Transcript Lake

Documentation / CLI

transcript-lake goal title

Distill caller-supplied text into a short goal title with the qualified local GGUF.

Invocation#

sh
transcript-lake [--data-dir <path>] goal title (--text <text>|--stdin) [--json]

Purpose#

Distill caller-supplied text into a short goal title with the qualified local GGUF.

Inputs and options#

  • Exactly one of --text <text> or --stdin is required. Input is trimmed without a character cap; the model's declared context governs what Ster can accept.
  • --json emits goal, model filename, and model SHA-256.
  • Requires ster on PATH or named by TRANSCRIPT_LAKE_STER. The goal model's GGUF is run on Ster's own decoder from a checkpoint directory under the Lake data directory: the GGUF linked beside the config.json, tokenizer.json and tokenizer_config.json of the base model its published card declares (cardData.base_model at the pinned revision), fetched once at the base repository's revision of that moment and recorded in base-revision. The answer runs until the model ends it or fills the context the base config declares (max_position_embeddings), decoded by argmax. Model/prompt overrides are accepted only when their source validation succeeds.

Output and state effect#

Local model inference only; no text is sent to an inference service. The qualified GGUF may be downloaded and cached beneath the selected Lake data directory. Human output is the model's goal title or <goal/>; JSON uses null for no goal. The CLI imposes no title word or character count.

Exact refusals#

  • Input selection: goal title requires exactly one of --text <text> or --stdin.
  • Blank input: goal title input must not be empty.
  • Missing runtime: local goal model runs on Ster: install ster on PATH or name it with TRANSCRIPT_LAKE_STER.
  • No base model declared: <repository>@<revision> declares no base_model in its card (<path>), so Ster has no config or tokenizer to run it with.
  • Base model without a revision: base model <id> answered no revision (<path>); a base config without a context: <path>/config.json declares no max_position_embeddings.
  • Ster failed: Ster could not run the goal model (<ster> generate exited <status>): <Ster's own refusal>.
  • Malformed model output: local goal model returned no goal tag or local goal model returned no opening goal tag.

If a downloaded artifact cannot be published, the refusal names both paths and the operating-system cause: could not publish model artifact <temporary> as <destination>: <cause>. The partial download is removed; if removal also fails, the same refusal appends <temporary> was left behind: <cleanup cause>. A failed download similarly retains its exit status and any cleanup failure. These failures do not produce a goal title.

After digest verification, the validation stamp is published through a unique sibling file, flushed before replacement, followed by a directory flush. A failed stamp operation reports its path and cause; cleanup failures additionally name the sibling left behind. The command does not report successful validation when stamp publication fails.

The checkpoint's first recorded base revision uses the same durable publication path. A failed write cannot leave a partial revision that a later run would use to fetch the tokenizer or configuration.

All long flags are exact and single-use. A value flag never consumes another --flag. Global --data-dir may appear anywhere in the invocation, must appear once, and selects an absolute root.

Back to the complete CLI command tree.