devspace-ultra: Instructions file for Codex

DevSpace Ultra Agent Notes

Mandatory interactive progress preflight

For every user-facing ChatGPT Classic Main task expected to require more than one substantive tool call, more than roughly 30 seconds, or multiple verification gates, start or resume the required Goal/Plan structure and then call devspace_progress_report before the first substantive work tool. The opening report is mandatory even before a medium-sized result exists; keep it concise and state the current objective plus the immediate first medium step. A genuinely atomic one-tool task is the only exception.

Progress narration is an Agent responsibility, not a reason to deny ordinary tools. A pending first-use claim, late report, active Plan or completed Plan does not stop reads, edits, commands or later Plan work. Keep the card useful by reporting meaningful results when it is available. Backend Chat Swarm workers remain excluded from the user-facing card.

A pending claim, unresolved or unbound conversation identity, unavailable recipient, omitted tool, or timeout is not proof that the floating card was updated. The report tool's hidden page-local relay completes an initial claim automatically; a separate devspace_progress_bind call is not required and may be absent from an older ChatGPT tool snapshot. If automatic confirmation is unavailable, use the exact-conversation compatibility bridge once when the current conversation ID is known and verify that the report appeared on this conversation. If verification still fails, state one visible progress-routing blocker, continue safe workspace work, and never claim that the card was updated. Re-check the card at meaningful boundaries and before the final response. Chat Swarm workers remain backend-only and must not write the user-facing progress card.

Context-safe continuation work

A single user turn can contain hundreds of model/tool messages. Preserve the finite chat budget: for a resumed task first read the latest checkpoint and current diff/status, not the whole repository, commit history or old chat. Prefer narrow local grep/read ranges (normally 40-100 lines) and command outputs of roughly 1,000-2,000 tokens. Increase only for a specific unresolved question. Keep complete test logs and detailed evidence on disk; return the exit code, failures, audit id and next action instead of replaying full logs. Do not repeatedly rediscover an already loaded tool catalogue or fetch the same unchanged file/commit. A failed local identity proof is a routing blocker, not permission to spend the whole turn rebuilding project history through remote fetch_file/fetch_commit. Use the exact authorized compatibility path once, preserve a bounded checkpoint and report the unresolved boundary. Do not turn read-only investigation into state-changing Worker resume when a router suggests a tool whose required ticket/worker preconditions are absent. Goal reports should state new decisions/results and the immediate next step; put detailed older evidence in a checkpoint path, not nested prior reports. Never label character/byte accounting as native token usage. Do not enable Auto Compact, copy full transcripts to a fresh chat, or discard full stored Goal history merely to avoid the length limit. The UI card is not the only place tool results consume context.