- Ask once for a new film project's collaboration pace when it is unknown: discuss, execute directly, or agree creative checkpoints. Explicit self-direction reuses its existing scope. Professional groups remain independent from real subagent dispatch.
- Use native host questions only for creative decisions and the transition to a concrete production scope. Keep ordinary discussion conversational, deduplicate pending questions, reject stale or negated starts, and continue authorized production and repairs without per-stage approvals.
- Preserve current interaction state across tools and installed handoff replay without rewriting the original media request. Temporary analysis detours retain the project's choice.
- Expose root foundation candidates and role-specific editable preparation inputs through read-only CLI helpers. Validate character execution fields and their current authored facts before provider work; preserve content words while tolerating cosmetic punctuation. Coverage authoring can inspect current allowed values and an editable input shape before writing a large batch.
- Route reviewed local repairs through the same public preparer, with one bound source image and explicit edit arguments. Reject stale create packets posing as repairs and unused best-base flags. Keep unapproved aspect-defective candidates usable as repair inputs while final delivery remains strict.
- Align the entrypoint with batch generation and dependency-based review. Show useful completed content and saved image previews; keep real visual QA, five-view identity, parent evidence and full deliverable coverage.
- Routing, question lifecycle and production handoff tests cover declared behavior. Read-only input helpers do not generate media or approve image quality; timing improvements require observed forward runs.
- Private project transcripts, timing logs and generated test media are kept outside the release package. Synthetic fixtures are not visual approval.
- Preserve all host-path rejection branches after release privacy sanitization. The sanitizer could otherwise rewrite the literal Linux/private path alternatives inside the prompt-surface regular expression.
- Exercise the actual sanitized adapter against synthetic macOS, Linux and private paths, with an ordinary-prose control. Prompt methods and authored examples are unchanged from 0.11.0.
- Reworked video prompt authoring around scene intent, reference roles, stable design and evolving event state across action, performance, environment, product and graphic work. Preserve authored format and craft without a mandatory twelve-part cinematic prefix.
- Removed unrequested photoreal/skin, pause, palette ratio, composition formula, resolution, frame-rate and shutter defaults. Explicit quality overrides remain supported.
- Carry shot transitions, local continuity and native sound bridges into compiled prompts; preserve unit scope and locally rebased timing without exposing internal identifiers.
- Added source-backed prompt research and anonymized failure analysis covering reference-board leakage, held contact poses, state resets and conflicting visual authority. Private media and conversation evidence are not bundled.
- Compiler/fixture tests and installed-method exercises verify instruction and text behavior. The inspected negative video informs the diagnosis; no claim is made that a new video-model candidate has proven quality improvement.
- Bring Jingzao photography advice into existing Studio story and shot development before camera choices are locked. Connect composition and environment to story, blocking and shot information; preserve story ownership, Fast scope, existing budgets and separate image execution.
- Support reference-free and reference-backed composition work, with role-scoped advisory media and reference reads. Keep ordinary asset compilation on its original reference pack; test natural story-only entry, negative requests, missing/oversized providers and unchanged execution gates.
- Add an optional product-CG directing method and six source-image-free style capsules. Reuse existing film/production-design routing and hash-bound Jingzao style handoffs; preserve target identity, materials, scope and authorization.
- Connect product facts to material events, shot information, transitions and timing. Include powder-compact and precision-drive PromptIR examples with different product logic, and 11 official source cases with observation/inference boundaries.
- Preserve CG and explicitly excluded media in the video-quality prefix, without forcing live action or performer language into product-only scenes. Respect later cancellation and explicit restart of CG requests.
- Seal the installed Jingzao compiler's optional reference_profile module when present while retaining four-file provider compatibility and strict hash/path checks.
- Test 12 target scenarios with 14 native still outputs, retaining two initial powder failures and bounded independent visual conclusions. Actual raster size, cropping and material-scale limitations are recorded; no moving-video acceptance is claimed.
- Fix a missing JSON import in the combined dependency installer that caused a reporting failure after installation. Add complete combined-entrypoint tests for successful dependencies and preserved/conflicting local providers, including output and exit status.
- Add opt-in dependency installation, a licensed hash-locked humanizer-zh snapshot, and an official Jingzao latest-stable-release updater. Preserve unmanaged installations, local edits and newer or unorderable versions; verify files and protect rollback without running upstream scripts.
- Add Chinese and English product overviews, introduction artwork and static landing pages. Keep the detailed technical and verified-release guide, add dependency setup and provider availability documentation, and refresh search/social metadata and the machine-readable index.
- Validate release-versus-tag selection, annotated tag resolution, edited and malformed receipts, concurrent edits, version protection, archive extraction and actual isolated public installation. Marketing illustrations are explicitly illustrative, not runtime acceptance evidence.
- Add public asset preparation and batch review commands over the existing provider, design, evidence and execution contracts. Reuse authored design passes and selected style capsules without task-local adapter scripts; retain actionable provider review results and immutable input bindings.
- Reuse a valid substantive review instead of a duplicate executor checklist. Explicit executor reviews permit only generated candidates as internal draft references; locks, headless-safe authority and final adoption keep their stronger requirements. Bind authenticated downstream reuse to the source generation task.
- Record batch failures as concrete failed check observations that feed the existing repair path. Keep passed members independent, and scope new review hashes to their actual asset, use and dependencies rather than unrelated inventory administration.
- Preserve declared adult body proportions, garment cut/openings and portrait framing. Replace the invented approved A-pose default with natural garment-display stance while retaining explicit pose locks and all requested five views. Keep tested complete prompts as authored baselines without automatic translation or second style-layer expansion.
- Prefix every compiled video submission with twelve ordered cinematic quality sections, before shot content. Preserve explicit scene lighting, optical facts, continuity and audio routing; remove redundant old global sections. Optional quality overrides do not change provider parameters or claim actual 8K/24fps delivery.
- Document input/output image-block diagnostics and uncertainty boundaries. Real tests include a successful authored five-view adult costume, retained blocked variants, and a native product run through preparation, registration and one review. These are bounded cases, not a universal speed or generation-success claim.
- Separate local image-repair prompts from initial design and layout requirements. Keep original asset truth in the bound plan while compiling only the source-preserving edit intent and observed corrections; allow either valid side-profile order on new human masters.
- Produce independent assets/content before consolidated batch review. Remove plan-wide candidate-check blocking and the character-only serial restriction, while retaining actual parent checks and final acceptance. Candidate self-check no longer duplicates the character detector gate.
- Stop automatic recursive edits of a previously repaired image. Allow an explicitly bound best-quality same-truth base, require its known defects to be addressed, and preserve candidate lineage without promoting it to an approved source.
- Resolve headed-character detector disagreements within the existing visual review using the actual probe receipt and visible observations. Preserve blocked measurements, opacity checks, source hashes, formal review and host authorization; headless safeguards are unchanged.
- Fix first-output registration for standalone asset plans without falsely claiming completed media. Clarify first-pass reference/design preparation, conditional negative prompts, clean-image label policy and dependency-scoped rework; update only project-level maintenance guidance.
- Keep requested five-view coverage. Support explicit nonhuman character identity/state without human-anatomy probes, and allow source-bound authored 2D layout references without pretending they are measured spatial exports.
- Forward hash-bound Jingzao style capsules through asset and narrative-frame replay. Accept its two documented nonblocking review scopes without stripping style detail, while keeping exact compilation checks and detected contamination blocks.
- Keep handoff validators as hash-verified executable tools instead of mandatory model-reading material. Preserve full contract/schema reads and existing context budgets; validator code growth no longer blocks ordinary frame selection.
- These changes remove demonstrated workflow failures; structural checks do not establish universal first-pass image quality or pixel-perfect editing.
- Route natural storyboard-and-asset requests and independent character, state, scene and mechanical work into the existing design stages. Preserve mixed reuse/new-asset scope, negations, quoted story facts and bounded text/portrait work without granting generation from source text.
- Compile canonical asset-role requirements into the real Jingzao spec before execution. Keep the five-view human master, chosen artistic medium and valid create/ratio choices; replace stale DIR role/mode clauses when preparing a derivative.
- Prepare one current, validated native argument set and bind output registration to its exact plan hash. Record real candidates, require role-specific executor observations and the character structural probe before continuing, and retain failed candidates for focused repair. Self-check remains separate from independent QA, host adoption and user approval.
- Integrate the audited external
mr-li-seedance-252.0 body and four production references within the existing context budget. Preserve DIR source, camera, asset, state and authorization ownership; keep provider tutorials, historical rules and aggregate runtime out of the selected compilation context. - Ground the next unmade source cursor in the actual script, reject repeated consumed dialogue locators, and validate applicable voice bindings by speaker and generation unit. Keep the Seedance 2.0 surface independent of the 2.5 method contract.
- Carry forward source-continuity checks, medium-aware quality, planning/reference-role separation and requested-output reconciliation through existing DIR stages; do not replace these with metadata claims or a second controller/ledger.
- Supply the real Noto CJK font dependency in Linux CI so the annotated-storyboard regression runs with readable Chinese glyphs; preserve the missing-font runtime guard and all workflow tests.
- Publish the completed 0.9.1 workflow changes together with this CI repair. The already-pushed 0.9.1 tag is retained without a release archive.
- Connect natural full-preproduction requests and developed shot cards to the existing action, camera, panel-coverage and Jingzao frame-compilation stages. Selectors now discover configured host Skills by default and consume each applicable stage without adding a second router.
- Rehearse complex action with model-generated line drawings, semantic-color arrows, notes and legend. Preserve the original annotated board and explicit cell mapping; inspect it before supported storyboard-reference or required color-frame use. Reuse coverage IDs and Jingzao's rough styleboard path without counting sketches as production images or demanding finished assets before drawing.
- Declare supplied foundation PNGs through the existing inventory and read back only available assets. Preserve valid small planning crops through verified source pixels; combine corrected native panels by proportional document layout with replayable provenance and a new page review.
- Default to native files, inline images, copyable text and a short index. Routine ledgers and browsing do not require a website, dashboard, workbench or local server; focused spatial/camera interaction remains available when needed.
- Carry existing image authorization from film development into the exact asset-execution context; keep video authorization, asset packets and formal review requirements separate.
- Preserve action phases, motivated viewpoint changes, reference roles, impact intensity and environment consequences through shot development. Clean inputs follow the selected generation-unit strategy instead of a fixed per-shot quota.
- Require the actual Jingzao narrative-film profile before frame compilation and bind that profile in production handoffs. Review source states against all compiled fields so generic template defaults, wrong actor actions and stale light/camera descriptions cannot masquerade as reviewed cinematography.
- Require whole-film and sequence pre-video asset checks to validate panel coverage and the existing production handoff for required action images. Reject reuse of one image across different action phases while keeping visual-quality and user-acceptance claims unverified.
- Deduplicate shared scene state and projected objects in camera/phase previews, retaining the full interactive selection within the existing payload limits.
- Route natural-language blocking, camera and reverse-shot discussions into the current host visualization contract, with a complete text fallback and no mandatory stage approval for viewing.
- Derive interactive camera views and text-free layout references from the same revision-bound scene/shot state; reject stale sources, mismatched projections and unbound adoption.
- Bind layout controls to the existing Jingzao and video-prompt paths, preserve reference sovereignty, and return explicit upload order with copyable prompts.
- Keep bounded edits small and continue authorized full preproduction through the requested asset, storyboard, continuity and submission materials. Local rendering, real model generation, visual review and released installation remain separate evidence levels.
-
Fixes global installed-runtime acceptance to verify the actual target against an independent source checkout, adds source/installed entry checks to CI and release gates, and replaces the legacy operator prompt with v2 instructions.
-
Adds explicit creative/director collaboration intent with native subagent versus user-task boundaries; keeps primary routing authoritative through perspective selection and ADCO contexts.
-
Fixes negated side-effect routing, resolved-concept gates, one-shot scope, and single-page client-story expansion; adds isolated behavior regressions.
-
Restores bounded system and packaged-provider discovery without re-exposing internal Skills or reviving disabled/duplicate providers.
-
Removes mandatory stage confirmations and routine humanization cascades from active modules; preserves genre, source, asset, review and execution contracts.
-
Keeps controller IDs and hashes out of direct asset prompts while preserving their packet bindings, and scopes historical worktree checks to explicit removal evidence.
-
Preserves exact candidate Skill identity in isolated Codex text evaluations, requires an explicit model for live runs, and records requested identity separately from unattested actual identity.
- Defaults the primary character identity portrait to a frontal, level face shared by headed and headless variants. Neutral-background masters must be fully opaque; raster alpha facts are measured and bound to their review receipts without forbidding valid RGBA or flattening source images.
- Keeps Apple Vision identity/probe mocks inside portable test fixtures so Linux CI does not claim a real macOS backend; Pillow-free tests still exercise opacity and receipt integrity.
- Preserves the selected image craft owner during authorized execution, including clean-image and key-visual scenarios, while keeping missing packets and wrong media scope blocked.
- Counts selected provider metadata once across craft and isolated contexts, preventing boundary-budget fallback from dropping required validators or collaborators without raising context limits.
- Removes the first-image circular dependency: selected design stages can precede generated assets; complete asset/stress certification remains mandatory for video compilation. Adds from-zero coverage with no placeholder PNGs.
- Supports standalone product/character/prop/scene studies with an explicit
asset_onlyscope and no invented film timeline. Direct still prompts preserve the requested canvas and omit controller IDs, hashes and unsolicited page titles. - Applies current GPT-6 Astra guidance to scenario selection and repository instructions: bounded autonomy, explicit Skill conflicts, proportional verification, and model-independent controller ownership.
- Keeps short Chinese dialogue edits bounded without a document-level voice corpus, and uses Sepia for demonstrated structural/discourse needs while treating its older model tables as priors for Astra.
- Separates
planning_only,prompt_only, real pre-video image assets, and final-video generation. Natural-language image execution aliases route through Delivery, while questions, planning requests, authorization applications, and explicit no-media requests stay non-executing. - Adds a cross-task scope-consistency contract that preserves decision-relevant user wording and required methods without pretending local hashes prove user authorization.
- Connects character-master and every imagegen execution path to an active asset packet, host-bound project root, visual-plan truth, role contract, prompt hash, dependency DAG, saved evidence and scope-specific route context. Storyboard/clean prompts must be exact outputs of a sealed production Jingzao handoff bound to the complete active shot truth; multi-shot overview pages use deterministic assembly. Incorrect two-view character prompts, stale cross-project frames, identity/wardrobe/material drift, arbitrary dependencies, non-execution scenarios, and image/video authorization crossover fail before the media adapter is exposed.
- Adds a pre-video asset evidence gate: planned/prompt-ready rows remain incomplete; generated candidates need actual PNG evidence and visual-review binding before they are ready for user review, which still does not grant user acceptance.
- Adds a post-generation character-master vision gate. It detects the far-left portrait and four separate full bodies before any pass/downstream claim, preserves the remaining orientation/material review boundary, and blocks the exact two-body false-positive found by the isolated forward test.
- Supports the installed
mr-li-seedance-25v1.9.0 only as an isolated Seedance method provider: exact version/body/reference receipts, pre-write capacity, model-scoped segment limits, speaker-change cuts, dialogue permission, current-segment bindings, clean prompt surfaces and accurate next-start state. Handoffs remainapplied_unverifieduntil host adoption evidence exists; its asset layout and controller/ledger semantics are intentionally not copied. - Distinguishes humanization planning from execution. A ready plan starts
not_run; local before/after/provider receipts remainapplied_unverifiedand cannot self-grant completion. New-project register rules no longer masquerade as source-author voice evidence. - Binds Script-to-Seedance handoffs to the trusted capability-card registry, typed image/video/audio/total reference limits, an inode-stable project-root script file, and exact source-line dialogue locators. Forged cards, inflated limits, path escapes, symlinks, hardlinks, stale scripts and shortened/negated dialogue fragments fail closed.
- Adds compact-state 2.0 → 2.1 migration. Ambiguous legacy generation authorization becomes
scope_conflictwith both image and video permissions false until reconfirmed. - Replaces large host-trace accumulation with bounded streaming reads: 16 MiB sealed prefixes, 15 MiB records, a distinct 10 MiB inline-PNG limit and 64 KiB chunks, retaining common image events while avoiding the prior 512 MiB/64 MiB memory path.
- Pins project-relative ancestor directories while reading execution inputs, rejects symlink/hardlink/path escapes, bounds Prompt IR and handoff payloads, and revalidates Jingzao only from a private sealed snapshot of its content-addressed provider and project inputs.
- Adds a post-v0.7.1 product specification for a native Visual Workspace adjustment panel. The PRD remains a future implementation basis rather than part of the v0.8.0 execution runtime.
- Adds Studio
video_distillationwith an offline Workbench evidence bridge, twelve craft axes, candidate mechanisms and explicit effect-comparison records; media/binding validation remains separate from factual and creative-quality review. - Preserves generation and delivery gate precedence, rejects changed media properties and invalid evidence, and requires study/upgrade review before an original-production trial. Interface integration remains a later PRD.
- Adds a supplementary detailed-storyboard coverage contract for within-shot action panels and explicit reverse/eyeline pairs, preserving legacy representative-frame and clean-input contracts. Whole-film asset checks cannot pass with only planned images; model-specific grouping can retain connected multi-shot action.
- Adds a fail-closed scene/support truth gate to the existing Jingzao frame handoff: high-risk panel bindings require it, while scene attachments, role-specific reference sovereignty, visible support declarations, parent-frame invalidation, risk-selected constraint inputs, signed prompt-authority semantic preflight, and host-observed prompt/reference requests are hash-bound without changing legacy rough-planning packets. Every direct image-tool selection binds validated upstream risk; high risk also binds prompt, ordered attachments, truth revision, and structured output semantic review through the existing production ledger, while hash-scoped low-risk legacy selections stay lightweight.
- Revalidates the complete storyboard-to-Jingzao packet before ledger selection instead of accepting extracted truth wrappers. Reviewer public keys now default to persistent host configuration outside the installed Skill, with explicit status/provisioning CLIs and no private-key generation.
- Treats the validated packet target-frame prompt entry as the execution source of truth: prompt text/hash and normalized ordered attachments must match exactly before ledger selection.
- Requires an explicit handoff provider root when a ledger revalidates a production packet. Empty review registries now report
TOOL_BLOCKED; public-key provisioning verifies canonical public input externally before Python reads or stores bytes. - Restricts production handoff providers to resolved installed
jingzao-image-forgecatalog paths and rejects provider/artifact overlap in either direction, including symlink aliases. - Adds a bounded
human_language_revisionroute: explicit “说人话 / 去 AI 味 / 自然一点” requests selectshuorenhua, withhumanizer-zhas diagnostic-only validation and DIRcreative fallback when the optional provider is absent. - Makes source-derived authored voice a default requirement for audience-facing story, screenplay, storyboard-description, PPT, README and report copy while protecting facts, literal SUPER, names, coined terms, character diction and genre register.
- Preserves mythic, historical, literary and period language instead of flattening every natural-language pass into casual modern Chinese;
de-AI-writingremains the ordinary bounded fidelity cleanup rather than a stacked second owner. - Records the cross-project behavior study that exposed the routing gap without copying RTBC or Shanhaijing wording into the Skill.
- Updates the external
mr-li-seedance-25integration for the verified 1.8.2 contract: formal prompt compilation routes to Studio, binds visual-baseline/prompt/duration/format references, and uses a separate 64 KiB method context when paired with the existing compiler; bounded Fast edits do not inflate their context or falsely claim the method was loaded. - Adds a layered humanization planner and Sepia 0.5.0 routing by operation and text/venue profile. Architecture/venue, discourse and surface use source-bound findings; cluster and whitelist verdicts keep intentional voice, venue conventions and isolated hits out of the editable set; model identity is never inferred from prose.
- Makes refactor/recreate two-stage: the first Sepia call is findings-only, while edit authority requires the exact diagnosis/source hash, accepted finding IDs and content-addressed voice/venue/domain calibration. Document/profile mismatches and evidence-free direct selector calls fail closed.
- Requires host-read calibration sources before sampled VOICE/venue contexts can unlock edit authority; unresolved references return
waiting_for_host_readback. Inline sources are capped at 64 KiB and the complete humanization evidence payload is counted against a separate 128 KiB Studio budget instead of disappearing from context accounting. - Replaces self-asserted preservation booleans with actual source text, exact entry spans, coverage and set contracts for recreation. The selector binds document type and a hashed narrative guard, so direct Sepia calls cannot bypass protections or invent corpus-derived subplots, nonlinear time, delayed reveals, named references or rarity moves.
- Keeps
shuorenhua,de-AI-writing, andhumanizer-zhas conditional bounded or diagnostic passes rather than stacking rewriting owners.de-AI-writingis reachable only as an explicit bounded Chinese fidelity route; short English work uses DIR's bounded method; final validation is diagnosis-only and cannot rewrite the prior result. External provider bodies remain host-installed and hash-read at selection time; this repository contains only routing and distilled method contracts. - Adds an answer-first README definition, AI-video use cases, workflow FAQ, a compact
llms.txtdiscovery map, and a 60-second release-candidate regression report. The project-level regression correctly blocks two S07 images with false scene/support geometry while preserving the other 28 panels and both model plans.
- Replaces the split face/wardrobe default for new recurring clothed humans with a unified headed character master sheet generated coherently in one pass: one large three-quarter portrait plus matched full-body front, left profile, right profile, and back views in one physical image.
- Requires a single wide horizontal turnaround row, a declared 75% saved-canvas height floor, and signed external visual review; the validator checks policy/evidence binding but does not claim automatic pixel measurement.
- Adds an explicit
headless_safederivative mode that preserves the approved master body and wardrobe, retains one readable portrait, and makes all four full-body views headless; headed and headless masters cannot be active together. - Adds conditional character detail sheets for left/right asymmetry, accessories, logos/emblems/text, closures, pockets, hardware, lining, directional materials, and macro-shot needs. Exact graphics require verified source bytes rather than generated approximations.
- Lets Seedance bind one active canonical character sheet plus a shot-needed supporting detail reference from the same stress report, while exact-graphic sources remain planning-only and cannot attach directly.
- Introduces
ai_film_asset_stress_test_v3with headed/headless/state-derivative validation, mutually exclusive full-source or identity-plus-wardrobe generation provenance, approved-master hash binding, horizontal-layout and external readable-scale review contracts, headless hand/wrist/rear-collar preservation, detail-callout coverage, exact-graphic source binding, 14/16-case matrices, and 50 negative regression cases while preserving v1/v2 verification. - Adds a focused DIR character-master-sheet workflow and updates reference planning, asset foundation, Seedance handoff, and prompt binding without creating another controller or modifying external provider Skills.
- Fixes the formal release gate so archive-install parity independently verifies the canonical annotated remote tag instead of rejecting valid
CANONICAL_REMOTE_TAGmetadata after installation.
- Makes
mr-li-seedance-25the default method owner for direct Seedance 2.5 prompts, applies it before emotion-specialist or deterministic conversion passes, recognizes declared 2.5 targets before exact capability resolution, and keeps exact 2.0 cards authoritative against alias conflicts. - Consolidates overlapping Seedance responsibilities inside the DIR matrix without merging external Skills: emotion work retains its specialist owner,
convert-script-to-seedanceretains the compile-only handoff, and preflight remains independently selected. - Adds explicit face-identity and headless-wardrobe reference roles plus an eleven-case character stress matrix, preventing multiple readable face sources while retaining versioned appearance-state, full-body, lighting, group, handedness, and scale checks.
- Maps the reviewed HELL GRIND workflow to existing asset-foundation, stress-test, GEO, prompt, ledger, and edit surfaces without adding another controller.
- Adds an independently sourced Seedance 2.5 capability card with a 30-second upper bound, typed 30-image/10-video/10-audio limits, native audio-video intent, editing/extension evidence, and fail-closed manual execution status while retaining the complete Seedance 2.0 card.
- Integrates the globally discoverable
mr-li-seedance-25Skill as a bounded method collaborator only for exact Seedance 2.5script_to_seedance_v1work; DIRcreative, the existing converter, asset-foundation gate, prompt validator, and legacy fallback ownership remain unchanged. - Makes the Seedance adapter accept the union of known exact-card surfaces while the prompt compiler enforces version-specific duration and typed reference limits, with 2.0/2.5 positive and negative regression coverage.
- Makes macOS tagged reproducibility builds use the same annotated-tag restore and
CANONICAL_REMOTE_TAGrelease mode as the Linux matrix, restoring byte-identical cross-platform artifacts.
- Completes annotated-tag CI by requiring remote-tag verification when parity-checking a
CANONICAL_REMOTE_TAGformal install.
- Publishes the 0.5 AI-film asset-foundation, stress-test, production-ledger, Seedance, and Jingzao workflow after closing source/installed Skill-entry layout gaps.
- Makes annotated-tag CI restore the verified tag commit onto a local
mainbranch before the formal release preflight, preserving the exact commit/tag binding without detached-HEAD false failures.
- Adds a dynamic whole-film visual asset contract covering every scene, approved shot, director-storyboard page, and scene-coherent video-generation unit; replaces the former small vertical smoke case with a 60-second 16:9 TVC acceptance fixture, 24 frame-aligned formal shots, and 48 required assets.
- Upgrades the visual plan to v2.2: source-inventory and approved-shot-card binding, continuous frame-aligned timecodes, full per-shot creative truth and per-asset truth hashes, exact storyboard/direct-input dependencies, model-strategy input ranges, canonical dependency-free PNG evidence, role-appropriate delivery aspect checks, full raster decoding, decoder-portable normalized pixel hashes, package-root containment, separate technical receipts, atomically read external visual-review manifests, metadata-resistant duplicate controls, and a stdlib schema fallback.
user_lockedcannot bypass review, and technical stamping cannot grant visual QA. The highest claim isvisual_assets_complete; finished-master acceptance is intentionally out of scope. - Makes Fast and Studio content-first by default with zero unconditional protocol reads, one Route Card, bounded context, and measured warm/cold routing budgets.
- Requires candidate-exact Codex event-log prefixes, a sealed explicit user invocation, real imagegen/view event IDs, hash-bound prompts/references/outputs, OpenAI C2PA, and a sealed-input separate-task visual review for global-install media claims; unsigned host trace and reviewer judgment remain explicit, and video remains unverified.
- Adds Specialist Exchange v2 while retaining tested v1 read compatibility, and binds v2 activation to ADCO's exact exchange registration, descriptor snapshot, locked inputs, isolated outputs, and no-overwrite receipt path.
- Hardens v1/v2 receipt creation against ancestor-directory races with project-root directory FDs, no-follow traversal, atomic no-replace rename, inode readback, and open-FD tombstoning without name-based cleanup.
- Makes formal installation a same-process exact-source verification and install: artifact/checksum/commit/tag binding, reproducible rebuild, complete archive manifest, same-filesystem staging, post-swap readback, and rollback on mismatch; extracted installers and metadata-only claims cannot authorize a formal install.
- Seals the pinned C2PA verifier into a private exact-byte snapshot for the whole audit, preventing validation-time pathname replacement from changing the trust root.
- Seals one commit across media, preflight, build, verification, and final HEAD readback so a long release gate cannot mix evidence from two revisions.
- Adds a six-stage, byte-bound asset-foundation pass plus the callable
ai-film-asset-stress-testandai-film-production-ledgerSkills. Per-asset stress coverage, combination/evidence isolation, canonical/planning provenance, signed host trust, append-only genesis/attempt lineage, exact asset/prompt/output bytes, review, delivery, and verified billing all fail closed. - Gates
script_to_seedance_v1on project-bound foundation/stress evidence and exact attached asset versions while preservingjingzao-image-forgeas the cinematic storyboard-frame owner. The Studio selector routes the new validator/ledger without adding a controller or increasing context budgets.
- Adds interactive, customer-facing chat visualizations for creative choices, story curves, script timing, shot rhythm, visual directions, reference graphs, prompt handoff, and generation QA.
- Adds responsive light/dark/mobile layouts, synchronized recommendations and actions, real image review, explicit no-media and illustrative-placeholder states, and read-only confirmation echoes.
- Hardens project-relative image and evidence handling with physical SHA-256 verification, current lifecycle checks, JSON Pointer validation, confirmed source/authorization/channel evidence, bounded media, and path/symlink escape rejection.
- Adds 14-page, 84-scenario browser regression coverage and dependency-free fail-closed schema validation for clean Python installations and CI.
- Fixes tagged GitHub Actions release validation by fetching the annotated tag object before exact-release packaging.
- Fixes CI portability for the prompt-system negative fixture by matching the stable offending field instead of environment-specific JSON Schema wording.
- Adds the executable prompt production system: Prompt IR schema, image/video prompt manifests, prompt compiler, fixture audits, prompt QA taxonomy, and Seedance prompt fixtures.
- Adds client-film hard gates for voiceover timing, shot timing, asset contracts, internal-term hygiene, and direct-prompt rejection.
- Strengthens professional image/video prompt authoring around reference binding, multi-character/multi-asset continuity, timeline beats, audio, look layers, and single-variable retry.
- Fixes installer self-test behavior when validation is run from the formal installed SkillHub directory.
- Fixes GitHub Actions Ruby/Psych YAML alias compatibility by explicitly enabling fixture aliases in audit loaders.
- Fixes the release artifact installed-validation negative fixture for legal
RELEASE-METADATA.jsonpackages.
- Fixes formal release package validation so installed packages correctly treat
.dircreativeas lifecycle sentinels only, while source checkouts still require current-first runtime state. - Aligns installed-package behavior audits with the sanitized
INTERNAL_SKILL.mdlayout. - Hardens release staging target validation for package roots inside temporary directories.
- Automatically routes substantial film and advertising work into a bounded Director Room while preserving a lightweight path for small edits and factual questions.
- Adds executable role harnesses, typed disagreements, arbitration, handoffs, retries, and fail-closed council evidence.
- Adds source-bound creative strategy, brand-voice copy development, de-AI fidelity review, and editable presentation routing.
- Replaces static model folklore with versioned capability cards, official-source tiers, rights and authorization gates, and image/video behavior evaluations.
- Adds current-first standalone state, semantic migration, thread/disk reconciliation, supersession, storage lifecycle, and adversarial self-tests.
- Hardens packaging, exact-release verification, atomic installation, checksums, CI, and ADCO bilateral compatibility gates.
- Initial prompt-only DIRcreative release with film-preproduction skills, local installation, technical validation, and ADCO specialist exchange v1.