Role-template compression¶

bernstein templates compress shrinks the role prompt templates that are prepended to every worker spawn, using an LLM rewrite that is mechanically validated, backed up out of tree, and receipted on the audit chain. bernstein templates restore reverses it byte-identically.

Compression is never automatic - it runs only through this explicit, operator-gated command.

Source:

src/bernstein/core/tokens/template_compression.py - compress/restore
src/bernstein/core/tokens/template_compress_validate.py - validators
src/bernstein/cli/commands/templates_cmd.py - compress / restore

Why a token cut here compounds¶

A role template is fixed prose that ships on every spawn for that role. Trimming it once reduces the input tokens of every subsequent worker of that role, so the saving is per-spawn and recurring rather than one-off.

`bernstein templates compress`¶

bernstein templates compress reviewer
bernstein templates compress --all
bernstein templates compress backend --model <model> --provider openrouter --yes

Argument / flag	Default	Meaning
`ROLE`	-	Compress one role's templates. Mutually exclusive with `--all`.
`--all`	off	Compress every role template.
`--workdir DIR`	`.`	Project root the role templates are resolved from.
`--model NAME`	`anthropic/claude-haiku-4-5`	Model used for the rewrite.
`--provider NAME`	`openrouter`	Adapter/provider for the rewrite (`openrouter`, `openai`, ...).
`--yes`	off	Skip the confirmation prompt.

You must provide exactly one of ROLE or --all.

The safety pipeline¶

Every rewrite passes through the same fixed pipeline before it lands:

LLM rewrite through the configured adapter.
Mechanical validators - the base compaction validators plus template-specific checks: fenced code blocks, headings, URLs, inline code, placeholders, and the completion-contract block must all survive. At most two targeted fix-only retries are allowed; otherwise the compression aborts and the template is left untouched.
Out-of-tree backup of the originals, keyed by content hash, with readback verification.
Chained receipt appended to the audit chain.

The command prints only the template token delta (role: template reduced <pre> -> <post> tokens). Per-spawn dollar savings come from the ledger on subsequent runs, via bernstein cost --by role. Compression itself never reports a dollar figure it cannot yet observe.

Backups¶

Originals are stored under $XDG_DATA_HOME/bernstein/template-backups (falling back to ~/.local/share/bernstein/template-backups), out of tree by design and keyed by content hash. The backup location is printed in the confirmation prompt.

The receipt¶

The audit-chain event is template.compression.receipt and carries {role, pre_sha256, post_sha256, pre_tokens, post_tokens, validators, adapter, model}, with the previous chain digest embedded so the receipt is position-locked. A row is also appended to a templates.lock file mapping the role's pre and post directory digests.

Interaction with team drift¶

That templates.lock row is what lets bernstein team drift classify a compressed template as an intentional change rather than reporting spurious drift against a pinned team manifest. A compression explained by a receipted lock entry is labelled receipted template compression (intentional, not drift).

`bernstein templates restore`¶

bernstein templates restore reviewer

Argument / flag	Default	Meaning
`ROLE`	-	Role whose templates to restore.
`--workdir DIR`	`.`	Project root the role templates are resolved from.

Restore reads the out-of-tree backups keyed by content hash, verifies every hash on the way in and the role directory digest on the way out, and records the reversal on the audit chain. The result is byte-identical to the pre-compression originals.