agentvision 0.3.0__tar.gz → 0.3.2__tar.gz

{agentvision-0.3.0 → agentvision-0.3.2}/CHANGELOG.md

@@ -55,6 +55,52 @@ clean, provider-agnostic handoff — it perceives and hands off, it does not dec
   brain's sense adapter can keep it for provenance yet exclude it from gating (this honors the
   contract Verel's `verel.senses.sight` already documents and tests).
 
+## [0.3.2] — 2026-06-19
+
+### Fixed — v0.3.1 retest residuals (canvas/WebGL visual path)
+Two residual findings from re-grading the live WebGL dashboard, both on the vision path for
+**visual (non-text) elements**:
+
+- **Freeze no longer blanks canvas/WebGL scenes that build inside `requestAnimationFrame`.**
+  Freeze is now *canvas-aware*: CSS animations/transitions are still paused, but when a
+  `<canvas>` is present rAF is left running and the renderer waits `canvas_settle_ms`
+  (default 1500) so the scene draws before capture ("settle-then-freeze"). Heavy scenes can
+  still raise `--settle-ms`. (Static pages keep the full rAF freeze.)
+- **Vision "missing" claims about visual elements are overruled by the DOM.** The renderer
+  reports sizable `canvas`/`svg`/`img`/`video` elements (`RenderResult.visual_tags`); a vision
+  `missing_element`/`intent_mismatch` that names a visual (canvas, chart, 3D scene, image…)
+  is suppressed when the matching element is actually present — so a downscaled/blurry full
+  page no longer false-fails a chart or canvas that demonstrably exists.
+
+## [0.3.1] — 2026-06-19
+
+### Fixed / changed — live-dashboard field report (rhtrader)
+Hardening from grading a live, continuously-polling WebGL dashboard. The biggest wins:
+don't default to `networkidle`, settle/freeze before capture, and never let a vision
+"missing" claim survive when DOM/OCR proves the element is present.
+
+- **`nav_wait` default is now `load`** (was `networkidle`). Polling/websocket pages never
+  go idle and used to hang to `RenderTimeout`. When `networkidle` *is* requested it is now
+  **bounded** (≤5s) and falls through instead of blocking. `--nav-wait` is a CLI flag.
+- **Settle + freeze before capture.** A short post-load settle (`settle_ms`, default 400)
+  lets client-rendered data populate (fixes false "blank/missing" on the shell-then-fill
+  frame); `freeze_animations` (default on) pauses CSS animations + the `requestAnimationFrame`
+  loop and the screenshot uses `animations="disabled"`, so canvas/WebGL/animation-heavy pages
+  (incl. **`--full-page`**) capture deterministically instead of timing out. `--settle-ms`,
+  `--freeze/--no-freeze`, and `--wait-for <selector>` are CLI flags.
+- **Vision claims cross-checked against DOM/OCR ground truth.** An advisory vision
+  `missing_element` / `intent_mismatch` finding is suppressed when the quoted element text is
+  actually present in the DOM/OCR — killing the false-fail class entirely.
+- **Intent text claims grade against DOM *and* OCR**, so `--expect 'must: "…"'` is
+  deterministic even with no OCR/no key (the text is read from the DOM).
+- **Oversized screenshots are downscaled** before the vision LLM (`vision_max_edge_px`,
+  default 2000) — large/dense images made models return lazy/generic critiques.
+- **`--allow-local`** CLI flag for localhost / LAN dev servers (clearer than the env var),
+  and the `UnsafeSourceError` now names it.
+- **`--quiet` machine mode**: only the JSON object on stdout (logs on stderr), errors as a
+  JSON `{"error": …}`, stable exit codes (0 pass/warn, 2 fail, 3 error).
+- **Default `render_timeout_s` raised to 60s** and exposed as `--render-timeout`.
+
 ## [0.2.0] — 2026-06-18
 
 ### Added

{agentvision-0.3.0 → agentvision-0.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentvision
-Version: 0.3.0
+Version: 0.3.2
 Summary: Eyes for AI Agents — a machine-graded visual feedback loop coding agents consume to self-correct before claiming done.
 Project-URL: Homepage, https://github.com/amitpatole/agent-vision
 Project-URL: Repository, https://github.com/amitpatole/agent-vision
@@ -161,6 +161,23 @@ agentvision analyze ./page.html --handoff
 `report.to_handoff()`, the MCP `perceive_handoff` tool, `POST /handoff`, and a `handoff.json`
 per loop iteration — provider- and brain-agnostic. See [docs/handoff.md](docs/handoff.md).
 
+## Eyes & Brain — AgentVision × Verel
+
+AgentVision is the **eyes**. It pairs with **[Verel](https://github.com/amitpatole/verel)**,
+the **brain** — an agent framework where *nothing is "done" until a grader returns a verdict.*
+The eyes perceive and grade intent; the brain decides with attestation and **compounds only
+verified work** into memory; then the eyes look again.
+
+<p align="center">
+  <img src="docs/unified-architecture.png" alt="Eyes & Brain — AgentVision perceives and grades intent; Verel decides and compounds verified work into memory" width="100%">
+</p>
+
+They ship and version independently (`pip install agentvision`, `pip install verel`) yet work
+in sync: AgentVision plugs into Verel as its `verel.senses` perception organ — mapped onto a
+unified verdict bus (vision *alongside* tests, lint and types), with intent conformance
+recorded in the brain's memory each iteration. AgentVision stays brain-agnostic; Verel is the
+reference brain. See [docs/handoff.md](docs/handoff.md).
+
 ## Many faces, one core
 
 | Surface | Who it's for |
@@ -221,6 +238,18 @@ agentvision baseline ./index.html --name home
 agentvision regress  ./index.html --name home
 ```
 
+**Live pages, SPAs & dashboards** (polling, websockets, canvas/WebGL):
+
+```bash
+# localhost dev server, wait for the data to render, freeze animation, machine output
+agentvision analyze http://localhost:5173 --allow-local \
+  --wait-for "#dashboard" --settle-ms 800 --quiet
+```
+
+`--nav-wait` defaults to `load` (polling pages never go idle); `--freeze` (default on) pauses
+animations + `requestAnimationFrame` so canvas/WebGL pages capture without hanging; `--quiet`
+prints only JSON (logs to stderr, exit codes 0 pass/warn · 2 fail · 3 error).
+
 Library:
 
 ```python

{agentvision-0.3.0 → agentvision-0.3.2}/README.md

@@ -97,6 +97,23 @@ agentvision analyze ./page.html --handoff
 `report.to_handoff()`, the MCP `perceive_handoff` tool, `POST /handoff`, and a `handoff.json`
 per loop iteration — provider- and brain-agnostic. See [docs/handoff.md](docs/handoff.md).
 
+## Eyes & Brain — AgentVision × Verel
+
+AgentVision is the **eyes**. It pairs with **[Verel](https://github.com/amitpatole/verel)**,
+the **brain** — an agent framework where *nothing is "done" until a grader returns a verdict.*
+The eyes perceive and grade intent; the brain decides with attestation and **compounds only
+verified work** into memory; then the eyes look again.
+
+<p align="center">
+  <img src="docs/unified-architecture.png" alt="Eyes & Brain — AgentVision perceives and grades intent; Verel decides and compounds verified work into memory" width="100%">
+</p>
+
+They ship and version independently (`pip install agentvision`, `pip install verel`) yet work
+in sync: AgentVision plugs into Verel as its `verel.senses` perception organ — mapped onto a
+unified verdict bus (vision *alongside* tests, lint and types), with intent conformance
+recorded in the brain's memory each iteration. AgentVision stays brain-agnostic; Verel is the
+reference brain. See [docs/handoff.md](docs/handoff.md).
+
 ## Many faces, one core
 
 | Surface | Who it's for |
@@ -157,6 +174,18 @@ agentvision baseline ./index.html --name home
 agentvision regress  ./index.html --name home
 ```
 
+**Live pages, SPAs & dashboards** (polling, websockets, canvas/WebGL):
+
+```bash
+# localhost dev server, wait for the data to render, freeze animation, machine output
+agentvision analyze http://localhost:5173 --allow-local \
+  --wait-for "#dashboard" --settle-ms 800 --quiet
+```
+
+`--nav-wait` defaults to `load` (polling pages never go idle); `--freeze` (default on) pauses
+animations + `requestAnimationFrame` so canvas/WebGL pages capture without hanging; `--quiet`
+prints only JSON (logs to stderr, exit codes 0 pass/warn · 2 fail · 3 error).
+
 Library:
 
 ```python

agentvision-0.3.2/docs/field-report-rhtrader.md

@@ -0,0 +1,180 @@
+# Field report: AgentVision on a live trading dashboard (rhtrader)
+
+> **Status: all 8 addressed in v0.3.1.** Summary of fixes:
+> 1. `nav_wait` defaults to `load`; `networkidle` is now bounded (≤5s) — no more hangs.
+> 2. Freeze (pause CSS animations + rAF) + `screenshot(animations="disabled")` → full-page
+>    capture works on canvas/WebGL/animation-heavy pages (`--freeze`, default on).
+> 3. Post-load settle (`settle_ms`, default 400) + discoverable `--wait-for <selector>` →
+>    no false "blank/missing" on the shell-then-fill frame.
+> 4. Oversized screenshots downscaled before the vision LLM (`--vision_max_edge_px`/2000);
+>    **vision `missing_element` claims cross-checked against DOM/OCR and suppressed if the
+>    element is provably present.**
+> 5. Text claims grade against DOM **and** OCR — deterministic even with no OCR/no key.
+> 6. `--allow-local` CLI flag (and the error message names it).
+> 7. `--quiet` machine mode: JSON-only stdout, logs to stderr, stable exit codes (0/2/3).
+> 8. Default `render_timeout_s` raised to 60s; exposed as `--render-timeout`.
+
+
+Real drawbacks hit while using AgentVision (`agentvision analyze`, `local` + `openai`
+backends) to grade and fix a live, continuously-polling, WebGL dashboard
+(`http://localhost:8910/` — a stdlib HTTP server rendering a cockpit that polls
+`/state.json` every 3s and runs a three.js scene with a perpetual `requestAnimationFrame`
+loop). Dated 2026-06-19. Ordered by impact. Each item: **what happened → impact →
+suggested enhancement.**
+
+## 1. `networkidle` (the default `nav_wait`) hangs forever on polling pages
+- **What:** the dashboard polls `/state.json` every 3s, so the network is never idle.
+  `analyze <url>` hit `RenderTimeout: Render exceeded 30.0s (hanging page?)`. Only
+  `AGENTVISION_NAV_WAIT=load` unblocked it.
+- **Impact:** out-of-the-box failure on *any* live dashboard / SPA with polling or
+  websockets — arguably the most common "real" target.
+- **Suggestion:** don't default to `networkidle`; default to `load` + a short fixed
+  settle, OR cap the networkidle wait (e.g. 3–5s of quiet) and fall back to `load`
+  automatically. At minimum, when `RenderTimeout` fires, the message should suggest
+  `nav_wait=load`.
+
+## 2. Continuously-animating pages never "settle" → full-page capture times out
+- **What:** `--full-page` timed out even with `nav_wait=load` and
+  `AGENTVISION_RENDER_TIMEOUT_S=75`. Root cause: the three.js scene animates forever
+  (rAF), and full-page scroll-and-stitch waits for stability that never comes.
+  Viewport-only worked; full-page did not.
+- **Impact:** can't grade the whole page on any canvas/WebGL/animation-heavy app.
+- **Suggestion:** a "freeze before capture" mode — pause rAF (override
+  `requestAnimationFrame`), set `* { animation-play-state: paused !important;
+  transition: none !important }`, and/or `prefers-reduced-motion` — then snapshot.
+  Or a hard `--settle-ms N` that snapshots regardless of stability.
+
+## 3. Captures at `load` BEFORE client-rendered data populates → false "blank/missing"
+- **What:** the page renders its shell at `load`, then JS fills the metric cards/data a
+  beat later. AgentVision captured the in-between frame; the vision backend then
+  reported plainly-present elements ("TODAY'S NET", regime, equity curve…) as **missing**
+  (`matches_intent: false`, 10+ false `[intent/must]` items). The actual screenshot
+  (artifact) showed a fully-populated page — the capture was just early.
+- **Impact:** systematically wrong verdicts on client-rendered UIs; erodes trust.
+- **Suggestion:** make `--wait-for <selector>` / `--settle-ms` first-class in
+  `analyze --help` (the selector path exists in the renderer but isn't discoverable);
+  consider a default small post-`load` settle (e.g. 800ms); or wait for DOM mutations
+  to quiesce.
+
+## 4. Vision-LLM gives lazy/hallucinated output on large/dense images
+- **What:** with `--backend openai` on a `1440x1600` screenshot, the model returned a
+  generic checklist-echo ("everything missing"), clearly not reading the image. The
+  **same** model on a focused `1280x720` viewport produced a genuinely useful, specific
+  critique (low-contrast amber, icon crowding, "booting…" confusion, missing section
+  grouping). The failure correlated with image size/density.
+- **Impact:** large pages silently get garbage critiques.
+- **Suggestion:** (a) downscale/tile oversized screenshots before sending and/or warn
+  when an image exceeds a model-friendly size; (b) **cross-check vision
+  `missing_element` claims against DOM/OCR ground truth and suppress contradicted
+  ones** — if the DOM/OCR clearly contains the element, don't emit a vision "missing"
+  finding. This single guard would have prevented the entire false-fail above.
+
+## 5. Intent-conformance wasn't actually deterministic in practice
+- **What:** docs say text claims grade deterministically via OCR, but with the `openai`
+  backend, `--expect 'must: shows "TODAY'S NET"'` came back failed even though that exact
+  text is large and OCR-able and present.
+- **Impact:** `--expect`/`--brief` conformance can't be trusted when a vision backend is
+  selected.
+- **Suggestion:** grade `--expect` text claims via OCR **independently** of the vision
+  model's holistic judgment, and report per-claim which grader decided it (ocr vs
+  vision) so the user can tell deterministic from advisory.
+
+## 6. SSRF guard blocks localhost/LAN with only an env-var escape hatch
+- **What:** `analyze http://localhost:8910/` →
+  `UnsafeSourceError: Refusing to render 'localhost' … (SSRF protection)`. Fixed only via
+  `AGENTVISION_BLOCK_PRIVATE_NETWORKS=false`.
+- **Impact:** the most common dev target (a localhost dev server) is blocked by default
+  with no obvious CLI flag.
+- **Suggestion:** a `--allow-local` / `--lan` CLI flag (clearer than the env var), and
+  let the `UnsafeSourceError` message name that flag.
+
+## 7. Machine-readable output is hard to capture reliably
+- **What:** `--handoff` / `--json` output was intermittently empty on stdout, sometimes
+  appeared mixed with Rich/console formatting, and some runs exited non-zero with no
+  parseable body. We had to regex-extract the JSON object (`{…}`) from mixed output and
+  retry.
+- **Impact:** brittle to script around (and scripting is the whole point for agents/CI).
+- **Suggestion:** a strict `--quiet`/`--machine` mode that writes **only** the JSON
+  object to stdout (all logs/Rich to stderr), with documented, stable exit codes.
+
+## 8. Default `render_timeout_s` (30s) is low for heavy SPA/WebGL
+- **What:** 30s wasn't enough once the page was non-trivial; bumping via
+  `AGENTVISION_RENDER_TIMEOUT_S` was necessary (though for full-page the real fix is #2,
+  not a bigger timeout).
+- **Suggestion:** raise the default a bit and/or expose `--render-timeout` as a CLI flag
+  (currently env-only).
+
+---
+
+### What worked well (keep)
+- The **`local` backend** (DOM geometry + WCAG contrast + OCR/typo) was reliable and
+  trustworthy throughout — it correctly found **106 real WCAG-AA contrast violations** on
+  first run and a clean **PASS / 0 issues** after the fix. The deterministic path is the
+  star; the advisory vision path needs the guardrails above.
+- The **handoff shape** (`perceived` / `next_action` / `todo` / `open_questions`) mapped
+  cleanly into a Verel verdict-bus gate.
+
+### One-line meta-takeaway
+The biggest wins are **(a) don't default to `networkidle`, (b) settle/freeze before
+capture, and (c) never let a vision "missing_element" claim survive when DOM/OCR proves
+the element is present.** Those three remove ~all the false fails we hit.
+
+---
+
+## v0.3.1 retest (2026-06-19) — confirmed fixes + TWO residual findings
+
+Re-ran against the same dashboard on **v0.3.1**. Confirmed working: full-page no longer
+times out (freeze), `--allow-local`, `--settle-ms`, `--quiet` (JSON-only, exit 0/2/3),
+console-error capture (it caught our WebGL GPU-stall warnings), and — the big one —
+**OCR-literal `--expect` claims (`"TODAY'S NET"`, `"EQUITY CURVE"`, `"OPEN POSITIONS"`)
+are now correctly recognized as present** (previously all false-missing). Two new things
+surfaced:
+
+### 9. `--freeze` (default ON) pauses rAF → WebGL/canvas scenes that *build* via rAF never render
+- **What:** our three.js scene builds its first frame *inside* the `requestAnimationFrame`
+  loop. With `--freeze` (default), rAF is paused before that first build, so the canvas
+  stays empty and the vision backend reports **"no 3D scene"** — a false negative. With
+  `--no-freeze` + `--settle-ms 6000`, the full scene rendered correctly (verified: floor,
+  figures, ticker ring, bins all present).
+- **Impact:** freeze (great for CSS-animation determinism) silently breaks grading of any
+  WebGL/canvas/game UI that lazy-builds in rAF.
+- **Suggestion:** freeze CSS animations/transitions by default but **let rAF run for N ms
+  (or until the canvas has drawn a non-blank frame) BEFORE pausing it** — i.e.
+  "settle-then-freeze," not "freeze-then-settle." Or auto-detect a `<canvas>` and switch
+  to no-freeze. At minimum, document that WebGL targets need `--no-freeze`.
+
+### 10. Full-page downscale makes small-but-present *visual* elements read as "missing"
+- **What:** the full page is ~3350px tall; downscaled to the vision max edge (~2000px),
+  the 3D-scene region becomes small/blurry, and the model reported the (clearly present)
+  scene as missing. The DOM/OCR cross-check (fix #4) saves **text** claims but not
+  **visual-only** regions like a `<canvas>`.
+- **Impact:** intent claims about non-text visuals (charts, canvases, images) still
+  false-fail on tall pages.
+- **Suggestion:** for a visual intent claim, resolve the relevant DOM node's bbox and send
+  the model a **focused, full-res crop** of that region instead of (or in addition to) the
+  downscaled whole page. And extend the cross-check: if a `<canvas>`/`<img>` of meaningful
+  size exists where the claim expects a visual, suppress the vision "missing" finding (or
+  drop it to `warn`/advisory) just like the text path.
+
+### Net for v0.3.1
+Infra drawbacks (#1–#8) are resolved — full-page works, localhost works, output is
+clean/scriptable, and text-conformance is trustworthy. The remaining gaps are both about
+the **vision-LLM path on visual (non-OCR) elements**: don't freeze rAF before a canvas
+draws, and crop-to-bbox for visual intent claims rather than judging a downscaled whole
+page.
+
+---
+
+## v0.3.2 — residuals #9 & #10 addressed
+
+- **#9 (rAF-built canvas blanked by freeze):** freeze is now *canvas-aware*. CSS animations
+  are still paused, but when a `<canvas>` is present rAF is left running and the renderer
+  waits `canvas_settle_ms` (default 1500) so the scene builds before capture
+  ("settle-then-freeze"). Verified: a canvas that draws only inside rAF now renders non-blank
+  with `--freeze` on (default). Heavy scenes can still bump `--settle-ms`.
+- **#10 (visual element read as "missing" on a downscaled full page):** the renderer now
+  reports present `canvas`/`svg`/`img`/`video` (`RenderResult.visual_tags`), and a vision
+  "missing" claim naming a visual (canvas, chart, 3D scene, image…) is **suppressed when the
+  matching element is actually in the DOM** — extending the text cross-check to the visual
+  path. (A focused full-res crop-to-bbox per visual claim remains a future quality
+  enhancement; the false-fail itself is now prevented.)

agentvision-0.3.2/docs/unified-architecture.svg

@@ -0,0 +1,143 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="1600" height="1060" viewBox="0 0 1600 1060" font-family="Inter, Segoe UI, Helvetica, Arial, sans-serif">
+  <defs>
+    <linearGradient id="bg" x1="0" y1="0" x2="1" y2="1">
+      <stop offset="0" stop-color="#0c0f17"/>
+      <stop offset="1" stop-color="#0f1320"/>
+    </linearGradient>
+    <linearGradient id="eyeGlow" x1="0" y1="0" x2="1" y2="1">
+      <stop offset="0" stop-color="#5ad1e6"/>
+      <stop offset="1" stop-color="#3a9fd6"/>
+    </linearGradient>
+    <linearGradient id="brainGlow" x1="0" y1="0" x2="1" y2="1">
+      <stop offset="0" stop-color="#8b7cff"/>
+      <stop offset="1" stop-color="#b06cff"/>
+    </linearGradient>
+    <linearGradient id="nerve" x1="0" y1="0" x2="1" y2="0">
+      <stop offset="0" stop-color="#5ad1e6"/>
+      <stop offset="1" stop-color="#8b7cff"/>
+    </linearGradient>
+    <marker id="ah" markerWidth="12" markerHeight="12" refX="8" refY="4" orient="auto">
+      <path d="M0,0 L8,4 L0,8 Z" fill="#8b7cff"/>
+    </marker>
+    <marker id="ahb" markerWidth="14" markerHeight="14" refX="9" refY="5" orient="auto">
+      <path d="M0,0 L10,5 L0,10 Z" fill="#46d39a"/>
+    </marker>
+  </defs>
+
+  <rect width="1600" height="1060" fill="url(#bg)"/>
+
+  <!-- Title -->
+  <text x="800" y="66" text-anchor="middle" font-size="48" font-weight="800" fill="#eef2fb">Eyes &amp; Brain</text>
+  <text x="800" y="104" text-anchor="middle" font-size="22" fill="#9aa4b8">AgentVision <tspan fill="#5ad1e6">sees</tspan>  ·  Verel <tspan fill="#8b7cff">decides</tspan>  ·  only <tspan fill="#46d39a">verified</tspan> work ships</text>
+
+  <!-- ================= EYES PANEL ================= -->
+  <rect x="56" y="150" width="610" height="700" rx="22" fill="#12161f" stroke="#274050" stroke-width="1.5"/>
+  <text x="92" y="200" font-size="15" letter-spacing="3" fill="#5ad1e6" font-weight="700">THE EYES · PERCEPTION</text>
+  <text x="92" y="246" font-size="38" font-weight="800" fill="#eef2fb">AgentVision</text>
+  <!-- eye glyph -->
+  <g transform="translate(566,210)">
+    <ellipse cx="0" cy="0" rx="44" ry="26" fill="none" stroke="url(#eyeGlow)" stroke-width="4"/>
+    <circle cx="0" cy="0" r="13" fill="#5ad1e6"/>
+    <circle cx="5" cy="-5" r="4" fill="#0c0f17"/>
+  </g>
+  <text x="92" y="282" font-size="17" fill="#9aa4b8">It writes a UI, chart, SVG, PDF or image — and actually looks at the result.</text>
+
+  <!-- pipeline -->
+  <g font-size="16" font-weight="700" text-anchor="middle">
+    <rect x="92" y="312" width="118" height="46" rx="10" fill="#14202a" stroke="#2f5163"/>
+    <text x="151" y="341" fill="#cfe9f2">RENDER</text>
+    <rect x="232" y="312" width="118" height="46" rx="10" fill="#14202a" stroke="#2f5163"/>
+    <text x="291" y="341" fill="#cfe9f2">PERCEIVE</text>
+    <rect x="372" y="312" width="118" height="46" rx="10" fill="#14202a" stroke="#2f5163"/>
+    <text x="431" y="341" fill="#cfe9f2">REPORT</text>
+    <rect x="512" y="312" width="118" height="46" rx="10" fill="#14202a" stroke="#2f5163"/>
+    <text x="571" y="341" fill="#cfe9f2">DIFF</text>
+    <text x="221" y="341" fill="#5ad1e6">›</text>
+    <text x="361" y="341" fill="#5ad1e6">›</text>
+    <text x="501" y="341" fill="#5ad1e6">›</text>
+  </g>
+
+  <!-- trust card -->
+  <rect x="92" y="386" width="538" height="146" rx="14" fill="#101a22" stroke="#23323d"/>
+  <text x="112" y="416" font-size="18" font-weight="700" fill="#eef2fb">Grounded truth, not vibes</text>
+  <text x="112" y="446" font-size="16" fill="#bfe7f0">▪ DOM geometry + computed-style WCAG contrast — <tspan fill="#46d39a">precise</tspan></text>
+  <text x="112" y="474" font-size="16" fill="#bfe7f0">▪ OCR word boxes + spelling/typo check — <tspan fill="#46d39a">precise</tspan></text>
+  <text x="112" y="502" font-size="16" fill="#bfe7f0">▪ Vision-LLM critique (Claude/GPT/Gemini/Ollama) — <tspan fill="#f5b14c">advisory</tspan></text>
+
+  <!-- conformance card -->
+  <rect x="92" y="548" width="538" height="150" rx="14" fill="#101a22" stroke="#23323d"/>
+  <text x="112" y="578" font-size="18" font-weight="700" fill="#eef2fb">Matches intent, not just defects</text>
+  <text x="112" y="608" font-size="16" fill="#bfe7f0">brief / checklist / reference  →  a render is graded per requirement</text>
+  <text x="112" y="636" font-size="16" fill="#bfe7f0">must · should · nice  →  <tspan fill="#46d39a">satisfied</tspan> · <tspan fill="#ff7a7a">violated</tspan> · <tspan fill="#f5b14c">uncertain</tspan></text>
+  <text x="112" y="664" font-size="16" fill="#9aa4b8">PASS means “it’s what we set out to build,” not “nothing’s broken.”</text>
+
+  <!-- handoff output pill -->
+  <rect x="92" y="716" width="538" height="104" rx="14" fill="#0e2230" stroke="#2f6f7e"/>
+  <text x="112" y="752" font-size="18" font-weight="800" fill="#5ad1e6">The handoff →</text>
+  <text x="112" y="782" font-size="15" fill="#cfe9f2">verdict · next_action · matches_intent · todo[] · open_questions[]</text>
+  <text x="112" y="806" font-size="14" fill="#9aa4b8">a clean afferent signal — the eyes report, they don’t decide</text>
+
+  <!-- ================= OPTIC NERVE ================= -->
+  <text x="800" y="430" text-anchor="middle" font-size="14" letter-spacing="2" fill="#9aa4b8">THE OPTIC NERVE</text>
+  <line x1="676" y1="470" x2="924" y2="470" stroke="url(#nerve)" stroke-width="7" marker-end="url(#ah)"/>
+  <g text-anchor="middle" font-size="14" font-weight="700">
+    <rect x="690" y="492" width="220" height="34" rx="17" fill="#13202b" stroke="#2f5163"/>
+    <text x="800" y="514" fill="#bfe7f0">perception → decision</text>
+  </g>
+  <text x="800" y="560" text-anchor="middle" font-size="14" fill="#9aa4b8">grounded observations + a verdict,</text>
+  <text x="800" y="580" text-anchor="middle" font-size="14" fill="#9aa4b8">mapped onto one bus the brain trusts</text>
+
+  <!-- ================= BRAIN PANEL ================= -->
+  <rect x="934" y="150" width="610" height="700" rx="22" fill="#15131f" stroke="#3a2f5e" stroke-width="1.5"/>
+  <text x="970" y="200" font-size="15" letter-spacing="3" fill="#b6a6ff" font-weight="700">THE BRAIN · DECISION + MEMORY</text>
+  <text x="970" y="246" font-size="38" font-weight="800" fill="#eef2fb">Verel</text>
+  <!-- brain glyph -->
+  <g transform="translate(1444,212)">
+    <circle cx="0" cy="0" r="30" fill="none" stroke="url(#brainGlow)" stroke-width="4"/>
+    <circle cx="-12" cy="-6" r="4" fill="#8b7cff"/><circle cx="8" cy="-12" r="4" fill="#8b7cff"/>
+    <circle cx="12" cy="8" r="4" fill="#8b7cff"/><circle cx="-8" cy="10" r="4" fill="#8b7cff"/>
+    <path d="M-12,-6 L8,-12 M8,-12 L12,8 M12,8 L-8,10 M-8,10 L-12,-6" stroke="#6a5acf" stroke-width="1.5" fill="none"/>
+  </g>
+  <text x="970" y="282" font-size="17" fill="#9aa4b8">Nothing is “done” until a grader returns a verdict.</text>
+
+  <!-- five organs -->
+  <g font-size="16">
+    <rect x="970" y="312" width="258" height="76" rx="12" fill="#1a1730" stroke="#3a2f5e"/>
+    <text x="986" y="340" font-weight="700" fill="#eef2fb">⚖  Verdict bus</text>
+    <text x="986" y="366" font-size="13" fill="#b6a6ff">vision + tests + lint + types → one pass/warn/fail</text>
+
+    <rect x="1250" y="312" width="258" height="76" rx="12" fill="#1a1730" stroke="#3a2f5e"/>
+    <text x="1266" y="340" font-weight="700" fill="#eef2fb">Memory that compounds</text>
+    <text x="1266" y="366" font-size="13" fill="#b6a6ff">only verified facts/skills graduate</text>
+
+    <rect x="970" y="402" width="258" height="76" rx="12" fill="#16202a" stroke="#2f5163"/>
+    <text x="986" y="430" font-weight="700" fill="#cfe9f2">Eyes = AgentVision</text>
+    <text x="986" y="456" font-size="13" fill="#7fd3e6">perception as one sense among many</text>
+
+    <rect x="1250" y="402" width="258" height="76" rx="12" fill="#1a1730" stroke="#3a2f5e"/>
+    <text x="1266" y="430" font-weight="700" fill="#eef2fb">Fleet</text>
+    <text x="1266" y="456" font-size="13" fill="#b6a6ff">agents managing agents, gated by the bus</text>
+
+    <rect x="970" y="492" width="538" height="76" rx="12" fill="#1a1730" stroke="#3a2f5e"/>
+    <text x="986" y="520" font-weight="700" fill="#eef2fb">Tool-smith</text>
+    <text x="986" y="546" font-size="13" fill="#b6a6ff">agents build, test &amp; register their own tools — sandboxed, admitted only on a passing eval</text>
+  </g>
+
+  <!-- compounds pill -->
+  <rect x="970" y="588" width="538" height="110" rx="14" fill="#10231b" stroke="#2f6f55"/>
+  <text x="990" y="624" font-size="18" font-weight="800" fill="#46d39a">Only verified work compounds</text>
+  <text x="990" y="654" font-size="15" fill="#c8e9da">the brain decides with grader attestation — a hollow check can’t mint green</text>
+  <text x="990" y="680" font-size="14" fill="#9aa4b8">records “did it match what we set out to build?” across every iteration</text>
+
+  <!-- brain decision output -->
+  <rect x="970" y="716" width="538" height="104" rx="14" fill="#1a1730" stroke="#4a3f7e"/>
+  <text x="990" y="752" font-size="18" font-weight="800" fill="#b6a6ff">The decision →</text>
+  <text x="990" y="782" font-size="15" fill="#d7ccff">done · revise (act on the to-do) · review</text>
+  <text x="990" y="806" font-size="14" fill="#9aa4b8">a full brain runs its own gate; a simpler brain can use the handoff verbatim</text>
+
+  <!-- ================= EFFERENT LOOP ================= -->
+  <path d="M1230,866 C 1100,948 520,948 372,894" fill="none" stroke="#46d39a" stroke-width="4" stroke-dasharray="2 0" marker-end="url(#ahb)"/>
+  <rect x="612" y="906" width="376" height="44" rx="22" fill="#10231b" stroke="#2f6f55"/>
+  <text x="800" y="934" text-anchor="middle" font-size="16" font-weight="700" fill="#c8e9da">decide → act (edit code / refine prompt) → look again</text>
+  <text x="800" y="1000" text-anchor="middle" font-size="18" fill="#9aa4b8">render → perceive → gate → fix → re-render → <tspan fill="#46d39a" font-weight="700">pass</tspan> — the loop where “done” becomes a verdict, not an opinion</text>
+</svg>

{agentvision-0.3.0 → agentvision-0.3.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "agentvision"
-version = "0.3.0"
+version = "0.3.2"
 description = "Eyes for AI Agents — a machine-graded visual feedback loop coding agents consume to self-correct before claiming done."
 readme = "README.md"
 requires-python = ">=3.11"

{agentvision-0.3.0 → agentvision-0.3.2}/src/agentvision/__init__.py

@@ -42,7 +42,7 @@ from .models import (
     Viewport,
 )
 
-__version__ = "0.3.0"
+__version__ = "0.3.2"
 
 __all__ = [
     "__version__",