v0.7.0

0.7.0 (2026-06-20)

Features

• onboarding: agent-aware first-run welcome (#67) (0de0b46)

v0.6.1

0.6.1 (2026-06-17)

Dependencies

• bump charm.land/lipgloss/v2 from 2.0.0 to 2.0.4 (#61) (87404d8)
• bump github.com/mattn/go-sqlite3 from 1.14.32 to 1.14.45 (#64) (3c78ed8)
• bump github.com/pelletier/go-toml/v2 from 2.2.4 to 2.4.0 (#63) (e2ac17f)
• bump golang.org/x/term from 0.41.0 to 0.44.0 (#62) (a30840a)

v0.6.0

0.6.0 (2026-06-12)

Features

• turn-by-turn agent conversations, model selection, and zsh dialect (#59) (5d41fec)

v0.5.1

0.5.1 (2026-06-05)

Bug Fixes

• completion: stabilize autocomplete interactions (#57) (0ab97a4)

v0.5.0

Hash 0.5.0 Release Notes

Hash 0.5.0 is the tool-calling release. This release moves Hash from an agent-assisted shell that suggests commands toward an agent-assisted shell that can mediate tool execution with explicit user permission.

Headline: Agent Tool Calls

Hash now supports ACP agent tool permission requests.

When an ACP agent wants to run a tool, Hash receives the permission request, renders a terminal-native prompt, waits for the user's decision, and sends the result back to the agent. This makes tool use part of the shell workflow instead of a separate agent UI.

Permission prompts support three decisions:

Key	Decision	Effect
y	Allow once	Approves this one tool request.
n / Esc	Deny	Rejects this tool request.
a	Always allow	Approves this exact command and stores it in the configured allowlist scope.

Always-allowed commands are exact command matches. Allowing git status does
not also allow git status --short or git.

Approval Scopes

Tool approval persistence is controlled by:

[agent]
allowed_commands_scope = "project"

Supported scopes:

Scope	Storage	Behavior
project	~/.config/hash/project_allowlists/<project-hash>.json	Default. Persists approvals per project directory, outside the worktree.
global	~/.config/hash/allowed_commands.json	Persists approvals across all projects.
session	Memory only	Clears approvals when the shell exits.

Project-scoped approvals reload when the working directory changes. This keeps approval decisions isolated between projects while still allowing commands to be remembered inside a project.

Permission UI and Streaming

Agent output is now coordinated through a shared output path so streaming text, permission prompts, confirmation hints, and cancellation cleanup do not write over each other.

This release includes:

• Permission prompt rendering with optional tool-name context.
• Stream buffering while a permission prompt is active.
• Prompt cleanup after allow/deny decisions.
• Cancellation cleanup for visible permission prompts.
• Race coverage for concurrent agent-output operations.

Agent and ACP Reliability

0.5.0 hardens the ACP transport around real shell usage:

• Bidirectional ACP request handling for session/request_permission.
• Tool call ID and tool-name extraction from ACP request payloads.
• Permission requests are serialized so prompts do not overlap.
• ACP timeout handling avoids firing while Hash is waiting on a local permission decision.
• ACP reconnect behavior is hardened after broken pipes, cancellation, and startup failures.
• Plain explanatory agent responses are filtered before command confirmation to reduce accidental command classification.
• Bare ?? after a failed command can use the last error as context.

Agent Configuration

Hash now supports named agent profiles.

Flat config still works:

[agent]
transport = "stdio"
command = "claude-code-acp"

Named agents can be selected with [agent].default:

[agent]
default = "ollama"
timeout = "120s"
allowed_commands_scope = "project"

[agent.claude]
transport = "stdio"
command = "claude-code-acp"

[agent.gemini]
transport = "stdio"
command = "gemini"
args = ["--experimental-acp"]

[agent.ollama]
transport = "http"
url = "http://localhost:11434/api/generate"
model = "codellama:13b"
headers = { Authorization = "Bearer token" }

The HTTP transport uses an Ollama-style /api/generate request and response
shape. It is not generic OpenAI-compatible API support in 0.5.0.

Daily-Driver Polish

0.5.0 also includes a broad set of shell quality improvements.

Migration and startup

• First-run migration asks before loading detected shell settings.
• hash migrate can still import detected bash/zsh settings with compatibility filtering.
• hash migrate status reports imported and skipped items.
• Generated migration comments no longer mention the removed hash migrate generate command.
• Startup tips are shown once and remain available via tips.

History

• Builtin commands and agent prompts are recorded in history.
• History can be disabled with history.enabled = false.
• The history database path can be configured with history.path.
• Ctrl+R search handles space input and avoids terminal response leakage.

Completion

• VCS-aware git/jj branch and revision completion.
• Semantic completions for common tools including SSH, kill, make, man, npm, pip, and brew.
• Directory-only completion for commands such as cd, pushd, popd, and rmdir.
• Better shell-word escaping for completed file paths.
• Completion behavior around drill-down directories and single matches is more predictable.

Editing and prompt behavior

• Fish-style inline history suggestions.
• Shift+Arrow text selection in insert mode.
• Improved shell-token word movement and deletion.
• Prompt and Starship subprocess work is reduced.
• Rendering fixes for wrapping, cursor placement, and terminal escape leakage.

Integrations

• setup-zoxide configures zoxide integration from Hash.
• chpwd hooks can run commands after directory changes.
• Zoxide, direnv, fzf, and Starship documentation has been updated.

Pipeline and PTY fixes

• Pipeline producers no longer run behind a PTY when stdout is a pipe.
• This fixes commands such as git branch | grep <branch> hanging because the
  upstream command incorrectly believed it was writing to a terminal.
• PTY output handling and command cancellation paths are more robust.

Behavior Changes

• First-run migration no longer auto-loads detected shell files.
• hash migrate generate is not included in 0.5.0.
• HTTP agents are Ollama-style, not generic OpenAI-compatible endpoints.
• History retention fields are reserved but not enforced in 0.5.x.
• Tool approvals are exact-command approvals; prefixes and substrings do not match.

Known Limitations

• Tool calling depends on ACP agent support.
• HTTP agents do not get the same ACP session and tool-permission behavior.
• Tool approval persistence stores exact command strings, not parsed command policies.
• History retention settings are documented as reserved but are not enforced.
• Hash remains pre-1.0 and may still make breaking changes.

v0.4.3

0.4.3 (2026-01-31)

Bug Fixes

• agent: broken pipe on context cancelation (0bcef00)
• lint errors (gofmt, gosec, staticcheck) (9b39a3d)
• resource leaks and resilience improvements (b32e0fe)

v0.4.2

0.4.2 (2026-01-31)

Bug Fixes

• race condition in spinner (ed5608d)

v0.4.1

0.4.1 (2026-01-29)

Bug Fixes

• resolve release CI failures (#13) (0c5f775)

v0.4.0

0.4.0 (2026-01-29)

Features

• add bash/zsh migration system with compatibility layer (8474089)
• alias/function tab-completion (ec68874)
• completion: mask sensitive env var values in completion preview (09d2d88)

Bug Fixes

• add environment variable tab-completion with value preview (f07c236)
• copy agent response to system clipboard with feedback (9110225)
• disable ONLCR on PTY when stdout is a pipe (8b7fc9c)
• handle agent startup failures gracefully (67959bd)
• improved handling of complex aliases when migrating from zsh (#12) (43ec193)
• preserve cursor position when rendering completion menu (fa7f0f3)
• use graceful degradation for source/eval compatibility (7101c49)

v0.3.0

0.3.0 (2026-01-26)

Features

• delightful ux (c36611e)