Aider Nightly Pipeline & Student Kit¶
Chapter 11 of the ResearchTools manual. Back to table of contents. Skill row: 06-skills.md. Local-delegation counterpart: 05-agents.md.
Overview¶
A second, independent local-coding pipeline that needs no interactive harness: two local Ollama models
(a writer that codes and tests, a reviewer that never edits, gated by measured token budgets)
run one aider process per plan overnight, driven by aider-plan.ps1 / aider-night.ps1. Owns
the packaging pipeline that assembles the student-facing aider-kit.zip from this skill's own
canonical sources.
Full manual, Ollama tuning reference, worked example, and continuity notes: docs/aider-setup.md.
local-coder and the aider-setup pipeline - one harness at two scales¶
Since 2026-11-06 local-coder no longer has a driver of its own. It is the aider kit run on one
plan: the plan is mounted as the only plan of a small kit project outside the repository (new-project.ps1), aider-plan.ps1 runs it with the kit's writer
model and its required reviewer audit, and the kit starts without a wrapper. The per-stage skills
(write, audit, reopen) are mandatory, for a single plan as for a night: there is no switch to turn them off. aider-night.ps1 is
the same engine for several plans or an unattended night: it adds the working branch, the wake
lock and the optional push. A plan opens with a header (executor, tier, bounds) validated by
plan_contract.py. The R35 review is done by the cloud session before the pull request, and a
finding becomes an - [ ] audit: line that the kit runs again. The old text driver
(local_coder_driver.py) is retired: it asked a small model for tool-like output in plain text and
its replies were refused.
Plans and the local coding contract. Since 2026-10-17 every plan, for aider or for
local-coder, opens with a header naming executor, tier and the review bounds. The aider
driver neither validates nor uses those three keys: it only skips a closed leading front matter
block when it lists the files a plan names (Get-PlanTargets in aider-plan-core.ps1, proven by
verify-aider-plan.ps1). The header is validated by plan_contract.py in the harness that
writes the plan. The size limits the driver enforces come from context-budget.json, which is
also the single source for every project:
| File | Ceiling in tokens |
|---|---|
spec.md |
4000 |
plan.md |
4000 |
progress.md |
2048 |
audit.md |
4096 |
The tier-to-role table lives in plan-tiers.json and applies to local-coder, not to aider,
whose writer and reviewer models are chosen in its own configuration. The review-loop bounds of
R35 do not govern aider's internal writer-reviewer loop. They govern the /code-review and
/security-review run by the calling harness on the branch the night pushes, before any pull
request, and the night never merges. During a run the Real-Time tab of the dashboard shows the
run's fixed cost and task budget (chapter 10). The choice between
the two executors is described in chapter 11.
The nightly run¶
aider-setup is not a wrapper around local-coder, and does not go through an interactive
harness at all once installed - the point, since it is built to run when every interactive
harness is closed. Two local Ollama tags do the work: a writer that codes its own tests,
and a reviewer, sandboxed and content-hash checked, that can never edit code however its
prompt is answered. The driver (aider-plan.ps1) runs one aider process per plan file,
which is what drops the context window between plans; after each plan it runs the test suite,
then the reviewer, which writes audit.md and may reopen the plan for one more bounded round.
A plain cmd entry point (aider-night.ps1/.bat) starts a night, holding a wake lock so
Windows' Modern Standby does not throttle an unattended run, and pushes a branch at the end -
merging to a protected branch stays a human decision, made the next morning after reading
audit.md. The rules every model follows (config/rules.md) are 26 coding rules curated by hand,
R0 to R25, in plain words and with no reference to this toolkit: constants and configuration, errors,
tests, function headers, naming, comments and layout. The file is the only copy and is edited in place;
aider-rules-check.py validates it before every run. Rules about documents, agents or process (the review bounds,
the plan header, the skill-first rule) are left out on purpose, because a model that writes code
cannot apply them. The file measures 1618 tokens
(token_count.py, 2026-11-07) against a ceiling of 2000. Setup, tuning, and rebuilding the student-facing kit: see
.claude/skills/aider-setup/SKILL.md and its MANUAL.md.
← 13 3D CAD design, CNC, 3D printing | Table of contents | 15 Security audit and guards →