v0.7.6

Patch release: bundles the implement-review review-depth and reviewer-channel work (on main since v0.7.5) with two anywhere-agents issue fixes (#12 composer crash on =1, #13 router extension point for consuming repos).

Added

• Code-review lens gains depth dimensions. The implement-review code lens in references/review-lenses.md now covers complexity and over-engineering (YAGNI, speculative configurability, single-implementation indirection), long-term maintainability for durable code (hidden coupling, magic constants, implicit invariants, debuggability), and temporary-artifact hygiene (debug prints, commented-out code, dead scaffolding, scratch and backup files, stray TODOs). The maintainability dimension is down-weighted for short-lived code.
• Reviewers may parallelize and web-verify (opt-in, capability-gated). The review prompt now invites a reviewer whose runtime supports it to parallelize a large multi-file review across sub-agents, and to use web search to confirm any finding that rests on a checkable external fact (citation, link, library or API behavior, version) before asserting it. The headless Claude reviewer backend (dispatch-claude) gains the built-in WebSearch and WebFetch tools; the empty-MCP isolation still holds because those are built-in tools, not MCP servers. The Copilot fallback backend stays offline by design and a contract test pins that.
• Router extension point for consuming repos (anywhere-agents#13). my-router now consults a bootstrap-proof consumer-local routing file (routing-table.local.md at the repo root, or a ## Routing section in AGENTS.local.md) and merges those rows on top of the shipped table, with local rows winning on conflict. The prior "extend it in your fork (or in consuming projects)" guidance pointed at on-disk router copies that every bootstrap and pack verify --fix reverts, so a consuming repo had no durable place to register a project-local skill with the router.

Fixed

• ANYWHERE_AGENTS_UPDATE=1 no longer aborts compose (anywhere-agents#12). compose_packs.py prompt_user_for_updates() now accepts common truthy spellings (1, true, yes, y, on, plus the unset or empty default) as apply, and falsy spellings (0, false, no, n, off) as skip, case-insensitively and whitespace-trimmed. Previously any value other than the exact words apply, skip, or fail raised a ValueError and stopped composition. The exact word fail still raises PackLockDriftAborted; no truthy or falsy alias maps to that fail-closed path. Genuine typos still raise ValueError, now naming every accepted spelling. The wheel-bundled composer mirror is updated byte-identically and six regression tests pin the behavior.
• Check 8 dispatch-tail scan cuts its near-100% false-positive rate. The implement-review Auto-terminal health-check gains a line-level echo classifier (it strips line-numbered source citations and literal regex-source lines while preserving every real failure line) plus an intrinsic and generic two-tier split. Intrinsic failure forms such as HTTP or status 429/5xx and the Windows 1312 sandbox-runner error count on their own; generic words such as a bare "rate limit" count only when an error-frame token sits on the same or an adjacent line.

Versions

• PyPI anywhere-agents __version__ and pyproject.toml version: 0.7.5 -> 0.7.6
• npm anywhere-agents package.json version: 0.7.5 -> 0.7.6

SemVer: 0.7.5 -> 0.7.6, released as a patch. No API or pack-manifest schema change.