Simplification Plan

A plan to make Argus smaller. It is a weight-loss problem, not a rewrite: the system works, and the goal is to remove what is not carrying weight.

This plan derives from the Argus Principles. Companion documents: system audit for the measurements this plan acts on, failure modes for the behaviour it is meant to change.

Chinese version: simplification-plan.zh-CN.md

---

The rule

Never fix a bug by adding a mechanism. Fix the cause or delete the code.

Everything below follows from it. This plan was reviewed by a second model from a different lab; where its reading of the tree differed from ours, we re-measured and the corrected numbers are the ones used here.

The core finding

The expensive thing is not any single gate. It is a loop:

gates produce artifacts → artifacts produce repair state → repair state is injected back into the prompt → retries spin.

skills/research_gates.py is that loop in one file: a single advisory check writes up to four artifacts per run (RESULT.json, REVIEW.md, FAILURES.json, REPAIR_TASKS.md), and REPAIR_TASKS.md is a set of instructions written back into the next round. Delete the loop and let the read-only Reviewer judge real evidence instead.

---

Ordered plan

Work top to bottom. Each step is verifiable before the next begins.

1. Delete the no-op Wiki lifecycle API — completed

The unused post-mission Wiki lifecycle hook was removed. Wiki lifecycle now exports only ensure_project_wiki(); Agents maintain pages and INDEX.md during the mission.

Verified: repo-wide reference search; pytest tests/test_wiki_bootstrap.py tests/test_minimal_skill_wiki.py.

2. Delete the 31 unreferenced event types

Of 129 EventType members, 31 are never referenced by symbol or by string value; 20 of those are SKILL_* and WIKI_* — the two knowledge surfaces the system is supposed to learn through. Remove them together with their payload schemas, frontend catalog entries, and generated types.

Handle separately: 6 events are emitted by raw string literal rather than through the enum (LIFE_MISSION_SKIPPED, LIFE_MISSION_REQUEUED, LIFE_VERTICAL_RESOLVED, LIFE_INBOX_QUEUED, SKILL_OUTCOME, OPERATOR_ALERT). These are not dead. Route them through the enum so the catalog becomes a reliable index again.

Risk: external consumers of old event logs. Verify: replay an old events.jsonl; pytest tests/core/test_event_catalog.py.

3. Cut the unconditional rendering block — best value per line removed

manager_rendering_prompt contributes 2,923 characters to every stage decision in every vertical, whether or not the decision concerns rendering (appended at manager/_stage_ops.py:779). It is larger than the median vertical banner (1,318 chars) and nearly as large as the decision prompt it accompanies.

This is the only prompt cost every vertical pays. Make it conditional on the decision actually involving presentation, or delete it.

Verify: stage decisions still produce valid verdicts across two verticals.

4. Remove prose-regex ownership escalation

core/role_handoff.py:12-79 decides who owns a decision with a regex over prose containing access, release, production, delete, pay. A summary that says "delete the temporary directory" routes to a human.

Delete _OPERATOR_AUTHORITY_RE, _REVIEW_ACTION_RE, and _runtime_owned_review_request(). Make NEXT_OWNER authoritative: an explicit reviewer means reviewer regardless of vocabulary; a legacy OPERATOR_QUESTION with no owner stays with the operator.

Real authority checks at irreversible-action boundaries stay. A word list is not an authority boundary.

Verify: pytest tests/test_structured_decisions_are_not_reparsed.py, plus handoffs whose text contains "release", "production", and "delete".

5. Delete the gate/repair ecology — highest behavioural value

Delete the advisory literature, theory, numerical, novelty, novelty-seeking, paper-type, and manuscript-package gates, then their shared skills/research_gates.py, physics/gate_feedback.py, the capability trace, the generated repair artifacts, and the role-banner injection that carries them (physics/stages.py).

physics/gates/novelty_seeking.py:40 is the clearest case: ten directions, eleven reasoning columns, six scores, four supporting files — 170 table cells to earn the right to make a claim — and it never evaluates novelty. It counts rows.

Keep only outcome checks on a real requested deliverable — "the paper compiles" when a paper was requested. Delete exact CSV headers, figure and reference counts, wording bans, and section quotas.

This directly restores the posture already written into the kernel Reviewer: "Ignore GROUND_TRUTH/gate/marker/status/provenance files … and artifact hygiene — the scorer's number is the only evidence."

Verify: on completed physics missions, compare Reviewer verdict quality, build success, and claim accuracy — not gate-pass rate.

6. Stop re-litigating clean Reviewer acceptances

Manager remains the sole stage-state writer, but should not be invoked to re-decide a Reviewer acceptance that raised no exception. Reserve it for strategy changes, conflicts, rollback, terminal completion, and authority.

This is the intended division doing its job: Manager strategy, Planner decomposition, Engineer local iteration, Reviewer independent judgement.

7. Make daemon self-maintenance an ordinary mission

Argus changes are ordinary engineering missions under the normal Engineer/Reviewer loop. Autonomous discovery and repair remain; reviewed changes cross one operator-approved deployment boundary instead of a resident promotion daemon.

8. Only then purge defensive handlers and spin loops

Do this last. Cleaning exception handling inside subsystems you are about to delete is wasted work.

---

A mechanical rule for the 2,277 try: blocks

Do not review them case by case. Keep a handler only when all four hold:

1. it surrounds one external boundary — network, subprocess, optional telemetry, cleanup;
2. it catches expected failures of that boundary, or is the single top-level mission boundary;
3. it returns an explicit non-success — blocked, a retryable failure, unchanged state — or affects only optional observability;
4. it records the failure and has one bounded retry owner.

Delete or propagate every handler that returns a success-shaped default ("", [], {}, False, pass, continue) for state, prompts, routing, checklists, authority, evidence, locks, budgets, or stage transitions.

	Example
Good	reviewer/_core.py catches backend failure and returns blocked
Bad	skills/stage_machine.py silently substitutes an empty checklist

A runtime that cannot fail is a runtime that cannot tell you it is broken.

---

What not to delete

• The independent read-only Reviewer. require_independent_review defaults to True (engineer/round_config.py) and the Reviewer runs read-only. This is the load-bearing wall that makes unattended operation possible.
• One canonical stage writer and atomic stage state.
• Durable event and backlog history.
• Real authority checks at irreversible-action boundaries — as opposed to regexes over summaries.
• Provider failure-to-block behaviour, cost accounting, GPU leases.
• The tolerant prose parsing in core/role_reply.py. It already makes this document's argument: "The harness is not smarter than the agent, and demanding a wire format is the harness deciding how the agent may speak."

---

The trap

The likeliest way this effort fails is by replacing what it deletes with a "unified gate system", a "unified event system", or a "resilience layer".

research_gates.py is already that mistake: one advisory check grew into four artifacts, a stall tracker, repair prompts, and extra control flow.

Use the owners that already exist. Delete callers and formats rather than consolidating them. If a simplification exposes a failure, surface it to the Reviewer — do not add a layer.