Use a coding assistant with research guardrails

16 GB RAM is a practical minimum for a 7B local code model; 32 GB provides more context headroom. GPU or Apple Silicon improves latency.

Component	Role	Price	Link
Ollama 0.32.5	local model runtime	Free / 免费local runtime; electricity and hardware costs are separatechecked 2026-08-02	https://github.com/ollama/ollama/releases/tag/v0.32.5
Continue CLI 1.5.47	source-controlled AI checks and review workflows	Free / 免费Apache-2.0 open-source CLI; model provider cost depends on configurationchecked 2026-08-02	https://registry.npmjs.org/@continuedev%2fcli/1.5.47

Save as .continue/config.yaml

name: Research Local Coding
version: 1.0.0
schema: v1
models:
  - name: Qwen 2.5 Coder 7B Local
    provider: ollama
    model: qwen2.5-coder:7b
    roles:
      - chat
      - edit
      - apply
    defaultCompletionOptions:
      temperature: 0.1
      contextLength: 8192
rules:
  - Never modify raw research data.
  - Run repository tests after each accepted change.
  - Report assumptions and failed checks.

Install local runtime and Continue CLI

ollama pull qwen2.5-coder:7b
npm install --global '@continuedev/cli@1.5.47'
ollama list
curl --fail-with-body -sS http://localhost:11434/api/tags >/dev/null
cn --version
cn --config .continue/config.yaml
# Keep the repository test command ready and run it after every accepted patch.

Research-code change contract

Before editing, restate the scientific invariant, input schema, expected output and acceptance test. Make the smallest change that satisfies the task. Never modify raw data, suppress a failed assertion, fabricate fixture rows, silently drop missing values or change a statistical method without calling it out. After editing, show changed files, tests run, test results, unresolved assumptions and a rollback command. If the requested result cannot be supported by the data, stop and explain the gap.

Known pitfalls

• A small local model may miss cross-file contracts and statistical errors.
• Generated tests can merely encode the generated bug; include independent invariants.
• Never expose secrets or restricted data in prompts.
• Review dependency additions and licenses before installation.

Any supported editor or GitHub interface. Repository content and prompts may be processed by cloud services.

Component	Role	Price	Link
GitHub Copilot Pro Pro plan current 2026-08-02	editor completion, chat, agent mode and code review	$10/mo$10 per user/month with $15 monthly total credits shown on the official plans pagechecked 2026-08-02	https://github.com/features/copilot/plans
GitHub Actions hosted CI current 2026-08-02	independent tests and checks on generated changes	Usage-based / 按用量计费included minutes and overage depend on repository visibility and account plan; unverified for a specific userchecked 2026-08-02	https://docs.github.com/en/billing/reference/actions-runner-pricing

Pull-request acceptance gate

required:
  - human_diff_review
  - unit_tests
  - data_contract_tests
  - lint_or_static_check
  - no_secret_scan_findings
  - dependency_review_for_new_packages
research_specific:
  - raw_data_unchanged
  - random_seed_policy_preserved
  - exclusion_rules_documented
  - output_schema_unchanged_or_migrated
forbidden:
  - merge_on_failing_checks
  - generated_citations_without_verification
  - silent_test_skip

Known pitfalls

• Copilot usage beyond included credits can incur additional charges when paid usage is enabled.
• Individual-plan interaction data settings require review and can change.
• Agent-produced pull requests still need domain review.
• CI pricing and included minutes are account-specific, so usd_month is null.