Glue

Changelog ​

The canonical changelog lives in the CLI package — cli/CHANGELOG.md — and is rendered below. Entries under Unreleased are shipping on the main branch and not yet tagged.

All notable changes to Glue CLI will be documented in this file.

[Unreleased] ​

Added ​

  • MCP OAuth that actually works on real servers. Glue now follows the current MCP authorization spec end-to-end (RFC 9728 protected resource metadata + RFC 8414 authorization server metadata + RFC 7591 DCR + PKCE), with the bits that make it usable in practice:
    • Auto-detect on 401. A remote MCP server returning 401 Unauthorized with a WWW-Authenticate: Bearer … challenge triggers OAuth automatically — you no longer need to set auth: {kind: oauth} in ~/.glue/config.yaml up front. The first successful login writes that back for you.
    • Silent token refresh. When the access token expires, Glue runs the refresh-token grant transparently and retries the request. The user only sees the OAuth flow when the refresh token itself has expired or been revoked.
    • McpAwaitingAuth connection state. Servers that fail with 401 are parked in a new dedicated state — distinct from McpDead — that does not consume the reconnect budget or arm a retry timer. Status bar reads MCP:1🔑 when the only unhealthy servers are waiting on auth, vs MCP:1⚠ for mixed/dead.
    • Context-aware /mcp panel actions. Selecting an HTTP/WS server in the /mcp panel surfaces Authenticate, Re-authenticate, or Sign out depending on credential state. stdio servers never get auth actions.
    • Auto-reconnect after sign-in. Once the loopback callback captures the authorization code and tokens are stored, Glue immediately reconnects the parked server without any extra command.
    • Cached discovery URLs. After the first successful login the server's resource_metadata_url and authorization_server are written into ~/.glue/config.yaml so subsequent sessions skip the discovery probe.
    • Granular credential invalidation. A new invalidateMcpAuth helper drops just the part of the credential that failed (tokens, client, or all), so a refresh-token failure doesn't force a re-DCR.
    • McpAuthFlowRunner shared across the CLI command, the slash command, and the auto-triggered flow — one place that owns discovery → DCR → loopback → token exchange → persist.
    • Bind-then-register order so DCR registers the exact loopback redirect URL with the bound port. Strict auth servers (SmartBear) that require an exact redirect_uri match — RFC 8252 says they shouldn't, many do — now work. With DCR available, Glue re-registers on every login to keep the URI in sync; the cached client_id path is reserved for pre-registered (no-DCR) servers.
  • AgentNotice event in glue_core — provider-neutral soft- degradation signal carrying a message + kind (info / warning). Rendered as a dim system line in the TUI, sent as a marker-prefixed AgentMessageChunkUpdate over ACP, and printed to stderr with a glyph prefix in --print mode. First user: the tool-less Ollama soft fallback below.
  • ToolsNotSupportedException in glue_core — typed exception thrown by OllamaClient when the daemon returns 400 with a body matching does not support tools, replacing the previous raw Exception. Other adapters can throw the same shape if they ever expose tool-less models.
  • web_browser.navigate accepts a wait_until parameter so the agent can self-correct: "load" (default), "domcontentloaded" (fastest, DOM parsed), "networkalmostidle" (≤2 in-flight for 500 ms), or "networkidle" (strict). Invalid values are rejected before the browser is provisioned. Implemented in packages/glue_harness/lib/src/tools/web_browser_tool.dart.

Changed ​

  • glue serve renamed to glue acp — the ACP server subcommand now matches the convention used by every other ACP-speaking agent (opencode acp, goose acp, …). No backward-compat alias: editors with "args": ["serve"] in ~/.config/zed/settings.json, ~/.jetbrains/acp.json, or equivalent must update to "args": ["acp"]. Help text, the WebSocket startup banner, and docs/advanced/acp-server all reflect the new name.
  • Dart SDK floor bumped to 3.12.0 across every package (cli, glue_core, glue_strategies, glue_runtimes, glue_harness, glue_server). Enables the 3.12 private-named-parameter syntax and is required for pub workspaces.
  • Monorepo is now a pub workspace. A root pubspec.yaml lists all six packages and each member opts in with resolution: workspace. dart pub get from the repo root resolves the whole tree into a single shared pubspec.lock and .dart_tool/. Per-package dart pub get still works.
  • Analyzer config consolidated. One canonical analysis_options.yaml at the repo root; each package's analysis_options.yaml is reduced to include: ../../analysis_options.yaml. Picks the looser unawaited_futures: false everywhere (was mixed) and adds unnecessary_to_list_in_spreads: true to every package (previously cli-only).
  • Build output relocated to <repo>/dist/glue (was cli/dist/bundle/bin/glue). just cli::build produces a single AOT binary at the repo root and the intermediate bundle/ is cleaned up. CI release workflow renames per-platform (dist/glue-linux-x64, etc.) and uploads from there. just cli::install now symlinks <repo>/dist/glue → ~/.local/bin/glue.
  • Constructor boilerplate trimmed via Dart 3.12 private initialising formals. ~30 constructors across cli/, glue_harness, glue_strategies, glue_runtimes, and glue_server migrated from : _field = field initializer lists to this._field parameters. Call-site API is unchanged (the public parameter name is the field name without the underscore).
  • unawaited(...) wrappers removed in lib/. With unawaited_futures: false the wrappers were pure noise; 17 call sites across cli/lib, glue_harness, glue_server, and glue_strategies were unwrapped. Test files were left alone (intent-signaling use is still useful there).

Fixed ​

  • Tool-less Ollama models no longer crash the agent. When the active model lacks function-calling support (e.g. ollama/deepseek-coder-v2:16b, ollama/qwen2.5:7b), Glue used to surface Ollama's 400 "does not support tools" as a raw exception mid-session. Now the agent loop catches this case, disables tools for the rest of the session, emits a one-line notice into the transcript, and continues in chat-only mode. The same fallback fires at boot when the catalog declares the model without tools capability, skipping the wasted first turn. Subagent spawns with a tool-less model override degrade rather than killing the parent session. glue doctor adds a new info finding ("Agent model") when the active model is catalogued without tool support.
  • web_browser navigate no longer times out on ad/tracker-heavy sites. Default wait condition changed from Puppeteer's strictest Until.networkIdle (requires the page to make zero network requests for 500 ms — never resolves on sites with continuous beacons like nrk.no, GA, etc.) to Until.load (the page load event fires). Matches Playwright's default and the de-facto MCP browser-server convention; Playwright's own docs flag networkidle as "DISCOURAGED". Previously, navigating to such sites surfaced a TimeoutException: Navigation Timeout Exceeded: 30000ms exceeded to the agent even though the page had visibly loaded and was fully usable.
  • bin/glue serve tests no longer break when run from the workspace root.cli/test/bin/glue_serve_test.dart and glue_serve_ws_test.dart previously hardcoded Directory.current.path as the working dir for the spawned glue process, which fails when the test runner is launched from the repo root via dart test cli. A _cliPackageRoot() helper now resolves the cli package from either cwd.

Removed ​

  • Dormant OllamaShow + ToolCapabilityProbe scaffolding — a 7-lens swarm review of the original hard-fail-with-probe design concluded the runtime catch approach is strictly better (matches Continue.dev, Aider/LiteLLM, Codex, OpenCode; no per-boot /api/show round-trip; no second source of capability truth). The probe code is replaced by the soft-fallback path above.
  • Dead win32 direct dependency dropped from cli/pubspec.yaml — it was declared but never imported. (The four code_assets/hooks/mason_logger/ native_toolchain_c pub-outdated warnings remain: they're pulled in transitively via cli_completion → mason_logger → win32 and can't move until those packages bump upstream.)

[0.6.0] - 2026-05-20 ​

Added ​

  • TTY/NO_COLOR-aware brand markers + colorized MCP / session output — every brand glyph (●, ✓, !, ✗, ·) and every styled string from command output now routes through a new styledOrPlain() helper that collapses ANSI to plain text when stdout is not a terminal or when NO_COLOR is set. glue mcp list, glue mcp tools, glue mcp auth status, glue mcp add, glue session list, glue session show, and glue session apply/export now render with the same brand-dot header, bold ids, and severity-marker status lines as glue catalog and glue doctor. Existing surfaces (--where, catalog *, doctor, serve) had their inline .styled chains migrated to styledOrPlain so the same piping safety applies everywhere. Output of glue mcp list | grep enabled is now grep-friendly with no ANSI noise.
  • CLI output formatting style guide — new docs/design/cli-output-formatting.md codifies the brand vocabulary (● for headers, ✓/!/✗/· for status), the four-layer rendering pipeline (Command → *_format.dart → styledOrPlain → Styled), output shape per command class (diagnostic vs action vs pure-config), stdout/stderr discipline, JSON/--print/NO_COLOR semantics, and how to test the formatter. Includes a "drift register" table tracking which surfaces are fully on the spec. Linked from CLAUDE.md so new command work picks it up.
  • OTLP session.id resource attribute — every Glue invocation now emits a stable per-process session.id (format glue-<base36-ts>-<base36-rand>) on every OTLP/HTTP trace export, so observability backends that follow the OpenInference convention (llmflow, Phoenix, Langfuse, Helicone, Opik) can group multiple traces from the same Glue session under one "session" view. Implemented in packages/glue_harness/lib/src/observability/otlp_http_trace_sink.dart as a late final field initialised once per process.
  • /mcp tools and glue mcp tools now list every server when no argument is given — output is grouped by server with per-server status annotations (connected is unmarked; connecting, reconnecting, disconnected, dead, and disabled get a tag in parentheses), and a friendly per-server reason when a server has no tools to show (e.g. disabled; enable to list tools). The single-server form (glue mcp tools <id>) still exits 1 when the named server is disabled, so existing scripts that gate on the exit code keep working. CLI and slash share a single formatter in cli/lib/src/commands/mcp_tools_format.dart, with formatMcpToolsByServer taking pure value objects so the renderer is unit-testable without spinning up a real McpClientPool. CLI variant now waits for all selected servers to settle (with a 10 s cap), so the no-arg listing prints a single coherent snapshot rather than only the first server to respond.
  • glue mcp add --help ships five worked examples — Playwright via npx, GitHub via docker with a PAT env var, Context7 hosted HTTP with no auth, GitHub Copilot hosted HTTP with bearer, and a generic OAuth-via-DCR server. Discoverable via the command's description in the help output.
  • Interactive /mcp panel — pressing Enter on a server row in the status panel now opens an action submenu with Reconnect, Enable/Disable for this session, View tools, Copy server ID, and Show last error (when a failure is recorded). Actions reuse the pool's existing reconnect() / toggle() methods, so the panel and the slash subcommands stay interchangeable. Tools and error views render as scrollable read-only modals.
  • Tab completion for /mcp subcommands and server IDs — /mcp <TAB> enumerates subcommands, /mcp reconnect|toggle|tools <TAB> enumerates configured server IDs (case-insensitive prefix), and /mcp auth login|logout <TAB> filters down to HTTP/WebSocket servers (stdio can't OAuth). Adds mcpSubcommandCandidates, mcpAuthSubcommandCandidates, and mcpServerIdCandidates to arg_completers.dart, wired through the standard SlashArgCompleter override on McpSlashCommand.
  • Automatic MCP reconnect with backoff — when an MCP server's initial handshake or transport fails, the pool now transitions to McpReconnecting and schedules a retry via the existing mcpBackoff helper (exponential with jitter, 500ms→30s over 10 attempts by default; tunable via mcp.reconnect.* in config.yaml). Only after max_attempts consecutive failures does the server land in McpDead. McpPoolServerDisconnectedEvent now carries the populated reconnectAttempt + nextAttemptIn fields it always had room for, so the status bar and panel can render the retry countdown. Manual /mcp reconnect <id> and /mcp toggle <id> cancel any pending retry timer and reset the attempt counter — pool tests cover both paths. Closes the docs-vs-reality gap flagged in #26.
  • Editor Integration (ACP) docs + branded glue serve output — new website/docs/advanced/acp-server.md page covers what ACP is, the two glue serve transports (stdio for editors, WebSocket for browser/notebook clients), the full flag reference, and copy-pasteable configs for Zed (official), JetBrains AI Assistant 2025.3+ (official), VS Code (formulahendry.acp-client), Neovim (agentic.nvim + alternatives), Emacs (agent-shell), and marimo/use-acp/agent-client-kernel. glue serve --help now ends with a usageFooter pointing at that page. glue serve --port prints a brand-styled startup banner (● glue serve + indented url/auth/docs/stop rows) instead of the old single-line [glue serve] log, matching the shape of glue catalog show and glue doctor. Wired into the Advanced sidebar between Web Tools and MCP Servers.
  • glue mcp add | remove | enable | disable — manage MCP server entries from the shell instead of hand-editing ~/.glue/config.yaml. add takes --transport stdio|http|ws, accepts stdio commands after a -- separator (glue mcp add foo --transport stdio -- node srv.js), HTTP/WS servers via --url, env vars via -e KEY=value, and a --disabled flag to park a server until you run glue mcp enable <id>. remove clears stored bearer/OAuth credentials by default (opt out with --keep-credentials). Mutations go through the new McpConfigWriter so comments, key order, and formatting in config.yaml are preserved. Verb set mirrors Gemini CLI and Copilot CLI; transport grammar mirrors Claude Code / Amp.
  • glue catalog open and glue catalog edit — open launches the configured catalog.remote_url (or the canonical GitHub raw URL) in the default browser via open/xdg-open/rundll32; --print emits the URL only for piping. edit opens ~/.glue/cache/models.yaml (or $GLUE_CATALOG_CACHE) in $EDITOR with inherited stdio, warns and hints at glue catalog refresh when the cache is missing, and errors out when $EDITOR is unset. Both reuse the brand-dot header and ✓ · ! ✗ severity markers from refresh/show/path so the catalog surface stays visually consistent.

Fixed ​

  • glue catalog refresh now writes YAML, not single-line JSON. The remote-catalog sanitizer used to round-trip the upstream document through jsonEncode after stripping credential-leak vectors — technically valid YAML, but unreadable in glue catalog edit. It now uses YamlEditor to surgically remove disallowed provider fields and clamp auth.api_key: none in place, preserving the upstream's block structure and comments.
  • glue mcp tools <server> no longer hangs for ~10 s when the server is disabled. The transient pool skips disabled servers during connectAll, so the command was waiting for the connect/error event that never came. It now short-circuits with a clear warning pointing at glue mcp enable <id>.
  • In-app transcript selection and copy — drag-select text in the output zone, release to copy to the clipboard. Selection is anchored to (blockId, plain-text offset) so it survives streaming chunks and terminal resize without pointing at the wrong text. Drag is reported via xterm ?1002 button-event tracking; SGR modifiers/motion are surfaced on MouseEvent. Shift-drag is passed through to the terminal as a native-selection escape hatch.
  • Double-click selects a word, triple-click selects a line — click-chain detection (300 ms window, cell-exact) mirrors the convention used by VS Code / Zed / token-editor. Word boundaries use the same three-class model as token-editor (whitespace / word / punctuation), so foo_bar selects as one identifier, a lone . between identifiers selects just the dot, and CJK / emoji / combining marks stay atomic. Both gestures auto-copy.
  • Ctrl+Shift+C copies the current selection — Ctrl+C is left alone so it still cancels in-flight agent work (you often select text because the agent is misbehaving). Esc clears an active selection without falling through to autocomplete-dismiss or cancel-agent.
  • OSC52 clipboard transport — when running under tmux or SSH, clipboard writes go through OSC52 (with the tmux passthrough envelope when TMUX is set) instead of relying on host pbcopy/clip/wl-copy. Host commands stay the default outside multiplexers; OSC52 also acts as the fallback when host commands fail. Payloads >74 KB skip OSC52 cleanly.
  • Transient copy-confirmation toast — successful copies surface a narrow charcoal chip at the top-right of the output viewport (✓ Copied 3 lines, yellow glyph on dim text) for ~1.8 s. Failures show a red-glyph variant for ~3.5 s. Painted directly into the viewport as a content-sized rect so it doesn't blank the row of transcript behind it, and never written to _blocks or the session log — the transcript stays clean.

Internal ​

  • New cli/lib/src/terminal/brand.dart centralises the brand dot (● in RGB 250,204,21) and the ✓ · ! ✗ severity markers used by glue catalog, glue doctor, glue serve, and glue --where. Replaces the private copies that had been duplicated across catalog_command.dart and doctor.dart, so all branded surfaces share a single source.
  • New cli/lib/src/app/transcript_selection.dart houses the coordinate model (TranscriptPosition, TranscriptSelection), drag gesture state, char-class helpers (classify, findClassRange), and the ClickChain synthesiser. Render pipeline keeps a per-frame plainOutputLines shadow + line→block anchor list to support hit-testing and plain-text extraction without touching Layout.paintOutputViewport.
  • applySelectionHighlight (in ansi_utils.dart) is a new cell-aware ANSI splicer used to wrap selected ranges with reverse-video while preserving wide glyphs and combining marks.

[0.4.1] - 2026-05-19 ​

Added ​

  • Bootstrap error classification — BootstrapException now carries a typed kind (auth, network, saml, missingBinary, prep, upload, cloneBundle, clone, checkout, unknown) and a remediationHint. The clone path pattern-matches git stderr to assign the right kind, so failures surface as "sandbox couldn't authenticate to the remote — switch to bundle bootstrap or inject an HTTPS token" instead of BootstrapException(stage: clone, exit: 128).

  • Bare-repo refusal — buildHostBundle refuses bare/mirror clones with a clear message instead of producing a silently empty bundle.

  • Submodule warning — bundle bootstrap warns when host has .gitmodules since submodule contents are not transferred (gitlink pointer only). Recursive submodule fetch would require re-introducing sandbox-side auth, so this stays a warning until a full solution lands.

  • glue doctor host-git check — surfaces a warning at the Runtime section when host git is missing, since that disables bundle bootstrap and forces all cloud sessions through the clone-from-remote fallback.

  • glue session CLI subcommand surface — list, show, diff, apply, export. apply defaults to creating a branch glue/<session-id> from current HEAD and runs git am --3way (falls back to git apply --3way for working-tree-only patches); pass --in-place to apply on the current branch (Q6 default resolved). Refuses to apply truncated patches.

  • Session meta now persists runtime info — runtime_id, sandbox_id, runtime_bootstrap_sha, runtime_remote_url, runtime_patch_path, runtime_closed_at. Lets /session, the glue session … commands, and a future cleanup sweep reason about prior cloud sessions without scanning the filesystem.

  • /session shows cloud runtime info when the session is running in a cloud sandbox (runtime id, sandbox id, where the patch will land on close).

  • findOrphanedRuntimeSessions helper — detects cloud sessions whose runtime_closed_at is null and start time is > 24h old (likely leaked sandboxes), exposed for a future glue runtime cleanup command.

  • Cloud runtime bootstrap captures the host working tree via a git bundle, replacing the clone-from-remote-only path that lost uncommitted edits, unpushed commits, untracked files, and locked out non-git workspaces entirely. New host-side helper builds a single-commit bundle in a temp --git-dir overlay (host's actual .git is never touched), uploads it to the sandbox via the runtime's existing writeFile primitive, and clones from the bundle inside. Bundle SHA becomes the bootstrapSha so the diff layer is unchanged. Per-runtime upload caps: Daytona 200 MB, Modal 30 MB, Sprites 3 MB. Hosts without git, or bundles exceeding the cap, fall back to clone-from-remote. Resolves W1–W5, T1–T4, A1–A4 from the correctness plan + Q4 default (no .glueignore, respect host .gitignore via git add -A).

  • Runtime diff outcomes are typed — RuntimeSession.diffSinceBootstrap now returns a sealed RuntimeDiffOutcome (Success / Empty / Unavailable(reason)) instead of a nullable string. Surfaces show a warning at session shutdown when the diff couldn't be captured (Sprites resumed without a baseline, Modal sandbox auto-terminated, runtime workspace isn't a git repo, etc.) — no more silent nulls.

  • runtime.patch.meta.json sidecar — every saved runtime patch now has a metadata sidecar with runtime_id, sandbox_id, bootstrap_sha, remote_url, runtime_cwd, format, captured_at, size_bytes, and truncated. Apply tools and glue session … (forthcoming) read this instead of re-inferring context from the patch body.

  • Patch size cap — runtime patches are capped at 50 MB by default; a larger diff is written to runtime.patch.truncated with a visible warning so the user can investigate without flooding the session directory.

  • Sprites dirty-resume refusal — resuming a sprite whose /workspace has uncommitted changes from a previous session now refuses with a remediation message instead of silently producing a null baseline (which dropped every subsequent diff). Resolves Q1 in docs/plans/2026-05-19-cloud-runtimes-correctness-plan.md.

  • Modal sandbox death detection — diffSinceBootstrap preflight-checks the sidecar before attempting git diff, so an auto-terminated Modal sandbox produces a clear executorDead warning instead of a generic transport exception.

  • Runtime command events — every executor (host / docker / daytona / sprites / modal) now emits RuntimeCommandStarted / RuntimeCommandCompleted / RuntimeCommandFailed / RuntimeCommandCancelled when constructed with a RuntimeEventSink. Threaded through RuntimeFactory.create({eventSink}). Opt-in — null sink is free.

  • End-of-session workspace diff for cloud runtimes — cloud RuntimeSessions implement diffSinceBootstrap() by running git -C /workspace diff <bootstrapSha> inside the sandbox on session shutdown; the result is saved to <session-dir>/runtime.patch so the user can review (or apply) the agent's edits after a cloud run.

Changed ​

  • Runtime workspace diff is now an mbox (runtime.mbox, not runtime.patch) produced by git format-patch --binary -M -C plus a working-tree git diff --binary -M -C HEAD. An add -N preamble guarantees untracked files survive. Result: agent commits keep their authorship + message (apply with git am --3way), binary files round-trip byte-for-byte, renames stay as renames, and files the agent created but didn't git add no longer vanish. Round-trip integration test verifies all three (packages/glue_runtimes/test/common/diff_roundtrip_test.dart). Resolves Q3 default.

[0.4.0] - 2026-05-18 ​

Added ​

  • Cloud runtimes — three remote sandbox adapters, transparent to the agent (same bash, file tools, background jobs as host/Docker).
    • Daytona (runtime: daytona) — REST over the control plane; per-sandbox toolboxProxyUrl discovered automatically; US + EU regions. Workspace bootstrapped via git clone or tarball into /workspace. DAYTONA_API_KEY for auth.
    • Sprites (runtime: sprites) — persistent Fly.io sandbox via the sprite CLI. Resumes by name; auto-sleeps when idle. Authenticates through sprite login.
    • Modal (runtime: modal) — Modal sandbox via an embedded Python sidecar speaking JSON-RPC over stdin/stdout (Modal's sandbox primitive is Python-SDK-only). Sandbox auto-terminates on sandbox_timeout_seconds to cap runaway billing.
  • Runtime selection via runtime: YAML key or GLUE_RUNTIME env var (precedence: env → YAML → legacy docker.enabled fallback).
  • /runtime slash command — shows the active runtime, its key config (image / sandbox name / API base URL), and registered cloud adapters.
  • glue doctor runtime checks — per-runtime block for host / docker / daytona / sprites / modal; reports API-key presence, CLI install/auth status, and config sanity.
  • RuntimeFactory.register(name, adapter) — pluggable cloud adapter registration. cli/bin/glue.dart registers daytona / sprites / modal at startup; downstream forks can register their own.
  • RuntimeSession umbrella type — bundles executor + workspace
    • sandbox metadata (id, sandboxId, bootstrapSha, resumed) + close() lifecycle hook used to stop cloud sandboxes on session end.

Changed ​

  • Architecture refactor to support cloud runtimes:
    • File tools (ReadFileTool, WriteFileTool, EditFileTool, ListDirectoryTool) now route through Workspace instead of dart:io. GrepTool routes through CommandExecutor.
    • ShellJobManager operates on RunningCommandHandle instead of dart:io.Process directly, so background jobs work uniformly across runtimes.
    • New shared TransportWorkspace + RuntimeFsTransport lets each cloud adapter implement just a thin transport — workspace logic (path translation, WorkspaceAccessError, list anchoring) is centralized.
    • Workspace bootstrap (git clone or tarball + SHA recording) lives in shared WorkspaceBootstrap used by all three cloud adapters.
  • Monorepo split: glue_runtimes/ package houses the cloud adapters and shared cloud-runtime utilities. The cli depends on it only for register*Runtime() — no cloud SDK leaks into glue_harness.
  • RuntimeApiException unified the previous per-adapter *ApiException classes; carries runtimeId + endpoint for observability.

[0.3.0] - 2026-05-15 ​

Added ​

  • MCP (Model Context Protocol) client — connect to any MCP server and surface its tools to the agent alongside Glue's built-ins.
    • stdio transport with scrubbed environment (allowlist + explicit env: only) so user secrets don't leak into spawned servers.
    • Streamable HTTP transport (2025-03-26 spec): single POST, server picks JSON vs SSE response. Bearer auth, captured Mcp-Session-Id echoed back on subsequent calls.
    • WebSocket transport for ws:// / wss:// servers.
    • OAuth 2.1 with discovery (RFC 8414), Dynamic Client Registration (RFC 7591), PKCE, loopback redirect. Tokens stored encrypted in CredentialStore. Run glue mcp auth login <server>.
    • McpClientPool with eager non-blocking connect, exponential backoff with jitter for reconnect, crash-loop detection, and live tools/list_changed reactivity.
    • Namespaced tools: <serverId>__<bareName> (e.g. playwright__browser_navigate). Native tool names always win on conflict.
  • glue mcp CLI — list, tools <server>, auth set --bearer, auth login, auth logout, auth status.
  • /mcp slash commands — status panel, list, tools <server>, reconnect <server>, toggle <server>, auth login|logout|status, help. Status bar shows an MCP:N⚠ badge when servers are unhealthy.
  • tool_policy for MCP — glob auto_approve / deny patterns scoped to the namespaced tool name; routes through the same PermissionGate as native tools.

Changed ​

  • App decomposition — the cli/lib/src/app/ part-of split has been collapsed back into a single app.dart. The part-of files were cosmetic (extension methods on App with full access to its private state), so the split was hiding coupling rather than reducing it. app.dart is now one honest 2100-line class.
  • Slash command refactor finished — SlashCommandContext lost the last domain-leaking callbacks (forkSession, resumeFromMeta). HistoryCommand and ResumeCommand now compose primitives through ctx.session.fork/resume + ctx.conversation.resetForReplay() + appendReplayEntries(...) directly. New ConversationView.resetForReplay and appendReplayEntries methods own the transcript-shape state.
  • PanelController renamed to ModalSurface across the codebase.

[0.2.0] - 2026-05-04 ​

Added ​

  • Harness-layers architecture — Glue is now a four-layer monorepo: cli/ (surface) → packages/glue_harness/ (orchestration) → packages/glue_strategies/ (provider/shell/web adapters) → packages/glue_core/ (pure data types). A separate packages/glue_server/ ships the ACP-over-stdio/WS daemon. See docs/plans/2026-04-29-harness-layers.md for the full split.
  • glue serve — ACP (Agent Client Protocol) server with stdio and WebSocket transports (--stdio, --port N). Typed content blocks, resource_links, diff support, image inputs, OAuth device flow, and session/usage_summary.
  • Prompt caching across Anthropic + OpenAI + OpenRouter — the adapters now opt into provider-specific cache directives, surfaced via /usage and the OTel spans.
  • End-to-end token usage tracking — main agent, subagents, and title generator each report into UsageStats; /usage shows the per-role breakdown; resume carries totals over instead of restarting at zero.
  • Thinking-token streaming for reasoning-mode models (Anthropic thinking, OpenAI o-series).
  • Native Gemini provider — first-class GeminiProvider adapter in packages/glue_strategies/ talking to the Gemini Developer API (generativelanguage.googleapis.com) via streamGenerateContent SSE. Function calling, image inlineData, system prompts, and thinking-mode thoughtSignature round-trip are all wired in. Auth via GEMINI_API_KEY only (Vertex / Google-account login intentionally out of scope).
  • Two-press SIGINT in --print / --json mode — first Ctrl+C cancels the in-flight agent stream so the JSON envelope can include cancelled: true and the OTel span records the cancellation; the process exits 130. A second Ctrl+C during teardown hard-exits 130. Replaces the previous loop with StreamIterator so cancellation flows from the SIGINT handler into the agent's async* generator. See docs/reference/sigint-handling.md.
  • /copy slash command — copies the most recent assistant block to the system clipboard via the existing fallback chain (pbcopy / clip / wl-copy / xclip / xsel). Reports byte count on success; mid-turn /copy qualifies the message as "partial in-flight response".
  • /model aliases /models — single command surface; both names open the picker, and /model <query> switches directly.
  • Catalog refresh (2026-04) — adds OpenAI o3 and o4-mini reasoning models, Mistral magistral-medium-latest plus a Small 4 promotion, Groq Llama 4 Scout (default) and Maverick, Ollama qwen3-coder-next:80b, Ollama llama4:8b, and a new "reasoning" profile.
  • enabled flag on ModelDef — surfaces that aren't yet wired up at runtime can hide entries via the picker without removing them from the catalog.
  • Native OllamaAdapter — Ollama now has a dedicated ProviderAdapter + OllamaClient that talks to /api/chat directly instead of riding the OpenAI-compat adapter at /v1/chat/completions. Errors correctly attribute to Ollama (no more "OpenAI API error 404" on missing Ollama models), and Ollama-specific request options now have a home in the native client.
  • options.num_ctx injection — OllamaClient now sets options.num_ctx = min(ModelDef.contextWindow, 131072) on every request when the catalog knows the model's context window. Fixes the silent agent-loop truncation when Ollama falls back to its 2048-token default. Clamped at 128K so catalog entries that claim 1M contexts don't OOM mid-range GPUs. Uncatalogued passthrough models (user-typed tags) omit options and get Ollama's default.
  • Exact-match model resolver — new lib/src/catalog/model_resolver.dart is the single source of truth for turning a user-typed identifier into a ModelRef. Explicit <provider>/<id> inputs are never fuzzy-matched: catalogued refs return the catalog entry, uncatalogued refs pass through to the provider verbatim. Bare inputs require an exact match against id or display name; ambiguous bare inputs return the candidate list and unknown bare inputs return a clear error. The previous substring fallback silently rewrote gemma4 into gemma4:26b; it's gone.
  • Status bar + /info show the wire address — the right-hand status segment now reads <provider> · <apiId> (what the provider actually receives), not the internal catalog key. /info expands to <name> — <provider>/<apiId> for catalogued models and falls back to <provider>/<modelId> for uncatalogued passthrough. Surfaces apiId drift (e.g. Groq's gpt-oss-120b goes out as openai/gpt-oss-120b).
  • Ollama dynamic discovery — lib/src/providers/ollama_discovery.dart merges Ollama's /api/tags into the /model picker. Catalogued+pulled entries render normally; catalogued but not-pulled show [pull]; tag-only locally-installed models show [local] and are synthesised into picker rows. 2 s timeout, 30 s in-memory cache. Fail-soft: picker falls back to the bundled catalog silently when the daemon is down.
  • "Pull this model?" confirmation flow — selecting a not-installed Ollama tag (via the /model picker or by typing ollama/<tag> directly) opens a ConfirmModal. On Yes, Glue streams POST /api/pull progress as system messages and only switches the active model after a {"status":"success"} frame. On No or on pull failure, the active model stays put.
  • Clean ConfigError surface at startup — bin/glue.dart now catches ConfigError and ModelRefParseException, writes a single-line Error: … to stderr, and exits with code 78 (EX_CONFIG) instead of dumping a Dart "Unhandled exception" stack trace.

Changed ​

  • glue --resume now mirrors the common CLI pattern used by Claude and Copilot. Bare glue --resume opens the resume panel. Passing an argument (glue --resume <id> or glue --resume=<id>) resumes that session directly. Trailing positional text is preserved as the next prompt, so glue --resume <id> "continue here" resumes and immediately submits continue here. Print mode still rejects bare --resume because the panel is interactive-only.
  • Ollama no longer masquerades as OpenAI-compat. The ollama provider in docs/reference/models.yaml now uses adapter: ollama (not adapter: openai + compatibility: ollama), and base_url drops the legacy /v1 suffix. Catalog regenerated.
  • /model picker picks the user's intent, not a substring match. See "Exact-match model resolver" under Added. Errors list candidates for ambiguous inputs instead of silently choosing one.

Fixed ​

  • Docker executor test now skips when the daemon is down. The skip guard changed from docker --version (CLI-only check) to docker info (requires the daemon), so the test stops failing on machines that have Docker installed but not running.

Removed ​

  • CompatibilityProfile.ollama — dead after the native adapter move. CompatibilityProfile.fromString('ollama') now falls through to openai (kept for a forgiving parse; the native adapter handles the real wiring).

[0.1.1] — 2026-04-20 ​

Added ​

  • Build metadata injection — just build now passes GLUE_BUILD_TIME, GLUE_GIT_SHA, GLUE_GIT_DIRTY, and GLUE_BUILT_BY to dart compile via --define flags. BuildInfo reads them at startup; glue --version prints a compact summary (e.g. glue v0.1.1 (a1b2c3d, 2026-04-20T…)), glue --version --debug prints a detailed block, and --debug emits a banner to stderr before the TUI launches. Dev builds (dart run) fall back to (dev). just release now builds after the tag commit so the binary carries a clean SHA.
  • glue config init — non-interactive config initializer that writes an annotated v2 config.yaml template to the resolved Glue home (~/.glue/config.yaml or $GLUE_HOME/config.yaml). Supports --force overwrite/reset behavior, and /config init now delegates to the same real config writer instead of creating an empty ./config.yaml.
  • glue config path and glue config validate — scriptable config utilities for printing the resolved config.yaml path and validating the active config/provider credential setup.
  • glue doctor — read-only install/config diagnostic command that reports resolved paths, parse errors for config/preferences/credentials/catalog files, active config validation, malformed session files, orphaned temp files, and returns non-zero when errors are found. Output uses the Glue brand header (yellow ●), bold section headings, and coloured severity glyphs (✓/·/!/✗). Informational findings (e.g., empty session directories missing conversation.jsonl) are hidden by default — pass --verbose/-v to surface them.
  • DuckDuckGo search provider — zero-config duckduckgo search backend that scrapes the HTML endpoint (html.duckduckgo.com/html/), decodes the uddg redirect parameter to surface clean result URLs, and requires no API key. Registered in the default search provider chain. Includes unit tests for the HTML parser and an opt-in live integration test (dart test --run-skipped -t integration test/integration/duckduckgo_search_integration_test.dart).
  • Hyperbrowser backend for web_browser — new hyperbrowser browser backend provisions Hyperbrowser cloud sessions (POST /api/session), connects over the returned wsEndpoint CDP WebSocket URL, surfaces the live view URL in browser tool output, and stops the remote session on disposal (PUT /api/session/{id}/stop). Configured with web.browser.backend: hyperbrowser plus HYPERBROWSER_API_KEY or web.browser.hyperbrowser.api_key. Includes docs/config examples and an opt-in live smoke test: dart test --run-skipped -t hyperbrowser test/integration/hyperbrowser_e2e_test.dart.
  • Anchor Browser backend for web_browser — new anchor browser backend provisions Anchor Browser cloud sessions, connects over the returned CDP WebSocket URL, includes the live view URL in browser tool output, and stops the remote session on disposal. Configured with web.browser.backend: anchor plus ANCHOR_API_KEY or web.browser.anchor.api_key. Includes docs/config examples and an opt-in live smoke test: dart test --run-skipped -t anchor_browser test/integration/anchor_browser_e2e_test.dart.
  • ModelDef.apiId — optional field in docs/reference/models.yaml decoupling the stable catalog key from the mutable upstream identifier. Defaults to the YAML key when omitted (zero churn for simple entries). Adapters send model.apiId on the wire; the YAML key stays as the URL-safe, user-facing identifier in configs, sessions, and the /model picker. Motivation: upstream renames (e.g. Groq dropping Qwen3-Coder) no longer invalidate user configs. Matches precedent from OpenRouter, LiteLLM, Continue.dev, MLflow, and Docker.
  • Ollama catalog registry integration test — test/integration/ollama_catalog_registry_test.dart verifies every Ollama catalog tag resolves via registry.ollama.ai. Opt-in via dart test --run-skipped -t ollama_registry; skipped by default to stay hermetic. Catches future drift when model authors rename tags.
  • glue --where — prints GLUE_HOME and every resolved path (config.yaml, preferences.json, credentials.json, models.yaml, sessions/, logs/, cache/, skills/, plans/) with an exists / not-yet marker. Useful on a fresh install to see where Glue will read and write.
  • $GLUE_HOME environment variable — overrides the default ~/.glue root. Every on-disk path (sessions, logs, cache, credentials, config) moves with it. Documented in docs/reference/config-yaml.md and the Installation / Configuration docs.
  • Installer script at getglue.dev/install.sh — POSIX sh, detects linux|macos|windows × x64|arm64, pulls the matching binary from the latest GitHub Release, verifies the .sha256, drops it in ~/.local/bin (overridable via --dir or $GLUE_INSTALL_DIR). Supports --version vX.Y.Z for pinned installs.
  • ToolCallPhase.cancelled — distinct from denied (never ran) and error (ran but failed). Agent cancel now marks every in-flight tool — including ones awaiting approval — as cancelled instead of overloading error.
  • TUI behavior contract — docs/reference/tui-behavior.md codifies the alt-screen, scrollback, resize, tool-phase, spinner, focus-priority, wrapping, and glyph rules the TUI follows today.
  • CLI prompt arguments and print mode — pass prompts directly from the command line for non-interactive use:
    • glue "review my code" — positional prompt argument
    • glue -p "query" — print mode, streams response to stdout without TUI
    • glue -p --json "query" — JSON output for scripting
    • glue -r <session-id> — resume a specific session by ID
    • Model aliases: glue -m opus resolves to claude-opus-4-6
  • DevTools instrumentation — consolidated observability with dart:developer integration:
    • DevToolsSink bridges observability spans to developer.log() and developer.postEvent() for Dart DevTools visibility
    • GlueDev utility: CPU profiler tags (tagRender, tagLlmStream, tagToolExec, tagAgentLoop), Timeline helpers, DevTools URL helper
    • TTFB (time to first byte) tracking in ObservedLlmClient
    • Tool call name and response preview tracking in observability spans
  • Session replay improvements — proper tool_call/tool_result grouping during session resume and fork:
    • Assistant messages now include their associated tool calls
    • Tool results are properly paired with their tool_use IDs
    • Orphaned tool_result messages are filtered from conversation history, preventing Anthropic API 400 errors
  • Subagent output improvements — richer grouped output display:
    • _SubagentEntry class with optional JSON pretty-printing for expandable tool result content
    • _SubagentGroup shows current tool name during execution
    • Expanded view renders indented JSON for structured results
  • Multimodal tool results — tools can now return images (e.g. browser screenshots) as native content blocks instead of base64 text. This reduces token usage from ~738K text tokens to ~1,600 vision tokens per screenshot.
    • ContentPart sealed class hierarchy (TextPart, ImagePart) for structured content in tool results.
    • Tool.execute() returns List<ContentPart> — single dispatch method for all content types (text and images).
    • ForwardingTool base class for decorators (Go embedding pattern) — new Tool methods auto-forward to all decorators.
    • Provider-native image formats: Anthropic inline image blocks in tool_result, OpenAI/Ollama follow-up user messages with image_url/images arrays.
    • WebBrowserTool screenshots now return ImagePart instead of base64 text strings.
  • Observability & debug system — pluggable telemetry with zero changes to business logic. Three wrapper layers instrument all activity without polluting core code:
    • LoggingHttpClient — wraps http.Client, logs every outbound HTTP request (method, URL, status, duration).
    • ObservedLlmClient — wraps LlmClient, tracks provider, model, message count, token usage, tool calls, and latency per generation.
    • ObservedTool — wraps each Tool, tracks execution time, args, and output size.
  • /debug slash command — toggles verbose debug mode at runtime. Also available via --debug / -d CLI flag or GLUE_DEBUG=1 env var.
  • File-based debug logging — daily-rotating log files at ~/.glue/logs/glue-debug-YYYY-MM-DD.log with timestamped entries for HTTP calls, LLM generations, tool executions, and span lifecycle.
  • Observability facade — composite dispatcher that fans out to multiple sinks. Sinks are independently enabled/disabled. The facade provides log(), startSpan(), flush(), and close() methods.
  • ObservabilityConfig — configuration model for debug and telemetry settings, parsed from config file and env vars following the existing resolution chain.
  • Terminology standardization — established canonical glossary (docs/architecture/glossary.md). Two-level hierarchy: Project (registered directory) → Session (resumable agent conversation). Sessions are independent and carry their own model, worktree, branch, and conversation history. No intermediate "workspace" layer — matches the data model used by Claude Code, Codex CLI, and Cline.
  • Enhanced SessionMeta (schema v2) — sessions now store rich metadata: project_path, worktree_path, branch, base_branch, repo_remote, head_sha, title, tags, pr_url, pr_status, token_count, cost, summary. All new fields are optional; schema v1 files are read with permissive defaults. Added SessionMeta.fromJson factory for consistent deserialization and SessionStore.updateMeta() for mid-session metadata writes. Timestamps now consistently use UTC.
  • Model registry & picker — curated ModelRegistry catalog of 10 models across Anthropic, OpenAI, Mistral, and Ollama with capability, cost, and speed metadata. /model with no args opens a selectable panel picker grouped by provider; /model <name> does fuzzy lookup by ID or display name. Only models with configured API keys are shown.
  • GlueConfig.copyWith — immutable config update for provider/model switching.
  • LlmClientFactory.createFromEntry — create an LLM client directly from a ModelEntry.
  • Spinner animation in status bar during LLM streaming — braille dot pattern cycles at 80ms instead of static ● indicator.
  • Collapsible subagent output — subagent activity is now grouped by task into compact summary lines (e.g. ↳ [1/3] task… (5 steps…)) that update in-place. Click a group to expand/collapse its full step log.
  • Alt+Backspace deletes previous word (same as Ctrl+W).
  • Alt+Left / Alt+Right word-level cursor navigation in input editor.
  • KeyEvent and CharEvent carry an alt flag for modifier-aware input handling.
  • web_fetch tool — fetches a URL and returns clean markdown for the LLM. Three-stage pipeline: (1) try Accept: text/markdown header, (2) HTML fetch → Readability-style content extraction → HTML-to-markdown conversion, (3) optional Jina Reader API fallback. Configurable timeout, max bytes, and token budget. Auto-approved (read-only).
  • web_search tool — searches the web via configurable providers (Brave, Tavily, Firecrawl) with unified result model. Auto-detects provider from available API keys (priority: Brave → Tavily → Firecrawl) with automatic fallback on error. Supports explicit provider selection via parameter. Configured via web.search.* in config.yaml or BRAVE_API_KEY/TAVILY_API_KEY/FIRECRAWL_API_KEY env vars.
  • WebConfig — web tool configuration model with WebFetchConfig and WebSearchConfig, wired into GlueConfig with env var and config file resolution following existing patterns.
  • Hidden aliases for slash commands — SlashCommand now supports hiddenAliases that resolve on execution but are excluded from autocomplete and /help. /q is now a hidden alias for /exit.
  • Multi-shell support — unified CommandExecutor abstraction with HostExecutor that respects the user's shell via $SHELL, GLUE_SHELL/GLUE_SHELL_MODE env vars, or shell.* in config.yaml. Supports bash, zsh, fish, pwsh with correct flag mapping for interactive/login/non-interactive modes.
  • Docker sandbox — DockerExecutor runs agent commands in ephemeral docker run --rm containers with bind-mounted directories. Uses cidfile-based container termination with retry for race conditions. ExecutorFactory handles Docker availability detection with automatic host fallback. Configurable via docker.* in config.yaml or GLUE_DOCKER_* env vars.
  • Session-scoped Docker mounts — SessionState persists directory whitelist additions in state.json per session, merged with config mounts at executor creation.
  • /models command — lists available models from the current provider (Ollama /api/tags, OpenAI /v1/models, Anthropic /v1/models). Shows model name, size (Ollama), and marks current.
  • E2E integration tests — headless agent loop tests via AgentRunner with real Ollama (qwen3:1.7b). Tagged @e2e, skipped by default, run with dart test --run-skipped -t e2e. Retry wrapper handles small-model non-determinism.
  • Tool call intent indicator — the UI now shows ▶ Tool: write_file (preparing…) as soon as the LLM begins generating a tool call, rather than waiting for the full arguments to stream. Tool calls progress through visible phases: preparing → running → done (or denied/error). Eliminates the "hanging spinner" feel during large tool argument generation.
  • ToolCallStart LLM chunk — Anthropic and OpenAI clients now yield a ToolCallStart chunk at content_block_start / first tool delta, surfacing the tool name before arguments finish streaming.
  • Mistral LLM provider — fourth provider alongside Anthropic, OpenAI, and Ollama. Uses OpenAI-compatible API with Mistral-specific base URL. Configurable via MISTRAL_API_KEY env var or mistral.api_key in config.yaml.
  • Agent Skills (agentskills.io spec) — discover and activate reusable skill definitions from .glue/skills/ (project-local), ~/.glue/skills/ (global), and extra paths via config/env. Skill parser validates YAML frontmatter. skill tool lists or activates skills. /skills slash command opens a two-pane SplitPanelModal browser. Configurable via skills.paths in config.yaml or GLUE_SKILLS_PATHS env var.
  • Browser tool infrastructure — BrowserManager with pluggable BrowserProvider abstraction for Chrome DevTools Protocol connections. Five provider implementations: local Chrome, Docker container, Browserbase (cloud), Browserless (cloud), and Steel (cloud). BrowserConfig with auto-detection priority chain.
  • PDF text extraction — web_fetch now handles PDF URLs with a two-stage pipeline: (1) direct text extraction from PDF bytes, (2) OCR fallback via Mistral Pixtral or OpenAI GPT-4o vision models for scanned/image-heavy PDFs.
  • GitHub Actions CI/CD — six workflows: Dart checks (analyze, format, test), multi-OS matrix (Ubuntu/macOS/Windows), docs build validation, nightly e2e integration tests, release tag builds, and auto-labeling. Dependabot configured for Dart and GitHub Actions dependency updates.

Changed ​

  • Web search and browser automation support are constructed lazily during startup. ServiceLocator.create() now wires memoized lazy factories into WebSearchTool and WebBrowserTool, so SearchRouter, browser provider selection, and BrowserManager are only built on first valid tool use.
  • Model catalog refreshed against live provider availability (April 2026).
    • Mistral default → devstral-latest (agentic coding model). Added mistral-medium-latest. Kept mistral-large-latest, mistral-small-latest.
    • Ollama default → qwen3-coder:30b (community consensus, 256K context, dedicated tool-call parser). Recommended list: qwen3-coder:30b, qwen3.6:35b, gemma4:26b, devstral-small-2:24b, qwen2.5-coder:32b, qwen3:8b. Six popular non-agentic families (mistral:7b, gemma3:12b, codellama:13b, codegemma:7b, starcoder2:15b, deepseek-coder:33b) included as recommended: false with notes explaining why they're not suitable for tool loops.
    • Groq default → openai/gpt-oss-120b (reasoning + coding). Added gpt-oss-20b, llama-3.1-8b-instant. Kept llama-3.3-70b-versatile.
    • Slash-bearing catalog keys slugified: groq/openai/gpt-oss-120b → groq/gpt-oss-120b (upstream openai/gpt-oss-120b now lives in api_id). Same treatment for Groq gpt-oss-20b and OpenRouter's three slash-keyed entries.
  • Resize preserves scroll position. UserResize no longer snaps the transcript back to the tail; the render pipeline clamps any out-of-range offset after the viewport changes.
  • Ctrl+End jumps to the bottom and resumes follow-tail. Plain End stays reserved for the line editor (cursor-to-end-of-line).
  • /status is now a hidden alias of /info. One command in help + autocomplete; muscle memory for /status keeps working.
  • Release workflow overhaul — five-way matrix (linux-x64 / linux-arm64 / macos-x64 / macos-arm64 / windows-x64.exe), per-asset .sha256, aggregated SHA256SUMS, smoke-test of every binary, prerelease auto-flag for tags containing -.
  • GH Action version audit. actions/cache v4 → v5 (Node 24); everything else already on current major.
  • Status bar reformatted: bold mode indicator on left, model/mode/ cwd as right-aligned segments with │ separators.
  • GlueConfig.load() no longer accepts cliProvider parameter (provider inferred from model). Removed --provider flag (frees -p for print mode).
  • Terminal parser now decodes CSI modifier parameters (;3 = Alt, ;5 = Ctrl) from extended arrow key sequences.
  • Terminal parser handles ESC + byte sequences for Alt+char and Alt+Backspace (macOS Terminal convention: ESC prefix = Alt modifier).
  • Eager tool call emission — AgentToolCall events are emitted during LLM streaming (not after stream end), so auto-approved tools can start executing while the model may still be finishing the response.
  • Dart analyzer hardening — expanded analysis_options.yaml with always_use_package_imports, strict-casts, strict-raw-types, avoid_dynamic_calls, prefer_const_constructors, unawaited_futures, discarded_futures, and other safety/style rules on top of package:lints/recommended.yaml. Converted all relative imports to package:glue/ imports across 51 files. Applied dart fix auto-fixes for const correctness, unnecessary lambdas, and parentheses.
  • Default model strings removed from GlueConfig — _defaultModel() now delegates to ModelRegistry.defaultModelId().
  • Subagent updates use a grouped data model (_SubagentGroup) instead of individual conversation entries — reduces output noise during multi-agent orchestration.

Fixed ​

  • Spinner no longer stuck after cancel. _cancelAgentImpl now stops the spinner before flipping mode; _cancelBashImpl mirrors the pattern defensively.
  • Tool phase no longer stuck on awaiting approval after cancel — approval-modal-open cancels now transition to cancelled too.
  • Cancel no longer corrupts conversation — cancelling (Escape) while a tool was executing left the conversation with tool_use blocks but no matching tool_result messages, causing the next API call to fail with a 400 error. ensureToolResultsComplete() now injects synthetic [cancelled by user] results for any unmatched tool calls.
  • Unused callId parameter in _ConversationEntry.toolResult — was accepted but silently discarded; removed the dead parameter.
  • /model switch now updates _config (provider + model) via copyWith, fixing stale config bug where session metadata and subagent spawning read outdated values.
  • ANSI codes in split panel highlight — reverse-video selection highlight in SplitPanelModal now strips ANSI escape sequences before applying highlight.
  • Skill discovery trailing-slash safety — skill directory paths ending with / no longer produce empty skill names.
  • Exit message styling — exit prompt now shows the yellow diamond brand mark with session ID.
  • Sema scripting skill documentation examples corrected (8 runtime errors fixed in documented code).

Removed ​

  • codestral-latest from Mistral catalog — FIM-lineage model; enumerates tools but narrates instead of calling them in agent loops. Replaced by devstral-latest.
  • qwen/qwen3-coder from Groq catalog — Groq no longer serves this model; the entry would 404. Replaced by openai/gpt-oss-120b.
  • llama3.2:latest from Ollama catalog — general chat model, not coding/tool-focused.
  • devstral:latest from Ollama catalog — replaced by the pinned devstral-small-2:24b (known version, known SWE-bench scores).
  • Interaction modes (code / architect / ask), plan-mode UI, GLUE_INTERACTION_MODE env var.
  • OTEL / Langfuse / DevTools observability sinks — ~/.glue/logs/spans-YYYY-MM-DD.jsonl is the single local trace log now.

[0.1.0] — Initial development ​

Added ​

  • TUI application shell with raw-mode terminal, layout zones (output/overlay/status/input), and 60fps render throttling.
  • Readline-style line editor with history, Emacs keybindings (Ctrl+A/E/K/U/W), and cursor movement.
  • AgentCore ReAct loop with streaming LLM ↔ tool execution, parallel tool calls, and token counting.
  • LLM providers: Anthropic Messages API (SSE), OpenAI Chat Completions (SSE), Ollama (NDJSON) — each with per-provider tool schema encoding and message mapping.
  • SSE stream decoder for chunked event parsing.
  • LlmClientFactory for provider instantiation from config.
  • GlueConfig resolution chain: CLI args → env vars → ~/.glue/config.yaml → defaults.
  • Headless AgentRunner with configurable approval policies (allowlist, always-approve, always-deny).
  • AgentManager for subagent orchestration with depth-limited recursive spawning.
  • spawn_subagent and spawn_parallel_subagents tools.
  • Built-in tools: read_file, write_file, edit_file (multi-line find-and-replace), bash (with configurable timeout), grep, list_directory.
  • Slash command system with /help, /model, /clear, /resume, /info, and tab-completing autocomplete overlay.
  • @file reference expansion — type @ to get fuzzy-matched file hints (recursive, with directory browsing), expanded inline on submit.
  • Inline confirmation modal for tool approval (hotkeys + arrow navigation).
  • Full-screen panel modal with scrolling, barrier rendering, and dismiss.
  • Session persistence — ~/.glue/sessions/ with conversation logging, listing, and /resume command.
  • ConfigStore with mtime-based caching and atomic saves.
  • DebugLogger for ~/.glue/debug.log.
  • Auto-loads AGENTS.md and CLAUDE.md into system prompt.
  • Bash mode — ! prefix for shell passthrough with background job lifecycle management (ShellJobManager).
  • bashMaxLines config setting and box-drawn bash output renderer.
  • Markdown table rendering with box-drawing characters.
  • Animated mascot splash screen with liquid simulation.
  • Mascot explodes into goo particles when clicked repeatedly.
  • Status bar with mode indicator, model name, working directory, scroll position, and token count.
  • Status bar padding accounts for ANSI escape sequence lengths.
  • Scroll/resize events routed through event bus.
  • Render throttling at ~60fps to reduce flicker during streaming.

Fixed ​

  • Ollama tool results use tool name instead of call ID.
  • OpenAI tool call arguments serialized as JSON instead of Dart toString().
  • /model command now actually switches the LLM client.
  • BashTool uses Process.start with proper kill on timeout.
  • ConfigStore uses mtime-based caching with atomic saves and update() API.
  • AtFileHint caches directory listing to avoid blocking UI on every keystroke.
  • Expanded @file content stored in conversation entries for session logging.
  • Fire-and-forget .then() replaced with proper async error handling.

Released under the MIT License.

© 2026 Glue