<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>harness-atlas — data snapshots</title>
  <id>https://orvii.github.io/harness-atlas/</id>
  <link href="https://orvii.github.io/harness-atlas/feed.xml" rel="self"/>
  <link href="https://orvii.github.io/harness-atlas/"/>
  <updated>2026-10-06T00:00:00Z</updated>
  <author><name>Orvii</name></author>
  <entry>
    <title>v3.5: thirty-seven harnesses — the platform vendors arrive</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-06/v3-5-thirty-seven-harnesses-the-platform-vendors</id>
    <updated>2026-10-06T00:00:00Z</updated>
    <content type="html">Added Seven rows: **Warp** (terminal-native agent; client is AGPL-3.0 and public at `warpdotdev/warp`; per-action autonomy profiles; orchestration explicitly one level deep; Agent Memory research-preview), **Trae** (ByteDance&#x27;s VS Code-fork IDE plus a separate VS Code/JetBrains plugin product; opt-in sandbox whose allowlist bypasses it, and an Auto-Run mode the docs say &quot;bypasses all security checks&quot;; SOLO Spec/Plan workflows pause for confirmation), **Google Antigravity** (default-on OS sandbox from Linux kernel namespaces and macOS Seatbelt/SBPL, no VMs or containers; `~/.ssh` and `.env` blocked; Deny &gt; Ask &gt; Allow precedence engine; plugin.json marketplace), **Qoder** (Alibaba; 14/14 capabilities documented as yes — the first row to do so without a partial; three access-permission modes, selectable execution environments incl. isolated git worktrees), **Lovable** and **v0** (web-native builders joining Bolt and Replit — IDEs connect *to* them via MCP rather than the reverse, so `ide_integration` is `no` on the Zed precedent; v0&#x27;s VM-backed chats run inside a Vercel Sandbox), **Void** (open source, archived, README opens &quot;Void is now deprecated&quot;; below the capability floor because the project stopped, not because it was designed narrow). `scripts/fetch-trae-docs.py` — Trae&#x27;s docs are a client-side SPA with no `llms.txt`, no `sitemap.xml`, no `.md` variants and no content API: all four return the same 287 KB app shell. The script reads the server-rendered payload the SPA embeds (`window._ROUTER_DATA…docDetail.content`, a component tree of Quill deltas) and flattens it to text, so the row stays reproducible by anyone. `SYNTHESIS.md` §11 — the floor is what a platform vendor ships on day one; documentation architecture became a research variable; Warp breaks the closed-source expectation; the web-native cluster inverts the `ide_integration` axis. `GATES.md` — Trae&#x27;s Auto-Run added to the eighth kind (the gate that opens under automation), now five rows deep. `TRUST.md` — seven rows; Antigravity and Qoder recorded as *partial* on the model-review column with the reason stated (their docs do not say whether the reviewer is a model or a rule engine).</content>
  </entry>
  <entry>
    <title>full-audit link-rot pass: Qodo repaired, cited-as-dead class recognized</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-06/full-audit-link-rot-pass-qodo-repaired-cited-as-</id>
    <updated>2026-10-06T00:00:00Z</updated>
    <content type="html">Modified `journals/wf_v34-incoming.jsonl` — the first full two-class audit (636 claim-bearing URLs across all six waves, previously only markdown-linked forms were checked) found two genuinely dead Qodo prose citations: `code-review/trigger-a-code-review` → `install-qodo/install` (its installation matrix carries the platform list the sentence claims, and the sentence now names the platforms the live page names: Bitbucket Cloud and Data Center, Azure DevOps) and `agentic-toolbox/manage-standards` → `…/manage-standards-skill`. Verdicts unchanged. `scripts/link-rot.sh` — third URL class: **cited-as-dead**. A note may cite a URL precisely because it 404s (&quot;the architecture page returned HTTP 404&quot; — Aider; &quot;docs frozen&quot; — Continue; &quot;legacy slug&quot; — Sourcegraph Amp). The checker detects the 404 predication in the clause immediately following each URL (bounded by the next URL, a newline, or 60 characters), skips the request, and reports the citation separately. Without this, the gate would flag the atlas&#x27;s most honest cells — deprecation documentation — as contract violations. The locality bound matters: one Amp sentence cites the dead legacy slug *and* the live replacement docs home in the same breath; window- and sentence-wide heuristics both misclassified the live URL, suppressing its check (caught by comparing run 2&#x27;s 84 cited-as-dead against the 3 URLs any note actually calls dead). Added `reports/link-rot.md` — first full-audit run: 1116 unique URLs, claim/provenance/cited-as-dead separated.</content>
  </entry>
  <entry>
    <title>wave 6 link-rot repair: twelve moved evidence URLs re-pointed</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-06/wave-6-link-rot-repair-twelve-moved-evidence-url</id>
    <updated>2026-10-06T00:00:00Z</updated>
    <content type="html">Added `scripts/link-rot.sh` rewrite — checks **every** URL on every page (markdown-linked and bare), split into two classes: claim-bearing (evidence links, prose citations, quote attributions — a dead one exits non-zero and fails the monthly drift CI, report still commits first) and provenance (&quot;Sources fetched&quot; navigation logs — dead entries reported, not failed). The old markdown-only regex could not see parenthetical prose citations; the wave-6 audit found a dead one (`replitai/plan-vs-build-mode`) in exactly that blind spot. `drift.yml` — gate step after the report commit: dead claim-bearing URLs now turn the monthly run red instead of accumulating silently in the report. Modified `journals/wf_v34-incoming.jsonl` (and the gitignored `journals-incoming/` sources) — eleven vendor doc pages that moved between the research run and publication, plus one prose citation, re-pointed to their live canonical URLs: Augment `cli/agents`→`cli/subagents`, `cli/integrations-and-mcp`→`cli/integrations`, `cosmos/sessions`→`cosmos/sessions-overview`, `cosmos/understanding-automation`→`cosmos/automations`, `setup-augment/install-augment-for-{vscode,jetbrains}`→ the live IDE-setup paths; Replit `features/agent/task-system`→`core-concepts/agent/task-system`, `replitai/build-in-parallel`→`learn/build-in-parallel`, `replitai/mcp`→`chat/connect-through-mcp`, `billing/managing-your-spend`→`billing/managing-spend`, `replitai/plan-vs-build-mode`→`learn/plan-vs-build-mode`; Bolt `integrations/github`→`integrations/git`. `harnesses/{augment-code,replit-agent,bolt}.md` regenerated — verdicts and notes unchanged, only the citation URLs moved.</content>
  </entry>
  <entry>
    <title>v3.4: thirty harnesses — the capability-vendor cluster</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-06/v3-4-thirty-harnesses-the-capability-vendor-clus</id>
    <updated>2026-10-06T00:00:00Z</updated>
    <content type="html">Added Five rows: **Augment Code** (Auggie CLI + Cosmos: subagents, trigger automations, Expert memory, Git-marketplace plugins), **Tabnine** (broad CLI agent surface on a CLI the vendor calls maintenance-mode, deprecation after 2026-12-31), **Replit Agent** (cloud task system with plan-gated parallelism, no published version), **Qodo** (Agentic Toolbox: review agents and standards that plug into other agents — a capability layer with no execution layer), **Bolt** (web-native agent; plan review without a documented hard gate). `SYNTHESIS.md` §10 — the stack is splitting: capability vendors that execute nothing; web-native rows where the environment is the product; deprecation as a row property; unknowns concentrating in hosted vendors. `GATES.md` eighth kind — the gate that opens under automation: Tabnine&#x27;s headless YOLO switch, Bolt&#x27;s plan-after-first-build, Replit&#x27;s billable review, Augment&#x27;s CLI-only enforcement. `TRUST.md` — five rows; Qodo recorded as n/a on the sandbox axis (it executes nothing); new bullet on review-layer vendors riding host agents&#x27; trust models. `journals/wf_v34-incoming.jsonl` — wave 6, five independent researchers, merged via the new `scripts/merge-incoming.py` (contract-validated on entry).</content>
  </entry>
  <entry>
    <title>enforced reproducibility: public journals + regeneration gate</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-06/enforced-reproducibility-public-journals-regener</id>
    <updated>2026-10-06T00:00:00Z</updated>
    <content type="html">Added `journals/` — sanitized researcher results (one JSON line each; head + enrichment lines, wave-ordered; MANIFEST.md records merge order). `python3 scripts/generate.py journals/*.jsonl` in wave order reproduces every committed page byte-identically. `scripts/export-journals.py` — exports sanitized results from workflow journals in the generator&#x27;s line shape. `.github/workflows/regen-check.yml` — regenerates pages from journals/ on data-path pushes and fails if any page or the harness set differs. Modified `scripts/generate.py` — enrichment-only results may enrich an existing row but never start one; `as_of` carried forward from the committed matrix.yaml so regeneration is byte-stable.</content>
  </entry>
  <entry>
    <title>site redesign: specimen-plate layout, plate-mode grid, deep links</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-06/site-redesign-specimen-plate-layout-plate-mode-g</id>
    <updated>2026-10-06T00:00:00Z</updated>
    <content type="html">Modified `scripts/build_site.py` — index.html rebuilt around a broadsheet masthead: two-column head with a framed &quot;reading the plate&quot; specimen (symbol key, live cell sample, snapshot ledger) replacing the chip row; legend moved out of the toolbar; toolbar gains verdict filter (rows containing ●/◐/○/?) and row sort (name a–z, most documented, most contested); compare view gains a disagreement-count summary line; grid gains a per-column census bar (yes/partial/no/unknown proportions + documented count) as a second sticky head row and two-digit row indices; every cell is deep-linkable (`#&lt;harness&gt;/&lt;capability&gt;`) with a copy-link button in the drawer; `&lt;noscript&gt;` fallback points at matrix.md. `scripts/build_site.py` — sticky-head fix: the grid is now its own scroll plate on desktop (`max-height` + `overflow: auto`), so the head pins to the plate top; on ≤980px the head goes static and the name column stays pinned. Root cause: an `overflow-x` scroll container becomes the sticky reference, which previously floated the head mid-grid on small screens. `scripts/build_site.py` — shared CSS keeps `.chips`/`.chip` for method.html; entry animation shortened to 350ms.</content>
  </entry>
  <entry>
    <title>v3.3: twenty-five harnesses — the cloud-autonomous cluster</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-05/v3-3-twenty-five-harnesses-the-cloud-autonomous-</id>
    <updated>2026-10-05T00:00:00Z</updated>
    <content type="html">Added `harnesses/cursor.md`, `windsurf.md`, `google-jules.md`, `amazon-kiro.md`, `devin.md` — five closed-source agents incl. the first three whose loop runs server-side (Jules, Devin, Kiro cloud sessions). Version pins from npm dist-tags, vendor changelogs and API self-labels; each row says which. `SYNTHESIS.md` §9 &quot;Execution locality is now a first-class axis&quot; — the human gate relocates when the loop leaves the machine (Jules: plan approval instead of prompts; Windsurf/Devin: fail-closed sandbox); where the human left, a model was hired (Cursor Auto-review default, Jules&#x27; critic agent); `background_tasks` was conflating async with remote. `SYNTHESIS.md` §6 retitled: the compatibility chain no longer has one direction — v3.3 rows cross-read Claude/Codex/Cursor directories. `TRUST.md` five rows; prompts-floor now has two exceptions (Amp by choice, Jules by structure); cloud rows are VM-isolated by construction while local sandboxes stay opt-in. `GATES.md` seventh gate kind (locality) plus the new env/tier/OS/maturity/deprecation entries.</content>
  </entry>
  <entry>
    <title>v3.2: twenty harnesses — the closed-source cluster</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-05/v3-2-twenty-harnesses-the-closed-source-cluster</id>
    <updated>2026-10-05T00:00:00Z</updated>
    <content type="html">Added `harnesses/sourcegraph-amp.md`, `harnesses/factory-droid.md`, `harnesses/jetbrains-junie.md` — three closed-source commercial agents, researched from vendor docs with the same evidence contract (per-cell URL + verbatim quote). Version pins come from npm `dist-tag latest`, the vendor changelog page, and the vendor what&#x27;s-new page respectively, since no release tags exist; each row says which. `SYNTHESIS.md` §8 &quot;The closed-source cluster answers differently&quot; — what changes when a pin is not a git tag, when deprecation is a policy rather than a commit, and when a trust claim is a vendor security page. `GATES.md` &quot;A sixth kind: the closed-source gate&quot; plus the new harnesses&#x27; env/tier/OS/maturity/deprecation gates in the existing five lists. `TRUST.md` rows for all three, including **Sourcegraph Amp as the first harness in the set that documents no local permission prompt at all** (&quot;By default, Amp does not ask for approval before running tools&quot;) — the universal-prompt floor now has one hole. `bench-notes`: `the-benchmark-code-is-the-result` (note 14). `equivalence-notes`: `environment-is-part-of-the-program` (note 11).</content>
  </entry>
  <entry>
    <title>v3.1: gates, coverage honesty, reading guides</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-05/v3-1-gates-coverage-honesty-reading-guides</id>
    <updated>2026-10-05T00:00:00Z</updated>
    <content type="html">Added `GATES.md` — the layer behind documented: env/flag gates, tier gates, OS gates, maturity labels, deprecation gates README &quot;How to read a cell&quot; — partial is a verdict with a note; unknown is absence of evidence</content>
  </entry>
  <entry>
    <title>v3: 17 harnesses, deep sections, trust + precedence + citation</title>
    <id>tag:github.com,2026:Orvii/harness-atlas/2026-10-05/v3-17-harnesses-deep-sections-trust-precedence-c</id>
    <updated>2026-10-05T00:00:00Z</updated>
    <content type="html">Added `harnesses/` pages for crush, amazon-q-developer-cli, vibe (batch 3, full capability + deep pass) Deep-dive sections on all pages: architecture, context management, ecosystem, governance, limitations, quotes, sources (research runs wf_15cd0c06, wf_3decfb86) `TRUST.md` — three-layer trust model (OS sandbox / prompts / model review), default postures per harness `scripts/drift-check.sh` — mechanical version-pin freshness report; multi-stream release tags handled (Qwen CLI vs sdk), nightlies excluded `scripts/link-rot.sh` — evidence-URL rot checker; dead (404/410) vs bot-gated (403/429) separated; first run: 235 URLs, 232 ok, 0 dead</content>
  </entry>
</feed>
