Compress

Phase 0 — Register and validate

0.1 Register session per session-protocol.md and verbose-progress.md. Generate SESSION_ID, log skill_start to .cc-sessions/activity-feed.jsonl.

0.2 Validate inputs. For each target:

• File exists and is readable.
• Extension is allowed (reject code/config files with a clear message).
• No existing <file>.original sibling (if present, it means the file was already compressed — skip with a notice, do not re-compress).
• File size ≤ 500KB (soft safety limit).

0.3 If any target is rejected, list the rejected files and continue with accepted ones. If zero targets remain, exit with a summary.

Phase 1 — Backup

1.1 For each accepted target, write a backup to <file>.original (preserving the original extension order: reference.md → reference.md.original). Use Read + Write. Never skip this step.

1.2 Verify the backup: its byte count must match the source. If not, abort with error.

Phase 2 — Compress

2.1 Read the target file. Apply the Terse Output Protocol at full intensity:

• Drop articles, fillers, pleasantries, hedging (per /_shared/terse-output.md).
• Rewrite verbose constructions to fragments where grammatical loss is acceptable.
• Preserve all elements in the "Preservation boundary" list verbatim: code fences, inline code, URLs, file paths, commands, grep patterns, YAML/JSON blocks, table cells that contain patterns or identifiers, headings (exact text), dates, version numbers, error codes.

2.2 Specific preservation rules, enforced mechanically:

• Every line starting with ``` in the original MUST appear unchanged in the output at a corresponding position.
• Every http(s)://... URL in the original MUST appear in the output (set equality).
• Every heading line (^#+ ) in the original MUST appear in the output with identical text.
• Every table row (lines starting with |) MUST remain a table row in the output; row count must match.
• Any text inside backticks ` must remain verbatim.
• Any value that looks like a grep pattern (contains regex metacharacters or is referenced as "pattern" / "regex" nearby) MUST remain verbatim — when in doubt, preserve.

2.3 If the source file contains any of these markers anywhere, ABORT compression with a clear explanation: "File contains exact-match agent prompt or grep-pattern content — classified UNSAFE, refusing to compress. See caveman-compress input-side research for classification."

• Literal string agent prompt template (heading or bold)
• Heading text Grep Patterns by Check
• Frontmatter field output_style: exact (opt-out signal)

2.4 Write the compressed content back to the target file path.

Phase 3 — Validate

3.1 Run the structural validator on the compressed pair:

bash hooks/scripts/reference-compression-validate.sh

3.2 If validation fails:

• Print the drift report from the validator.
• Restore the original: mv <file>.original <file> (delete the backup to return to pre-compress state).
• Mark this target as FAILED in the summary.
• Continue to the next target.

3.3 If validation passes, target is COMPLETE.

Phase 4 — Report

4.1 Write a summary to the user: