Session-scoped goal seeking for piclaw.

Inspired by the /goal loop described in Simon Willison’s write-up of Codex CLI 0.128.0, this add-on lets a chat keep working toward a saved objective until it is completed, paused, or the configured token budget is exhausted.

Install

Open Settings → Add-Ons and install goal from the catalog.

What it does

• adds a Goal settings pane
• stores editable goal prompt templates in extension KV
• stores a per-chat goal state (objective, enabled flag, token budget, usage, latest progress phase)
• adds /goal commands for starting, resuming, pausing, clearing, and inspecting a goal run
• auto-continues the active goal after each turn while goal seeking is enabled
• updates the standard Pi progress UI live with phase labels and a Braille glyph bar showing remaining tokens for the run
• provides an internal update_goal tool so the model can mark a goal complete after verification

/goal command

• /goal — post help text and current goal state to the timeline
• /goal <objective> — start or replace the active goal run in the current chat
• /goal status — show the current goal state
• /goal on or /goal resume — resume the saved objective
• /goal off — pause goal seeking in this chat
• /goal clear or /goal reset — clear the saved goal state

Settings pane

Open Settings → Goal to:

• turn goal seeking on or off for the current chat session
• edit the saved objective and token budget for the current chat
• edit the global prompt templates used for goal seeking
• review current token usage and remaining budget

Prompt placeholders

The editable prompt templates support these placeholders:

• {{ objective }}
• {{ time_used_seconds }}
• {{ tokens_used }}
• {{ token_budget }}
• {{ remaining_tokens }}
• {{ status }}
• {{ chat_jid }}
• {{ completion_summary }}

Storage model

Notes

• Goal seeking is scoped to the current chat/session, not globally across all chats.
• The add-on uses a token-budget heuristic based on assistant message usage.
• Goal execution keeps detailed phase/token progress in native/web streaming progress UI rather than emitting extra timeline status messages.
• The default token budget is 400k tokens.
• The native Pi progress indicator uses a Braille token-availability bar, e.g. [⣿⣿⣦⣀], where filled cells are remaining budget; visible token counts use friendly units such as 20k or 1.25m.
• Live progress phases include starting, waiting for model, working, receiving response, tool use, usage updated, continuing, budget-limited, and complete.
• Goal progress uses only the core Pi progress indicator (no extra goal overlay strip in the web UI).
• When the objective is truly done, the model should call update_goal with status complete.