fallback-mode

Fallback Mode

Model

• Preferred: claude-sonnet-4.6 (explicit effort-capable fallback lane for sticky --effort xhigh)
• Premium-exhausted fallback: n/a (already at fallback tier)
• Source of truth: Model Routing Matrix in .github/skills/dev-session/SKILL.md

Posture: thoroughness over speed

Fallback mode trades latency for accuracy. The founder has explicitly stated: prefer slower, verified work over fast, plausible-sounding work.

• It is correct to take +N turns, +1 rubber-duck pass, or to stop and defer rather than guess.
• It is incorrect to compress steps to "save time" when premium quota is the bottleneck.
• If a task feels rushed, slow down. Add a verification turn instead of skipping one.

When to invoke

• Session received a quota-exhausted / HTTP 429 response (auto-surfaced by the fallback-detector extension)
• Founder explicitly wants to conserve premium requests
• Starting a long batch of routine work where Opus is overkill

Activation Checklist (self-verify before continuing)

Do not proceed with any task until all three steps are confirmed.

• Model switched: typed /model claude-sonnet-4.6 and the session confirmed the switch
• Effort set: will start the next CLI invocation with --effort xhigh; if /model cannot set effort directly, pass --effort xhigh on the command line when launching the CLI
• Recipe read: read all 6 steps in The Recipe below before writing any code or plan

If you cannot check step 1 right now, stop here and type /model claude-sonnet-4.6 before continuing.

The Recipe (6 steps — follow all of them)

1. Switch model: /model claude-sonnet-4.6 for reasoning tasks. Use /model gpt-5.3-codex explicitly for pure code edits.
2. Crank reasoning: start the next CLI invocation with --effort xhigh. Sonnet 4.6 accepts this flag and gains ~3-6 points on agentic benchmarks.
3. Mandatory rubber-duck pass: call 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.
4. Decompose aggressively + savepoint-preferred: break the task into ≤2-task atomic chunks. Between chunks, prefer the savepoint pattern (write RESUME.md → /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).
5. Pin code generation to GPT-5.3-Codex: during fallback, route all non-trivial code edits through task(model="gpt-5.3-codex", ...).
6. Defer-don't-substitute for the list below — stop and wait for quota reset, don't attempt on Sonnet.

Anti-hallucination amplifiers (fallback-specific)

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:

• No-cite = no-claim is HARD here, not soft. For any "must cite" claim (per claim tiers in copilot-instructions.md), if you don't have evidence in this session, run the tool call before responding — do not promise to verify "next turn."
• Read source files even when you "remember" them. Sonnet's recall of file contents from prior context degrades faster than Opus's. View the file this turn before quoting it.
• Grep before naming any symbol, file, or path. Do not type a path or function name from memory in fallback mode without a glob or grep confirmation in this session.
• Verify identifiers via tool, not memory. Issue numbers, PR numbers, commit SHAs, version pins, dates: re-fetch via gh or view this session, even if a memory states the value.
• Run the rubber-duck before any factual assertion in these high-risk classes (per universal rules): architecture claims, tool/model capability claims, issue closure logic, release state, default-branch assertions, side-effecting founder asks.
• When in doubt, defer to the next premium-quota window and add a Stale-risk: or Needs founder input: marker rather than guessing.

Defer-don't-substitute categories

Stop and open a follow-up issue instead of attempting on standard-tier:

• Swift 6 concurrency debugging (actor isolation, data-race sendable violations)
• Multi-file refactors touching ≥5 files (Opus lead on multi-file agentic coding is ~12 pts on SWE-bench Verified)
• Image-input analysis at >2× resolution (Opus-only capability)
• Deep architectural decisions requiring wide codebase context
• Any claim that will drive a side-effecting founder action (merge, deploy, publish, money/legal) where evidence cannot be gathered with tool calls in this session

Rules while active

• Do not invoke task(model="claude-opus-*") — the extension hook will reject it; this skill states the rule so human agents honor it too.
• Do not pair sticky --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.
• Do not switch off --effort xhigh until /fallback-mode-off or the founder explicitly clears it.
• Every handoff produced under this mode MUST prefix its header with: FALLBACK_MODE: active — quality may be reduced, founder review weight ↑
• Every handoff produced under this mode MUST include an Evidence: block (per copilot-instructions.md compact evidence pattern) — no implicit "trust me" claims.

Worked example (BUILDER task)

Task: "Add a 2-line logging statement to .github/skills/status/SKILL.md and update the Matrix row."

• Default (Sonnet 4.6): 1 turn, ~3 tool calls.
• Under fallback-mode: /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.

Exit

• /fallback-mode-off (founder only) — clears the flag, resumes normal model routing.
• Quota reset — the fallback-detector extension auto-surfaces a "premium restored" notice; founder confirms before clearing.