From orqit
Degraded-but-disciplined mode for when Claude Opus 4.7 premium requests are exhausted or must be conserved. Codifies the recipe that recovers ~85-92% of Opus quality on standard-tier models via extra process discipline.
How this skill is triggered — by the user, by Claude, or both
Slash command
/orqit:fallback-modeThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
- **Preferred:** `claude-sonnet-4.6` (explicit effort-capable fallback lane for sticky `--effort xhigh`)
claude-sonnet-4.6 (explicit effort-capable fallback lane for sticky --effort xhigh).github/skills/dev-session/SKILL.mdFallback mode trades latency for accuracy. The founder has explicitly stated: prefer slower, verified work over fast, plausible-sounding work.
fallback-detector extension)Do not proceed with any task until all three steps are confirmed.
/model claude-sonnet-4.6 and the session confirmed the switch--effort xhigh; if /model cannot set effort directly, pass --effort xhigh on the command line when launching the CLIIf you cannot check step 1 right now, stop here and type /model claude-sonnet-4.6 before continuing.
/model claude-sonnet-4.6 for reasoning tasks. Use /model gpt-5.3-codex explicitly for pure code edits.--effort xhigh. Sonnet 4.6 accepts this flag and gains ~3-6 points on agentic benchmarks.task(agent_type="rubber-duck", ...) after each plan and after each implementation. This pattern catches ~70% of bugs that Sonnet misses but Opus would catch. Rubber-duck passes do NOT recursively trigger another rubber-duck./clear → reload) over /compact — savepoint provides 4–8× more context headroom and Sonnet degrades faster on residual summaries than Opus does. Use /compact only when one small step remains or savepoint setup cost exceeds benefit (see dev-session/SKILL.md § Savepoint pattern for the full decision rule).task(model="gpt-5.3-codex", ...).The universal Anti-Hallucination Rules in .github/copilot-instructions.md apply at all times. In fallback mode, these are tightened — Sonnet hallucinates more than Opus on identifier recall, API surface area, and "do I remember this file?" questions. Mitigations:
glob or grep confirmation in this session.gh or view this session, even if a memory states the value.Stale-risk: or Needs founder input: marker rather than guessing.Stop and open a follow-up issue instead of attempting on standard-tier:
task(model="claude-opus-*") — the extension hook will reject it; this skill states the rule so human agents honor it too.--effort xhigh with /model auto or claude-haiku-4.5. Auto can route to Haiku 4.5, which rejects reasoning effort. If you need Haiku, start a fresh session or clear effort first.--effort xhigh until /fallback-mode-off or the founder explicitly clears it.FALLBACK_MODE: active — quality may be reduced, founder review weight ↑Evidence: block (per copilot-instructions.md compact evidence pattern) — no implicit "trust me" claims.Task: "Add a 2-line logging statement to .github/skills/status/SKILL.md and update the Matrix row."
/model claude-sonnet-4.6 → view target file (don't trust memory) → plan → rubber-duck plan → apply → view diff → rubber-duck diff → verify Matrix row name via grep → commit; handoff prefixed with FALLBACK_MODE: active and an Evidence block.Overhead: +2 rubber-duck turns, +1 verification view, +1 grep. Value: catches a typo, a stale path memory, or a logic miss Sonnet alone would have let through.
/fallback-mode-off (founder only) — clears the flag, resumes normal model routing.fallback-detector extension auto-surfaces a "premium restored" notice; founder confirms before clearing.npx claudepluginhub foculoom/plugins --plugin orqitGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.