5CypZ747c3c87onpmy3sfbhVvCCdUkTNecGWECwAViNsjoGj

[shallowtensr]

Sub-agent fanout architecture + supporting context-quality levers. Net +1412/-221 in agent.py, single file, no contract / sampling / hostname changes. Companion-test gate and other base infrastructure preserved unchanged.

---

Core mechanism: <parallel_edits> XML protocol fans out independent file edits to child _solve_attempt calls via ThreadPoolExecutor. Children run in lean mode (skip preload) with the brief carrying target-file content + siblings + issue. Per-batch wall cap 200s, size cap _RALPH_MAX_PER_BATCH=5, _RALPH_MAX_WORKERS=4, child max_steps=_RALPH_CHILD_MAX_STEPS=10. Suppressed for single-file batches.

Safety invariants:

• Recursion-depth guard: children get _lean_mode=True + _recursion_depth=1. Triple-checked: (a) _solve_attempt entry asserts _recursion_depth < 2, (b) the <parallel_edits> extraction site sets _is_child and skips parsing, (c) _ralph_subagent dispatch site asserts child kwargs match the invariant before spawn.
• Worst-case inference: _RALPH_MAX_WORKERS × _RALPH_CHILD_MAX_STEPS = 4 × 10 = 40 chat calls per round, bounded further by WALL_CLOCK_BUDGET_SECONDS=248.
• Validator routing: every child + critic chat_completion threads validator-supplied (api_base, api_key, model). No new endpoints, no env reads.

Context-quality levers:

• Distinctive-phrase preload — grep issue-quoted strings + multilingual tokens.
• Read-cache short-circuit + pagination-loop detector — close pagination-evasion paths.
• Path-segment vs substring token matching.
• Canonical-file content grep — boost files containing issue-quoted snake_case keys; guides edits to where keys already live (avoids orphan-duplicate creation in parallel locations).

Defensive passes:

• _revert_syntactically_broken_files — ast.parse + brace-balance for JS/TS/Rust/Go/Java/C/Kotlin/C#/PHP. Reverts to HEAD on partial-edit breakage.
• _revert_docker_mount_mode_artifacts — restores HEAD mode on files whose diff is mode-only and content-empty (docker volume mounts emit spurious 100755 → 100644 on shell scripts the agent never touched). Validator generates the submitted diff from working tree, so the on-disk restore is required.

Prompt work: Style-task gate suppresses _strip_low_signal_hunks on formatting sweeps (those whitespace-only hunks are the work product).

---

Base preserved unchanged: solve() signature + return shape; _resolve_inference_config; all DEFAULT_MODEL / API_BASE / API_KEY constants; every refinement gate including companion-test execution; multi-shot wrapper; DANGEROUS_PATTERNS; # MINER-EDITABLE / # VALIDATOR CONTRACT markers; _EDGECASE_GUARDRAIL; extract_command / _read_context_file helpers. Stdlib only (concurrent.futures). No sampling/hostname/env changes.

[github-actions]

OpenRouter PR Judge

Verdict: WARN
Model: anthropic/claude-opus-4.7
Threshold: 70

Score	Value
Overall	72
Real edit	80
Safety	85
Scope	75
Contract	90

Summary

Large (~1600 line) but mechanically substantive PR. Headline change is a new <parallel_edits> fanout that parses an <edit file=...> block from the model and dispatches per-file sub-agents via ThreadPoolExecutor as recursive _solve_attempt calls in 'lean mode' with a recursion-depth guard (depth>=2 disallowed, children cannot fan out further). Also adds a cat-cache short-circuit for repeated reads, a pagination-loop detector that forces a write, a style-task gate that suppresses low-signal-hunk stripping for formatting issues, a distinctive-phrase / snake_case-quoted-key grep for context ranking, parallelized git grep, a parallel-nudge refinement turn, a syntax-revert post-pass that restores HEAD on files whose patch is syntactically broken, and a docker-mount mode-only diff revert. solve() signature, return shape, validator contract, DANGEROUS_PATTERNS, and stdlib-only constraint are preserved. _EDGECASE_GUARDRAIL scrubbing is retained.

Static Checks

• Large patch with 1630 changed lines; judge should inspect for churn.

Judge Reasons

• Diff is genuinely mechanical: new fanout architecture, new caching, new detectors — not a rename/reorder/Goodhart sweep over the base file.
• Only stdlib import added is concurrent.futures.ThreadPoolExecutor / as_completed.
• Recursion is bounded: child _solve_attempt sets _lean_mode=True and _recursion_depth=1; fanout path is gated on _is_child to forbid grandchildren and asserts the invariant at dispatch and at function entry.
• All child solves pass through the validator-supplied model/api_base/api_key — no new endpoints, no hardcoded secrets, no env-var additions.
• solve() public signature and return-dict shape preserved; get_patch gains an optional issue= kwarg (internal helper, not contract).

• MINER-EDITABLE / # VALIDATOR CONTRACT markers preserved; identifiers remain readable; docstrings retained on new helpers.

• Style-task gate (_issue_is_style_task) is conservative and only suppresses an existing hygiene step when issue language strongly indicates formatting work — plausible behavior change, not scoring-game shaping.
• _EDGECASE_GUARDRAIL list is retained and still applied; no weakening of DANGEROUS_PATTERNS.

Risks

• scope-drift: parallel fanout can burst up to ~4 workers × 10 child steps = 40 inference calls per round, considerably more than the single-thread baseline; legitimate but high-impact on validator proxy load.
• goodhart (minor): build_parallel_nudge_prompt and revised budget-pressure prompts heavily push the model toward writes regardless of confidence, which could increase 'just-edit-something' outputs on tasks where exploration was actually warranted.
• obfuscation (minor): the patch parser accepts both <parallel_edits> and [parallel_edits] bracket forms — slightly odd surface area but justified in the comment; not used to hide anything.

Required Changes

• No required changes returned.

[github-actions]

Ninja PR Scope Guard

Verdict: PASS
Author: shallowtensr
External contributor file allowlist: agent.py

This PR satisfies the external contributor file-scope and agent.py contract rules.

Changed Files

• agent.py