longer omits the mcpServers key when no servers are configured.
  • Rollback no longer silently wipes .agent-layer/tmp/ content that was excluded from the snapshot.
  • Grouped tmp-unknowns deletion fallback fixed when DeleteUnknownTmpAllFunc is not wired in PromptFuncs.
  • Env preview redaction edge cases in al wizard hardened.
  • Doctor and wizard previews hardened for additional edge cases.

Removed​

  • tools.allowed field from generated .gemini/settings.json. The next al sync rewrites the file in the new shape, removing the deprecated key.
  • Per-client skill directories (.codex/skills/, .gemini/skills/, .agent/skills/, .vscode/prompts/, .github/skills/) retired; al sync cleans these paths and writes to .agents/skills/ instead.

v0.9.2 - 2026-03-21​

Adds GitHub Copilot CLI as a supported agent client, introduces context files for plan/task artifacts, supports max reasoning effort for Claude Opus, and consolidates internal abstractions for cleaner dependency injection. Instructions and skills are improved for better autonomy and tradeoff handling.

Added​

  • GitHub Copilot CLI integration: new al copilot command, sync support, config fields ([agents.copilot_cli]), wizard catalog entries, doctor checks, and v0.9.2 migration manifest. Stale Copilot artifacts (.copilot/mcp-config.json, managed skill dirs under .github/skills/) are cleaned when the agent is disabled.
  • Context file (.context.md) for plan/task artifact system. Captures key file paths, current state, constraints, and an entry point so implementing agents can orient immediately without re-discovering what the planner found. Produced by plan-work (Phase 3b), consumed by implement-plan, review-plan, verify-against-plan, and other plan-aware skills.
  • max as a valid reasoning_effort value for Claude Opus models. Since Claude Code only supports max as a session-scoped CLI flag, all effort values are now passed via --effort and max is excluded from settings.json sync.
  • Memory hygiene improvements in instruction and doc templates: "What NOT to store" section in 02_memory.md, character-budget awareness rule (~8k chars / ~2k tokens per file), completed-phase archival guidance in ROADMAP.md template, and entry-ID placeholder update from "abcdef" to "short-slug".
  • Explicit "Upgrade successful." message when al upgrade completes, resolving ambiguous output on no-op completions.
  • Nil guard for sys.HTTPClient() in downloadHTTPClientWithSystem, falling back to defaultHTTPClient when the System implementation returns nil.
  • INSTRUCTION-DESIGN.md internal reference document for instruction authoring principles.

Changed​

  • --effort CLI flag now respects agent_specific.effortLevel override: when the override is set, the managed --effort arg is skipped so the user's setting takes precedence.
  • Model catalogs updated: added gemini-3.1-flash-lite, opus[1m], gpt-5.4; removed deprecated gemini-2.0-*, gpt-5/5.1-*, claude-sonnet-4.5.
  • Instruction quality improvements: removed rules that duplicate baseline model behavior, strengthened tradeoff protocol to require at least two options with pros/cons, added memory pruning rule for DECISIONS.md and CONTEXT.md.
  • audit-documentation and audit-tests skills rewritten to fix autonomously with human checkpoints for genuine tradeoffs instead of dual report-only/fix-mode pattern. audit-tests now autonomously deletes rubber-stamp tests, consolidates duplicates, and combats agent-caused test bloat. All 22 skills receive a tradeoff checkpoint for standalone distribution.
  • ship-pr skill fixes two edge cases: on a non-default branch with no uncommitted changes, proceeds to create the PR; on the default branch with uncommitted changes, creates a new branch before committing.
  • Dispatch System interface expanded: replaced package-level mutable function stubs (osStat, osChmod, osRename, lockFileFn, flockFn, dispatchSleep, httpClient, etc.) with methods on the System interface for proper dependency injection.
  • Duplicate abstractions consolidated: approval mode constants unified into config.ApprovalMode*, agent-enabled helpers unified into config.IsAgentEnabled.
  • Skills docs page restructured: orchestrator/primary/supporting tiers with recommended workflow section, separated universal skill standard from Agent Layer-specific features.
  • Gitignore template updated to include open-vscode.sh and compiled al binary.

Fixed​

  • --effort flag no longer shadows user's agent_specific.effortLevel override in Claude Code.
  • ship-pr no longer stalls on a non-default branch with no uncommitted changes and no longer tries to commit on the default branch without creating a feature branch first.
  • Nil panic prevented when System.HTTPClient() returns nil in download paths.
  • Security: Updated github.com/modelcontextprotocol/go-sdk from v1.4.0 to v1.4.1 to fix improper handling of null Unicode character when parsing JSON (high severity).

Improved​

  • Expanded automated test coverage across Copilot CLI sync, dispatch System interface, model catalogs, gitignore templates, effort flag paths, and instruction template assertions.

v0.9.1 - 2026-03-07​

Overhauls the built-in skill library from 10 to 22 structured, workflow-driven skills and introduces a user-managed conventions file so you can tailor project-specific rules without losing them on upgrade. Instructions are reordered, deduplicated, and compressed for better agent compliance.

Added​

  • New user-managed 04_conventions.md instruction template for project-specific conventions (architecture, code quality, data safety, time/data, environment). Seeded on al init, never overwritten on al upgrade; future convention updates delivered via append_to_file migrations with duplicate detection.
  • New append_to_file migration kind for delivering content to user-managed files without overwriting edits. Supports duplicate-detection via match string, automatic file creation, and atomic writes.
  • New delete_file migration operations for removing deprecated skill directories during upgrade.
  • 17 new built-in skills following a normalized workflow-driven structure with explicit phases, global constraints, guardrails, and human checkpoints: address-pr-comments, audit-and-fix-uncommitted-changes, audit-memory, audit-tests, complete-current-phase, debug-issue, fix-ci, implement-plan, improve-codebase, plan-work, repair-checks, resolve-findings, review-scope, schedule-backlog, ship-pr, simplify-code, verify-against-plan.
  • Site documentation page for built-in skills (site/docs/skills.mdx) covering the full 22-skill library organized by category with usage examples and customization guidance.
  • Site documentation page for evidence-based skill design (site/pages/skill-design.mdx) with 5 core design principles and 18 academic/industry citations.
  • Internal skill authoring reference (docs/SKILL-DESIGN.md) and audit workflow specification (docs/SKILL-AUDIT.md).

Changed​

  • Instruction files reordered for primacy effect: 02_rules.md → 00_rules.md, 00_base.md → 01_base.md, 01_memory.md → 02_memory.md. Hard constraints now load first to improve model compliance. al upgrade migrates existing repos via rename_file operations.
  • Cross-file instruction duplicates removed: 6 items that appeared in multiple instruction files consolidated to a single canonical location, reducing ~57 instructions to ~50.
  • Verbose instruction sections compressed (~50% fewer tokens in Critical Protocol, Workflow & Safety, and Tools sections) without removing guidance.
  • UTC-only internals and No system Python rules moved from 02_rules.md to 04_conventions.md as project-specific conventions (delivered to existing users via append_to_file migrations).
  • 5 existing built-in skills restructured to normalized workflow pattern: audit-documentation, boost-coverage, finish-task, fix-issues, review-plan.
  • Decision hygiene guidance updated: superseded decisions should be replaced (not accumulated), and entries that become self-evident from the codebase should be removed.
  • Skill loader error message for directories missing a skill file shortened from "missing SKILL.md or skill.md" to "has no SKILL.md".
  • Site docs sidebar positions updated to accommodate new Skills page; upgrade-checklist shell examples consolidated.

Removed​

  • 5 built-in skills replaced by normalized workflow equivalents: cleanup-code (→ simplify-code), continue-roadmap (→ complete-current-phase), find-issues (→ resolve-findings), fix-tests (→ repair-checks), update-roadmap (→ schedule-backlog). Migration manifest deletes these (plus mechanical-cleanup) from user directories during upgrade.

Improved​

  • Expanded automated test coverage for user-owned instruction files (seed-on-init, no-overwrite-on-upgrade, excluded-from-diffs), append_to_file migration paths (apply, no-op, file creation, rollback), normalized workflow skill structure validation, artifact naming conventions, and skill deletion migration scenarios.

v0.9.0 - 2026-03-01​

Added​

  • Added a reusable internal/skillvalidator package with parse/validate separation and deterministic findings for agentskills.io-aligned skill validation.
  • Added al doctor skills diagnostics for standards checks (unknown frontmatter keys, name/path mismatches, and non-canonical directory filenames such as skill.md).
  • Added release manifests for v0.9.0: internal/templates/migrations/0.9.0.json and internal/templates/manifests/0.9.0.json.
  • Added al doctor check for stale flat-format skill files (.md at skills root) with guidance to run al upgrade.
  • Added CONTEXT.md memory file template for general-purpose project context, domain concepts, naming conventions, and lessons learned.
  • Added data-driven breaking-change display: migration manifests now carry breaking, breaking_notice, and breaking_details fields, and the upgrade report renders them generically instead of hardcoding per-kind display logic.
  • Added yellow highlighting for readiness warnings, file-removal counts, and review-needed items in upgrade plan and upgrade output for improved scanability.
  • Upgrade snapshot rollback now accepts snapshots in created status, enabling recovery from interrupted upgrades that failed before reaching applied status.

Changed​

  • Renamed legacy "slash command" source and output terminology to "skills" across config, sync pipelines, templates, and docs.
  • Canonicalized source layout to .agent-layer/skills/, with migrations that rename legacy .agent-layer/slash-commands/ and embedded skill template paths.
  • Breaking: Flat-format skills (<name>.md) are no longer supported by the skill loader. All skills must use directory format (<name>/SKILL.md). al upgrade migrates both built-in and user-authored skills automatically via a single migrate_skills_format operation with pre-flight conflict detection and user confirmation.
  • Skill frontmatter parsing/generation now uses YAML (go.yaml.in/yaml/v3) with support for name, description, license, compatibility, metadata, and allowed-tools, while keeping unknown fields parse-tolerant for portability.
  • Increased skill parser/validator single-line scanner caps to 8 MiB to reduce token-limit failures on large single-line skill content.
  • Documentation now explicitly states that missing or empty skill description is load-enforced (fail-loud), while missing name remains backward-compatible with doctor warnings.
  • Skills migration user-facing copy updated to "Slash-commands renamed to skills" with data-driven breaking-change notices sourced from the migration manifest.
  • Unknown-file scanning now covers both .agent-layer/ and docs/agent-layer/, with a fresh post-migration re-scan so the unknown-file prompt reflects actual post-migration state instead of stale pre-migration paths.
  • Migration-covered diff suppression now uses ancestor-directory matching, so migrations that own an entire directory (e.g., skills format migration) suppress noisy per-file template diffs in plan output.
  • Git safety instruction now clarifies that commit/push authorization applies only to the specific request and does not carry forward.

Fixed​

  • al doctor lenient-config fallback now best-effort loads skills, preventing false "No skills configured" results when strict config validation fails.
  • Skill name handling is now Unicode NFKC-aware across loading and validation paths, preventing false duplicates/mismatches for normalization-equivalent names.
  • Skill metadata and text limits now use rune counts (not bytes), and validation now rejects empty names and non-ASCII digit forms in slug normalization.
  • Directory-format loading now accepts lowercase skill.md as a compatibility fallback while preserving canonical SKILL.md precedence.

Removed​

  • Removed docs/agent-layer/SKILLS_WORKFLOWS.md; workflow guidance is now provided by individual skill sources.

Improved​

  • Expanded automated coverage across skill loading/validation, upgrade migrations, prompt generation, install ownership/readiness checks, and docs surfaces touched by the skills-standard migration.
  • Updated default embedded skills to canonical agentskills directory format (skills/<name>/SKILL.md) and aligned memory/workflow docs for Phase 15 completion.

v0.8.8 - 2026-02-25​

Added​

  • Added embedded upgrade migration manifest support for v0.8.8 to canonicalize the Claude VS Code agent key from agents.claude-vscode.enabled to agents.claude_vscode.enabled and enforce a default for the canonical key.
  • Expanded branch-coverage suites across upgrade/install, dispatch, wizard PTY interactions, and command paths to harden release confidence.

Changed​

  • Canonical config/docs naming now uses [agents.claude_vscode] (snake_case). Legacy [agents.claude-vscode] remains accepted at load time and is migrated during upgrade.
  • Release and upgrade-doc validation checks are stricter, including stronger release workflow/docs contract assertions and stable-tag publishing guardrails.
  • cmd/publish-site release version-retention and versioned-doc pruning behavior is more robust for deterministic stable release publication.

Fixed​

  • Made dispatch cache sync-error coverage deterministic by injecting sync failures in tests instead of relying on platform-specific /dev/null behavior.
  • Aligned upgrade/readiness messaging and docs references to the canonical claude_vscode key to avoid mixed-key guidance.

Improved​

  • Removed obsolete Claude-specific GitHub workflow files in favor of the current release/verification workflow set.
  • Updated README and site docs for key-name consistency and release guidance clarity.

v0.8.7 - 2026-02-24​

Added​

  • al <client> commands now support --quiet / -q for one-off quiet runs that suppress Agent Layer informational output while preserving client output and error exit behavior.
  • Added end-to-end coverage for quiet Claude runs to ensure al claude --quiet emits no Agent Layer output and still launches correctly.
  • Wizard back-navigation support: pressing Esc now moves to the previous step, with explicit first-step exit confirmation and deterministic state rollback behavior for partial selections.
  • Wizard Ctrl+C now exits immediately without saving, distinct from Esc (back). Both keys are shown as hints in the bottom navigation bar (esc back • ctrl+c exit).
  • Config guardrail test for required fields: automated enforcement now checks that newly required config fields have matching config_set_default migration coverage (with explicit legacy baseline allowlist).
  • Claude reasoning-effort support in wizard/config/sync paths (low|medium|high), including projection into .claude/settings.json as effortLevel.

Changed​

  • warnings.noise_mode now supports quiet in addition to default and reduce.
  • Quiet handling is now applied consistently across dispatch and client launch paths, including argument forwarding and no-sync execution.
  • Wizard-managed config.toml writes now use an explicit preferred section order policy (approvals, enabled agents, mcp, warnings) instead of implicitly coupling ordering to template parse order.
  • Upgrade diff output is now colorized in interactive terminals for better scanability (adds/removes/hunks), with plain-text fallback preserved for non-interactive and no-color environments.
  • Wizard profile apply flow now warns when replacing an existing TOML-corrupt .agent-layer/config.toml.

Fixed​

  • al sync warning-only outcomes in quiet mode now preserve non-zero exit behavior without printing warning text.
  • Quiet-mode behavior now avoids leaking dispatch/update-check banners when quiet is enabled via flag or config.
  • Prevented hidden behavior drift in multiline TOML patching by refactoring duplicated state tracking into a shared iterator with regression coverage across call sites.
  • Gemini reasoning_effort is now rejected explicitly with a clear validation error instead of allowing ambiguous/unsupported behavior.
  • Claude reasoning-effort validation now fails loudly for invalid option values and unsupported model selections.

Improved​

  • Documentation and default config comments now describe quiet-mode behavior and its interaction with al doctor (which always prints warnings).
  • Validation and warning messaging now includes quiet as a first-class supported noise mode.
  • Expanded test coverage for wizard back-navigation state transitions, profile corruption warning paths, reasoning-effort capability validation, Claude sync projection behavior, and upgrade diff color/no-color rendering.
  • Added PTY integration tests for wizard Esc/Ctrl+C keystroke classification, validating the full chain from raw terminal bytes through bubbletea to error classification.
  • Updated project memory/docs (ISSUES.md, BACKLOG.md, ROADMAP.md, DECISIONS.md, README.md) to reflect completed sprint scope and release-facing behavior.

v0.8.6 - 2026-02-23​

Added​

  • al wizard now prompts to enable per-repo Claude settings and caches isolation (local_config_dir) when Claude or Claude VS Code is enabled. Default is false (shared global config); selecting true sets CLAUDE_CONFIG_DIR to a repo-local directory for separate settings and caches per repository.

Fixed​

  • .gitignore template inline comments on /.claude/ and /.claude-config/ patterns were treated by Git as part of the literal pattern, causing both directories to not be gitignored. Comments moved to their own lines. Affected users (v0.8.5 installs): run al upgrade or al sync to pick up the corrected template.

Improved​

  • Per-repo isolation is now documented as a core feature in the README comparison table, key properties, and a dedicated Per-repo credential isolation section in the site concepts page (Codex auth isolation plus Claude settings and caches isolation; Claude auth remains shared due to an upstream limitation).

v0.8.5 - 2026-02-23​

Added​

  • Optional agent-specific passthrough configuration for Claude and Codex via agents.claude.agent_specific and agents.codex.agent_specific.
  • Optional agents.claude.local_config_dir support for repo-local Claude config isolation (.claude-config). al vscode sets CLAUDE_CONFIG_DIR only when both local_config_dir = true and agents.claude-vscode.enabled = true.
  • al upgrade rollback --list support to inspect available snapshot IDs and statuses before executing rollback.
  • al sync now auto-adds the repository root to ~/.gemini/trustedFolders.json when Gemini is enabled so Gemini CLI reliably loads project-level .gemini/settings.json. If this write fails, sync still succeeds and emits a non-fatal warning with manual remediation guidance.

Changed​

  • Agent-specific passthrough keys intentionally override Agent Layer-managed keys when they collide, with sync warnings to keep overrides explicit.
  • al vscode now clears CLAUDE_CONFIG_DIR only when it points at the repo-local .claude-config; user-defined non-repo values are preserved.
  • Codex sync now writes agents.codex.agent_specific keys before managed MCP tables so top-level overrides remain at the TOML root.

Fixed​

  • Config parsing now rejects unrecognized keys during strict decode instead of silently ignoring them, with actionable validation guidance in the returned error.
  • .env parsing now correctly handles quoted values, escaped newline/carriage-return sequences, and invalid trailing characters after quoted values.
  • MCP tool-name collision warnings are deterministic: warning subjects and per-tool server lists are sorted for stable output.
  • Upgrade config_set_default prompts no longer mark one choice as "recommended"; migration manifest values are still pre-selected but users must make an explicit choice.
  • al upgrade now accepts zero-byte snapshot file entries, fixing failures where unknown empty files previously produced requires content_base64 errors.
  • al gemini no longer fails MCP discovery for the internal agent-layer server when PATH al is a non-runnable repo-pin shim. Prompt-server command resolution now prefers local source execution (go run <repo>/cmd/al mcp-prompts) when available, al mcp-prompts bypasses repo-pin dispatch, and prompt-server source roots are validated as the Agent Layer module.

Improved​

  • Expanded automated coverage for root/upgrade command paths, Gemini trust flows, upgrade-readiness helpers, warning policy branches, prompt-server root resolution, dispatch behavior, and codex/claude agent-specific config rendering.
  • Added e2e coverage for upgrade flows that include empty unknown files and for wizard MCP sanitization behavior when profile defaults disable an injected server block.
  • Website documentation now includes al upgrade rollback --list guidance, Gemini trusted-folder remediation notes, and expanded agent-specific and repo-local Claude configuration guidance.

v0.8.4 - 2026-02-20​

Fixed​

  • Version dispatch no longer prints the version-source diagnostic twice when a pinned or environment-overridden version dispatches to a cached binary. The diagnostic now prints only from the binary that actually runs the command.

Changed​

  • Config validation now silently strips transport-incompatible MCP server fields instead of rejecting the config with an error. For example, headers on a stdio server or command/args on an HTTP server are removed during validation rather than causing a load failure. This makes configs more resilient to leftover fields from transport changes or manual editing.

Improved​

  • al wizard config patching now removes dotted sub-key lines (e.g., headers.Authorization = "Bearer ...") when stripping a parent key like headers. Previously, only inline-table syntax (headers = { ... }) was handled.
  • Scenario-based end-to-end test framework with 26 scenarios and 436 assertions replaces the previous monolithic e2e script. Covers fresh install, wizard profiles, upgrade paths, error propagation, agent launch, and rollback workflows with mock agent binaries.

v0.8.3 - 2026-02-19​

Fixed​

  • al doctor lenient config fallback now injects built-in environment variables (e.g., AL_REPO_ROOT), fixing false "missing environment variables" warnings for MCP servers like filesystem that reference ${AL_REPO_ROOT} in their args.
  • al wizard now sanitizes transport-incompatible MCP server fields during config patching. For example, leftover headers on a stdio server or leftover command/args on an HTTP server are automatically removed. Previously, the wizard would complete successfully but al sync would fail with a validation error, creating a circular "run wizard to fix" loop.

v0.8.2 - 2026-02-18​

Added​

  • Migration manifest chaining: al upgrade now loads all intermediate migration manifests between the source and target versions during multi-version jumps. Users upgrading from 0.8.0 to 0.8.2 will receive migrations introduced in intermediate releases.
  • Config resilience: al wizard, al doctor, and al upgrade now use lenient config parsing so they always work even on broken or incomplete configs. Runtime commands remain strict with actionable guidance.

Changed​

  • The agents.claude-vscode.enabled config migration has been moved from the v0.8.1 manifest to v0.8.2. This ensures all users (including those who installed v0.8.1 before the migration was added) receive the prompt during upgrade.

Fixed​

  • Users jumping multiple versions (e.g., 0.8.0 to 0.8.2) no longer miss intermediate migration operations.

Removed​

  • Slash command auto-approve frontmatter. Approval permissions are controlled entirely by approvals.mode.

Improved​

  • Slash command frontmatter now rejects unrecognized keys with a clear error, catching typos and unsupported fields early.

v0.8.1 - 2026-02-18​

Added​

  • [agents.claude-vscode] config section for Claude Code VS Code Extension support. al vscode is the single command for launching VS Code with both Codex and Claude extension settings based on which agents are enabled.
  • approvals.mode = "yolo" for maximum agent autonomy: skips all permission prompts (Claude --dangerously-skip-permissions, Gemini --approval-mode=yolo, Codex approval_policy=never + sandbox_mode=danger-full-access, VS Code chat.tools.global.autoApprove). Intended for sandboxed/ephemeral environments.
  • Slash command auto-approve frontmatter: auto-approve: true auto-approves MCP prompt retrieval for that skill in Claude clients. It does not auto-approve agent actions after reading the prompt; those remain governed by approvals.mode.
  • Upgrade readiness checks now detect stale disabled-agent artifacts for .claude/settings.json.

Changed​

  • al vscode now launches when either [agents.vscode] or [agents.claude-vscode] is enabled in config.toml, unifying both extensions under a single launch command.
  • CODEX_HOME is now cleared from the VS Code process environment when [agents.vscode] is disabled, preventing stale Codex configuration leakage when only Claude VS Code is enabled.
  • Release process now includes a preflight gate (make release-preflight RELEASE_TAG=vX.Y.Z) that validates upgrade-contract documentation before tagging.

Fixed​

  • al vscode no longer appends . when pass-through arguments include a positional path or file argument. (#51)
  • Sync warning exit behavior restored: non-suppressible warnings correctly propagate exit status regardless of noise_mode setting.

v0.8.0 - 2026-02-16​

Added​

  • New upgrade command surface centered on al upgrade: al upgrade plan (dry-run preview), al upgrade rollback <snapshot-id> (manual restore), al upgrade prefetch --version X.Y.Z (cache warm-up), and al upgrade repair-gitignore-block (managed block repair).
  • al upgrade plan now produces plain-language categorized upgrade previews with readiness checks and line-level diff previews (default 40 lines per file, configurable via --diff-lines).
  • Upgrade planning ownership inference is now deterministic and offline using committed release manifests (internal/templates/manifests/*.json) plus repo baseline state (.agent-layer/state/managed-baseline.json).
  • al upgrade now creates managed-file snapshots under .agent-layer/state/upgrade-snapshots/ and automatically rolls back transactional changes when an upgrade step fails.
  • Embedded per-release migration engine: al upgrade executes migration manifests (internal/templates/migrations/<target>.json) before template writes and emits deterministic migration reports.
  • al wizard now supports non-interactive profile mode (--profile, optional --yes) and backup cleanup (--cleanup-backups).
  • Warning noise control added via warnings.noise_mode (default or reduce) to reduce non-critical suppressible warning output when desired.

Changed​

  • Breaking: al init is now one-time scaffolding only. If .agent-layer/ already exists, al init errors and directs users to al upgrade plan + al upgrade.
  • Breaking: al upgrade non-interactive execution now requires --yes plus explicit apply flags (--apply-managed-updates, --apply-memory-updates, --apply-deletions) to make mutation intent explicit.
  • .agent-layer/.env and .agent-layer/config.toml are now strictly user-owned: seeded only when missing and never overwritten by init/upgrade operations.
  • .agent-layer/.gitignore is treated as internal agent-owned state: always rewritten from templates and excluded from upgrade plans/diff prompts.
  • Default MCP server templates now pin concrete tool versions by default, with inline floating/latest opt-in examples for teams that want automatic upstream updates.
  • Update-check handling now degrades gracefully on GitHub API rate limits so init/doctor flows continue with actionable warning output.
  • Release process now requires generating and committing a per-tag template ownership manifest before tagging (./scripts/generate-template-manifest.sh --tag vX.Y.Z).

Fixed​

  • al upgrade rollback <snapshot-id> now rejects path separators in snapshot IDs, preventing path traversal attempts during manual rollback resolution.
  • al vscode launch preflight now fails fast with explicit guidance when the code CLI is missing on PATH or when .vscode/settings.json has a managed-block marker conflict.

Removed​

  • Breaking: al init --overwrite and al init --force have been removed. Use al upgrade plan and al upgrade for upgrades/repairs.
  • Breaking: al upgrade --force has been removed. Use explicit apply flags (plus --yes for non-interactive runs) to select mutation categories.
  • Breaking: al upgrade plan --json has been removed; text output is now the only supported plan interface.

v0.7.0 - 2026-02-07​

Added​

  • Upgrade contract published at site/docs/upgrades.mdx: defines upgrade event categories (safe auto, needs review, breaking/manual), sequential compatibility guarantees (N-1 to N), release-versioned migration rules, and OS/shell capability matrix.
  • Release gate validates upgrade documentation for each release tag (make docs-upgrade-check), ensuring migration table rows exist and placeholder text is replaced when changelog notes breaking changes.
  • al init --version latest resolves the latest GitHub release to a semver pin before writing .agent-layer/al.version.
  • al init --version X.Y.Z validates the release exists on GitHub before writing the pin file, failing with a clear "release not found" message instead of writing a pin that 404s on next use.
  • al init auto-recovers from empty or corrupt .agent-layer/al.version pin files with a warning instead of blocking all commands.
  • Binary download progress indicator: ensureCachedBinary emits "Downloading al vX.Y.Z..." / "Downloaded al vX.Y.Z" to stderr.
  • Actionable error messages for binary download failures (404 not-found and timeout scenarios).

Changed​

  • Breaking: Windows support removed. Deleted al-install.ps1 installer, Windows release target, open-vscode.bat launcher, and all Windows-specific code paths in dispatch, cache, exec, and lock packages. Windows was never tested and best-effort support eroded trust. macOS and Linux remain fully supported.
  • al init now bypasses repo-pin binary dispatch and always executes on the invoking CLI binary, preventing older pinned versions from running upgrade operations.
  • Launcher template writes refactored for reliability with proper macOS path escaping.
  • Codex MCP header projection order corrected.
  • CI workflow caches pinned tools in GitHub Actions for faster builds.
  • Upgrade contract linked from README, site docs, DEVELOPMENT.md, and RELEASE.md.

Removed​

  • al-install.ps1 (Windows PowerShell installer).
  • open-vscode.bat (Windows VS Code launcher).
  • Windows release targets (windows/amd64) from build scripts.
  • Windows-specific dispatch, cache, exec, and lock code paths.

v0.6.1 - 2026-02-06​

Added​

  • CLI argument forwarding: al <client> now forwards extra arguments to the underlying client. Use -- to separate Agent Layer flags from client arguments (e.g., al claude -- --help or al vscode --no-sync -- --reuse-window).
  • VS Code launchers are now created during al init in addition to al sync, so launchers are available immediately after initialization.
  • .gitignore managed block is now updated during both al init and al sync operations for consistency.

Fixed​

  • AL_SHIM_ACTIVE environment variable no longer leaks into VS Code's integrated terminal when launching via al vscode. Previously, this caused subsequent al commands in the terminal to fail with "version dispatch already active" errors. (#46)
  • Wizard now rewrites config.toml sections in the template-defined canonical order, preventing section ordering drift after multiple wizard runs.

Changed​

  • Launcher code moved to internal/launchers package with exported EnsureGitignore for cross-package use.
  • Documentation updated with clearer guidance on gitignore template format, wizard behavior, and troubleshooting MCP server startup on macOS.

v0.6.0 - 2026-02-03​

Added​

  • Documentation website with comprehensive guides covering getting started, concepts (approvals, MCP servers, project memory, version pinning), reference (CLI, configuration, environment variables), and troubleshooting.
  • Website publishing pipeline (cmd/publish-site) with automated deployment in the release workflow.
  • Playwright MCP server template in default config.toml for browser automation workflows.
  • Descriptive comments for all default MCP server templates explaining purpose and required credentials.
  • Claude Code VS Code Extension added to supported clients table in README.

Changed​

  • README rewritten with clearer value proposition, comparison table (manual vs Agent Layer), and improved quick start flow.
  • Default MCP server examples in README now use generic example-api instead of GitHub-specific config for clarity.
  • Documentation structure consolidated from nested pages to flat MDX files for better navigation.

v0.5.8 - 2026-01-30​

Changed​

  • Breaking: Environment variables now require AL_ prefix to avoid conflicts with shell environment (e.g., GITHUB_PERSONAL_ACCESS_TOKEN → AL_GITHUB_PERSONAL_ACCESS_TOKEN). This ensures Agent Layer variables don't override existing environment variables when VS Code terminals inherit the process environment.

Fixed​

  • VS Code open-vscode.app launcher now uses osascript with a login shell (zsh -l) instead of hardcoded VS Code CLI paths, fixing launch failures when VS Code is installed via Homebrew, in ~/Applications, or other non-standard locations. This also fixes MCP server failures where VS Code couldn't find node because Finder-launched apps have a minimal PATH.
  • All VS Code launchers (.app, .command, .bat, .desktop) now delegate to al vscode for loading .agent-layer/.env, ensuring consistent parsing (KEY=VALUE data, not sourced) across platforms. Only AL_* variables with non-empty values are loaded, and existing environment variables take precedence—matching the documented behavior for al commands.
  • VS Code .app launcher now shows a descriptive alert when the code command is not found, instead of silently failing.
  • Linux .desktop launcher simplified to delegate to .command script for consistent behavior and maintainability.

v0.5.7 - 2026-01-29​

Added​

  • Custom HTTP header support for Codex MCP servers: bearer_token_env_var for Authorization: Bearer ${VAR}, env_http_headers for other env-var-sourced headers, and http_headers for static literals.
  • X-MCP-Tools header in default GitHub MCP server template for server-side tool filtering, reducing projected tool count.
  • Detailed per-tool token breakdown in al doctor MCP schema bloat warnings, showing top contributors by token count.
  • Documentation for MCP HTTP header projection across all supported clients (docs/MCP_HEADERS_SUPPORT.md).

Changed​

  • Default MCP schema token thresholds increased to accommodate larger MCP servers (server: 7500→20000 tokens, total: 10000→30000 tokens).
  • Doctor command now shows real-time discovery progress when checking MCP servers.
  • Large internal modules (install, dispatch, config) split into smaller, focused files for maintainability.
  • golangci-lint upgraded to v2.8.0 with additional linting rules enabled.

v0.5.6 - 2026-01-27​

Added​

  • http_transport config option for HTTP MCP servers to specify transport mode (streamable or sse).
  • Three new MCP server templates in default config.toml: fetch (mcp-server-fetch), ripgrep (mcp-ripgrep), and filesystem (server-filesystem with repo-scoped access).
  • ${AL_REPO_ROOT} built-in variable for resolving repository root path in MCP server args.
  • VS Code settings sync now preserves existing user settings and comments using JSONC-aware block insertion instead of overwriting the entire file.
  • Memory file templates (BACKLOG.md, COMMANDS.md, DECISIONS.md, ISSUES.md, ROADMAP.md) now include detailed formatting guidelines and entry templates.

Changed​

  • MCP projection refactored: new internal/projection/resolvers.go module centralizes server resolution logic, used by both sync and warning checks.
  • Update-available warning now includes full upgrade instructions for Homebrew, macOS/Linux shell script, and Windows PowerShell.
  • Instruction templates consolidated and shortened to reduce token count while preserving key guidelines.
  • Terminal detection moved to canonical internal/terminal package with IsInteractive() function.
  • Default MCP server templates no longer specify clients filter (servers are projected to all clients by default).

Fixed​

  • MCP server health checks now properly handle HTTP transport timeout scenarios.

v0.5.5 - 2026-01-25​

Added​

  • New 03_tools.md instruction template with comprehensive tool usage guidelines: time-sensitive information handling, Context7 documentation lookups, MCP tool constraints, approval workflows, and error handling.
  • New fix-tests slash command runs repo-defined checks (lint/format/pre-commit/tests) in a loop, fixing failures until all checks pass or max iterations reached.

Changed​

  • Temporary artifact location moved from tmp/agent-layer/runs/ to .agent-layer/tmp/runs/, keeping all agent artifacts within .agent-layer/.
  • Slash command artifact naming standardized across workflows: .agent-layer/tmp/<workflow>.<run-id>.<type>.md with run-id = YYYYMMDD-HHMMSS-<short-rand>. User path overrides removed for consistency.
  • finish-task workflow now delegates to fix-tests when available before falling back to manual repo-defined commands.
  • README updated with new artifact naming convention and VS Code reauthentication note for new CODEX_HOME environments.

v0.5.4 - 2026-01-24​

Changed​

  • Memory file FEATURES.md renamed to BACKLOG.md to better reflect its purpose (unscheduled user-visible features and tasks vs deferred issues).
  • al init --overwrite now detects and prompts to delete unknown files under .agent-layer that are not tracked by Agent Layer templates.
  • al init --force now deletes unknown files under .agent-layer in addition to overwriting existing files without prompts.
  • Memory instruction templates improved with clearer formatting rules and entry layouts.
  • Slash command templates (continue-roadmap.md, update-roadmap.md) simplified and clarified.
  • VS Code launcher paths centralized in internal/launchers package, consumed by sync and install to prevent drift.
  • Sync package refactored with system abstraction layer for improved test isolation and reliability.

v0.5.3 - 2026-01-24​

Changed​

  • User-facing strings consolidated into internal/messages/ package for consistency and maintainability.
  • Python release tools (extract-checksum.py, update-formula.py) replaced with Go implementations in internal/tools/.
  • Release test script reorganized into modular components (scripts/test-release/release_tests.sh, scripts/test-release/tool_tests.sh).
  • Slash command templates (find-issues.md, finish-task.md) simplified to reduce duplication with base instructions; formatting rules now delegate to individual memory file templates.

v0.5.2 - 2026-01-24​

Added​

  • Automated Homebrew tap updates: release workflow now opens a PR against conn-castle/homebrew-tap to update the formula with the new tarball URL and SHA256.

v0.5.1 - 2026-01-23​

Added​

  • Source tarball (agent-layer-<version>.tar.gz) published with releases for Homebrew formula support.

Changed​

  • Release scripts now generate and verify the source tarball via git archive + gzip -n.
  • Documentation cleanup: simplified release process, corrected make dev description.

v0.5.0 - 2026-01-23​

Major shift from repo-local binary to globally installed CLI with per-repo version pinning.

Added​

  • Global CLI installation via Homebrew (brew install conn-castle/tap/agent-layer), shell script (macOS/Linux), or PowerShell (Windows).
  • al init command initializes .agent-layer/ and docs/agent-layer/ in any repo.
  • Per-repo version pinning via .agent-layer/al.version; global CLI dispatches to the pinned version automatically.
  • Cached binary downloads with SHA-256 verification; cached binaries stored in ~/.cache/agent-layer/versions/.
  • Shell completion for bash, zsh, and fish (al completion <shell> with optional --install flag).
  • Update checking: al init and al doctor warn when a newer release is available.
  • Linux desktop entry launcher (.agent-layer/open-vscode.desktop).
  • E2E test suite (scripts/test-e2e.sh) and release test script (scripts/test-release.sh).
  • Environment variables: AL_CACHE_DIR (override cache location), AL_VERSION (force version), AL_NO_NETWORK (disable downloads).

Changed​

  • Breaking: Repo-local ./al executable replaced with globally installed al CLI.
  • Breaking: al install renamed to al init.
  • Breaking: Repository moved from nicholasjconn/agent-layer to conn-castle/agent-layer.
  • Install script renamed from agent-layer-install.sh to al-install.sh.
  • al init --overwrite now prompts before each overwrite; use --force to skip prompts.
  • al init --version <tag> pins the repo to a specific release version.
  • Commands run from any subdirectory now resolve the repo root automatically.
  • .agent-layer/.gitignore added to ignore launchers, template copies, and backups.

Removed​

  • Repo-local ./al binary; global al dispatches to pinned versions as needed.
  • agent-layer-install.sh (replaced by al-install.sh).

v0.4.0 - 2026-01-21​

Added​

  • al doctor command reports missing secrets, disabled servers, and common misconfigurations.
  • al wizard command provides interactive setup for approval modes, agent enablement, model selection, MCP servers, secrets, and warning thresholds.
  • Configurable warning system with thresholds for instruction token count, MCP server/tool counts, and schema token sizes.
  • Antigravity slash commands now generate skills in .agent/skills/<command>/SKILL.md.
  • VS Code launchers: macOS .app bundle (no Terminal window), macOS .command script, and Windows .bat file, all with CODEX_HOME support.
  • al install --no-wizard flag skips the post-install wizard prompt.
  • Atomic file writes across all sync operations prevent partial file corruption.

Changed​

  • al install now prompts to run the wizard after seeding files (interactive terminals only).
  • Gitignore patterns use root-anchored paths (/AGENTS.md instead of AGENTS.md) for precision.
  • Default Codex reasoning effort changed from xhigh to high.
  • Codex config header now warns about potential secrets in generated files.
  • Environment variable loading: process environment takes precedence; .agent-layer/.env fills missing keys only; empty values in .env are ignored.
  • Improved instruction and slash-command templates.

Fixed​

  • VS Code launcher now works correctly with proper error messages for missing code command.
  • MCP configuration for Codex HTTP servers now handles bearer token environment variables correctly.

v0.3.1 - 2026-01-19​

Added​

  • Installer failure output now includes clear, actionable error messages.

Fixed​

  • Installer checksum verification now handles SHA256SUMS entries with "./" prefixes.

Changed​

  • Quick start documentation no longer suggests manual install fallback when only ./al is present.

v0.3.0 - 2026-01-18​

Complete rewrite in Go for simpler installation and fewer moving parts.

Added​

  • Single repo-local Go binary (./al) replaces the Node.js codebase.
  • al install command for repository initialization with template seeding.
  • al install --overwrite flag to reset templates to defaults.
  • al sync command to regenerate client configs without launching.
  • Support for five clients: Gemini CLI, Claude Code CLI, VS Code/Copilot Chat, Codex CLI, and Antigravity.
  • Unified [[mcp.servers]] configuration in config.toml for both HTTP and stdio transports.
  • Approval modes (all, mcp, commands, none) with per-client projection.
  • ${ENV_VAR} substitution from .agent-layer/.env with client-specific placeholder syntax preservation.
  • Internal MCP prompt server for slash command discovery (auto-wired into client configs).
  • Golden-file tests for deterministic output validation.
  • Managed .gitignore block with customizable template (.agent-layer/gitignore.block).

Changed​

  • Breaking: Complete rewrite from Node.js to Go.
  • Breaking: Configuration moved from config/agents.json to .agent-layer/config.toml (TOML format).
  • Breaking: MCP servers now configured via [[mcp.servers]] arrays in config.toml.
  • CLI simplified: ./al <client> always syncs then launches.
  • Instructions now in .agent-layer/instructions/ (numbered markdown files, lexicographic order).
  • Slash commands now in .agent-layer/slash-commands/ (one markdown file per command).
  • Approved commands now in .agent-layer/commands.allow (one prefix per line).
  • Project memory standardized in docs/agent-layer/ (ISSUES.md, FEATURES.md, ROADMAP.md, DECISIONS.md, COMMANDS.md).

Removed​

  • Node.js codebase (src/lib/*.mjs, test files, package.json).
  • config/agents.json and separate MCP server configuration files.
  • Built-in Tavily MCP server (now configurable as external server in config.toml).

v0.2.0 - 2026-01-17​

Major architectural overhaul moving core logic from shell to Node.js.

Added​

  • Per-agent opt-in configuration via config/agents.json with interactive setup prompt.
  • HTTP transport support for MCP servers.
  • Tavily MCP server for web search capabilities.
  • ./al --version flag with dirty suffix for non-tagged commits.
  • User config preservation and backup during upgrades.

Changed​

  • Breaking: CLI entrypoint is now .agent-layer/agent-layer; ./al remains as the launcher wrapper in the parent root.
  • Root resolution, environment loading, and cleanup moved from shell to Node.js (src/lib/roots.mjs, src/lib/env.mjs, src/lib/cleanup.mjs).
  • Test framework migrated from Bats (shell) to Node.js native test runner.
  • GitHub MCP server switched to hosted HTTP endpoint with PAT authentication.
  • Architecture documentation updated to reflect new layer boundaries.

Removed​

  • Shell scripts: al, run.sh, setup.sh, clean.sh, check-updates.sh, open-vscode.command.
  • Shell-based root resolution: src/lib/parent-root.sh, src/lib/temp-parent-root.sh.

v0.1.0 - 2026-01-12​

Initial release.

Added​

  • Installer for per-project setup that pins .agent-layer/ to tagged releases, with upgrade, version, and dev-branch options.
  • Repo-local ./al launcher with sync and environment modes plus local update checks.
  • Sync pipeline that generates client configs from .agent-layer/config sources.
  • MCP prompt server that exposes workflows as prompts.
  • Project memory templates and setup/bootstrap helpers.