The harness's tool floor

What this harness assumes a host provides, what it does without, and where the line between "degrade and note it" and "do not proceed" sits. Written for an agent that has never seen this repo — and for the person deciding whether an alternate agent can be swapped in.

The machine-readable form is tests/unit/fixtures/toolchain_capabilities.json; the probe is scripts/improve_toolchain.py. This page explains them.

uv run python scripts/improve_toolchain.py --status          # what this host has
uv run python scripts/improve_toolchain.py --require playwright   # gate, exits 1 if missing
uv run python scripts/improve_toolchain.py --json            # machine-readable

The floor

Capability	Needed for	Without it
uv	everything — the repo is uv-managed (python-preference = only-managed)	nothing works; make dev-install is the setup
git	ship discipline, gates that diff against origin/main	stop: the gates cannot attest anything
gh, authenticated	issue triage, CI badge snapshot, PR merge, push-gate	the work may proceed, but log github: unavailable — never invent a green badge
make	the tier vocabulary (ci-fast, ci-changed, preflight-surface, push-gate)	substitute the underlying command; note the degradation
postgres	any runtime path: serve, migrate, integration tests, live walks	DSL/IR work is unaffected; every live-path task is BLOCKED
playwright and its browser binary	qa smoke-crawl/smoke-dig, journey/story walks, visual probes	those tasks are BLOCKED — never a stamped PASS. The package alone is not enough: playwright install chromium is a separate download and the step that gets missed
a served app (served-app)	every live walk (smoke, journey, acceptance panel, story walk)	same — BLOCKED with the remedy

Capability mapping: features vs binaries

AGENTS.md's Capability Mapping table maps harness features (ask-user-choice, task-list, subagent-dispatch, scheduled-loop, …) across named hosts. Features are what the harness asks the host to do; binaries are what the harness runs. This page is the binary floor. A host can satisfy every row of that table and still be unable to do the work — for instance a CLI with no gh authentication can hold a perfectly good task list and cannot read the issue inbox.

For a host that lacks part of the floor

1. Run the probe and record what is missing.
2. Prefer work whose capabilities you have: DSL/IR work, gate authoring, docs — none need a browser.
3. When a task needs something absent, report BLOCKED with the remedy.
4. Do not substitute a different tool silently. A substitute is a different claim.