Night Shift: local coding worker

Night Shift takes a bounded software task, edits an explicitly selected repository, runs its verification contract, and returns local changes with diagnostic evidence. The coordinator owns planning, task selection, human review, publication and deployment. This worker is not a general planning assistant, model-training system or evaluation platform.

Supported local version

Use /Users/nick/git/night-shift-agent, branch main. The supported worker was merged into main on 2026-10-05. Older worktrees are evidence, not alternative launch locations. This branch starts from card-17's combined worker at 47a050f and preserves cards 50 and 16. Lane removal shipped on Control Board on September 16. Its optional legacy inputs remain compatible with this worker.

Install and verify

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python scripts/run-gate.py worker_tests
.venv/bin/python scripts/run-gate.py worker_syntax

Run one board card

Use a clean, isolated target checkout whose directory name matches the card's repo. It must contain verification.json and scripts/run-gate.py, with the card's declared gates. Test/build commands, interpreter, SDK selection and coverage floors belong to that repository. There are no implicit Gradle commands. KMP repositories declare Gradle/Xcode gates; other repositories declare their own commands.

cd /Users/nick/git/night-shift-agent
FORCE_PROVIDER=ollama OLLAMA_MODEL='<already-installed-model>' \
NIGHT_SHIFT_RECORD_DIR='/Users/nick/git/.night-shift-records' \
.venv/bin/python agent_night_shift.py \
  --project-dir '/absolute/path/to/isolated/repository-name' \
  --repos repository-name --max-runs 1 --until-empty

Choose the model explicitly. ollama uses only local inference. Other existing headless adapters may fall back to Ollama; consult ProviderManager before choosing a chain. No quota scheduler is supported. Model-tier escalation is opt-in, see below. Board credentials use the existing Secret Manager/identity configuration. Never put credentials into URLs or task files. Scope the worker identity to software repos; non-software cards received accidentally are returned blocked for a coordinator.

The runner does not automatically push, create a PR or monitor remote CI. Shell tools are not a security sandbox: run only trusted bounded tasks in an appropriate local execution environment. Publication and credential operations are outside this worker's supported purpose.

Records and completion

Each board attempt writes an independent record directory with manifest, ordered redacted diagnostic events and patch. Default capture is metadata-only; payloads are not retained. Use NIGHT_SHIFT_CAPTURE_MODE=payload only when full context capture is appropriate for the target's data. Missing acceptance evidence is not a quality pass.

.venv/bin/python scripts/run-record.py validate /path/to/record
.venv/bin/python scripts/run-record.py review /path/to/record

A passing gate is build/test evidence. A no-change success additionally requires explicit goal-specific acceptance criteria. Stalls return blocked rather than repeatedly consuming attempts. Inspect the record and diff before accepting a result. A lost release acknowledgement is recorded as unknown; reconcile the board before retrying. Never overwrite prior records to make a failed attempt look successful.

Model tier escalation

Off by default. Set NIGHT_SHIFT_MODEL_LADDER to a ladder of Antigravity models, lowest tier first, to retry a card on a stronger model after it uses its whole iteration budget:

NIGHT_SHIFT_MODEL_LADDER='{"antigravity": ["gemini-3.8-flash-high", "gemini-3.1-pro-high"]}'

• The first tier must be the configured ANTIGRAVITY_MODEL, so turning the ladder on never changes what a first attempt runs on. Every model must appear in agy models. A ladder that fails either check is not used, and each run record says why.
• The previous attempt is read from the run records, newest first, matched on card id. A card with no earlier attempt, and every independent card, starts at the first tier.
• A failed attempt that ended on the iteration cap moves the card up one tier. At the top tier it stays there, and the board's attempt limit ends the climb.
• Any other failure keeps the tier: no provider response, an unhandled error, or a run that failed on a model outside the ladder, such as the local fallback.
• A card that stalled without writing is blocked and is never retried or escalated on its own. Once an operator acts on the card after the stall, the next attempt moves up one tier.
• A step into another vendor's model, for example from gemini- to claude-, draws on another quota pool. It is refused unless NIGHT_SHIFT_LADDER_ALLOW_POOL_CHANGE=1.
• Each manifest records tier_decision (provider, model, tier, decision, reason, prior termination) when a ladder is set, and execution_summary.termination_reason always: completed, iteration_cap, stall_without_write, no_response or error.

A run that is still failing tests when it hits the cap ends as iteration_cap, because the worker keeps iterating until the cap. Failing tests inside the budget cannot be told apart from running out of it.

Archived experiments

Context decision probes, board-reconciliation benchmarks and frozen research datasets are outside the supported checkout, in: /Users/nick/git/.priority-intake/2026-09-16-night-shift-consolidation/experiments/. The adjacent pre-consolidation.bundle preserves their full original source/history. Existing run-record fixtures remain regression tests of recording and coding safeguards; they are not a benchmark or a training dataset program.

See docs/CONSOLIDATION.md for provenance, evidence and outstanding integration decisions.