agents-remember

Cost-aware Bootstrap

Repository bootstrap can be token-heavy because the agent has to inspect broad source areas before it can write useful onboarding. This guide is for operators choosing the harness, model, and run shape before starting c-03-repo-bootstrap.

The runtime does not try to optimize spend on its own. Treat model choice as an operator decision, then give the agent a clear bootstrap scope.

Strategy

Use a cost-efficient, large-context coding or reasoning model for the broad, evidence-bound parts of bootstrap, and reserve more expensive models for review or for areas where the first pass gets stuck.

Good candidates for the first pass are models that can:

• read large source files and route overviews without aggressive truncation
• use tools reliably over many turns
• follow path and write-boundary instructions
• produce structured Markdown consistently
• run long enough to finish one wave without harness session limits

DeepSeek V4 Pro is one example of this kind of model when used through a harness that supports its reasoning and tool-call behavior directly.

Check Cost During And After Bootstrap

Before running every file onboarding wave, run one representative wave and check the provider usage dashboard. Bootstrap workloads are usually input-heavy: the agent reads much more source and onboarding context than it writes. Output tokens matter, but repeated source/context reads usually dominate the token count.

As a dated example, on 2026-05-17 a completed TensorFlow memory bootstrap using DeepSeek V4 Pro through OpenClaw used 252 API requests and processed about 19.57 million total tokens: about 18.51 million input cache-hit tokens, 896.6 thousand input cache-miss tokens, and 160.5 thousand output tokens. The exported provider cost for that run was about $0.60, or roughly $0.0305 per 1 million blended tokens. The low blended rate depended on the workload shape and DeepSeek pricing/cache behavior at that time.

Always check current provider pricing before extrapolating from an example. For DeepSeek, see the official Models & Pricing page. On 2026-05-17, DeepSeek listed V4 Pro prices per 1 million tokens and noted a temporary discount through 2026-05-31 15:59 UTC; future prices may differ.

Suggested Phase Budget

Bootstrap work	Model choice	Notes
Source inventory	fast or cost-efficient model	Mostly classification and path discovery.
Scout report	cost-efficient large-context model	Broad reading matters more than polished prose.
Root overview	cost-efficient model, then review if needed	High-leverage artifact; review if confidence is mixed.
Route overviews	cost-efficient model	Template-driven, evidence-bound work.
File onboarding waves	cost-efficient model	Repetitive work; checkpoint between waves.
Curator or final review	strongest practical model or human review	Use this to catch unsupported durable claims.

Wave Size

Large waves can waste tokens when a run times out, hits provider limits, or drifts off task. Prefer checkpointable chunks:

• route overviews: 1-3 routes per run
• file onboarding: 5-10 files per run
• commit or otherwise checkpoint after each completed wave
• keep source reads and memory writes in separate, explicit roots

For very large repositories, start with root and route overviews before asking for dozens of file-level onboarding artifacts. That gives later runs a better map and reduces repeated rediscovery.

Escalation Triggers

Switch to a stronger model or stop for human review when:

• source and documentation disagree
• a route overview would record low-confidence architecture as durable fact
• cross-repo boundaries or branch-sensitive memory are involved
• the model invents invariants without source evidence
• repeated runs fail to keep the source-read and memory-write boundaries
• the harness starts summarizing away important context or hitting session limits

Harness Notes

Prefer a harness that lets the chosen model use the provider API directly and can handle the repository layout you are bootstrapping. For external-memory layouts, the harness must be able to read the source repository and write the memory repository without flattening or copying the Agents Remember runtime in a way that breaks script paths.

For OpenClaw, see Install For OpenClaw for the nested skill layout, allowSymlinkTargets, and long-running turn timeout settings.

Example Prompt Shape

Keep the prompt scoped to one checkpointable unit:

Read the resolved workspace instructions first.

Read source files from <source-repo>.
Write onboarding files only in <memory-repo>.
Do not modify <source-repo>.

Read onboarding/bootstrap/handoff.md and
onboarding/bootstrap/governing-route-map.md.
Continue Wave 1 only.
Match the existing onboarding style.
Commit only after Wave 1 is complete.

This keeps cost control outside the agent’s reasoning loop while still giving it enough structure to finish useful bootstrap work.

This site is open source. Improve this page.