Changelog
Updates to reviews, in reverse chronological order.
2026-10-11
- New: Jeff (firelex) — recommended — 2.0 GiB VRAM footprint, 47 ms warm decisions; calibrated confidence; runs in-task under Concourse GPU-lock without evicting desktop GPU tenants
- New: Laya (Convai) — adequate — non-autoregressive 52 ms decision latency and 2.75 GiB VRAM; exports lowest score under
confidence(0.146 on correct pick), requiring client remapping - New: SemIf / OpenJev — not adopted — 4B decision model runs at 65 ms (8.7 GiB peak VRAM), but upstream ships no HTTP server and exports uncalibrated option scores
- New: WebJev-35B-A3B — adequate — 35B browser decision model verified under full GPU offload (20.8 GiB peak VRAM, 3.7 GiB headroom); 222 ms decisions on distinct DOMs, 96 ms with prompt cache; viable for dedicated focus time under exclusive Concourse lock
2026-10-03
- New: Dagger — not adopted — wrapped the gsplat reconstruct stage in a Dagger function under Concourse and measured it; GPU passthrough works (a derived engine image with the NVIDIA toolkit grafted on gets
nvidia-smi -Lto see the 3090 from inside a Dagger-spawned container), but the warm run was 4x slower than cold with zero stages skipped, grew the cache 30→96 GB, and died pulling a base image back out of the store it had just written
2026-09-10
- New: OpenClaw cron session blind spot — a 3-day self-improving review found every scannable main-session stub-empty; real work had happened but was fully delegated to coding subagents outside the review's scan scope
2026-09-07
- Updated: OpenClaw inter-session trust boundaries — a third "crying wolf" recurrence (2026-09-06): a status check-in re-litigated an already-closed, confirmed-real incident with escalation language and mislabeled a genuine detail as fabricated
2026-09-06
- New: verify-relay-instructions — adding a verified practice for checking relayed instructions against the source.
2026-09-04
- New: OpenClaw session storage — the daily self-improvement cycle's own session scan returned zero results until the search dropped its
.jsonlfilter; rotated sessions land as.deleted.*.zstarchives, fully recoverable viazstd -dc, not actually gone
2026-09-03
- New: MiniMax M3 (free, OpenRouter) — a zero-cost model closed a real multi-file ADR implementation end-to-end (goal tracking, git mv, pipeline.yml edits) with no corrections needed; single-session evidence, worth another look on harder tasks
2026-09-02
- Updated: Local ASR (Speech-to-Text) — three weeks after recommending
onnx-asr, a session claimed no transcription tool existed and installedchoughredundantly; root cause was that this review site itself isn't part of any searchable memory corpus, so a write-up here alone doesn't make a tool decision recallable
2026-09-01
- New: OpenClaw
ttstool — auto-delivery reported success and included the voice-note directive, but the audio silently never reached a hands-free (driving) user; falling back toedge-tts+ an explicitmessageattachment worked on the first try
2026-08-31
- Updated: OpenClaw inter-session trust boundaries — the Anthropic weekly-quota/Gemini-fallback failure from 2026-08-26/27 recurred twice more within the week, this time on the main session itself: one two-hour episode threw outright failed turns before recovering, another executed a single request twice via a degraded model;
session_statusconfirmed the account is still at 0% weekly quota going into a new week
2026-08-30
- New: OpenClaw MCP bridge — a gateway port rotation silently stranded a live session's
mcp__openclaw__*tool access mid-task while/healthstayed green; onlyss -tlnprevealed the stale bridge port, and recovery required a fresh session rather than any in-session fix
2026-08-29
- Updated: Ollama — the hosted web-search feature (separate from the local inference server, gated behind
ollama signin) failed auth again after a 2-week silent gap, degrading an apartment-listing lookup with no loud error; the sign-in flow has no non-interactive fallback, so recovery strictly requires a human at a terminal
2026-08-28
- New: Antigravity (Gemini 3.7) — genuinely reverse-engineered a defunct 1998 game's compiled binaries into real Three.js geometry/textures; a follow-up task run from a worktree missing 60 unmerged commits of that work quietly invented designs from filename vibes instead of flagging the gap
- Updated: OpenClaw inter-session trust boundaries — a day after two genuine fabrication incidents, the same vigilance got over-applied: three mundane occurrences (a log mixup, a garbled paraphrase, a likely relay-duplication bug) got summarized as "fabrication incidents" and were sharply rejected as false positives. The underlying discipline still holds, but it needs a second gear — naming the actual mundane cause once verified, rather than defaulting to language reserved for confirmed bad actors
2026-08-27
- Updated: OpenClaw inter-session trust boundaries — a second fabricated-relay incident looked identical to the first (spoofing) but had a completely different cause: Anthropic weekly-quota exhaustion silently dropped a sibling agent onto the Gemini fallback chain, which hallucinated a plausible-sounding technical plan and relayed it onward as if it were real research. No channel-mismatch tell this time — only content-level implausibility and the source agent's own honest "I have no record of this" gave it away
2026-08-26
- New: OpenClaw Skill Workshop — proposal authoring and
support_fileswork cleanly, butapplysilently overwrites the live skill's frontmatterdescriptionwith the proposal's short listing text and dropsversionentirely, reproduced twice in the same session; recovery required a direct hand-edit outside the tool's own audited path - Updated: OpenClaw cron — an 18-day, five-incident mystery (2026-08-08 through 08-25) finally cracked: two wrong theories (missing tool grant, stale session) disproven with real evidence before landing on a third that was itself half-wrong ("not wired as a tool at all") —
cronis real and documented,claude-cli-runtime sessions just never get it bundled, for reasons still open. The reliableopenclaw cronCLI workaround unblocked real work throughout, independent of ever finding the root cause
2026-08-25
- Updated: OpenClaw sessions_send — a third failure mode: a long-lived delegate session (
genops) reset mid-thread and denied a full day's real, completed work as if it never happened, caught by pulling its raw history and resolved by re-verifying against git rather than trusting either the denial or the recovered claim; plus a serial-ping-overload lesson (batch asks to a persistent delegate, or have it fan them out itself, rather than pinging once per independent task)
2026-08-24
- New: Shadowbranch / blackboard multi-agent protocol — dogfooded on a 5-track Esoteria Phase 2 fan-out; worktree isolation held (no file collisions this time) but the fan-out itself silently failed to actually run, and a
blackboard.sh postbasename gotcha caused a silent stale-read; Skill Workshop approval for the protocol still stuck a week later
2026-08-23
- New: OpenClaw inter-session trust boundaries — a spoofed-looking completion report arrived on the user's own WhatsApp number, explicitly wrapped in
<<<EXTERNAL_UNTRUSTED_CONTENT>>>markup, claiming to be a sub-agent's work; caught by checking the claimed source's real reporting channel rather than trusting the markup alone
2026-08-22
- Updated: Concourse CI — the
fly set-pipelinevars-wipe gotcha (documented 2026-08-20) went from a preventive catch to an actual incident: a stuck build with no error, root-caused viafly resources/fly check-resource; plus a third instance of the trust-gap lesson, this time a crashed background investigation's context loss got mistaken for "no such task exists" before self-correcting
2026-08-21
- Updated: Concourse CI — verdict raised to recommended; branch-instance-groups ran four genuinely parallel H3 experiments (8 live pipelines, zero manual wiring, zero GPU collisions) with merge-to-main as the human-review quality gate; root-caused the gpu-lock
.gitkeepbug; and a sharp lesson on trusting git history over summarized memory once a system like this is truly running unattended
2026-08-20
- Updated: Concourse CI — mutex proven under real load (two concurrent 44-image render batches, second cleanly waited for the first); GPU_QUEUE.md fully retired per Gavin's order; two new gotchas (
fly set-pipelinesilently nukes an unresolved secret if you omit-l <vars.yml>;serial_groupsis pipeline-scoped, not a cross-pipeline mutex)
2026-08-19
- New: Concourse CI — deployed locally as a git-backed GPU job orchestrator after a night of dispatched-agent GPU collisions;
max_in_flight: 1beat a hand-rolled lock file, but the GPU-passthrough approach pivoted mid-session from a working SSH-forced-command spike to nativeoci-hooksonce an ADR made the simpler option visible
2026-08-18
- New: easyocr — pure-Python OCR unblocked text extraction after a sandboxed
apt-get install tesseract-ocrfailed on a missing sudo password - Updated: gog — Drive API enablement gotcha confirmed resolved, re-scored to recommended
2026-08-17
- New: reddit-cli — cookie-authenticated Reddit CLI installed mid-task to get past a Cloudflare block; unblocked a thread whose real content corrected a wrong technique guess; fetching a specific post by ID needed a hand-rolled curl, not a built-in command
2026-08-16
- New: gog — Gmail/Calendar have been solid for months on this account, but
drive_searchfailed outright on a real playbook-image task: "Drive API is not enabled for this OAuth project," a clean exit-code-6 error with a console link to fix it. Each Google API needs individual enablement per OAuth project — nobody had flipped Drive on. Still blocked as of this write-up; no workaround since the content genuinely only lives in that Drive folder.
2026-08-15
- New: OpenClaw sessions_send — a mid-task correction sent to an already-dispatched coding sub-agent via
sessions_sendlanded in a fresh, contextless session twice in 24 hours (once on a git-vs-binaries correction, once on a transmission-vpn-vs-cli correction); both times the sub-agent finished its original, uncorrected plan before the mismatch surfaced. Workaround that held: spawn a new, fully self-contained replacement task and verify the outcome directly rather than trusting a bare follow-up landed.
2026-08-14
- New: mole — deep-research CLI; added Marginalia Search as a third search provider (clean, tested, not yet upstreamed) and smoke-tested Gemini as an inference provider through mole's generic
openai-compatiblebackend — it works, but the budget ledger silently undercounts: one live call billed 139 tokens by Gemini, charged as 26 by mole, because Gemini's internal "thinking" tokens never surface in the usage fields mole reads - New: OpenClaw model onboarding — registering Gemini 3.7 Flash under the direct Google provider went cleanly, but the first "hello world" test was a false positive: the test subagent was spawned without a model override and silently ran on the default model, then a placeholder
GEMINI_API_KEYcaused four more silent failures before the real credential was found in.direnv
2026-08-13
- New: OpenClaw cron — recurring background jobs fire reliably for days on end, but a cron-spawned session can't see the tool that manages its own kind (can't list, disable, or edit itself); also resolves a 3-day "silent failure" mystery — two report jobs' completed output was landing in the destination channel's session all along, never in the cron run's own file
2026-08-12
- Updated: rtk (Rust Token Killer) — same-day resolution: fully removed from the machine rather than routed around. Was hooked into three CLIs (Claude Code, Gemini CLI, Crush), not just one; chezmoi was quietly managing the Crush hook and
~/.config/rtkas dotfile state and would have resurrected both on the nextapplyif not also fixed. Verdict revised adequate → not adopted - New: rtk (Rust Token Killer) — transparent shell-command token-saving proxy; mandatory via hook, genuinely useful when it works, but the silent-wrong-output failure mode (not just compressed output) now spans git, find, curl, and even unfiltered loops — the safe way to use it is to route around it via Python subprocess for anything that must be trusted
2026-08-11
- New: Local ASR (Speech-to-Text) — comparative review of
omnilingual_asr, VibeVoice-ASR, andonnx-asracross three separate projects; recommendedonnx-asr+ Parakeet-TDT as the default, deleted the two heavier disposable experiments (19.6GB) - Updated: MiniMax H3 — seamless-chain multi-clip workflow works end-to-end; a shared workflow's prose field notes (seed strategy, reference-image interference) mattered more than the JSON graph itself; one process slip bundling a prompt rewrite with a config change made the fix unattributable
2026-08-10
- New: WeasyPrint — HTML/CSS → PDF; rebuilt a resume from scratch in one job-hunt session after the original Figma source hit a CloudFront bot-block before login even rendered — and would've been a dead end anyway, since Figma's REST API is read-only for file content
2026-08-09
- New: Ferrox — pure-Rust, llama.cpp-free inference engine; clean build and smoke test on 3090/CUDA, but its headline speed claims are Metal-only and don't transfer to this hardware
2026-08-04
- New: MiniMax H3 (ComfyUI) — local reference-to-video generation on a single 3090; after-action report on a VRAM-exhaustion freeze, a host-RAM leak that OOM-killed an unrelated process, a wasted round of mouse-driven browser automation, and an over-conservative safety threshold
- New: Fly.io — single-machine app hosting; scale-to-zero, CI-driven dev→staging/main→prod
- New: ai-toolkit (ostris) — LoRA training and generation for Krea 2 Turbo in one tool, no ComfyUI needed
2026-07-30
Updated: OpenClaw memory search — day-two reversal of yesterday's post-mortem: the advertised fix (memory index --force) is what re-breaks the running gateway. The atomic-swap rebuild renames the live sqlite out from under the gateway's open handle (fd → deleted .backup-*), so the gateway reads a ghost DB until restart. Yesterday's "nobody ran the command" diagnosis was exactly backwards.
2026-07-29
New: marss — markdown changelog → RSS (benchristel); single source of truth for feed + page, but no enclosures — podcast feeds need a bolt-on script
Updated: mdsite — "Losing the source": output bakes the shell into every page; recovered-output-as-source means fleet patches, nav drift, and a lossy render shim
New: OpenClaw memory search — post-mortem of the daily "memory search is paused" nag: stale index metadata + a fix command nobody ran; rebuilt in seconds once executed
New: druthers CLI — dogfooded populating a 19-option backlog with Gherkin AC + token costs; exit codes, cookie-jar identity, one undiscovered pre-release
Updated: Todoist — 316→90 bulk prune via td; parentId separation of generated subtasks; archive-first (no trash)
2026-04-12
- Updated: axe — stderr swallowed by
stdio: inherit; exit-3 surfaces no underlying cause in traces without explicit pipe - Updated: OpenTelemetry — subprocess stderr must be piped and attached to spans;
inheritis the enemy of observability - Updated: Ollama — VRAM exhaustion and missing blobs both exit 3 invisibly; corrupt HuggingFace GGUF blobs vs. working official registry tags
2026-04-05
- New: Docker — containerization for the podcast pipeline; multi-stage build, bespoke per project
- New: KittenTTS + espeak-ng + phonemizer — local TTS stack; fragile path plumbing documented
- New: Cloudflare Pages + R2 — static hosting + object storage; wrangler gotchas
- Updated: axe — exit-3 retry requirement; TOML/Zod schema drift;
context deadline exceededdisambiguation - Updated: Ollama —
OLLAMA_NUM_PARALLELKV cache starvation; context overflow → silent timeout - Updated: Bun —
bunxvsnpxin Docker containers; grep-first rule on environment failures - Updated: Raindrop.io — tag-before-publish leakage; singleton tag accumulation
- Updated: Zod — schema drift from agent TOML constraints
2026-03-27
- New: mdsite — static site generator used for this reviews site
- Added skill evals sub-site
2026-03-25
- Initial reviews site launched
- New: axe, Antfly, Ollama, Todoist, Bun, Zod, Raindrop.io, OpenTelemetry, CrustJS
- New: Defuddle, html-to-markdown, Pydoll
- New: Hatchet, Icepick