18 — Latest-versions refresh (2026-07-05 sweep)

This page records the 2026-07-05 latest-versions verification sweep across every tool in the hub: GitHub/registry API reads, release-note review, and targeted raw-file fetches for each tool covered in chapters 01–17 and 19–25, plus a new-entrants search over the 2026-06/07 window. The dated chapters 13 — Reuse surfaces and deltas (2026-07-03 source sweep) (2026-07-03 source sweep) and 15 — Verification pass and the extended matrix (2026-07-03 verification pass) stay frozen as-written; this chapter is the new dated layer on top, and where a delta invalidates a claim in chapters 01–16 or 19–25 the correction is registered here rather than edited in place. All sources in this chapter were accessed 2026-07-05; version/star/date facts are registry- and API-verified (T1-grade metadata), vendor claims quoted from READMEs and release notes remain T4 unless said otherwise. Timing note: three tools (caveman, headroom, sem) shipped releases on 2026-07-03 itself — the same day as the last sweep's clones — so several "current as of 07-03" statements were stale within hours of being written.

TL;DR

• caveman shipped v1.9.1 on 2026-07-03 ("65%, honestly"), hours after the last sweep recorded it frozen at v1.9.0 — and formally retired the ~75% headline the hub spends a full section debunking, replacing it project-wide with a flat "65% average output-token reduction (range 22–87%), output tokens only" plus a new docs/HONEST-NUMBERS.md that concedes when caveman costs tokens. The "frozen flagship, stars pumping" framing in chapters 01 — Caveman: design teardown and 14 — The caveman ecosystem: flagship verified, cavemem and caveman-code deep-dived is stale; the 75%-vs-65% debunk now describes a claim the project no longer makes (release).
• cavemem is no longer unmaintained. Revived 2026-07-03: issues mass-triaged 44 → 4, npm packaging bug fixed, opt-in web-enrichment MCP tool added, in-repo version bumped to 0.3.0 (npm still 0.2.1). 14 — The caveman ecosystem's "dead since 2026-05-17 / 44 unanswered issues" is now false; the measurement-based "refuted as shipped" critique stands until re-measured — the 07-03 commits are enrichment/Windows/packaging, not compression-engine changes (commits).
• codedb's version narrative in 15 — Verification pass and the extended matrix is inverted: v0.2.5827 (native Windows support, 2026-06-23) was already the latest GitHub release at the 07-03 sweep, and page 15's explicit "current is 0.2.5826, not 5827" correction is false against GitHub releases; the "history squashed to one commit; no changes since 2026-06-23" clause is also wrong (main shows 100+ commits of history; repo pushed 2026-07-04). Mechanism evidence holds — main HEAD is still the exact audited commit c12e1e2 — and 0.2.5828 is in active preparation on a release branch.
• sem released v0.17.0 and v0.18.0 within ~19 hours of the last sweep and retired both hero claims (2.3× accuracy, 75% fewer tokens) for paired-run SWE-bench Verified A/B speed numbers with token parity and unchanged solve rate stated plainly. Chapter 25 — sem and mq: entity-level retrieval and structural markdown queries's =0.16.2 pin is two breaking 0.x minors behind and its T4 marketing caveat now describes retired marketing.
• headroom shipped v0.29.0 and v0.30.0 both on 2026-07-03 — --lossless graduated from Unreleased to a tagged release, and a new --force-kompress-all opt-in slightly softens the "ML being demoted" framing (gated, with an explicit force switch, not retired). RTK is unchanged at v0.43.0 but chapter 03 — RTK: design teardown's header table is still pre-0.43 and the uv run transparent prefix listed in 13 — Reuse surfaces and deltas as an incoming 0.44 feature was reverted on develop. lean-ctx has v3.9.0 fully cut in-repo (AI Gateway team mode, grammar-addon dylibs, agents-only hook default) but not yet tagged or published — the largest single release since the sweeps began, imminent.
• Where nothing moved: fff (last push 2026-07-02; =0.9.6 pin current; nushell still pins it; 0.9.7 nightlies signal an imminent bump), mq (one weekly patch behind: v0.6.4), the observability set (minor drift; token-optimizer restructured its savings claims, token-dashboard now ~2.5 months dormant), and the thinking-token blind spot the hub flags as a jackin❯ gap still holds — no surveyed tool added a count_tokens-based thinking estimator. The new-tools sweep found exactly one genuine arrival (Recall) plus watch-list-grade entrants; pxpipe joins the comparison set with its own chapter, 17 — pxpipe: design teardown.

Summary table

Recorded = the snapshot the hub's most-recent dated layer carries (chapter and date noted). Latest, pushed_at, and stars/watchers are 2026-07-05 API reads.

caveman family — deltas

Three load-bearing deltas, all post-dating the 2026-07-03 sweep by hours. Confirmations that need no section: issue #484 (over-shrinking PR/issue titles) still open — not fixed in v1.9.1, though the release exempts build files from /caveman-compress (#600, a sibling fix); MIT licenses unchanged; no new caveman-family repos under JuliusBrussee since the 07-03 sweep (newest repos are Loop-Factory and skills, both 2026-06-08, not family members — negative result, repos list).

Fix-worthy detail: v1.9.1's release notes confirm that several defects the hub treated as steady-state (broken caveman-shrink tarball, NL-trigger misfires, one-shot modes sticking) are now fixed. The 65% number caveman now claims matches the shape of the hub's own per-task-mean correction; the dossier's deeper corrections — the 58.5% Claude-tokenizer replication and the ~17%-of-dollars output-bucket ceiling — remain the load-bearing critique and are unaffected.

headroom — deltas

Core recorded architecture facts all hold: org move to headroomlabs-ai complete at repo level (badges still chopratejas, as 13 — Reuse surfaces and deltas already notes), Apache-2.0 unchanged, Phase H (Python-proxy retirement) not executed (headroom/proxy/server.py still present at 219,879 bytes; REALIGNMENT/10-phase-H-python-retirement.md still lists the 3-PR deletion as pending), no crates.io publication (headroom-core does not exist; the bare crate name headroom is squatted by an unrelated audio-loudness tool v2.1.1 — the same squat pattern the hub recorded for rtk), and kompress-v2-base still on the founder's personal HF account (lastModified 2026-06-10, 1,761 downloads, ungated). The README-vs-PyPI headline inconsistency (H-K5: "60–95%" vs "50-90%") stands as written. No new independent benchmarks surfaced; the "47.5% was self-report" correction from 15 — Verification pass and the extended matrix stands unchallenged.

New mechanisms in the two releases worth a line each: sk-ant-oat OAuth token-prefix auth fix (#1672); AVX2 gating for ONNX embedding + Magika backends to avoid SIGILL on old x86 (#1765/#1162); cache key now includes system/tools/sampling (#1473); unit-router path rejects lossy unmarked tool output (#1479); adaptive Otsu KEEP/DROP relevance threshold (#1726); npm headroom-ai clarified as TS-SDK-only, CLI pip-only (#1585); agent-savings workload personas for coding + general (#1732); Claude 5 family pricing (#1767). Post-0.30 main continues daily (opencode provider-config injection fix #1716, HEAD e8151f0 2026-07-03T22:20Z). Nothing in either release contradicts the hub's verdicts.

RTK — deltas

The 2026-07-03 sweep state holds almost entirely: v0.43.0 remains latest stable (0.44 unshipped; develop RC train at dev-0.44.0-rc.308, 2026-07-03), exactly 63 .toml files under src/filters/ re-confirmed via the git tree API, RTK Cloud still waitlist-only (20 "waitlist" mentions on the homepage, no Cloud section in the develop README), issue #1155 (hook auto-allow permission bypass) still open, and the develop-branch rtk/Cargo.toml / rtk/CHANGELOG.md version drift persists — both still say 0.42.4, and the v0.43.0 release is absent from the in-repo changelog. Apache-2.0, rtk-ai org, default branch develop, and the 14-agent-integrations README claim all confirmed.

New develop mechanisms since the sweep clone b52e02b: TOML filters now matched inside the PreToolUse hook itself via a match-only RegexSet (PR #2748, merged 2026-07-03) — the declarative 63-filter corpus now participates in auto-rewrite command matching, not only direct rtk invocation; trust-before-load hardening extended (rtk trust reuses the init opt-in prompt, project- vs global-scoped labels, untrusted filters skipped silently on the command path but failing loudly when non-interactive, rtk gain flags untrusted custom filters); git stash show compressed to a compact diffstat (PR #2801); UTF-8 char-boundary panic fixes closing a CJK/emoji crash class (PR #2789 + extract_json_object byte-offset fix). Note an internal doc inconsistency this sweep could not resolve from the API alone: 15 — Verification pass and the extended matrix says 77 commands / 62 modules / 63 filters / 86 rules while 13 — Reuse surfaces and deltas says 42 native modules + 64 TOML filters + 85 rewrite rules — only the 63-filter figure was re-verifiable remotely; the rest need a clone audit to reconcile.

lean-ctx — deltas

Docs are accurate on today's load-bearing facts: latest released version is v3.8.18 on both GitHub and crates.io, Apache-2.0, yvgude/lean-ctx, very active (~30 commits in the last 48 h). The recorded tokbench R1 numbers are verbatim-confirmed in the live README ("pilot, N=1 per arm — replication runs in progress"); tokbench itself last pushed 2026-06-15 with no R2. The time-sensitive item: v3.9.0 is fully cut in-repo (CHANGELOG section dated 2026-07-04 plus a version-bump commit) but not yet tagged, released, or published — and it is the largest single release since the sweeps began.

Other v3.9.0/unreleased mechanisms relevant to the engine spec: /v1/compress wire-compatibility with LiteLLM ≥ v1.92 pre_call prompt-compression guardrail; ctx_read repo=<alias> multi-repo parity (read-only by design); MCP roots/list retry re-arming; lost-update fix routing all knowledge writers through a single locked mutator; recovery-first refusal doctrine extended to heredoc-to-interpreter refusals; PowerShell signatures (#679).

codedb — deltas

The mechanism and install-surface evidence in 15 — Verification pass and the extended matrix holds by proxy: default branch main is frozen at the exact audited commit c12e1e2, and the main..release/0.2.5828 diff touches no install.sh/telemetry/hook files. What is wrong is the version narrative, in both directions.

Confirmed unchanged: BSD-3-Clause license; default-on telemetry upload to codedb.codegraff.com, global PreToolUse hook blocking 10 shell commands, ~5.16k-token MCP schema, 1.7 GB RSS figures — all by proxy via the frozen main HEAD, not re-measured live. Curiosity: an anomalous tag v0.2.58181 points at the same sha as v0.2.5819 — a typo tag, harmless but it pollutes tag listings.

sem — deltas

Structural picture unchanged: dual MIT/Apache-2.0, Ataraxy-Labs org, daily solo-founder velocity, telemetry default-on with SEM_NO_TELEMETRY/DO_NOT_TRACK opt-out, cloud tier present (now framed as scale/team/CI upgrade). The cli/mcp crates.io publishing wedge persists exactly as recorded — sem-cli and sem-mcp frozen at 0.14.1 (2026-06-23) while sem-core ships 0.18.0; sem-cloud-client still unpublished (404). The attention-ledger mechanism description stands (no ledger changes in 0.17/0.18 changelog sections).

New mechanisms relevant to the engine spec's read-stub/prefetch layer: sem orient --pack <tokens> — turn-zero packed briefings from raw issue/task text (code-ish term extraction, body-term-convergence ranking; IDF weighting, <details> stripping, name-echo slot in 0.18.0); unique-method-name call edges for Python/Ruby (a method name with exactly one repo-wide definition resolves to an edge, ambiguity yields none); Parent::child qualifiers accepted everywhere Parent.child works; sem impact --tests lexical fallback labeled as weaker evidence; git reftable ref-storage support.

mq — deltas

25 — sem and mq was researched 2026-07-03 — the same day v0.6.4 shipped — so it is exactly one weekly patch release behind. Nothing recorded is false: MIT license, solo maintenance (v0.6.4 changelog entirely harehare + dependabot; releases 6/14, 6/21, 6/24, 7/03), no format gating at 0.6.4 (csv, hcl-rs, toml, yaml-rust2, quick-xml, toon-format all non-optional in mq-lang's Cargo.toml), the non-optional nushell mq-markdown dependency (now 84,131 downloads), and mq-db still a 0.1.x prototype (0.1.3, 51 downloads) all reconfirmed.

New in v0.6.4, worth noting for the dossier's token-budgeted retrieval pattern: CLI --count/-c match-count flag and --skip/--limit result pagination; a new mq-lint crate (LintRule trait + six rules); jq-parity builtins (walk, type filters, entries conversion, regex helpers, plus semver_satisfies/uuid/rand families); mq-check pattern-matching against markdown node kinds; @parametrize in mq-test; a live playground embedded on mqlang.org (PR #1966). The hub's local measurements (53× TOC, 43× section extraction) were made with the v0.6.3 binary; no reason to believe 0.6.4 changes them, but they are now against a superseded release.

Observability set — deltas

Core recorded facts in chapter 23 — Token observability and session visualization (token-optimizer and peers) all hold: PolyForm Noncommercial 1.0.0 license (the GitHub API reports SPDX NOASSERTION because PolyForm is not auto-detected; the LICENSE file confirms), JSONL-to-SQLite architecture, localhost dashboard, keep-warm pinger still opt-in and API-billed-scoped. Most load-bearing: the thinking blind spot holds — the README mentions no thinking-vs-visible split and GitHub code search returns 0 hits for count_tokens in token-optimizer and claude-usage and no thinking-estimator code in token-dashboard, so the hub's "gap worth a jackin❯ tool" claim stands (caveat: code search may under-index small repos; corroborated by README content but not an exhaustive source-tree audit).

New mechanisms in v5.11.35: Total Recall coordination (defers low-relevance continuity injection to the external memory tool when detected, relevance bar raised to 0.6 — an anti-double-memory-layer guard, a pattern the hub's stacking chapter cares about); WSL-root Windows home recovery for Copilot/OpenCode installs; README-level progressive disclosure (archiving large tool results), decision extraction, activity-mode detection, and a post-compaction "context intel digest" — extending the smart-compaction checkpoint mechanism already mapped in the records ledger.

No material change

• fff (dmtrKovalenko/fff): last push 2026-07-02 — nothing moved since the 07-03 sweep. All load-bearing facts checked: v0.9.6 latest stable and crates.io max_stable_version for fff-search/fff-query-parser; nushell still pins fff-search = "=0.9.6" (live Cargo.toml check); the agent-formatting/MCP layer still bin-only (crates fff-mcp and fff-core do not exist on crates.io); author-controlled forks (neo_frizbee 0.10.3, fff-notify-debouncer-full 0.9.4, zlob =1.6.0-dev.7), vendored libgit2, heed 0.22, and the notify RC pin (rc.3 pinned; upstream now at rc.4, stable still 8.2.0 — the docs' notify = "8" recommendation holds) all confirmed in the workspace manifest; the crates.io fff name-squat caveat unchanged; T1-latency/T4-token verdict in chapter 19 — Code-intelligence tools: codedb, fff, CodeGraff, and alternatives stands. Two footnotes: a third published crate exists — fff-grep 0.9.6, a workspace dependency resolved transitively, so the "two crates" framing is slightly incomplete (minor, not an error) — and 0.9.7 is imminent (daily 0.9.7-nightly.* prereleases through 2026-07-02; PR #645 rewrites exactly the per-platform file walker and gitignore matcher that the "default-features off → pure-Rust walker" guidance depends on, so the =0.9.6 pin needs a deliberate re-check at 0.9.7). Stats: 9,524★ / 20 watchers / 367 forks / 45 open issues (repo API, fff-grep).
• tokbench (Entelligentsia/tokbench): no R2 — last pushed 2026-06-15, no releases, README still "pilot, N=1 per arm — replication runs in progress"; the R1 numbers the hub quotes are confirmed verbatim in the live README (repo API).
• cavegemma, issue rtk#1155, RTK Cloud waitlist, headroom Phase H, sem cli/mcp publishing wedge: all explicitly re-checked and unchanged (sources in the sections above).

Corrections register — chapters 01–16 and 19–25

Chapters 13 and 15 stay frozen as dated records; this table is the authoritative correction layer on top of them. "Chapter" names the page carrying the now-stale claim.

New-tools sweep (2026-06/07 window)

Method: GitHub search over token/context/claude/prompt-cache keyword combinations restricted to repos created after 2026-05, plus web search over HN, Show HN, and June–July 2026 roundups; watcher/fork ratios verified via the repos API since search results only echo stargazers. Market context worth one sentence in the hub's market chapter: Anthropic's 2026-06-15 subscription split (separate API-credit pool for Agent SDK / headless usage) is visibly driving a fresh wave of these tools. Net result: only one genuine major arrival (Recall); the rest are watch-list or skeptic-exhibit grade.

Rejected below note-level: iKislay/copium (8★, "65–90% savings zero quality loss" hype, no traction), tokenline (16★, cache-TTL statusline), the Graphify ecosystem (underlying repo predates the window; the 820★ memory-setup repo is April-era), MohammedBasioni/scalpel (cache-aware context-surgery proxy, 2★ — interesting idea, zero traction). Third-party roundups (Pinggy "8 tools", firecrawl "12 ways") surfaced nothing outside existing hub coverage except Graphify and prompt-caching basics — an explicit negative result for the "did we miss a major tool" question.

pxpipe joins the comparison set

pxpipe (teamchong/pxpipe, MIT, TypeScript; v0.8.0 on npm as pxpipe-proxy, published 2026-07-04; 1,728★ / 9 watchers / 100 forks, created 2026-05-20) enters the hub as a new category: cache-aligned prompt re-encoding ("optical context compression") — a local proxy that rewrites the static system/tool-docs slab, collapsed older history, and large tool_results into dense PNG pages to exploit flat vision-token pricing, while relocating rather than adding cache markers so the prefix cache stays warm. It is the first covered tool that directly attacks the cache-read bucket, and its per-request count_tokens-counterfactual accounting is the anti-K1 gold standard; it is also lossy with a silent-confabulation failure shape and its economics are hostage to provider pricing/vision behavior. Full dossier, claims table, risk register, and engine recommendation (PORT the measurement kernel and cache-alignment invariants; do not adopt the imaging mechanism yet): 17 — pxpipe design. Naming footnote for future sweeps: the npm package is pxpipe-proxy; the bare npm name pxpipe is unclaimed (typo-install squat risk), and GitHub Releases lag the git tags (only v0.7.0/v0.7.1 have release objects) — npm is the authoritative artifact stream.

Verification limits

Negative results and fetch failures this sweep, recorded so the next sweep does not re-derive them:

• cavemem 0.3.0 exists only in the repo package.json; the npm registry still serves 0.2.1 as latest — the publish is pending, not verifiable. caveman-code and cavegemma have no GitHub releases or tags (404 on releases/latest) — expected; their versioning is npm-only/none.
• huggingface.co API for headroomlabs-ai/kompress-v2-base returned an auth error — interpreted as the model being absent under the org (the personal-account location was confirmed positively), but a private org mirror cannot be ruled out from the API alone.
• crates.io rejects the default curl User-Agent per its data-access policy; all crates.io reads in this sweep succeeded on retry with an identifying UA — noted so future sweeps do not misread the first-attempt failure as an outage.
• RTK's 77-command/62-module/86-rule counts were not independently recounted (they need a clone-level audit like the 07-03 sweep); only the 63-TOML-filter count was checked via the git tree API. No contrary evidence found.
• lean-ctx issue #625 ("recovery-first turn"): GitHub yvgude/lean-ctx#625 is an unrelated closed hook bug; the changelog cites two trackers (GH and GL numbering) and the recorded #625 is likely GitLab-side with no discoverable URL — the recovery-first-turn citation could be confirmed neither way. The faithful-arm "minimal profile" note was likewise not checked again (not contradicted, just unchecked). No v3.9.0 artifact exists, so the 75.9 MiB binary figure could not be re-measured.
• codedb runtime claims (5.16k-token MCP schema, 1.7 GB RSS, telemetry upload) were confirmed only by proxy via the unchanged main HEAD, not re-measured live. token-optimizer ships no CHANGELOG file (404), so v5.x deltas were reconstructed from GitHub release notes only.
• mq's ~291-crate dependency-tree count was not re-resolved (needs cargo metadata); direction verified via Cargo.toml showing format deps still non-optional. sem's docs/attention-architecture.md was not re-fetched; ledger status was inferred from changelog absence of ledger changes.

Ledger

All accessed 2026-07-05.