…LI (#275) * feat(cli): take a prompt, agent, account, model and effort from the CLI Closes #273 and #274. `fleet add` grew the session-shaping flags `fleet worktree` already had — `-p`/`--prompt` (including `-p -` for stdin), `--agent` and `--account` — parsed with the same either-side peel loop, since a path is not free-form prose the way `fleet send`'s message is. Its path is now optional and defaults to cwd, but a bare `fleet add` stays a usage error: it was one before, and a mistyped `fleet add` in the wrong tab should not launch an agent wherever that tab happened to be. A `-p` prompt titles the session via naming.GenerateTitle and sets TitleGenerated exactly as the TUI's own fallback does, so the agent's own title still overrides it later while the worker stops regenerating an identical one from FirstPrompt. Both commands also take `--model` and `--effort`. They are one-shot on exactly InitialPrompt's terms and never persisted, which is a decision rather than an omission: all three agents treat these as session-scoped and all three let the user change model mid-session, so a stored value would be re-imposed on the next restart and silently undo that choice. Effort is one agent-neutral flag over three spellings — `--effort` for Claude, `--variant` for OpenCode, and `-c model_reasoning_effort=` for Codex, which has no flag at all — which is the per-agent divergence internal/agent exists to own. Unlike the prompt these ride *in* the command string that tmux send-keys types into the pane's shell, so agent.ValidateLaunchValue rejects anything outside a bare-name shape at parse time, before anything is created. Three pre-existing bugs in runAdd fixed on the way, all from it never having been held to the same standard as runWorktree: - No migration.Run(). session.Open creates state.db, after which migrateConfigDir permanently bails out and still writes its marker — silently stranding a brizz-code user's sessions, pins and slot bindings if `fleet add` was their first fleet command. - No debuglog.Init(). Session.Start logs at Info and debuglog's fallback writes to stderr, so the command printed slog lines over the user's terminal. - No installAgentHooks, so a CLI-created session had no status detection at all, and `--agent codex` would have opened onto Codex's untrusted directory prompt. cmd/fleet/session_launch.go now owns what the two commands do identically. They had drifted: `fleet add` carried a shorter copy of the account logic with no explicit-account branch, no note when the strategy could not rank, and GuardConflictingAuth called from inside the wrong branch — a second, weaker implementation of a policy whose whole point is that a mistake bills the wrong subscription. A failed SaveSession also no longer leaves an orphaned live tmux pane nothing points at. The lift is verbatim so the diff reads as a move; the helpers still exit the process rather than returning errors, so the policy stays untestable and only review catches a change to it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(cli): never hand OpenCode an --effort flag it will refuse Review catches on #275, all four confirmed. OpenCode has the concept of a reasoning effort but no way to accept one on the command fleet launches. `--variant` is declared on the `run` subcommand, while fleet runs the default `$0 [project]` command, whose builder declares only project, prompt and the network options — and OpenCode's root parser is yargs in `.strict()` mode, so an unknown option is not ignored: it prints the command list and exits. The cost of getting that wrong is not a dropped setting. fleet would have reported the session as created, saved its row and pinned its repo, while the pane sat at a shell prompt with no agent in it — the failure promptArg's comment describes, which nothing in fleet detects at launch. The test pinned the wrong spelling, so it passed. So agent.SupportsEffort is false for OpenCode and --effort is refused there rather than dropped: at parse time when --agent says so, and again in guardEffortSupported once default_agent is resolved — the same two-place check --account needs, and for the same reason, since the parse-time guard cannot see default_agent. BuildLaunchCmd guards as well, so a caller that forgets drops the flag instead of launching a command the agent refuses. --model is unaffected: the default command's handler reads args.model. Also: - An explicitly empty positional path no longer falls back to cwd. TrimSpace could yield "", which runAdd read as "no path given" — so `fleet add "$REPO" -p "fix it"` with REPO unset or misspelled started an agent in whatever directory the tab happened to be in. That is the failed-substitution case -prompt already rejects, and folding it into the omitted case defeated the guard errMissingAddArgs exists for. - Dropped the unreachable --account branch in runWorktree. parseWorktreeArgs already rejects --account alongside --no-session, so it could never run, and its comment read as a live guard in the one policy block the PR asks reviewers to read closely. - Fixed stale references to consumeInitialPromptLocked in CLAUDE.md and internal/ui/app.go, and the CLAUDE.md bullet still naming `fleet worktree -p` as the only thing that sets InitialPrompt. - Trimmed implementation detail out of the migration changelog fragment. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * docs(changelog): name the OpenCode --effort exception The fragment claimed --model and --effort work for all three agents, which stopped being true when 9038ef8 made --effort a hard error for OpenCode. A reader would have tried it and been refused. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01URJdPo7wRF9TsB8FQwDqNw --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-authored-by: Yuval Hayke <yuval@brizz.ai>

Uh oh!

Support passing model and effort to fleet add and `fleet … #274

Description

Feature Request

Problem

Environment

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Support passing model and effort to `fleet add` and `fleet … #274