Skip to content

Agent context: slim CLAUDE.md, two skills, guardrails, latitude rule - #530

Merged
sam-aydlette merged 6 commits into
mainfrom
chore/agent-context
Nov 7, 2026
Merged

sam-aydlette merged 6 commits into
mainfrom
chore/agent-context

Conversation

@sam-aydlette

Copy link
Copy Markdown
Owner

SCN-Type: not-significant
SCN-Rationale: Agent working context, permission rules and context tests only; no change to how security posture is verified or published.

Changed

  • CLAUDE.md slimmed to the always-true rules (invariants, repo map, PR rules, a plan gate with a scope tripwire, how to respond). Procedures move into two project skills, site-content and compliance-pipeline, loaded only when a task needs them.
  • The Proof-of-concept latitude section is carried word for word, in the form amended on the rule-records stack (the looser Simulated criterion).
  • PR bodies start with SCN-Type / SCN-Rationale (and SCN-Record when not not-significant), and include an Effect section.
  • .claude/settings.json: deny rules for force-push, terraform destroy, aws s3 rm; ask rules for applies, make pipeline, pushes to main, gh pr merge, IAM and S3 writes.
  • .claude/hooks/stop-check.sh: a Stop hook that runs make check when the working tree has changes and blocks ending a turn while it fails.
  • website/sitemap.xml lists eight live pages it was missing.
  • New tests: the sitemap matches the pages, the homepage cards match the article list, and every path named in CLAUDE.md and the skills exists.
  • Stale references removed: scripts/staged/ (being removed on the rule-records stack); the nightly's write scope is described by its role policy, not a fixed prefix.

Effect

After deploy, sitemap.xml lists eight more pages. No /.well-known/ artifact, gate or record changes. Agent sessions in this repo load the new context, rules and Stop hook.

Verified

  • make check exit 0 on this branch.
  • CodeGuard over the diff: no credentials, account IDs or ARNs; the Stop hook reads no untrusted input and quotes its variables.

Residual / needs human

Couldn't verify

  • The Stop hook's behaviour inside a live Claude Code session in this repo (it was read, not exercised).

Follow-ups noticed

None.

🤖 Generated with Claude Code

sam-aydlette and others added 6 commits November 6, 2026 19:53
…ne skills

CLAUDE.md keeps the always-true rules: invariants, the repo map, PR rules,
and a plan gate with a scope tripwire in place of open-ended autonomy.
Procedures move into two project skills that load only when a task needs
them. Out-of-date references (ground-truth.md, the SSP, POA&M entries,
make targets that live in infrastructure/Makefile) are corrected.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EPMbZUHNXLjJxbN2HEPpVs
Articles 22, 24, 25 and 26, the contact page, and three research pages
(authorization-boundary, staying-in-the-loop, vulnerability-evidence) were
served but absent from sitemap.xml. lastmod is each article's index date;
for the contact and authorization-boundary pages it is the last commit date.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EPMbZUHNXLjJxbN2HEPpVs
.claude/settings.json denies force-pushes, terraform destroy and aws s3 rm,
and asks the operator before terraform apply, apply-infra/pipeline, pushes
to main, gh pr merge, IAM writes and S3 writes. Pushes to branches stay
unprompted, so small tasks still go straight to a PR.

.claude/hooks/stop-check.sh runs make check at the end of a turn when the
working tree has changes, and blocks the stop while it fails. Missing tools
produce a visible non-blocking notice instead of a block.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EPMbZUHNXLjJxbN2HEPpVs
Three hand-maintained copies had no check: sitemap.xml (eight live pages
missing), the homepage's Recent Articles cards, and the paths named in
CLAUDE.md and the skills (the old CLAUDE.md named a file that did not
exist). Each test was confirmed to fail on the drift it targets. CLAUDE.md
and the site-content skill now say what is enforced.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EPMbZUHNXLjJxbN2HEPpVs
Adds CLAUDE.md §6: lead with the outcome, keep verified separate from
assumed, push back before acting on a plan that looks wrong, and no
opening praise. Unprompted ideas are capped at three, each with why it
matters, and "none" is a valid answer; they are never built unasked.
When the operator asks for ideas, range widely.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EPMbZUHNXLjJxbN2HEPpVs
…d Effect in PR bodies

The slim CLAUDE.md keeps the operator's Proof-of-concept latitude section
word for word (the four statuses, the stand-in test, who approves stand-ins
and scope cuts, never choosing a status because a gate requires it), in the
form amended on the rule-records stack. The PR-body rule now starts with the
SCN lines the SCN gate reads and includes an Effect section. References to
scripts/staged/ go, since that directory is being removed, and the nightly's
write scope is described by its role policy rather than a fixed prefix.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@sam-aydlette
sam-aydlette merged commit 1322de6 into main Nov 7, 2026
8 checks passed
@sam-aydlette
sam-aydlette deleted the chore/agent-context branch November 7, 2026 01:51
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant