@oh-story/dsh 0.1.12 → 0.1.13

package/lib/index.js

@@ -164,7 +164,7 @@ var DSH_DRAMA_BRIDGE = [
   "DSH owns the workspace, model, preset, permissions, Session Log, tools, approvals, cancellation, resume, and Agent UI.",
   "The \u77ED\u5267 tab is the creator workspace. Never start dashboard_server.py, another web server, Dashboard, Agent runtime, session transport, or model configuration.",
   "Drama Skills uses a creator-first contract: each episode keeps only the requested documents, up to five creator-facing sources at \u5267\u96C6/<EP>/\u5267\u672C.md, \u89C6\u89C9\u8BBE\u5B9A.md, \u5206\u955C.md, \u56FE\u7247\u63D0\u793A\u8BCD.md, and \u89C6\u9891\u63D0\u793A\u8BCD.md. Never precreate empty documents, backfill nominal stages, or start work the creator did not request. Persisted reviews use creator-readable Markdown under \u5BA1\u67E5/; an oral review writes nothing.",
-  "\u5267\u96C6/<EP>/\u526A\u8F91\u5355.md is the v0.7 assembly record, not a sixth creative truth: the five creator documents stay the creative truth. It is written after footage exists and records what reaches the cut \u2014 which footage is kept or left unused and why, each cut's in/out timing, its subtitles, \u753B\u9762\u6587\u5B57 overlays and \u97F3\u6548 cues, post-processing such as shot-match correction or an external mix, and the delivery spec including optional \u9897\u7C92 grain and \u63A5\u955C\u5339\u914D. It never changes a line, a shot's job, or a declared duration; changing any of those means returning to the document that owns it.",
+  "\u5267\u96C6/<EP>/\u526A\u8F91\u5355.md is the v0.7 assembly record, not a sixth creative truth: the five creator documents stay the creative truth. It is written after readable media exists and records what reaches the cut \u2014 which footage or accepted static keyframes are kept or left unused and why, each cut's in/out timing, its subtitles, \u753B\u9762\u6587\u5B57 overlays, \u97F3\u6548 cues, \u914D\u97F3 and \u73AF\u5883\u58F0, post-processing such as shot-match correction or an external mix, and the delivery spec including optional \u9897\u7C92 grain and \u63A5\u955C\u5339\u914D. It never changes a line, a shot's job, or a declared duration; changing any of those means returning to the document that owns it.",
   "For a v0.6 project, never create a parallel JSON/JSONL lifecycle truth, indexes, fingerprints, coverage tables, or QA records merely because legacy maintenance scripts and templates remain bundled.",
   "Never upgrade a v0.5 structured project in place or mix both contracts in one project root. Keep legacy production/audit artifacts read-only and pinned to v0.5; migrate only into a new creator-first root with manual per-episode creator confirmation.",
   "Use only tools visible in the current DSH preset and preserve the upstream project ownership, freshness, review, and explicit production-confirmation contracts.",
@@ -174,14 +174,14 @@ var DSH_DRAMA_BRIDGE = [
 ].join("\n");
 var DSH_DRAMA_OVERRIDES = {
   "short-drama": "A dashboard request means focus or use the native \u77ED\u5267 tab in this DSH Session. Do not run the bundled standalone Dashboard script. New projects follow only the v0.6 creator-first contract and create only documents required by the current request.",
-  "short-drama-produce": (skillRoot) => `Use an upstream adapter only after the creator explicitly confirms the exact current job. Its source must be the current creator-first Markdown and its output belongs under \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/. DSH permissions and approval UI remain authoritative. When oh_story_production is visible, register a previewed image or video job with track_job after the confirmation is valid and before run \u2014 never at prepare time \u2014 passing the same job ID, target, modality and count the creator just approved. track_job's jobKind is image, video or composition (the assembled cut), so from this Skill it registers image and video jobs only: speech (tts) and music jobs pass through the same prepare \u2192 explicit creator confirmation \u2192 run gate but are never registered with track_job or relabelled as another kind, and their audio results stay under \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/. In this DSH integration audio is never bound as a video job's reference: upstream's creator-first path has no audio binding (\u8F93\u5165\u53C2\u8003\u56FE takes png/jpg/webp images only) and audio reaches the cut only at the edit stage, as \u97F3\u6548 lines that render mixes in or as an external mix. DeepSeek generates no media: every image, video, speech, or music result comes from a provider adapter. The bundled adapters are ${dramaAdapterSummary()}; they are registered for this DSH host in ${dramaAdapterConfigPath(skillRoot).path} \u2014 pass that file as --adapter-config unless the creator names another config (${DRAMA_ADAPTER_CONFIG_ENV}). Credentials are read only from the DSH host process environment. Before preparing a paid job, name the adapter that will run it; if its variables are missing, tell the creator exactly which ones to export before starting DSH instead of failing inside run. Never read, print, or write credential values. A submitted video job is already billed, so an interrupted run is recovered, not resubmitted: when audit reports orphaned_provider_job, recover it with collect, passing --job-id \u2014 the drama job the finding names, not its provider_job_id, which collect reads from the run handle itself. collect spends nothing and does not need the confirmation gate \u2014 the gate exists to stop unintended spending \u2014 but re-running the same job does, and would charge the creator twice.`,
+  "short-drama-produce": (skillRoot) => `Use an upstream adapter only after the creator explicitly confirms the exact current job. Its source must be the current creator-first Markdown and its output belongs under \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/. DSH permissions and approval UI remain authoritative. When oh_story_production is visible, register a previewed image or video job with track_job after the confirmation is valid and before run \u2014 never at prepare time \u2014 passing the same job ID, target, modality and count the creator just approved. track_job's jobKind is image, video or composition (the assembled cut), so from this Skill it registers image and video jobs only: speech (tts) and music jobs pass through the same prepare \u2192 explicit creator confirmation \u2192 run gate but are never registered with track_job or relabelled as another kind, and their audio results stay under \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/. Optional voice references belong on the MOTION entry's separate \u53C2\u8003\u97F3\u9891 line, never in \u8F93\u5165\u53C2\u8003\u56FE or the workbench image picker. Compile them as reference_audio bindings, preserve character exactly from \u89D2\u8272, and match the project-local audio path to that character's single \u58F0\u97F3\u53C2\u8003 in \u89C6\u89C9\u8BBE\u5B9A.md. \u4E0D\u5F97\u63A7\u5236 lists \u53F0\u8BCD\u3001\u8BED\u6C14\u3001\u60C5\u7EEA individually. The selected adapter profile must explicitly declare reference_audio in reference_roles; otherwise reject before paid submission, never drop or substitute the binding. The generated generic adapter profiles intentionally omit reference_audio because their model IDs are environment-selected; use a creator-owned OH_STORY_DRAMA_ADAPTER_CONFIG profile that declares audio support only for a known reference-audio-capable model. DeepSeek generates no media: every image, video, speech, or music result comes from a provider adapter. The bundled adapters are ${dramaAdapterSummary()}; they are registered for this DSH host in ${dramaAdapterConfigPath(skillRoot).path} \u2014 pass that file as --adapter-config unless the creator names another config (${DRAMA_ADAPTER_CONFIG_ENV}). Credentials are read only from the DSH host process environment. Before preparing a paid job, name the adapter that will run it; if its variables are missing, tell the creator exactly which ones to export before starting DSH instead of failing inside run. Never read, print, or write credential values. A submitted video job is already billed, so an interrupted run is recovered, not resubmitted: when audit reports orphaned_provider_job, recover it with collect, passing --job-id \u2014 the drama job the finding names, not its provider_job_id, which collect reads from the run handle itself. collect spends nothing and does not need the confirmation gate \u2014 the gate exists to stop unintended spending \u2014 but re-running the same job does, and would charge the creator twice.`,
   "short-drama-edit": [
     "Assembly runs inside this DSH Session. Write \u5267\u96C6/<EP>/\u526A\u8F91\u5355.md first, then run edit_tool.py check and render through the current DSH execution world; the rendered cut belongs under \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/\u6210\u7247/ and the \u77ED\u5267 tab shows it there. ffmpeg and ffprobe come from the DSH host \u2014 when one is missing, say so instead of approximating or reporting an untested measurement as a pass.",
     // edit_tool.py 0.7.1 `_unaccounted_shots`: `- 未采用镜头：` is read only outside CUT blocks, split on `；`, and each item must fullmatch `MOTION-…（理由：<non-empty>）`.
-    "check blocks until every `## MOTION-*` heading in \u89C6\u9891\u63D0\u793A\u8BCD.md either backs a CUT's \u6765\u6E90 or is listed in the cut list's opening block, before the first `## CUT-` heading, on one line: `- \u672A\u91C7\u7528\u955C\u5934\uFF1AMOTION-\u2026\uFF08\u7406\u7531\uFF1A\u2026\uFF09\uFF1BMOTION-\u2026\uFF08\u7406\u7531\uFF1A\u2026\uFF09`. Join items with a full-width `\uFF1B`; each reason is non-empty, contains no `\uFF1B`, and names \u6587\u4EF6\u7F3A\u5931, \u8D28\u91CF\u4E0D\u53EF\u7528 or \u53D9\u4E8B\u53D6\u820D.",
-    "Every cut must share one width, height and frame rate: render hard-joins the encoded segments and never scales or retimes them. When check reports a mismatch, normalise the odd clips to the delivery spec with ffmpeg in the DSH execution world, under approval, preserving the composition. Write them under \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/\u6210\u7247/\u89C4\u683C\u7EDF\u4E00/, an edit-owned intermediate, never beside the produce-stage originals and never over produced footage, and name them without the original's job-id token (the MOTION ID will do). Record the crop or pad in that CUT's block (not in its \u753B\u9762 line, which accepts only \u4EAE\u5EA6/\u9971\u548C/\u8272\u6E29), point \u6765\u6E90 at the new file, update \u5165\u70B9/\u51FA\u70B9 if the conversion moved them, and rerun check.",
+    "Video cuts use existing MOTION IDs; static-frame cuts use SHOT IDs from \u5206\u955C.md or IMG IDs from \u56FE\u7247\u63D0\u793A\u8BCD.md and accepted png/jpg/jpeg/webp files. Every unused MOTION and, when static cuts exist, every storyboard SHOT must be accounted for before the first `## CUT-` on one line: `- \u672A\u91C7\u7528\u955C\u5934\uFF1AMOTION-\u2026\uFF08\u7406\u7531\uFF1A\u2026\uFF09\uFF1BSHOT-\u2026\uFF08\u7406\u7531\uFF1A\u2026\uFF09`. Join items with a full-width `\uFF1B`; reasons name \u6587\u4EF6\u7F3A\u5931, \u8D28\u91CF\u4E0D\u53EF\u7528 or \u53D9\u4E8B\u53D6\u820D. A static CUT starts at 0.00, lasts at least 0.50 seconds, requires the delivery width/height/frame rate, and may use \u8FD0\u955C. The native production sequence accepts video and accepted keyframe images, keeps storyboard order and source identity, and dispatches /short-drama-edit in Chat; never relabel images as video or generate paid video unnecessarily.",
+    "Every video cut must share one width, height and frame rate: render hard-joins the encoded video segments and never scales or retimes them. Static cuts are rendered to the delivery specification with the declared crop/motion. When check reports a mismatch, normalise the odd clips to the delivery spec with ffmpeg in the DSH execution world, under approval, preserving the composition. Write them under \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/\u6210\u7247/\u89C4\u683C\u7EDF\u4E00/, an edit-owned intermediate, never beside the produce-stage originals and never over produced footage, and name them without the original's job-id token (the MOTION ID will do). Record the crop or pad in that CUT's block (not in its \u753B\u9762 line, which accepts only \u4EAE\u5EA6/\u9971\u548C/\u8272\u6E29), point \u6765\u6E90 at the new file, update \u5165\u70B9/\u51FA\u70B9 if the conversion moved them, and rerun check.",
     "The default burned-subtitle route needs an ffmpeg built with libass; when the host's ffmpeg lacks it, say so instead of silently dropping subtitles. Remotion renders every frame through a headless browser from a Node workspace outside the project, ~/.cache/short-drama-edit/remotion, which edit_tool.py rewrites on each run: the optional subtitle route uses it, and any `\u753B\u9762\u6587\u5B57` line in \u526A\u8F91\u5355.md requires it whatever the subtitle route. Install it (npm install in that workspace; an install from before Drama Skills 0.8 must run npm install again) only after the creator approves, through DSH approval. Without that approval write no \u753B\u9762\u6587\u5B57 lines and tell the creator the blank on-screen panels stay empty. Never treat Remotion as a second Agent runtime, dashboard, or session transport.",
-    "The \u58F0\u97F3 line in \u526A\u8F91\u5355.md is a record, not an instruction the built-in render executes. render trims, auto-matches adjacent cuts within one scene (a CUT's \u753B\u9762 correction replaces it, `- \u753B\u9762\uFF1A\u4E0D\u6821` skips it, and `- \u63A5\u955C\u5339\u914D\uFF1A\u65E0` in the delivery spec turns it off), hard-joins, mixes in the `\u97F3\u6548` lines, burns subtitles and \u753B\u9762\u6587\u5B57, applies optional \u9897\u7C92, and normalises loudness. A mix beyond the \u97F3\u6548 lines, crossfade, music bed, transition or format conversion is an external ffmpeg step, run under approval, that writes a temporary file, ends as render does with whole-film two-pass loudnorm to the declared \u4EA4\u4ED8\u54CD\u5EA6 (I=<\u4EA4\u4ED8\u54CD\u5EA6>:TP=-1.5:LRA=11, the second pass fed the first pass's measurements with linear=true, audio re-encoded as AAC 192k at 48 kHz), and only then replaces \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/\u6210\u7247/\u6210\u7247.mp4; record its command in \u526A\u8F91\u5355.md and repeat it after any re-render, which overwrites that file. Run edit_tool.py verify last, on that delivered file, and report every \u672A\u6D4B item as untested, never as passed.",
+    "The \u58F0\u97F3 line in \u526A\u8F91\u5355.md is a record, not an instruction the built-in render executes. render trims, auto-matches adjacent cuts within one scene (a CUT's \u753B\u9762 correction replaces it, `- \u753B\u9762\uFF1A\u4E0D\u6821` skips it, and `- \u63A5\u955C\u5339\u914D\uFF1A\u65E0` in the delivery spec turns it off), hard-joins, mixes in the `\u97F3\u6548` lines, cross-cut `\u914D\u97F3` and looping `\u73AF\u5883\u58F0`, burns subtitles and \u753B\u9762\u6587\u5B57, applies optional \u9897\u7C92, and normalises loudness. A mix beyond the declared \u97F3\u6548/\u914D\u97F3/\u73AF\u5883\u58F0 lines, crossfade, transition or format conversion is an external ffmpeg step, run under approval, that writes a temporary file, ends as render does with whole-film two-pass loudnorm to the declared \u4EA4\u4ED8\u54CD\u5EA6 (I=<\u4EA4\u4ED8\u54CD\u5EA6>:TP=-1.5:LRA=11, the second pass fed the first pass's measurements with linear=true, audio re-encoded as AAC 192k at 48 kHz), and only then replaces \u5267\u96C6/<EP>/\u5236\u4F5C\u6210\u679C/\u6210\u7247/\u6210\u7247.mp4; record its command in \u526A\u8F91\u5355.md and repeat it after any re-render, which overwrites that file. Run edit_tool.py verify last, on that delivered file, and report every \u672A\u6D4B item as untested, never as passed.",
     "This stage never generates footage and never edits \u5267\u672C.md, \u5206\u955C.md, or \u89C6\u9891\u63D0\u793A\u8BCD.md \u2014 route those back to their owning Skill."
   ].join(" "),
   "short-drama-assets": "Keep one stable `- ID\uFF1AVISUAL-*` line on every \u4EBA\u7269/\u9020\u578B/\u5730\u70B9/\u9053\u5177/\u72B6\u6001 entry in \u89C6\u89C9\u8BBE\u5B9A.md, and never change an existing ID when editing its heading or text. The \u751F\u4EA7 view identifies canvas nodes by that ID; entries without one still render, but the workbench reports their node identity as unstable."
@@ -203,7 +203,8 @@ var DSH_VIDEO_BRIDGE = [
   "DSH owns the workspace, model, preset, permissions, Session Log, tools, approvals, cancellation, resume, Todo, Chat, and the \u89C6\u9891 preview Studio.",
   "Never start a second Agent runtime, dashboard, session transport, web server, polling loop, or model configuration.",
   "A dashboard request means the native \u89C6\u9891 tab in this DSH Session: dashboard_server.py and its dashboard assets are not bundled, so never try to run them. To inspect a resource library, use the read-only scripts/library.py check, list, or show.",
-  "Keep the complete upstream six-Skill pipeline. Put each project under video-recaps/<project>/, copy or import source media under sources/, and use work/ as the upstream work_dir so the Studio can discover authoritative manifests and outputs. A project binding passed with --project belongs at video-recaps/<project>/recap_project.json.",
+  "Keep the complete upstream six-Skill production pipeline plus the on-demand video-reference Skill; video-reference extracts reusable production references only when requested and never enters the default production path. Put each project under video-recaps/<project>/, copy or import source media under sources/, and use work/ as the upstream work_dir so the Studio can discover authoritative manifests and outputs. A project binding passed with --project belongs at video-recaps/<project>/recap_project.json.",
+  "full-mode validation is lint-only: never truncate, drop, merge or reorder approved narration to satisfy a budget; report over_budget and request a revised script. final_qc.json uses schema 2 findings (code/message/blocking/evidence/next_action); do not read removed golden_eval.json, mimo_qc.json or preflight_qc.json. Do not pass removed --mimo-qc, --preserve-approved-text or --clip-padding flags, MIMO_QC* variables, CLIP_PADDING or SOURCE_VIDEO. Except dub, delivery uses tagged H.264 8-bit 4:2:0 with true peak <= -1 dBTP; delivered_true_peak_over_target blocks delivery. If libass is unavailable, deliver external .srt rather than dropping subtitles or aborting preflight.",
   "The Video Studio is a preview and artifact surface, not a nonlinear editor. Do not invent a second project-state format, timeline truth, or render queue; recap_run_manifest.json, recap_phase.json, timeline.json, assembly_manifest.json, and the upstream artifacts remain authoritative.",
   "MIMO_API_KEY, FISH_API_KEY, and voice credentials stay in the host environment. Never write secrets into project files, tool arguments shown to the browser, or chat output.",
   "Use only DSH-visible tools and approvals. Run Python and ffmpeg through the current DSH execution world, preserve cancellation, and do not install or upgrade system dependencies without explicit user approval.",
@@ -1126,6 +1127,9 @@ var VISIBLE_TEXT_FILES = /* @__PURE__ */ new Set([
   "recap_phase.json",
   "agent_narration_brief.md",
   "recap_story_plan.json",
+  "production_reference.json",
+  "reference_measurements.json",
+  "reference_breakdown.json",
   "visual_audio_board.json",
   "clip_plan.json",
   "clip_plan_validated.json",
@@ -1149,6 +1153,9 @@ var ARTIFACT_META = {
   "recap_phase.json": { label: "\u9636\u6BB5\u8D26\u672C", kind: "manifest" },
   "agent_narration_brief.md": { label: "\u89E3\u8BF4 Brief", kind: "plan" },
   "recap_story_plan.json": { label: "\u6545\u4E8B\u65B9\u6848", kind: "plan" },
+  "production_reference.json": { label: "\u5236\u4F5C\u53C2\u8003", kind: "plan" },
+  "reference_measurements.json": { label: "\u53C2\u8003\u6D4B\u91CF", kind: "quality" },
+  "reference_breakdown.json": { label: "\u53C2\u8003\u62C6\u89E3", kind: "plan" },
   "visual_audio_board.json": { label: "\u97F3\u753B\u65B9\u6848", kind: "plan" },
   "clip_plan.json": { label: "\u526A\u8F91\u8BA1\u5212", kind: "plan" },
   "clip_plan_validated.json": { label: "\u5DF2\u6821\u9A8C\u526A\u8F91\u8BA1\u5212", kind: "plan" },
@@ -1359,7 +1366,9 @@ var MEDIA_TYPES = /* @__PURE__ */ new Map([
   [".mkv", "video/x-matroska"],
   [".mp3", "audio/mpeg"],
   [".wav", "audio/wav"],
-  [".m4a", "audio/mp4"]
+  [".m4a", "audio/mp4"],
+  [".aac", "audio/aac"],
+  [".flac", "audio/flac"]
 ]);
 var MEDIA_MAX_BYTES = 256 * 1024 * 1024;
 var FILE_LIMIT = 1e3;