VMAFx/ffmpeg-patches

Caveman: internal register (caveman)

Adapted from Caveman by Julius Brussee and contributors; MIT terms at adaptation (2026-09-18).

Caveman = internal register of text-register policy (register: in .standards.yaml; "Text Register" section of AGENTS.md). Config value stays internal; this skill = its form. Agent traffic pays per token, per hop -> caveman line carries same facts, fewer tokens.

Scope

Caveman covers text another agent reads:

• briefs, task prompts to subagent or cheap lane;
• agent returns, workflow returns;
• research fan-out results;
• tool-call notes, progress lines between agents.

Human-facing text keeps its own register: forge text (issues, PR bodies, review comments, commit bodies, changelog titles) -> social-text; documentation tree (docs), READMEs, ADR bodies -> docs register; reply to human operator -> full prose.

Rules

1. Cut grammar with no fact.

   Drop	Examples
   articles	a, an, the
   pronouns	I, we, it
   copulas	is, are, was
   auxiliaries	have, will, did
   hedges	seems, probably, might
   politeness	please, thanks
   framing	I think, note that, it looks like, as requested

2. Fragments. One fact per line. Lists over sentences.

3. Symbols replace connective phrases:

   Symbol	Reads as	Example
   ->	causes, leads to, then	stale receipt -> push rejected
   =	is, equals, means	block = 15 lines
   x2	count, times	retry x3
   !	warning, risk	! gate not rerun after rebase
   ?	open question	? Windows path untested

4. Verbatim tokens. Code, paths, file:line, commands, flags, error text, ids, shas, version strings, numbers = copied exact, never abbreviated. Compression hits English around them, never them. Invented abbreviations (cfg ldr, ctx cmp) cost reader more decoding than they save; use full word or real identifier.

5. Start with answer. No restated brief, no step recap before result, no closing line. Verdict first, then evidence.

6. Tables only when more than 3 rows compare more than 2 fields. Otherwise: list.

7. Return shape follows register block: verdict, changed paths, commands run, evidence pointers, open questions.

8. Evidence follows register block, same wording: evidence above 58 lines or 1500 tokens -> file in the evidence directory the block names; return evidence: <path> sha256:<12 hex> lines:<n>, fetch only when decision needs it. Manifest may tighten both numbers; block in AGENTS.md prints values in force.

Static checker

praetorctl caveman check defaults to runtime message form. Exception: AGENTS.md + compiled vendor files at own path below --root (exact, case-sensitive) default to --kind=context, same verdict as gate. Explicit --kind wins. Structured traffic selects kind explicitly:

• brief: --kind=brief; fields = goal, inputs, return, evidence, task; goal first;
• return: --kind=return; fields = verdict, changed, ran, evidence, open; verdict first;
• policy text: --kind=context; C1-C8 heuristics only, uncovered skill rules reported advisory.

PASS summary lists mechanically checked and advisory skill-rule numbers. Runtime producer wiring separate; checker validates supplied text only.

Clarity floor

Caveman = shorter text, same information, never shorter text that makes reader guess. Before sending: check every line -> can reader act on it without asking back? When compression drops fact reader needs (which file, which count, new failure or pre-existing, check bypassed or not), put fact back. Keep negations explicit (no bypass, not rerun); dropped not inverts line.

Before and after

Real return from a praetor lane:

• Before: "First two pushes were rejected by the pre-push gate with 'state synchronization stale' because gate run wrote .standards-receipt.json after the last state sync; fixed by re-running state sync and pushing again. Not bypassed."

• After: push rejected x2: state stale (gate run wrote receipt after sync). fix: resync, push. no bypass.

Test return:

• Before: "I ran the tests for the repository with the race detector and all of them passed. go vet did not report any issues. The only file I changed was .agents/skills/caveman/SKILL.md, where I added an error for a missing end marker."

• After:

  verdict: pass
  changed: .agents/skills/caveman/SKILL.md (missing end marker -> error)
  ran: go test -race -count=1 ./... = pass; go vet = clean
  evidence: none
  open: none
  

Clarity floor in action:

• Before: "The dedupe scan reported two clones. Only one of them comes from this change; the other one is in .agents/skills/social-text/SKILL.md and was already present on main, so I left it alone."

• Too far: dedupe: 2 clones. Reader cannot tell whether change is blocked.
• After: dedupe scan: 2 clones. 1 new (this change), 1 pre-existing on main (.agents/skills/social-text/SKILL.md), left as is.