What does framechord do?
Use the framechord binary from any agent that can run a process. If it is missing, read installation for public installers, user-local installation and explicit updates, or explain cargo install --path /path/to/framechord for a source checkout. Do not switch media tools or make provider API calls inside the skill.
Existing mediaagent/ projects are supported in place. Read renaming and legacy projects before upgrading an older installation; use returned data_directory paths rather than assuming the folder name.
Read commands for every command and flag in this binary version; workflow for outputs, limits, recovery, and review; authoring for character and infographic templates. Machine-readable schemas are in references/schemas/. Exported templates are in templates/.
Read character continuity when reusing the same characters across scenes. Use selected identity references, supported fidelity controls, and comparison against the actual original images; API success does not establish likeness.
Read connection checks when checking setup or provider/key failures: doctor checks presence offline; doctor --live explicitly authenticates with read-only requests. It supports provider filtering, timeouts, and JSON. Read status/actions per provider; successful authentication does not establish generation/model/billing readiness.
Read voices when discovering voices, designing previews, accepting authorized recordings, or explicitly creating/cloning/binding a selected provider voice. Preview generation, provider-side creation, canonical binding, and acceptance of recordings are separate operations.
Read setup and current models for the looping terminal wizard, Google-first defaults and an explicit OpenAI alternative, preserving project .env setup, masked credentials and project skill installation, noninteractive setup readiness, model filters and explicit catalog refresh. For a single asset, use explicit current provider/model/input/output flags without initialization. For repeated work, agents use init PATH --preset recommended --json --non-interactive, inspect config status --json, and supply exact configuration with config setup --file where needed. Human terminal users run init to resume setup. No provider calls occur during initialization. Existing character model locks are preserved; retired-model identities need a new character version and human-accepted references before generating again.
Preserve requested dialogue, lyrics, numbers, and labels exactly. Use a separate explicit planning operation when adapting an article into a podcast. Supply article text in a local file. Characters are optional; when selected, reuse their accepted references and pinned voice/provider/model. Derived identity assets must use the original asset’s recorded provider/model; conflicting overrides fail. Preserve origin metadata instead of copying the project default into derived job YAML. Describe stable identity separately from pose, scene, delivery, or temporary clothing. Candidate references need the user's selection before acceptance.
Author files using templates, then validate. Inspect generate --job ... --dry-run --json before execution when inputs are new. Call generation only when it is part of the user's request. Use --json --non-interactive for automation. Report artifact paths returned by the binary and incomplete jobs honestly.
Use runs status, runs wait, runs reconcile, and runs resume for recorded runs. Retrieve supported provider results before considering a paid retry. Never silently retry an unknown paid request; --retry-unknown requires explicit user authorization. Reviews record quality separately from generation success. Review does not automatically revise, and revisions do not replace canonical character references.
Read remediation workflows for v2 casts, selected planner inputs, bounded verification, reference storage, recovered-result attachment, usage, and migrations. Jobs/config/runs accept v1 and v2; new casts and podcast assembly require v2. Other manifest versions remain unchanged.
Use plan --character MANIFEST --asset FILE to select dependencies explicitly; the binary copies them into _inputs/. For a scene with multiple saved characters, use cast entries and job-specific direction files. Require accepted references and compatible original model lineages; never rewrite profiles to make a conflict disappear.
For acceptance tests, preview verify --suite SUITE --json and obtain approval for the bounded paid batch before --execute --preview-hash HASH. Missing credentials are skipped, not passed. Voice creation/cloning use their separate explicit persona commands and approvals. review report compares generated artifacts and saved references; only actual user selections authorize canonical acceptance.
models --live reads provider catalogs; visibility is not generation entitlement. usage reports returned units and unknown costs honestly. storage plan --output cleanup.yaml prepares a conservative cleanup preview outside framechord/; storage clean --plan cleanup.yaml only removes unreferenced blobs after revalidation. migrate --kind job|config|run --file FILE previews; --apply preserves a v1 backup. For uncertain jobs, runs recover requires the recorded fingerprint and reconciliation evidence, then runs resume publishes the supplied response without regeneration.