Atlas adopts the runtime rails — thin

Atlas accepts the three runtime-control axes off agent_invoke.v1, declares its mode matrix, and maps the thinking_level onto its own existing model.json tiers. Wiring only — zero new architecture.

Story-1564 · EPIC-1559 · adopter (updates Atlas) · branch issue-1564-atlas-adopt-rails · 2026-06-21

pramaan-docs ADR-129 ratified three orthogonal axes — run_mode, thinking_level, authority_mode — and named Atlas the thin adopter (§D12): it reuses its existing model tiers, no rebuild. The contracts landed in Story-1561; the broker-resident resolvers, floors, and effective_* computation landed in Story-1562. This Story does exactly one thing: make Atlas consume that posture. Atlas owns no resolver and no floor — it reads what the broker resolved, selects which of its own DeepSeek tiers serves the level, and stays proposal-only (ADR-070; ADR-128 does not relax Atlas).

What to review

Area	Where	What it does
Axis reader	agents/atlas/runtime_axes.py	Closed `StrEnum`s mirroring the contract VERBATIM + `parse_posture()` — reads the broker-resolved axes off the envelope, defaults `authority_mode` to `review_exceptions`, fails closed on an illegal value. No resolver, no floor.
Mode matrix	agents/atlas/atlas.yaml `mode_matrix:`	Atlas's self-declared envelope: `supported_modes:[interactive]`, `default_mode:interactive`, `default_thinking.interactive:medium`, allowed capabilities, `pause_stop:stop_cancel_turn`. The broker copy is authoritative; a parity test keeps them in lock-step.
Thinking → tier map	agents/atlas/model.json `thinking_tiers`	Maps `low/medium→quick`, `high/extra_high→advisory` onto Atlas's existing tiers — no new tier system. The seam for a stronger advisory model later.
Per-invoke resolution	agents/atlas/config.py `AtlasConfig.for_invoke()`	Reads the posture, points `model_id` at the level's tier, carries `run_mode/thinking_level/authority_mode`. `ATLAS_MODEL_ID` override still wins.
Wire-in	agents/atlas/runtime.py	`handle_request` + `stream_invoke` swap `AtlasConfig.load()` → `AtlasConfig.for_invoke(request)`. One-line seam; everything downstream (spans, cost, provenance) flows unchanged.
Tests	agents/atlas/tests/test_atlas_runtime_axes.py, test_atlas_mode_matrix.py	Axis parity, default resolution, tier selection, env-override precedence, fail-closed on illegal value, matrix lock-step.

The thin seam

# Broker (Story-1562) resolves + floors the level, then invokes Atlas:
agent_invoke.v1 { …, run_mode:"interactive", thinking_level:"high", authority_mode:"review_exceptions" }

# Atlas (this Story) READS it — and only SELECTS its own tier. No resolver, no floor.
posture = parse_posture(request, default_run_mode=matrix.default, default_thinking=matrix.per_mode)
model_id = base.model_for_thinking(posture.thinking_level)   # high → advisory tier → DeepSeek
# authority_mode is carried, NEVER branched on — Atlas auto-applies nothing (proposal-only).

Why Atlas owns no resolver: ADR-129 §D2/§D3 put the level→model map, the safety floors, and effective_authority in one place — the broker rails (Story-1562) — so a cost-cutting matter default can't silently de-floor a privileged step for any agent. If Atlas re-derived the level it would be a second, drifting copy of that safety logic. So Atlas only maps the already-resolved level onto a tier it already had. That is the whole "thin" of thin adoption.

Why the default invoke is byte-identical: an omitted thinking_level falls back to the matrix default medium, which maps to the quick tier, which is deepseek/deepseek-v3.2 — exactly what Atlas ran before Story-1564. A legacy invoke that sends no axes at all parses unchanged and runs the same model. Behaviour unchanged at default settings, by test.

How to verify

# Atlas's own unit suite (the adopter)
cd agents/atlas && uv run --group dev python -m pytest -q        # 223 passed

# lint + brain-purity egress gate stay clean (no DB/ORM/boto leak in the brain)
ruff check agents/
python scripts/check_brain_egress.py                            # clean

# repo-governance lane unchanged
ruff check agent_contracts/ tests/ scripts/
python -m pytest tests/ -q -m unit                              # 101 passed

Scope & boundaries honoured

held Thin — Atlas re-implements no resolver, no floor, no effective_* (broker rails own them, Story-1562).
held Reuses the existing model.json tiers — no new tier system (ADR-129 §D12).
held Mode matrix declared in atlas.yaml; interactive-first, stop_cancel_turn (never pause/resume).
held Atlas stays proposal-only — authority_mode is accepted + defaulted but never branches behaviour (ADR-070).
held Back-compat: a legacy invoke (no axes) parses unchanged and runs the same model.
held Axis value sets mirror agent_contracts.enums VERBATIM, fenced by a parity test (no per-agent dialect, ADR-110 §9).
deferred Atlas auto mode (wrap the extraction-worker pattern); the "+" capability loader; run-control event emission — downstream of interactive-first.