Workbench Token Context Discipline

Context Budget Rules

• Start with indexes, file lists, IDs, summaries, and targeted searches.
• Read the smallest slice that can answer the current question.
• Prefer

  rg

  ,

  jq

  , concise table output, and explicit IDs over dumping entire histories.
• Summarize large sources before passing them to another agent.
• Attach only role-relevant skills; more skills are not automatically better.
• Stop and report if two consecutive reads are too large or mostly irrelevant.

SDD Handoff Compact Mode

1. Check whether the prior stage comment includes

   HANDOFF_SUMMARY

   .
2. If it exists, read the summary header and the listed

   SCOPED_EVIDENCE

   before reading any wider history.
3. If no handoff exists, read at most the latest three relevant comments, request or create a compact handoff, then stop broad scanning.
4. Never read full issue lists, full agent rosters, or unrelated repo docs for a single-stage task unless the stage explains why.
5. Output must fit in one reviewable comment. If it does not, compress rather than split.

Cache And Quota Awareness

• Distinguish fresh input, output, cache read, and cache write when usage is under investigation.
• Do not assume dashboard usage is wrong without local/request-level evidence.
• Avoid repeated full-context reruns when a comment ID, file path, or artifact can be cited instead.
• Keep Workbench Max and existing private agents out of broad binding changes unless explicitly assigned.

Output Contract

• CONTEXT_USED

  : exact files, comments, commands, or IDs read.

• SKIPPED

  : intentionally avoided broad reads and why.

• SUMMARY

  : compact facts needed for the next step.

• NEXT_SMALL_READ

  : the next targeted evidence query, if any.

• TOKEN_RISK

  : normal / elevated / blocked.