@event4u/agent-config 14.9.0 → 14.11.0

package/src/scripts/ai-video/adapters/sora.sh

@@ -204,4 +204,14 @@ aiv_cmd_fetch() {
   jq -n --arg p "${out}" '{video_path:$p, audio_embedded:true}'
 }
 
+# Capability: `capability --model <id>` answers from
+# lib/model-capabilities/sora.json (schema v2 — durations, modeled cost,
+# and the start_frame/end_frame continuity answers). A bare `capability`
+# keeps the adapter-level audio flag byte-identical to before the manifest.
+sub="${1:-}"
+if [ "${sub}" = "capability" ]; then
+  shift
+  aiv_manifest_capability "${ADAPTER_ID}" "native" "$@"
+  exit 0
+fi
 aiv_dispatch "${ADAPTER_ID}" "native" "$@"

package/src/scripts/ai-video/lib/calibrate-cost.sh

@@ -0,0 +1,145 @@
+#!/usr/bin/env bash
+# calibrate-cost.sh — compare the MODELED cost against what the provider
+# actually CHARGED, before committing a batch.
+#
+# Why this exists, and why it is smaller than it first looked. The read-back
+# already existed: `resume-scan.sh` reads `<project>/scenes/<id>/cost.json`
+# `.charged_usd` and sums it as `spent_usd`, and the from-song command writes
+# that file after every live scene. What did NOT exist was any path from that
+# number back into `manifest.cost_per_second_usd` — money was spent, recorded,
+# and never read into the model it contradicted. This is that path's front end.
+#
+# It runs at the operator-pick moment (the checkpoint the operator was already
+# stopping at) rather than adding a gate of its own, so the calibration line
+# appears where a human is already looking.
+#
+# Usage:
+#   calibrate-cost.sh <project-dir> <scene-id> --adapter <id> --model <id>
+#                     [--scenes <n>] [--ledger <path>] [--no-append]
+#
+# Output (stdout, one line):
+#   calibration: modeled $<m>/s · charged $<c>/s · <delta> · extrapolated batch $<x> (modeled $<y>)
+#
+# `charged: null` is NOT `0`. An unpriceable or not-yet-charged scene prints
+# `charged: null`, extrapolates nothing, and never re-confirms. A modeled figure
+# is only ever corrected by a MEASURED one — treating null as zero would make
+# every future estimate quietly cheaper, which is the one direction a costing
+# error must never drift.
+#
+# Exit codes:
+#   0   calibration printed; within tolerance (or nothing charged yet)
+#   2   usage
+#   3   jq missing
+#   7   project/scene inputs unreadable
+#   13  charged exceeds modeled by more than the tolerance — RE-CONFIRM
+#       before the batch. Not a failure: a deliberate halt for a human.
+set -euo pipefail
+
+_cc_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../../.." && pwd)"
+# shellcheck source=/dev/null
+. "${_cc_root}/src/scripts/media/lib/adapter-common.sh"
+
+# The re-confirmation threshold. 25 % is the roadmap's stated number and is a
+# STATED DEFAULT, not a measured optimum: it is the point past which "the model
+# is roughly right" stops being a fair description. Revisit-if a provider's
+# real-world variance is measured and sits above it, which would make the halt
+# fire on noise rather than on drift.
+CC_TOLERANCE_PCT="${AIV_CALIBRATE_TOLERANCE_PCT:-25}"
+
+project=""; scene=""; adapter=""; model=""; scenes=""; ledger=""; append=1
+[ "$#" -ge 2 ] || { printf 'calibrate-cost: usage: %s <project-dir> <scene-id> --adapter <id> --model <id> [--scenes <n>]\n' "$0" >&2; exit 2; }
+project="$1"; scene="$2"; shift 2
+while [ "$#" -gt 0 ]; do
+  case "$1" in
+    --adapter) adapter="${2:-}"; shift 2 ;;
+    --model)   model="${2:-}"; shift 2 ;;
+    --scenes)  scenes="${2:-}"; shift 2 ;;
+    --ledger)  ledger="${2:-}"; shift 2 ;;
+    --no-append) append=0; shift ;;
+    *) printf "calibrate-cost: unknown arg '%s'\n" "$1" >&2; exit 2 ;;
+  esac
+done
+[ -n "${adapter}" ] && [ -n "${model}" ] \
+  || { printf 'calibrate-cost: --adapter and --model are required\n' >&2; exit 2; }
+command -v jq >/dev/null 2>&1 || { printf 'calibrate-cost: jq required\n' >&2; exit 3; }
+
+scene_dir="${project}/scenes/${scene}"
+[ -d "${scene_dir}" ] || { printf 'calibrate-cost: scene dir not readable: %s\n' "${scene_dir}" >&2; exit 7; }
+
+# Duration: the scene's own render input, never a default. A guessed duration
+# would turn a per-second comparison into a fabricated one.
+duration=""
+if [ -r "${scene_dir}/prompt.json" ]; then
+  duration="$(jq -r '.duration // empty' "${scene_dir}/prompt.json" 2>/dev/null || true)"
+fi
+
+manifest="$(aiv_manifest_path "${adapter}")"
+modeled_per_s="null"
+if [ -r "${manifest}" ]; then
+  modeled_per_s="$(jq -r --arg m "${model}" '.models[$m].cost_per_second_usd // "null" | tostring' "${manifest}" 2>/dev/null || printf 'null')"
+fi
+
+# The existing spend record — the same file resume-scan.sh already sums.
+charged_total="null"
+if [ -r "${scene_dir}/cost.json" ]; then
+  charged_total="$(jq -r '.charged_usd // "null" | tostring' "${scene_dir}/cost.json" 2>/dev/null || printf 'null')"
+fi
+
+line="$(MODELED="${modeled_per_s}" CHARGED="${charged_total}" DUR="${duration}" \
+  SCENES="${scenes}" TOL="${CC_TOLERANCE_PCT}" node -e '
+  const num = (v) => (v === "" || v === "null" || v === undefined ? null : Number(v));
+  const modeled = num(process.env.MODELED);
+  const charged = num(process.env.CHARGED);
+  const dur = num(process.env.DUR);
+  const scenes = num(process.env.SCENES);
+  const tol = Number(process.env.TOL);
+  const fmt = (v) => (v === null || Number.isNaN(v) ? "null" : `$${v.toFixed(4)}`);
+
+  // Per-second charged needs a duration. Without one the charge is real but not
+  // comparable, and saying so beats dividing by a guess.
+  const chargedPerS = charged !== null && dur !== null && dur > 0 ? charged / dur : null;
+
+  let delta = "delta: n/a (nothing charged yet)";
+  let reconfirm = false;
+  if (modeled !== null && chargedPerS !== null) {
+    if (modeled === 0) {
+      delta = chargedPerS === 0 ? "+0.0 %" : "modeled 0 — any charge is a model defect";
+      reconfirm = chargedPerS > 0;
+    } else {
+      const pct = ((chargedPerS - modeled) / modeled) * 100;
+      delta = `${pct >= 0 ? "+" : ""}${pct.toFixed(1)} %`;
+      reconfirm = pct > tol;
+    }
+  }
+
+  let extra = "";
+  if (scenes !== null && dur !== null) {
+    const m = modeled !== null ? modeled * dur * scenes : null;
+    const c = chargedPerS !== null ? chargedPerS * dur * scenes : null;
+    extra = ` · extrapolated batch ${fmt(c)} (modeled ${fmt(m)})`;
+  }
+
+  process.stdout.write(
+    `calibration: modeled ${fmt(modeled)}/s · charged ${chargedPerS === null ? "null" : fmt(chargedPerS)}/s · ${delta}${extra}\n` +
+    `RECONFIRM=${reconfirm ? "1" : "0"}\n`,
+  );
+')"
+
+printf '%s\n' "${line%$'\n'RECONFIRM=*}"
+reconfirm="${line##*RECONFIRM=}"
+
+if [ "${append}" -eq 1 ]; then
+  bash "${_cc_root}/src/scripts/ai-video/smoke-trace.sh" cost-ledger append \
+    --adapter "${adapter}" --model "${model}" \
+    --modeled "${modeled_per_s}" \
+    --charged "$(printf '%s' "${line}" | sed -nE 's/.*charged (null|\$([0-9.]+))\/s.*/\2/p' | sed 's/^$/null/')" \
+    --note "calibration probe at scene ${scene}" \
+    ${ledger:+--ledger "${ledger}"} >/dev/null
+fi
+
+if [ "${reconfirm}" = "1" ]; then
+  printf 'calibrate-cost: charged exceeds modeled by more than %s %% — RE-CONFIRM before the batch (the cap in --max-spend-usd is not a substitute: it bounds total spend, not a wrong per-second model).\n' \
+    "${CC_TOLERANCE_PCT}" >&2
+  exit 13
+fi
+exit 0

package/src/scripts/ai-video/lib/model-capabilities/README.md

@@ -1,8 +1,12 @@
 # Model-capabilities manifests
 
-One JSON manifest per **multiplexer adapter** (`fal.json`,
-`replicate.json`). A multiplexer reaches many video models through one
-uniform provider API; the manifest is the single source of truth for
+One JSON manifest per adapter whose capabilities differ per model —
+the multiplexers (`fal.json`, `replicate.json`), the local engine
+(`comfyui.json`), the lip-sync pair (`syncso.json`, `musetalk.json`) and
+the direct video adapters (`higgsfield.json`, `kling.json`,
+`gemini-veo.json`, `sora.json`). A multiplexer reaches many video models
+through one uniform provider API; a direct adapter reaches one provider's
+own model set. Either way the manifest is the single source of truth for
 what each reachable model can do, consumed by:
 
 - the adapter's `capability --model <id>` subcommand (audio routing for
@@ -10,13 +14,17 @@ what each reachable model can do, consumed by:
 - the music-video pipeline's segment planner (variable-length plans must
   stay inside `min_duration..max_duration` or the plan is unbuildable),
 - the batch cost gate (`cost_per_second_usd` × planned seconds feeds the
-  preview estimate and the `--max-spend-usd` kill-switch).
+  preview estimate and the `--max-spend-usd` kill-switch),
+- the continuity selection rule (`start_frame` / `end_frame` decide
+  whether a model can open on a handed-off frame or close on a given one
+  — `motion-choreographer` Step 0 reads them before it proposes a
+  continuous take).
 
-## Schema (v1)
+## Schema (v2)
 
 ```json
 {
-  "schema": 1,
+  "schema": 2,
   "adapter": "<adapter-id>",
   "models": {
     "<model_id>": {
@@ -27,12 +35,24 @@ what each reachable model can do, consumed by:
       "aspect": ["16:9", "9:16"],
       "cost_per_second_usd": 0.28,
       "verified": false,
-      "input_notes": "per-model input schema differences, one line"
+      "input_notes": "per-model input schema differences, one line",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": { "probed_at": null, "psnr_frame0": null }
     }
   }
 }
 ```
 
+**What v2 adds.** Three keys — `start_frame`, `end_frame`, `frame_lock` —
+and one rule that is not a shape: **`null` means *unknown*, and unknown is
+never treated as `true`.** Nothing else changed, so a v1 file becomes a v2
+file once the three keys are present on every entry. The keys are a
+catch-up rather than a new axis: the `ref_images[0] → image /
+start_image` mapping row below, and `higgsfield.sh`'s hard requirement for
+`ref_images[0]`, already asserted start-frame behaviour that v1 had no
+field to express.
+
 - `min_duration` / `max_duration` — clip length bounds in **seconds**.
   Equal values mean the model renders fixed-length clips only.
 - `audio_sync` — `true` when the model returns a muxed MP4 with
@@ -49,6 +69,26 @@ what each reachable model can do, consumed by:
   MUST surface the flag rather than trust the numbers silently.
 - `input_notes` — the per-model input schema difference in one line
   (see below).
+- `start_frame` — `true` when the model accepts a first frame to open on
+  (`ref_images[0]`), `false` when it cannot, `null` when nobody has
+  probed it. A continuous take may only be chained through models that
+  answer `true`.
+- `end_frame` — the same three values for a *closing* frame (`end_image`
+  in the stdin contract). An adapter handed `end_image` for a model that
+  does not answer `true` **refuses by name** — it never drops the image;
+  see [`adapter-contract.md`](../../../media/lib/adapter-contract.md)
+  § `end_image`.
+- `frame_lock` — the evidence behind those two answers: `probed_at` (ISO
+  date of the live probe) and `psnr_frame0` (measured dB of the rendered
+  frame 0 against the submitted still). Both `null` until a probe writes
+  them; a dB value is a **measurement**, never an estimate, so it stays
+  `null` rather than being filled in from documentation. A `true` on
+  either frame key without a `probed_at` is a contract violation.
+- `end_frame: true` beside `start_frame: false` is **incoherent** — a
+  model that cannot open on a supplied frame cannot close on one either.
+  The shared reader in
+  [`adapter-common.sh`](../../../media/lib/adapter-common.sh) refuses such
+  an entry instead of handing it to a planner.
 
 ## Per-model input schema differences
 
@@ -70,6 +110,12 @@ provider docs) until the first live smoke trace; mismatches surface as
 provider-side 4xx errors with the body echoed to stderr, never as
 silent drops.
 
+`ref_images[0]` is the **start** frame, and the row above maps only its
+*field name*. Whether a given model honours it is the separate `start_frame`
+answer in the schema above; the closing counterpart — `end_image` in the
+stdin contract — is `end_frame`. A mapping row is not evidence that a model
+accepts the field: only `frame_lock.probed_at` is.
+
 ## Verification workflow
 
 1. Maintainer wires a live key and captures a smoke trace under

package/src/scripts/ai-video/lib/model-capabilities/comfyui.json

@@ -1,5 +1,5 @@
 {
-  "schema": 1,
+  "schema": 2,
   "adapter": "comfyui",
   "comment": "Local ComfyUI capability manifest (ADR-060). cost_per_second_usd is a KNOWN 0.0 — a local render has no provider charge (distinct from an omitted/unknown estimate; the cost gate still treats MISSING as unknown, never 0). Every entry is documented-best-effort until a maintainer captures a smoke trace against a real containerized ComfyUI — verified:false entries MUST NOT be trusted for planning-critical decisions without surfacing the flag. The template field names the package-shipped workflow under lib/comfyui-templates/ — operator-supplied workflow JSONs are rejected (ADR-060 §3).",
   "models": {
@@ -13,7 +13,13 @@
       "verified": false,
       "license": "Apache-2.0",
       "template": "comfyui-templates/wan22-ti2v-5b.json",
-      "input_notes": "ASSUMED ~121-frame ceiling at 24 fps (~5 s) for the 5B TI2V checkpoint; optional start image (ref_images[0]) uploaded via /upload/image; silent video — the orchestrator muxes audio at stitch"
+      "input_notes": "ASSUMED ~121-frame ceiling at 24 fps (~5 s) for the 5B TI2V checkpoint; optional start image (ref_images[0]) uploaded via /upload/image; silent video — the orchestrator muxes audio at stitch",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "ltx-2": {
       "label": "LTX-2 (Lightricks) — local audio-synced video",
@@ -25,7 +31,13 @@
       "verified": false,
       "license": "LTX community license (verify before commercial use)",
       "template": "comfyui-templates/ltx-2.json",
-      "input_notes": "ASSUMED duration ceiling and audio-sync behaviour from vendor docs; the LTX-2 graph + weights must be present in the container image; verify license terms before any commercial render"
+      "input_notes": "ASSUMED duration ceiling and audio-sync behaviour from vendor docs; the LTX-2 graph + weights must be present in the container image; verify license terms before any commercial render",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     }
   }
 }

package/src/scripts/ai-video/lib/model-capabilities/fal.json

@@ -1,5 +1,5 @@
 {
-  "schema": 1,
+  "schema": 2,
   "adapter": "fal",
   "comment": "Per-model capability manifest for the fal multiplexer. Every entry is documented-best-effort until a maintainer captures a real-API smoke trace for that model \u2014 verified:false entries MUST NOT be trusted for planning-critical decisions without surfacing the flag. Durations in seconds; cost_per_second_usd is a modeled estimate, never a quote.",
   "models": {
@@ -15,7 +15,13 @@
       ],
       "cost_per_second_usd": 0.28,
       "verified": false,
-      "input_notes": "duration is an enum string '5'|'10'; aspect_ratio; negative_prompt supported; cfg_scale optional"
+      "input_notes": "duration is an enum string '5'|'10'; aspect_ratio; negative_prompt supported; cfg_scale optional",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "fal-ai/wan/v2.2-a14b/text-to-video": {
       "label": "Wan 2.2 A14B (text-to-video)",
@@ -29,7 +35,13 @@
       "cost_per_second_usd": 0.08,
       "verified": true,
       "input_notes": "resolution enum ('480p'|'580p'|'720p') instead of free aspect; num_frames-driven duration; seed supported",
-      "smoke_trace": "fal-fal-ai_wan_v2.2-a14b_text-to-video-live-2026-06-10T12-38-11Z"
+      "smoke_trace": "fal-fal-ai_wan_v2.2-a14b_text-to-video-live-2026-06-10T12-38-11Z",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "fal-ai/ltx-2/text-to-video": {
       "label": "LTX-2 (text-to-video, synchronized audio)",
@@ -44,7 +56,13 @@
       "cost_per_second_usd": 0.16,
       "verified": true,
       "input_notes": "generates synchronized audio natively (audio_sync); resolution up to 4K; duration in seconds",
-      "smoke_trace": "fal-fal-ai_ltx-2_text-to-video-live-2026-06-10T12-36-49Z"
+      "smoke_trace": "fal-fal-ai_ltx-2_text-to-video-live-2026-06-10T12-36-49Z",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "fal-ai/veo3": {
       "label": "Google Veo 3 (text-to-video, native audio)",
@@ -57,7 +75,13 @@
       ],
       "cost_per_second_usd": 0.5,
       "verified": false,
-      "input_notes": "fixed 8s clips; dialogue/ambient rendered from the prompt text; aspect_ratio; generate_audio flag"
+      "input_notes": "fixed 8s clips; dialogue/ambient rendered from the prompt text; aspect_ratio; generate_audio flag",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "fal-ai/hunyuan-video": {
       "label": "Hunyuan Video (text-to-video)",
@@ -71,7 +95,13 @@
       "cost_per_second_usd": 0.08,
       "verified": true,
       "input_notes": "num_frames-driven duration; resolution enum; no negative_prompt",
-      "smoke_trace": "fal-fal-ai_hunyuan-video-live-2026-06-10T12-39-13Z"
+      "smoke_trace": "fal-fal-ai_hunyuan-video-live-2026-06-10T12-39-13Z",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     }
   }
 }

package/src/scripts/ai-video/lib/model-capabilities/gemini-veo.json

@@ -0,0 +1,23 @@
+{
+  "schema": 2,
+  "adapter": "gemini-veo",
+  "comment": "Per-model capability manifest for the direct gemini-veo adapter. Every value is READ OUT OF THE TREE and cited in input_notes. Only veo-3.0-generate-001 is an entry: it is the adapter default (adapters/gemini-veo.sh:129) and the XML <default-model> (agents/templates/.ai-video.xml.example:72). veo-3.0-fast* and veo-2.0* appear only as rate PREFIX PATTERNS (adapters/gemini-veo.sh:93-95) with no concrete model id in the tree, so no entries are invented for them. verified stays false: the adapter header records 10/10 live renders of this exact model (adapters/gemini-veo.sh:9-15) but names no trace id reachable from the tree, and the local traces are gitignored.",
+  "models": {
+    "veo-3.0-generate-001": {
+      "label": "Google Veo 3 (text-to-video, native audio)",
+      "min_duration": 4,
+      "max_duration": 8,
+      "audio_sync": true,
+      "aspect": null,
+      "cost_per_second_usd": 0.4,
+      "verified": false,
+      "input_notes": "model id from adapters/gemini-veo.sh:129 and agents/templates/.ai-video.xml.example:72; durationSeconds enum {4,6,8} is ASSUMED in the adapter header (:34-37) and is the clamp the adapter applies (:132-137), upper bound matching the template tuning <max-duration>8</max-duration> (xml:77); audio_sync true from the declared audio=native (:4), the dispatch flag (tail) and <audio-native>true</audio-native> (xml:78); aspect is passed through as aspectRatio with no enum check (:151) so the accepted set is unknown — the template's tuning default is 16:9 (xml:75); cost_per_second_usd from _veo_model_rate_usd_per_s veo-3.0* = 0.40 (:94), corroborated by the header's ~$1.60 per 4 s render at $0.40/s (:10)",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
+    }
+  }
+}

package/src/scripts/ai-video/lib/model-capabilities/higgsfield.json

@@ -0,0 +1,23 @@
+{
+  "schema": 2,
+  "adapter": "higgsfield",
+  "comment": "Per-model capability manifest for the direct higgsfield adapter. Every value is READ OUT OF THE TREE and cited in input_notes — no field is estimated from provider documentation this repo cannot show. Unknown is null, never a guess: higgsfield.sh sends no duration, aspect or cost parameter (adapters/higgsfield.sh:186-187), so those bounds are unknown rather than defaulted. The provider's named motion presets (mix, burst, dvd, cinematic, talk — adapters/higgsfield.sh:37-38,58) are NOT models and stay on `capability --preset`. higgsfield-soul (the XML <default-model>) is an image model the adapter explicitly refuses for image2video and rewrites to dop-turbo (adapters/higgsfield.sh:113-120), so it is not an entry here. verified stays false: the adapter header records a live 1/1 image2video round-trip (adapters/higgsfield.sh:15-20) but names no per-model trace id reachable from the tree, and `verified` is a per-model claim.",
+  "models": {
+    "dop-turbo": {
+      "label": "Higgsfield DoP Turbo (image-to-video)",
+      "min_duration": null,
+      "max_duration": null,
+      "audio_sync": false,
+      "aspect": null,
+      "cost_per_second_usd": null,
+      "verified": false,
+      "input_notes": "model id from the image2video fallback at adapters/higgsfield.sh:116 (accepted pattern *dop*|*turbo* at :115), route POST /v1/image2video/dop (:70,:189); requires ref_images[0] and dies exit 7 without it (:133-135); request body carries model+prompt+input_images only — no duration, aspect or seed (:186-187), so duration/aspect bounds are unknown; audio_sync false from the video-only fetch shape (:259); the only cost figure in the tree is the dry-run fixture's per-clip cost_estimate 0.1 (media/lib/fixtures/higgsfield/result.json), which is not a per-second rate; observed live output 5.4 s at 1168x768 (:16) and the template's operator tuning cap is <max-duration>5</max-duration> (agents/templates/.ai-video.xml.example:63) — an operator preference, not a provider bound",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
+    }
+  }
+}

package/src/scripts/ai-video/lib/model-capabilities/kling.json

@@ -0,0 +1,23 @@
+{
+  "schema": 2,
+  "adapter": "kling",
+  "comment": "Per-model capability manifest for the direct kling adapter. Every value is READ OUT OF THE TREE and cited in input_notes. Only kling-v2-master is an entry: it is the adapter default (adapters/kling.sh:153) and the XML <default-model> (agents/templates/.ai-video.xml.example:87). kling-v1* appears only as a rate PREFIX PATTERN (adapters/kling.sh:105) with no concrete model id anywhere in the tree, so no v1 entry is invented. verified stays false: the adapter header records a live 1/1 round-trip (adapters/kling.sh:9-13) but names no per-model trace id reachable from the tree.",
+  "models": {
+    "kling-v2-master": {
+      "label": "Kling v2 Master (text-to-video)",
+      "min_duration": 5,
+      "max_duration": 10,
+      "audio_sync": false,
+      "aspect": null,
+      "cost_per_second_usd": 0.28,
+      "verified": false,
+      "input_notes": "model id from adapters/kling.sh:153 and agents/templates/.ai-video.xml.example:87; duration is the enum {'5','10'} the adapter clamps to (:151) with KLING_MAX_DURATION default 10 (:52) matching the template tuning <max-duration>10</max-duration> (xml:93); audio_sync false from the adapter's declared audio=none (:4), the video-only fetch (:245), the dispatch flag (:248) and <audio-native>false</audio-native> (xml:94); aspect is passed straight through as aspect_ratio with no enum check (:167) so the accepted set is unknown — the template's tuning default is 16:9 (xml:91); cost_per_second_usd is DERIVED, not quoted: _kling_model_rate_usd gives kling-v2* = 1.40 USD per 5 s (:102-104) and the header records ~$1.40 for a 5 s std render (:12), so 1.40 / 5 = 0.28",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
+    }
+  }
+}

package/src/scripts/ai-video/lib/model-capabilities/musetalk.json

@@ -1,5 +1,5 @@
 {
-  "schema": 1,
+  "schema": 2,
   "adapter": "musetalk",
   "comment": "Local (MIT) lip-sync capability manifest — the offline alternative to the hosted syncso adapter. cost_per_second_usd is a KNOWN 0.0 (local inference, ADR-060 cost semantics). The lipsync_budget block mirrors the syncso sparse discipline: local renders are free but the QUALITY argument for sparseness is unchanged (singing lip-sync degrades fast off frontal close-ups) — the orchestrator enforces it BEFORE invoking the adapter. Every entry is documented-best-effort until a maintainer captures a trace against a real MuseTalk install (verified:false must be surfaced, never silently trusted).",
   "lipsync_budget": {
@@ -18,7 +18,13 @@
       "cost_per_second_usd": 0.0,
       "verified": false,
       "license": "MIT",
-      "input_notes": "local video_path + audio_path (local-source rule); ASSUMED CLI surface --video/--audio/--output via MUSETALK_CMD; GPU strongly recommended — CPU inference is minutes-per-second; quality degrades off frontal close-ups like every lip-sync model"
+      "input_notes": "local video_path + audio_path (local-source rule); ASSUMED CLI surface --video/--audio/--output via MUSETALK_CMD; GPU strongly recommended — CPU inference is minutes-per-second; quality degrades off frontal close-ups like every lip-sync model",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     }
   }
 }

package/src/scripts/ai-video/lib/model-capabilities/replicate.json

@@ -1,5 +1,5 @@
 {
-  "schema": 1,
+  "schema": 2,
   "adapter": "replicate",
   "comment": "Per-model capability manifest for the replicate multiplexer. Every entry is documented-best-effort until a maintainer captures a real-API smoke trace for that model \u2014 verified:false entries MUST NOT be trusted for planning-critical decisions without surfacing the flag. Durations in seconds; cost_per_second_usd is a modeled estimate, never a quote.",
   "models": {
@@ -15,7 +15,13 @@
       "cost_per_second_usd": 0.5,
       "verified": false,
       "input_notes": "fixed 8s clips; dialogue/ambient rendered from the prompt text; resolution flag",
-      "official": true
+      "official": true,
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "kwaivgi/kling-v2.1-master": {
       "label": "Kling v2.1 Master (text/image-to-video)",
@@ -30,7 +36,13 @@
       "cost_per_second_usd": 0.28,
       "verified": false,
       "input_notes": "duration enum 5|10; start_image for image-to-video; negative_prompt supported",
-      "official": true
+      "official": true,
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "wan-video/wan-2.2-t2v-fast": {
       "label": "Wan 2.2 (text-to-video, fast)",
@@ -45,7 +57,13 @@
       "verified": true,
       "input_notes": "num_frames-driven duration; resolution enum; seed supported",
       "official": true,
-      "smoke_trace": "replicate-wan-video_wan-2.2-t2v-fast-live-2026-06-10T10-43-52Z"
+      "smoke_trace": "replicate-wan-video_wan-2.2-t2v-fast-live-2026-06-10T10-43-52Z",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "lightricks/ltx-video": {
       "label": "LTX Video (text-to-video)",
@@ -61,7 +79,13 @@
       "verified": true,
       "input_notes": "length in frames at fixed fps; cfg + negative_prompt supported",
       "official": false,
-      "smoke_trace": "replicate-lightricks_ltx-video-live-2026-06-10T10-46-47Z"
+      "smoke_trace": "replicate-lightricks_ltx-video-live-2026-06-10T10-46-47Z",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     },
     "tencent/hunyuan-video": {
       "label": "Hunyuan Video (text-to-video)",
@@ -76,7 +100,13 @@
       "verified": true,
       "input_notes": "video_length in frames; embedded_guidance_scale; no negative_prompt",
       "official": false,
-      "smoke_trace": "replicate-tencent_hunyuan-video-live-2026-06-10T10-54-32Z"
+      "smoke_trace": "replicate-tencent_hunyuan-video-live-2026-06-10T10-54-32Z",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     }
   }
 }

package/src/scripts/ai-video/lib/model-capabilities/sora.json

@@ -0,0 +1,45 @@
+{
+  "schema": 2,
+  "adapter": "sora",
+  "comment": "Per-model capability manifest for the direct sora adapter. Every value is READ OUT OF THE TREE and cited in input_notes. Both entries are concrete model ids the adapter names (adapters/sora.sh:20,73-74); sora-2 is the adapter default (:111) and the XML <default-model> (agents/templates/.ai-video.xml.example:101). verified stays false: the adapter header records a live 1/1 round-trip on sora-2 (adapters/sora.sh:9-13) but names no trace id reachable from the tree, and sora-2-pro has never been rendered here at all.",
+  "models": {
+    "sora-2": {
+      "label": "OpenAI Sora 2 (text-to-video, native audio)",
+      "min_duration": 4,
+      "max_duration": 12,
+      "audio_sync": true,
+      "aspect": [
+        "16:9",
+        "9:16"
+      ],
+      "cost_per_second_usd": 0.1,
+      "verified": false,
+      "input_notes": "model id from adapters/sora.sh:20,74 and agents/templates/.ai-video.xml.example:101; seconds enum {'4','8','12'} verified empirically via validation-error probes with no spend (:12,:15-16,:21) and is the clamp the adapter applies (:114-119); audio_sync true from the declared audio=native (:4) and the fixture's audio_embedded:true (media/lib/fixtures/sora/result.json); aspect is the adapter's own two-branch size mapping — 9:16 to 720x1280, everything else to 1280x720 (:121-122) — out of the verified size enum {720x1280,1280x720,1024x1792,1792x1024} (:12,:22); cost_per_second_usd from _sora_model_rate_usd_per_s sora-2* = 0.10 (:74), corroborated by the header's ~$0.40 for a 4 s render (:11)",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
+    },
+    "sora-2-pro": {
+      "label": "OpenAI Sora 2 Pro (text-to-video, native audio)",
+      "min_duration": 4,
+      "max_duration": 12,
+      "audio_sync": true,
+      "aspect": [
+        "16:9",
+        "9:16"
+      ],
+      "cost_per_second_usd": 0.3,
+      "verified": false,
+      "input_notes": "model id from adapters/sora.sh:20,73; seconds and size enums are the same adapter-level surface as sora-2 (:114-122) — the adapter branches on the model only for the rate; audio_sync true from the declared audio=native (:4); cost_per_second_usd from _sora_model_rate_usd_per_s sora-2-pro* = 0.30 (:73); NEVER rendered from this repo — no live round-trip is recorded for this id anywhere in the tree",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
+    }
+  }
+}

package/src/scripts/ai-video/lib/model-capabilities/syncso.json

@@ -1,5 +1,5 @@
 {
-  "schema": 1,
+  "schema": 2,
   "adapter": "syncso",
   "comment": "Lip-sync post-process capability manifest. Every entry is documented-best-effort until a maintainer captures a real-API smoke trace — verified:false entries MUST NOT be trusted for planning-critical decisions without surfacing the flag. The lipsync_budget block is the MACHINE-READABLE sparse-lip-sync discipline (council 2026-06-06: singing lip-sync is genuinely hard — keep it sparse): the orchestrator enforces it BEFORE any submit; tests lock its shape.",
   "lipsync_budget": {
@@ -17,7 +17,13 @@
       "aspect": ["16:9", "9:16", "1:1"],
       "cost_per_second_usd": 0.05,
       "verified": false,
-      "input_notes": "input is [{type:video,url},{type:audio,url}] — both https-hosted; ASSUMED max clip length 60s; singing-capable per vendor docs, quality degrades off frontal close-ups"
+      "input_notes": "input is [{type:video,url},{type:audio,url}] — both https-hosted; ASSUMED max clip length 60s; singing-capable per vendor docs, quality degrades off frontal close-ups",
+      "start_frame": null,
+      "end_frame": null,
+      "frame_lock": {
+        "probed_at": null,
+        "psnr_frame0": null
+      }
     }
   }
 }