Update to fal
Snapshot Nov 14, 2026 · 00:10 UTC · version 1.3.2
Collection source: downloaded plugin package. These snapshots do not have a confirmed matching collection source. Differences in file lists alone do not establish changes to the package.
Payment or plan references changed
Skill instructions — Now says: Bundled guidance is fallback craft. For inputs, recommend_model's `input_contract` and live validation at submission are authoritative. See [the shared...
Observed in published text. Live prices and checkout terms have not been verified by this change.
Up to 3 plugins free · no card required.
All fields and original values.
Skill instructions
Sep 30, 2026 · 22:46 UTC
endpoint. Prefer an exact match over a similarly named model exposed by an - When no model is specified, call `recommend_model` for a current candidate, then judge the result against the requested artifact and constraints. - Inspect ...
Nov 14, 2026 · 00:10 UTC
Bundled guidance is fallback craft. For inputs, recommend_model's `input_contract` and live validation at submission are authoritative. See [the shared production contract](../shared/production-contract.md). endpoint. Copy the exact re...
Compare saved observations
Download comparison JSONSKILL.md text changes
− removed · + added
--- before +++ after @@ -9,6 +9,10 @@ # fal.ai Media +Bundled guidance is fallback craft. For inputs, recommend_model's +`input_contract` and live validation at submission are authoritative. See +[the shared production contract](../shared/production-contract.md). + Route media generation and processing through the fal.ai MCP tools exposed by this plugin. This is the provider-level entry point; load a more specialized Fal skill when the task needs its production guidance. @@ -18,23 +22,87 @@ - Preserve an explicitly requested provider. When the user names Fal, keep the operation on Fal. When the user names another provider, do not invoke Fal. - For a named model, search Fal's live catalog and verify its exact canonical - endpoint. Prefer an exact match over a similarly named model exposed by an + endpoint. Copy the exact returned endpoint ID into every call; never abbreviate IDs or invent preview/version suffixes. Prefer an exact match over a similarly named model exposed by an intermediary. Do not silently substitute a different model or provider. -- When no model is specified, call `recommend_model` for a current candidate, - then judge the result against the requested artifact and constraints. -- Inspect the selected endpoint schema before constructing its input. Check - pricing when duration, resolution, batching, training, audio, video, or 3D - can materially affect cost. +- When no model is specified, call `recommend_model`. When it selects a + model, submit its `next_call` right away + with an input built from `input_contract`. Submission validates the input + against the live schema before any charge. Read `get_model_schema` only for + a parameter the contract omits, and look up pricing only when the user asks + about cost or sets a budget. ## Execution +- On a catalog/schema/pricing 429, stop the current model sweep. Respect the + returned cooldown across discovery tools and reuse already verified schemas + and candidates while waiting. Don't rotate endpoints or tools to evade it. + An unresolved canonical ID needs a corrected live ID, not unchanged retries. +- A history 403 means access was refused, not that a job failed or credits ran + out. Stop until access changes. Don't reconnect automatically, switch accounts + or silently change semantic search to another mode/API to bypass the refusal. + +- Don't check the account before an upload or generation: both report a + billing block themselves, with recovery steps. +- fal is pay-as-you-go. A billing block means uploads and generations on that + account will fail until the user resolves it; model search, schemas and + pricing still work. + Finish the plan, including the price, then tell the user once in plain words + with the returned `recovery.message` link and the connected account. Stop the + remaining uploads and generations in the plan instead of trying each one. + Never purchase credits, repeatedly retry, poll, or change accounts. +- When the user says they added credits or chose an account, check readiness + again if the tool is available, then continue the prepared plan. + `unknown` does not mean insufficient credits; `ready` is not an affordability + guarantee. Preserve the task in the conversation and verify reference URLs + remain usable. A readiness check never resumes work itself. +- If a request ID already exists, continue that job on its original account. + Do not submit another generation to recover from a billing or result error. +- To find an unsaved previous result, use `search_request_history` when + available. For a requested recreate or variation, recover actual original + settings with `get_job_recipe`, preserving the saved `response_url` for its + original account. If retained input is unavailable, don't infer it from an + asset prompt. These reads never authorize another generation themselves. - Use `run_model` for bounded image and utility operations. - Use `submit_job` for video, audio, 3D, training, or other long-running work; - poll the same request with `check_job`, then fetch it with `get_job_result`. -- Upload local inputs before generation when the endpoint requires a URL. + poll the same request with `check_job`, which returns the result once the + job is `COMPLETED` (`get_job_result` also works). Read finished results + promptly: fal keeps them for a limited time. +- For several variations or comparisons, use one `submit_jobs` call, then + `wait_for_jobs`. +- Show results with `show_media` (`show_media_batch` for several jobs) where + the host renders MCP app UI, otherwise share the media URLs. Then offer a + follow-up, such as variations, an edit, animating an image or an upscale. +- Upload inputs before generation when the endpoint requires a URL. Reuse a + usable host attachment URL or an existing fal URL when the model accepts it. + Hosted MCP cannot read your computer's `file_path`. If the host can read a + file and make HTTP requests, call `upload_file` with `prepare_upload=true`, + exact `file_name` and `file_size`, then PUT the actual bytes to `upload_url`. + Preparation alone is not an upload: use `file_url` only after PUT succeeds. + Never send fal credentials to the signed upload URL. If the host cannot read + or PUT bytes, ask for an accessible file URL. Don't invent remote local paths. + Base64 is for small files encoded programmatically from complete real bytes; + never print, hand-copy, guess or truncate it. A download rejection or redirect + is not permission to bypass URL protections; use the host's original attachment + or a permitted direct file URL instead of repeatedly retrying the same URL. +- For saved asset/entity IDs or an existing job output, use + `resolve_model_inputs` when available to resolve explicit `$fal_ref` markers + in schema-supported media fields. Pass only its plain returned input to the + generation tool. Older clients can read `get_asset`, `get_entity`, or + `get_job_result` and use the exact URLs directly; never guess fields or + silently drop extra defining references. - Preserve continuity with image-to-video, reference-to-video, or frame extraction when an earlier result should anchor the next shot. -- Return the endpoint ID, request ID, output URL, and material limitations. +- Read saved references with `inspect_images` when preservation requires + pixels. Keep original sources separate from user-approved and rejected + outputs; inspection is not approval. If inspection cannot establish a + requirement, report `unverified` and stop dependent preservation work. +- `get_asset_lineage` recovers captured ancestors, not fabricated recipes. + `get_generation_defaults` is read-only preference fallback; explicit requests + and actual source recipes win. +- Return exact `output_path` selectors, private original-account job handles + and honest expected/observed delivery checks. Use `prepare_deliverables` when + a reusable checked manifest helps. Save only when requested, using + `save_job_outputs`; read uncertain writes before retrying. ## Related Skills
Full technical diff · 1 changed fields
changed /skill_md_contents
"---\nname: fal-media\ndescription: >\n Generate, edit, or process media through fal.ai. Use when the user names Fal\n or fal.ai, requests Fal model discovery, schema, pricing, or job handling, or\n asks to use a named media model available through Fal such as MiniMax H3 Max.\n Do not activate when the user explicitly chooses another provider.\n---\n\n# fal.ai Media\n\nRoute media generation and processing through the fal.ai MCP tools exposed by\nthis plugin. This is the provider-level entry point; load a more specialized\nFal skill when the task needs its production guidance.\n\n## Provider And Model Routing\n\n- Preserve an explicitly requested provider. When the user names Fal, keep the\n operation on Fal. When the user names another provider, do not invoke Fal.\n- For a named model, search Fal's live catalog and verify its exact canonical\n endpoint. Prefer an exact match over a similarly named model exposed by an\n intermediary. Do not silently substitute a different model or provider.\n- When no model is specified, call `recommend_model` for a current candidate,\n then judge the result against the requested artifact and constraints.\n- Inspect the selected endpoint schema before constructing its input. Check\n pricing when duration, resolution, batching, training, audio, video, or 3D\n can materially affect cost.\n\n## Execution\n\n- Use `run_model` for bounded image and utility operations.\n- Use `submit_job` for video, audio, 3D, training, or other long-running work;\n poll the same request with `check_job`, then fetch it with `get_job_result`.\n- Upload local inputs before generation when the endpoint requires a URL.\n- Preserve continuity with image-to-video, reference-to-video, or frame\n extraction when an earlier result should anchor the next shot.\n- Return the endpoint ID, request ID, output URL, and material limitations.\n\n## Related Skills\n\n- Load `model-routing` for model selection beyond a direct exact match.\n- Load `fal-models-catalog` for broader family or modality comparisons.\n- Load `fal-prompting` after selecting an endpoint that needs family-specific\n prompt construction.\n- Load `fal-workflow` for model chains, batches, or dependent generation steps.\n"
"---\nname: fal-media\ndescription: >\n Generate, edit, or process media through fal.ai. Use when the user names Fal\n or fal.ai, requests Fal model discovery, schema, pricing, or job handling, or\n asks to use a named media model available through Fal such as MiniMax H3 Max.\n Do not activate when the user explicitly chooses another provider.\n---\n\n# fal.ai Media\n\nBundled guidance is fallback craft. For inputs, recommend_model's\n`input_contract` and live validation at submission are authoritative. See\n[the shared production contract](../shared/production-contract.md).\n\nRoute media generation and processing through the fal.ai MCP tools exposed by\nthis plugin. This is the provider-level entry point; load a more specialized\nFal skill when the task needs its production guidance.\n\n## Provider And Model Routing\n\n- Preserve an explicitly requested provider. When the user names Fal, keep the\n operation on Fal. When the user names another provider, do not invoke Fal.\n- For a named model, search Fal's live catalog and verify its exact canonical\n endpoint. Copy the exact returned endpoint ID into every call; never abbreviate IDs or invent preview/version suffixes. Prefer an exact match over a similarly named model exposed by an\n intermediary. Do not silently substitute a different model or provider.\n- When no model is specified, call `recommend_model`. When it selects a\n model, submit its `next_call` right away\n with an input built from `input_contract`. Submission validates the input\n against the live schema before any charge. Read `get_model_schema` only for\n a parameter the contract omits, and look up pricing only when the user asks\n about cost or sets a budget.\n\n## Execution\n\n- On a catalog/schema/pricing 429, stop the current model sweep. Respect the\n returned cooldown across discovery tools and reuse already verified schemas\n and candidates while waiting. Don't rotate endpoints or tools to evade it.\n An unresolved canonical ID needs a corrected live ID, not unchanged retries.\n- A history 403 means access was refused, not that a job failed or credits ran\n out. Stop until access changes. Don't reconnect automatically, switch accounts\n or silently change semantic search to another mode/API to bypass the refusal.\n\n- Don't check the account before an upload or generation: both report a\n billing block themselves, with recovery steps.\n- fal is pay-as-you-go. A billing block means uploads and generations on that\n account will fail until the user resolves it; model search, schemas and\n pricing still work.\n Finish the plan, including the price, then tell the user once in plain words\n with the returned `recovery.message` link and the connected account. Stop the\n remaining uploads and generations in the plan instead of trying each one.\n Never purchase credits, repeatedly retry, poll, or change accounts.\n- When the user says they added credits or chose an account, check readiness\n again if the tool is available, then continue the prepared plan.\n `unknown` does not mean insufficient credits; `ready` is not an affordability\n guarantee. Preserve the task in the conversation and verify reference URLs\n remain usable. A readiness check never resumes work itself.\n- If a request ID already exists, continue that job on its original account.\n Do not submit another generation to recover from a billing or result error.\n- To find an unsaved previous result, use `search_request_history` when\n available. For a requested recreate or variation, recover actual original\n settings with `get_job_recipe`, preserving the saved `response_url` for its\n original account. If retained input is unavailable, don't infer it from an\n asset prompt. These reads never authorize another generation themselves.\n- Use `run_model` for bounded image and utility operations.\n- Use `submit_job` for video, audio, 3D, training, or other long-running work;\n poll the same request with `check_job`, which returns the result once the\n job is `COMPLETED` (`get_job_result` also works). Read finished results\n promptly: fal keeps them for a limited time.\n- For several variations or comparisons, use one `submit_jobs` call, then\n `wait_for_jobs`.\n- Show results with `show_media` (`show_media_batch` for several jobs) where\n the host renders MCP app UI, otherwise share the media URLs. Then offer a\n follow-up, such as variations, an edit, animating an image or an upscale.\n- Upload inputs before generation when the endpoint requires a URL. Reuse a\n usable host attachment URL or an existing fal URL when the model accepts it.\n Hosted MCP cannot read your computer's `file_path`. If the host can read a\n file and make HTTP requests, call `upload_file` with `prepare_upload=true`,\n exact `file_name` and `file_size`, then PUT the actual bytes to `upload_url`.\n Preparation alone is not an upload: use `file_url` only after PUT succeeds.\n Never send fal credentials to the signed upload URL. If the host cannot read\n or PUT bytes, ask for an accessible file URL. Don't invent remote local paths.\n Base64 is for small files encoded programmatically from complete real bytes;\n never print, hand-copy, guess or truncate it. A download rejection or redirect\n is not permission to bypass URL protections; use the host's original attachment\n or a permitted direct file URL instead of repeatedly retrying the same URL.\n- For saved asset/entity IDs or an existing job output, use\n `resolve_model_inputs` when available to resolve explicit `$fal_ref` markers\n in schema-supported media fields. Pass only its plain returned input to the\n generation tool. Older clients can read `get_asset`, `get_entity`, or\n `get_job_result` and use the exact URLs directly; never guess fields or\n silently drop extra defining references.\n- Preserve continuity with image-to-video, reference-to-video, or frame\n extraction when an earlier result should anchor the next shot.\n- Read saved references with `inspect_images` when preservation requires\n pixels. Keep original sources separate from user-approved and rejected\n outputs; inspection is not approval. If inspection cannot establish a\n requirement, report `unverified` and stop dependent preservation work.\n- `get_asset_lineage` recovers captured ancestors, not fabricated recipes.\n `get_generation_defaults` is read-only preference fallback; explicit requests\n and actual source recipes win.\n- Return exact `output_path` selectors, private original-account job handles\n and honest expected/observed delivery checks. Use `prepare_deliverables` when\n a reusable checked manifest helps. Save only when requested, using\n `save_job_outputs`; read uncertain writes before retrying.\n\n## Related Skills\n\n- Load `model-routing` for model selection beyond a direct exact match.\n- Load `fal-models-catalog` for broader family or modality comparisons.\n- Load `fal-prompting` after selecting an endpoint that needs family-specific\n prompt construction.\n- Load `fal-workflow` for model chains, batches, or dependent generation steps.\n"
Full snapshot data
{
"description": "Generate, edit, or process media through fal.ai. Use when the user names Fal or fal.ai, requests Fal model discovery, schema, pricing, or job handling, or asks to use a named media model available through Fal such as MiniMax H3 Max. Do not activate when the user explicitly chooses another provider.\n",
"included_files": [],
"name": "fal-media",
"skill_md_contents": "---\nname: fal-media\ndescription: >\n Generate, edit, or process media through fal.ai. Use when the user names Fal\n or fal.ai, requests Fal model discovery, schema, pricing, or job handling, or\n asks to use a named media model available through Fal such as MiniMax H3 Max.\n Do not activate when the user explicitly chooses another provider.\n---\n\n# fal.ai Media\n\nBundled guidance is fallback craft. For inputs, recommend_model's\n`input_contract` and live validation at submission are authoritative. See\n[the shared production contract](../shared/production-contract.md).\n\nRoute media generation and processing through the fal.ai MCP tools exposed by\nthis plugin. This is the provider-level entry point; load a more specialized\nFal skill when the task needs its production guidance.\n\n## Provider And Model Routing\n\n- Preserve an explicitly requested provider. When the user names Fal, keep the\n operation on Fal. When the user names another provider, do not invoke Fal.\n- For a named model, search Fal's live catalog and verify its exact canonical\n endpoint. Copy the exact returned endpoint ID into every call; never abbreviate IDs or invent preview/version suffixes. Prefer an exact match over a similarly named model exposed by an\n intermediary. Do not silently substitute a different model or provider.\n- When no model is specified, call `recommend_model`. When it selects a\n model, submit its `next_call` right away\n with an input built from `input_contract`. Submission validates the input\n against the live schema before any charge. Read `get_model_schema` only for\n a parameter the contract omits, and look up pricing only when the user asks\n about cost or sets a budget.\n\n## Execution\n\n- On a catalog/schema/pricing 429, stop the current model sweep. Respect the\n returned cooldown across discovery tools and reuse already verified schemas\n and candidates while waiting. Don't rotate endpoints or tools to evade it.\n An unresolved canonical ID needs a corrected live ID, not unchanged retries.\n- A history 403 means access was refused, not that a job failed or credits ran\n out. Stop until access changes. Don't reconnect automatically, switch accounts\n or silently change semantic search to another mode/API to bypass the refusal.\n\n- Don't check the account before an upload or generation: both report a\n billing block themselves, with recovery steps.\n- fal is pay-as-you-go. A billing block means uploads and generations on that\n account will fail until the user resolves it; model search, schemas and\n pricing still work.\n Finish the plan, including the price, then tell the user once in plain words\n with the returned `recovery.message` link and the connected account. Stop the\n remaining uploads and generations in the plan instead of trying each one.\n Never purchase credits, repeatedly retry, poll, or change accounts.\n- When the user says they added credits or chose an account, check readiness\n again if the tool is available, then continue the prepared plan.\n `unknown` does not mean insufficient credits; `ready` is not an affordability\n guarantee. Preserve the task in the conversation and verify reference URLs\n remain usable. A readiness check never resumes work itself.\n- If a request ID already exists, continue that job on its original account.\n Do not submit another generation to recover from a billing or result error.\n- To find an unsaved previous result, use `search_request_history` when\n available. For a requested recreate or variation, recover actual original\n settings with `get_job_recipe`, preserving the saved `response_url` for its\n original account. If retained input is unavailable, don't infer it from an\n asset prompt. These reads never authorize another generation themselves.\n- Use `run_model` for bounded image and utility operations.\n- Use `submit_job` for video, audio, 3D, training, or other long-running work;\n poll the same request with `check_job`, which returns the result once the\n job is `COMPLETED` (`get_job_result` also works). Read finished results\n promptly: fal keeps them for a limited time.\n- For several variations or comparisons, use one `submit_jobs` call, then\n `wait_for_jobs`.\n- Show results with `show_media` (`show_media_batch` for several jobs) where\n the host renders MCP app UI, otherwise share the media URLs. Then offer a\n follow-up, such as variations, an edit, animating an image or an upscale.\n- Upload inputs before generation when the endpoint requires a URL. Reuse a\n usable host attachment URL or an existing fal URL when the model accepts it.\n Hosted MCP cannot read your computer's `file_path`. If the host can read a\n file and make HTTP requests, call `upload_file` with `prepare_upload=true`,\n exact `file_name` and `file_size`, then PUT the actual bytes to `upload_url`.\n Preparation alone is not an upload: use `file_url` only after PUT succeeds.\n Never send fal credentials to the signed upload URL. If the host cannot read\n or PUT bytes, ask for an accessible file URL. Don't invent remote local paths.\n Base64 is for small files encoded programmatically from complete real bytes;\n never print, hand-copy, guess or truncate it. A download rejection or redirect\n is not permission to bypass URL protections; use the host's original attachment\n or a permitted direct file URL instead of repeatedly retrying the same URL.\n- For saved asset/entity IDs or an existing job output, use\n `resolve_model_inputs` when available to resolve explicit `$fal_ref` markers\n in schema-supported media fields. Pass only its plain returned input to the\n generation tool. Older clients can read `get_asset`, `get_entity`, or\n `get_job_result` and use the exact URLs directly; never guess fields or\n silently drop extra defining references.\n- Preserve continuity with image-to-video, reference-to-video, or frame\n extraction when an earlier result should anchor the next shot.\n- Read saved references with `inspect_images` when preservation requires\n pixels. Keep original sources separate from user-approved and rejected\n outputs; inspection is not approval. If inspection cannot establish a\n requirement, report `unverified` and stop dependent preservation work.\n- `get_asset_lineage` recovers captured ancestors, not fabricated recipes.\n `get_generation_defaults` is read-only preference fallback; explicit requests\n and actual source recipes win.\n- Return exact `output_path` selectors, private original-account job handles\n and honest expected/observed delivery checks. Use `prepare_deliverables` when\n a reusable checked manifest helps. Save only when requested, using\n `save_job_outputs`; read uncertain writes before retrying.\n\n## Related Skills\n\n- Load `model-routing` for model selection beyond a direct exact match.\n- Load `fal-models-catalog` for broader family or modality comparisons.\n- Load `fal-prompting` after selecting an endpoint that needs family-specific\n prompt construction.\n- Load `fal-workflow` for model chains, batches, or dependent generation steps.\n"
}SHA-256 of public snapshot: 17eaab50c97107ec0a282539b8d666c514ea0aeef6d631da2ebf73735d91b371