No Silent Fallback

Core Rule

Work quietly when the planned path works. When you switch to a weaker path, make the switch visible: expected path, actual path, cause, and impact.

Expected Capability

A capability is expected only when at least one is true:

• the user explicitly requested it,
• the agent plan promised it,
• the host runtime declares it active,
• project memory/MEMORY.md or CAN.md requires it for this task,
• a previous fallback/pending item explicitly made it part of the current work.

Spec-only assets are not expected unless installed by an adapter or explicitly requested.

Fallback Means

• hook expected but unavailable,
• subagent expected but unavailable,
• tool expected but unavailable,
• expected skill not used,
• expected write/search/close/TodoWrite done manually instead of the expected mechanism,
• verification weakened or skipped,
• expected runtime capability absent.

Not Fallback

• a hook stays silent because no capability is relevant,
• a skill is used quietly,
• a tool succeeds normally,
• a subagent is unnecessary for a small task.
• a spec-only hook/subagent is not installed and was never requested.

Report Format

Report the fallback to the user. Persist meaningful or repeated fallbacks with the backbone script so future sessions can learn from them:

scripts/fallback_log.sh --expected "<what should have run>" \
  --actual "<what ran instead>" --mechanism "<name>" \
  --cause "<why>" --impact "<impact>" --signal skill|tool|hook|subagent|test|rule|none \
  --memory-dir ./memory

The script is idempotent (same expected+actual+mechanism is never duplicated) and writes a canonical block to memory/fallbacks.md. In Claude Code it lives at ${CLAUDE_PLUGIN_ROOT}/scripts/fallback_log.sh.

Learning Signal

If repeated or meaningful, propose a new skill, tool, hook, subagent, test, or rule (--signal). A recurring --mechanism is the trigger for fmc-fallback-review. Do not turn small harmless adaptations into incident paperwork.