v0.2.2

Release numbering consolidated on the 0.2 line. This release replaces the former
0.4.0–0.5.1 releases and includes the changes that followed them. No features are
removed by the renumbering. Install this version explicitly with
go install github.com/rogeriojunior31/lazyagents@v0.2.2: Go module mirrors may
still serve the retired higher versions through @latest.

Added

• Docs: published at rogeriojunior31.github.io/docs/lazyagents, with Brazilian Portuguese translations of the index and getting-started guide in docs/pt-br/, and versioned website metadata in docs/site.json.
• Skills: lazyagents install <source> [skill...] installs only the named skills from a source; an unknown name installs nothing and lists the skills the source has. --all also enables what was installed in every installed agent, so one command takes a skill from a repository to all your agents.
• Crush support: detection (binary, ~/.config/crush or ~/.local/share/crush) and skills in ~/.config/crush/skills, or in CRUSH_SKILLS_DIR; Crush also reads ~/.agents/skills, ~/.config/agents/skills and ~/.claude/skills.
• Crush sessions in the Sessions tab and lazyagents sessions: every project Crush knows, the transcript, resume with crush --session, and delete through crush session delete after a JSON backup.
• Crush session cost in the Sessions tab: the cost Crush summed from its own prices, shown as exact for an API-key setup.
• Crush providers and hooks: provider apply --agent crush and hook installs write delimited blocks at the end of Crush's crushrc (values single-quoted, the rest of the file untouched); clearing removes them.
• Pi coding agent support: detection (~/.pi/agent or PI_CODING_AGENT_DIR) and skills in ~/.pi/agent/skills. Pi also reads ~/.agents/skills, so skills enabled for Codex show up in Pi too.
• Pi sessions in the Sessions tab and lazyagents sessions: list, search, read the transcript (active branch), resume with pi --session and delete with a backup.
• Pi token usage and cost in the Usage tab, lazyagents usage and the session detail. The cost is the one Pi records for each response; responses from a provider signed in with OAuth (a subscription) cost $0.
• Agents whose files were moved with their own variable are found there: CLAUDE_CONFIG_DIR, CODEX_HOME, XDG_CONFIG_HOME/XDG_DATA_HOME/OPENCODE_DB (OpenCode), HERMES_HOME, PI_CODING_AGENT_DIR/PI_CODING_AGENT_SESSION_DIR.
• Hermes Agent: the skill dirs its config.yaml adds (skills.external_dirs, skills.create_dir), the profile picked with hermes profile use, and its Windows home %LOCALAPPDATA%\hermes.
• Provider profiles for Pi: provider apply --agent pi writes a lazyagents provider to ~/.pi/agent/models.json and makes it Pi's default; provider clear gives the previous default back.

Changed

• README: shorter, in English and in Portuguese (README.pt-br.md): demo, install and a highlighted link to the online documentation, where everything else now lives.
• Docs: new demos recorded at 2x on a 136-column terminal. A framed GIF opens each tab guide, and the README shows a short hero plus one clip per tab. scripts/record-demo.sh records every tape in demos/ into docs/assets/demos/: a GIF for the README and guides, an MP4 with a WebP poster for the website. Recording runs under a memory cap.
• TUI: the layout holds up from a small laptop terminal to a 2K screen. Under 30 rows the gap under the header goes, giving the tabs two more rows. Beside a table the detail takes 2/5 of the width up to 88 columns, so on wide screens the table grows instead of the detail; the strip under a table grows to 12 lines on tall terminals. Usage bars stop at 40 columns, and the top agents, models and projects show whenever the table leaves room for them. The Sessions project column widens with the table, and long paths and URLs in detail panels wrap after a /, aligned under their value.
• Sessions: the transcript reader is a log instead of chat bubbles. Prompts are numbered headers, agent turns sit behind a bar in the agent's color, and each turn shows the messages written after its last command, with the work before them folded into one ⋯ line (e unfolds them). Markdown tables in replies are drawn as tables, and the header no longer cuts the counts.
• Sessions: for Claude Code, Codex and Pi the reader shows the time of each prompt and how long each turn took, marks failed commands with ✗ (Claude Code and Pi), and shows compactions, interruptions and model changes as lines between turns. A Claude Code compaction summary no longer appears as one of your prompts, and slash commands appear as typed (/plan …) instead of being dropped.
• Sessions: tool calls read by what they did: $ command, ✎ file +3 −1 for edits (Codex patches included), dimmed lookups, ⎇ for subagents. The folded line counts the files edited and lines changed, a Claude Code plan shows as a document, and a task list as its checklist. The export carries the line counts, plans and task lists.
• Sessions: reader navigation. From 150 columns a rail lists the prompts with their time and marks (✎ edits, ✗ failures, ⎇ subagents), following what you read; ]/[ pick a turn and enter unfolds it alone; m switches between the log, conversation (prompts and answers) and actions (every call, no messages) views.
• Sessions: OpenCode and Crush transcripts also show times and failed commands (✗), a Crush summary shows as a compaction, and OpenCode edits and reads show their file.
• Sessions: a Claude Code subagent call shows how many calls the subagent made, and enter on it opens the subagent's own transcript (esc returns). Pi sessions mark where the conversation forks and open the other branches the same way, instead of showing only the active one.
• Wide terminals: the Usage tab lays each agent's limits out as blocks side by side, so many agents no longer make one tall column, and from 180 columns shows the top agents, models and projects beside the table; its bars grow with the room. The Agents tab shows the detail beside the table instead of below it, and in every tab a table beside the detail stops at 150 columns, leaving the rest to the detail instead of blank space.
• Providers: "in use" is a table with each agent's provider, endpoint and file (the endpoint and file columns from 100 columns), so with no profile there is no longer a detail panel repeating the same files beside an empty list.
• 1–9 now jump to tab N, as in lazygit, in every tab. They no longer toggle a skill, install a hook or apply a provider in agent N: that wrote to an agent's config on a key that is easy to press by habit. Pick the agent with ←/→ and press space. A plugin tab gets digits only while it is capturing input.
• Cost estimates cover Claude Fable 5 and 5.1, Mythos 5 and 5.1 and Sonnet 5.5 (prices checked 2026-09-30).
• The transcript index format changed twice (1-hour cache writes and pricing tiers, Codex limits): the first launch after upgrading reads every transcript once more, then only what was appended.
• Skills: enabling a skill for one agent no longer shows it to the other agents that read ~/.agents/skills. Codex skills now go to ~/.codex/skills; a skill is linked in ~/.agents/skills only while it is enabled for every installed agent that reads it (Codex, Gemini CLI, OpenCode, Pi), and disabling it for one of them moves it to the own directory of the others. Links already in ~/.agents/skills keep working.
• Usage limits and doctor leave out agents with nothing to report (not installed, not signed in, API key account, never used) instead of showing a message or a failure for them.

Fixed

• TUI: while filtering Skills or Sessions, the detail now follows the filtered selection instead of showing the item selected before the filter, and the Filter: line fits the table column. It was wider and pushed the detail past the right edge of the screen.
• TUI: the Agents detail shortens the home folder to ~ in the detection line too (config in ~/.codex), as the other paths already did.
• Docs: all 19 public documentation pages now have reviewed Brazilian Portuguese translations, including guides, topics, architecture and CLI/key/agent/theme references. The Portuguese index links to translated pages, with stable section anchors. Commands and literal program messages remain unchanged.
• Docs: requiredTranslations now requires complete, current Portuguese content; docs tests and the website build reject missing or outdated translations instead of silently publishing English fallback pages.
• Skills: when lazyagents install partly failed (a name already in the library), it no longer hides the skills it did install.
• Sessions: the transcript header said "1 thoughts" and "1 prompts".
• Usage: limit bars of a window without a reset time (Codex) were longer than the others; they align now.
• Providers: the profile column header was cut (profi…) when every profile name was short.
• Agents tab and --json output: an agent's version is just the number (0.159.2, not codex-cli 0.159.2; 0.97.1, not crush version v0.97.1).
• Skills: the ▸ in the detail pane, which marks the agent space toggles, now moves with ←/→; it stayed on the previous agent until the selection changed.
• Codex limits: a window whose reset time passed since the last Codex session showed its old use with "resets now"; it now shows 0%, since that use is from before the reset.
• A change the agent (or another lazyagents) made to its config while lazyagents was writing it is no longer overwritten: Codex's config.toml is edited again on the new content, and a JSON config is left as it is with an error.
• Codex providers: a config.toml with a multiline string (developer_instructions = """…""") is edited instead of refused, and an array spread over several lines can no longer rec...

v0.2.1

Consolidates the former 0.3.0–0.3.2 releases at the same source revision as the
former v0.3.2 tag.

Added

• Documentation in docs/: getting started, one guide per module, configuration, themes, safety and data, troubleshooting and architecture. The CLI, key, agent and theme references are generated from the code and checked by go test ./docs.
• lazyagents help <command> explains every command, with its options.
• CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md and issue and pull request templates.
• Skills: d in the profile list deletes a profile.
• CI runs the test suite on macOS and Windows too (informational for now).
• Full SP Night palette, 11 community themes and your own themes in ~/.config/lazyagents/themes/.
• Tabs redesigned around a matrix (skill × agent, profile × agent, hook × agent) with the detail beside or below the table.
• Transcript reader by turns, with reasoning, tool calls and navigation between your prompts.
• Providers: create and edit a profile from the tab. Hooks: read and edit a hook's commands and scripts, and toggle the commands of an imported pack.

Changed

• provider list --json uses snake_case like every other command: base_url, has_token, env_key, wire_api, and applied only when a provider is applied. providers.json on disk keeps its format.
• Sessions shows an estimated cost, and sessions --json includes cost_usd, only for agents authenticated with an API key, the same rule as Usage.
• English is now the project language: the TUI, CLI text, error messages, themes and docs are in English.
• SP Night themes are renamed: sp-night (default), sp-night-garoa and sp-night-jaragua. The old ids noite, garoa and jaragua still work in theme: and in a user theme's extends:. Plugins now receive sp-night… as theme.id in init.
• Dates are shown in ISO order (2026-09-25, Tue 09-23) and relative times in English (3min ago, yesterday).
• usage --json: rate-limit window label values are now in English (session 5h, week, week · <model>). label is display text and may change again; scripts should match on kind (session, weekly, weekly_model). Other human-readable text in --json output is English too: auth (subscription, unknown), placeholder session titles ((no prompt), (untitled)), agent details and error messages.
• Codex config.toml: the comments that delimit the lazyagents managed blocks are now written in English. Blocks written by earlier versions are still recognized and are rewritten on the next provider apply or clear.
• The usage limits cache (usage-cache.json) has a new format; a cache from an earlier version is discarded once and fetched again.

Fixed

• lazyagents --version reports the real version when installed with go install …@latest instead of dev.
• macOS: the live badge on sessions now lights when the session path goes through a symlink (for example /var → /private/var).
• Windows: doctor no longer reports every hook script as not executable, a session started at a drive root no longer shows \ as its project, and ~\ expands to the home dir like ~/.
• CI runs the whole test suite on macOS and Windows, as a required check. Plugin tests use a fake plugin built in Go instead of shell scripts, so they run on Windows too.
• enable and disable without --agent now act on the same agents as the TUI a/x keys: installed agents with a skills directory. They no longer create skills directories for agents that are not installed, nor fail on agents without one. --agent together with --all is a usage error.
• Disabling a skill in every agent no longer fails when several agents see it through the same shared directory.
• Deleting an OpenCode session now saves opencode export to the backups dir first, and nothing is deleted if the export fails.
• The session delete confirmation names the backups folder the files actually go to.
• Providers: applying a Claude Code profile no longer removes an ANTHROPIC_AUTH_TOKEN, ANTHROPIC_MODEL or ANTHROPIC_BASE_URL you set yourself. lazyagents only removes the keys it wrote, and clearing puts back an endpoint or model a profile had replaced.
• Providers: the env block of Claude Code's settings.json keeps its key order instead of being rewritten sorted by name.
• hooks list: the agent table is aligned on its own instead of inheriting the hook table's column widths.
• Leftover Portuguese text in the Skills tab and in hook repair errors is now English.
• Skills: confirming a skill removal no longer runs the action of the previous confirmation (for example updating another skill or deleting a profile).
• enable with no installed agent that has a skills directory fails with an explanation instead of reporting success.
• Providers: clear without a lazyagents record (nothing applied yet, or applied by v0.2) no longer removes Claude Code ANTHROPIC_* keys you set yourself. The record is saved before settings.json, so a failed write never leaves values nobody recorded.
• Providers: clear removes a Claude Code settings.json that lazyagents itself created for the provider, instead of leaving {} behind. A file that also holds hooks or your own keys stays.
• doctor no longer reports a problem for an agent you have not signed in to or used yet; the Usage tab shows it as not available yet instead of "Refresh failed". usage --json marks it with no_data.

v0.2.0

Novidades

Aba Uso filtrável

• Filtros de período (p: hoje, 7, 30, 90 dias ou tudo), agente (a), visão por dia, agente, projeto ou modelo (←/→) e texto (/).
• Resumo do período com sparkline, bloco de 5h atual e participação de cada linha; a tabela encolhe com a largura do terminal.
• Filtros também na paleta (:usage period 30d, :usage view projects, :usage clear) e padrão na seção usage: do config.yaml.

CLI

• lazyagents usage virou um painel: limites com barra e "reseta em", bloco atual de 5h, sparkline do período, participação por agente e top projetos.
• Novas visões usage limits|daily|agents|projects|models, com --since 7d|24h|AAAA-MM-DD, --limit, --agent e --json estável (rows + total).
• Custo em USD somado resposta a resposta, pela tarifa de cada modelo, só para agente autenticado por API key.
• lazyagents help <comando>, ajuda em português e --agent inválido lista os ids válidos.
• doctor --json; sessions --agent/--here/--limit; skills list|enable|… agrupado; descrição truncada no list.
• Saída colorida no terminal e sem cores em pipe, arquivo ou com NO_COLOR.

TUI configurável

• Nova seção tui: no config.yaml: splash, splashSeconds, startTab, tabs (ordem) e hidden.
• Aba oculta continua carregando em segundo plano; plugin oculto nem é iniciado. Valor inválido vira aviso ao sair, nunca erro.

Desempenho com histórico grande

• Índice incremental dos transcripts: cada sessão é lida uma vez e depois só o que foi anexado.
• Medido com 2000 sessões (860 MB): sessions --json 2,3 s → 0,11 s, usage 3,3 s → 0,15 s, busca nos transcripts 8 s → 0,36 s; rolar a aba Uso custa 0,1 ms.
• Sessão em andamento detectada via /proc no Linux, só nas modificadas nas últimas 24h; deletar confere o arquivo exato na hora.

Compatibilidade

• A saída em texto de lazyagents usage mudou: sem argumento mostra o painel. A tabela anterior de limites está em usage limits; scripts devem usar --json.
• A aba Uso agora considera o histórico inteiro, e não só os últimos 7 dias (o padrão do filtro continua 7 dias).
• Novo cache ~/.local/share/lazyagents/transcript-index.gob: pode ser apagado a qualquer momento; a próxima abertura relê os transcripts uma vez.

Changelog completo: v0.1.0...v0.2.0

v0.1.0

Full Changelog: https://github.com/rogeriojunior31/lazyagents/commits/v0.1.0