tandem-tokens
CLAUDE.md1 starsChanged 4 months agoMITView on GitHub
This project is ollamaq: a harness that offloads bulk, mechanical, or generate-and-test work to local Ollama models so premium Claude tokens are spent only on orchestration and the final reduce/cleaning pass. Read PLAN.md for the architecture; README.md for commands. Worker output stays on disk. You read only results/<job_id>/digest.md (hard- capped at 8 KB: stats, samples, failures, cloud cost avoided when pricing is configured, and GPU energy Wh / $ / CO₂ when metered). Pull results/<job_id>/merged.jsonl into context only when you…
The file
We split this CLAUDE.md into the rules it gives an agent
2 rules. About 0 are common, 0 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 47
If using the daemon, make sure one is running: ollamaq daemon (one per machine).
- Very specific: line 69
Prefer the CLI for batch work.
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
ollamaq — delegation policy for Claude Code
This project is ollamaq: a harness that offloads bulk, mechanical, or
generate-and-test work to local Ollama models so premium Claude tokens are
spent only on orchestration and the final reduce/cleaning pass. Read
PLAN.md for the architecture; README.md for commands.
The core contract (do not break)
Worker output stays on disk. You read only results/<job_id>/digest.md (hard-
capped at 8 KB: stats, samples, failures, cloud cost avoided when pricing is
configured, and GPU energy Wh / $ / CO₂ when metered). Pull
results/<job_id>/merged.jsonl into context only when you are deliberately
running a reduce/clean pass over it, and prefer a slice. Never echo bulk worker
output back into the conversation. After harvest, prefer ollamaq score for
cumulative savings (cloud $ avoided − local energy $).
When to offload
Offload when the task is (a) ≥ ~20 similar independent items, AND one of:
- mechanical — clean / normalize / extract / classify / summarize / boilerplate
codegen over many rows, files, or records (use
mapmode), or - generate-and-test — produce candidates in volume and score them with a
scriptable verifier (use
gen_testmode), or - staged — a cheap model triages, an expensive model handles only the flagged
subset (use
pipelinemode).
When NOT to offload
- A single hard reasoning problem (no map phase to parallelize) — do it yourself.
- Anything needing repo-wide context the local model won't have.
- Anything where a wrong answer silently corrupts data and there is no verifier.
- Fewer than ~20 items — the setup/queue overhead exceeds the savings; just do it.
Routing (from config/models.yaml)
model: "auto" routes by task_tag: clean/extract/code → qwen2.5-14b-16k,
summarize → hermes3-8b-32k (long context), classify → gemma2:9b (fast),
generate → qwen3:8b. Default model is qwen2.5-14b-16k. Never target
gemma2:27b (spills to CPU on the 16 GB card) unless the user explicitly opts in.
Workflow
/offload— turn the request into a validated job spec, submit it, get a job id + rough ETA. For a small batch useollamaq run <spec> --wait; for a large or long one useollamaq submit <spec>and let the daemon drain it.- If using the daemon, make sure one is running:
ollamaq daemon(one per machine). After submitting, don't poll tightly — use/loopor a scheduled wake-up (≥20 min) to checkollamaq status. /harvest— when jobs aredone, read their digests, and run the Opus/Sonnet reduce/clean pass over the residue the local model got wrong (guided by the digest's failures/samples), then write the final artifact. Cross-client policy also lives inskills/hybrid/SKILL.mdandAGENTS.md(Codex/Grok/Cursor). Report energy lines from the digest when present.
Auto-offload nudge (hook)
A UserPromptSubmit hook (hooks/auto_offload.py, registered in
.claude/settings.json) scans each prompt and, when it looks like a bulk
mechanical data task, injects a [tandem-tokens] suggestion naming the likely
mode. It surfaces the opportunity — you still decide whether to /offload. It
stays silent on ordinary prompts. (A hook can't auto-execute a job: a prompt
isn't a validated job spec.)
MCP tools (alternative to the CLI)
An MCP server (.mcp.json) exposes ollama_quick (one-shot, output-capped —
for tiny delegations only), queue_submit, queue_status, queue_digest.
Prefer the CLI for batch work; ollama_quick is for a single small prompt whose
short result is fine to bring back into context.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

