CtrlK
BlogDocsLog inGet started
Tessl Logo

bulk-read

Delegate large-file reads to a cheap opencode2 worker and read only the summary. Trigger: bulk read, leer archivo grande.

68

Quality

81%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Bulk read with opencode2

A cold opencode2 worker reads the files and answers one question. The file contents land in the worker's context, not here. This session receives only the summary.

Load when a read would burn more context than the answer is worth.

When to delegate a read

SituationAction
One file over ~350 lines, one specific questiondelegate
Several large files, one question about all of themdelegate
Docs or references you only need summarizeddelegate
You need exact text to editdo not delegate — targeted Read with offset/limit
You need to debug or hunt a subtle bugdo not delegate — read inline
The file is small, or you already know the sectiondo not delegate — targeted read is faster

Invocation

Write the worker prompt to a file first. Name paths; never paste file contents.

cd <repo>
~/.opencode/bin/opencode2 run --auto --standalone --format json \
  --model opencode-admin/glm-5.3-flash --title "bulk-read" \
  "$(cat /tmp/opencode/bulk-read-prompt.md)"

Prompt file template:

Read these files only:
- <path>

Question: <one question>

Output rules:
- Structured bullets only. No prose, no greetings.
- Lead each bullet with the exact symbol, name, or line number.
- Skip anything the question does not ask for.
- Keep the answer under 400 words.

This is a read-only task. Do not modify any file.

Traps

  • Name paths, never paste contents. The worker reads them with its own cheap tokens. Pasting re-costs the corpus in this session and defeats the purpose.
  • --auto always. Without it the run blocks on a permission prompt nobody can answer and dies.
  • The binary is ~/.opencode/bin/opencode2. opencode is a different tool.
  • Run from the repo root. Relative paths in the prompt resolve from there.
  • Tell the worker read-only. A cold agent that spots a bug will fix it unless told not to.
  • Cap the answer (~400 words). The summary returns through this session's bash output.
  • Follow-ups: re-running with the same paths is the simple path. To reuse the worker's warm context for heavy follow-ups, pass --session <id> from the first run's JSON output instead.

Verify before acting

  • Worker summaries lack reliable line numbers. Spot-check any symbol you will act on with a targeted Read (offset/limit).
  • Never edit based on a summary alone.
  • If the summary reports something odd, read that section inline before you trust it.

Related

  • delegate-opencode2 — general delegation mechanics, verification rules, and model routing.
  • The 4-file rule in agents/delegations.json — understanding that needs 4+ files is a delegation, not a read.
Repository
YoizenSA/dev-ai-workflow
Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.