vetting-claude-internals
Skill2 starsChanged 46 days agoApache-2.0View on GitHub
Use when `claude --version` differs from the version recorded on any CONFIRMED spike in notes/spikes/, or before merging design work that depends on a spike's claim. Re-runs the documented procedures, diffs against prior results, and updates the spike files honestly.
In this repositoryOne other file it gives its agents.
- codex-workerSkill
The file
We split this skill into the rules it gives an agent
9 rules. About 0 are common, 5 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 15
When the running claude --version is newer than the stamp on any CONFIRMED spike, re-vet that spike before relying on the claim.
- Very specific: line 32
Run claude --version and record the result.
- Very specific: line 33
Read every file under notes/spikes/.
- Very specific: line 50
Behavior changed but design still holds: keep Status CONFIRMED, record the new observation under ## Result (do not overwrite the prior …
- Distinctive: line 57
Update the design doc's "Empirical anchors (verified)" section if any anchor moved.
- Distinctive: line 59
If any spike flipped to REFUTED:
- Distinctive: line 63
Stop downstream work that depended on the refuted claim until the replacement is itself CONFIRMED.
- Distinctive: line 72
Never silently downgrade a Status: Always record what changed and why.
- Distinctive: line 76
If the user did not actually re-execute the engine-touching steps, say so: "filesystem-only re-verification.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Vetting Claude Code internals
eigenform's design rests on a small set of empirically verified claims about
Claude Code's internals — the JSONL format, --fork-session semantics,
the billing flip, cold-load behavior, the bundled-skill gap, the plugin
layouts on disk. Each is recorded in notes/spikes/<NN>-<topic>.md with
a claude version: stamp and a Status: (CONFIRMED / PENDING /
REFUTED / INCONCLUSIVE).
When the running claude --version is newer than the stamp on any
CONFIRMED spike, re-vet that spike before relying on the claim. When
about to merge work that depends on a CONFIRMED spike, re-vet it even
if the version hasn't moved.
When to invoke
claude --versionno longer matches the version on a CONFIRMED spike.- About to merge or ship work that depends on a CONFIRMED spike (e.g. surgery changes that depend on spike 03, daemon work that depends on spike 01).
- Periodic audit on a long-lived branch — spikes can decay even at the same version because cache behavior, plugin layout, or filesystem conventions shifted on the server side.
Procedure
- Run
claude --versionand record the result. - Read every file under
notes/spikes/. Build a worklist of spikes whoseStatus:is CONFIRMED and whoseclaude version:stamp is older than the running version. - Also worklist any spike on which the current task directly depends, regardless of version drift.
- For each spike in the worklist, re-execute its
Proceduresection verbatim:- Engine-touching steps (
claude --fork-session,claude --resume,claude -p, SDK calls) require explicit user authorization for this specific run. Do not run them silently. The feedback rule "No unauthorized engine invocations" is load-bearing here. - Filesystem-only steps (
ls,find,cat, parsing JSONL withjq/python) run freely.
- Engine-touching steps (
- Record the outcome in the spike file:
- No change observed: bump the
claude version:stamp to the current version, add a single line under## Result:RE-VETTED <ISO-date> on <new-version>: unchanged. - Behavior changed but design still holds: keep Status CONFIRMED,
record the new observation under
## Result(do not overwrite the prior one), append an## Implicationparagraph naming the shift. - Behavior changed and the design assumption is now wrong: flip
Status to REFUTED, document the new behavior, and add an
## Implicationparagraph naming which parts ofdocs/plans/2026-06-02-eigen-foundation-design.mdneed to change.
- No change observed: bump the
- Update the design doc's "Empirical anchors (verified)" section if any anchor moved.
- If any spike flipped to REFUTED:
- Open a new spike note proposing a replacement claim and procedure.
- Surface the implication explicitly in the end-of-turn report — name the design rules that need re-examination, do not bury this.
- Stop downstream work that depended on the refuted claim until the replacement is itself CONFIRMED.
Hard rules
- Never run
claude -porclaude --printor SDK calls for spike work without explicit per-invocation user authorization. The point of spikes is to verify behavior the user cares about; quietly burning credits to verify them is worse than asking. - Never silently downgrade a Status. Always record what changed and why. A spike that quietly flips from CONFIRMED to "well, sort of" is worse than one whose change is loud.
- A spike that re-passes by accident is worse than one not re-run. If the user did not actually re-execute the engine-touching steps, say so: "filesystem-only re-verification; engine steps not re-run." False confidence is the failure mode.
- Read all of
notes/spikes/before declaring vetting complete. A spike you forgot to re-read is a spike you implicitly assumed away.
Related
notes/spikes/README.md— spike format and conventions.docs/plans/2026-06-02-eigen-foundation-design.md— the design these spikes ground.- Feedback memory: "No unauthorized engine invocations" — load-bearing for step 4.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

