Claude Code practice for Opus 5.5: 2026-09-27 review resolved, host read-back, weekly reference freshness lane - #453

…d, cacheTtl erratum, PERM-03 correction

- DS-3 and DISP-01: superseded model/effort text marked in place (max-effort
  default 2026-09-23; claude-harness-settings item 1), original rule quoted,
  following the WH-3 precedent.
- CT-7: timing half primary-endorsed (prompt-caching docs) and adopted as
  operator guidance; the constraint-loss claim keeps its test.
- ORCH-03: any CLAUDE_CODE_EFFORT_LEVEL value overrides child and stage
  effort; points to the complete never-set rule.
- Key-name gap and utr-cache-ttl: experimental.cacheTtl erratum
  (anthropics/claude-code@7779afb CHANGELOG 2.1.248; sub-agents,
  prompt-caching and workflows docs).
- PERM-03: dated correction of the stall rationale (headless and
  permission-modes docs); the rule stands.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

… cleanup record

The project CLAUDE.md gains the documented "Compact Instructions" section
(code.claude.com/docs/en/how-claude-code-works; memory: project-root
CLAUDE.md is re-read after /compact; the 2.1.283 summarizer prompt carries
the heading). The 2026-09-26 cleanup record's scope sentence is amended by a
dated addendum that records the precondition audit of both installed
PreCompact hooks (each prints {} and exits 0), the scope of the "only
web-documented way" claim (in-product /hooks text documents PreCompact
stdout) and the timing caveat for running sessions and #381 E2E runs.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…th measured sizes

The memory and features-overview docs state a 200-line target per
CLAUDE.md as guidance; the MI-7 rejection covers enforcing a line budget as
a rule. Lines and bytes of every always-loaded file and the SubagentStart
carrier are recorded after the Compact Instructions section landed (wc -l -c;
host copies named by their generic location only). Budgets count bytes or
tokens per A13 of the 2026-09-24 sweep.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…ers, proxy-arm control

- foundation-stack: behind a non-first-party ANTHROPIC_BASE_URL, MCP tool
  search is off and tools load into the prefix; cache invalidation scoped
  per the prompt-caching table (connect, deliberate disable or deny
  invalidate; an unexpected disconnect keeps it since 2.1.281,
  anthropics/claude-code@7779afb CHANGELOG.md:207); ENABLE_TOOL_SEARCH and
  CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS per the mcp and env-vars docs.
- token-practice: counters row for the /usage "Prompt cache (main)" line
  and the status-line prompt_cache object (costs, statusline docs; 2.1.251+,
  likely cause 2.1.260+; main conversation only), with child alternatives
  (child-usage.mjs, OTel claude_code.token.usage by query_source).
- token-practice: proxy-arm control for measured comparisons (Headroom
  v0.27.0 study scope from the 2026-09-26 landscape sweep; hold tool search
  and TTL variables fixed; report cache creation and reads per arm).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…integrity brief clause, background-wait ceiling, SOTA references

- Usage accounting: advisor calls are usage.iterations[] advisor_message
  entries outside top-level usage (platform advisor-tool docs, usage and
  billing), subagents inherit the configured advisor (advisor docs), and
  child-usage.mjs reads only top-level counters; complete usage from /usage;
  hold the advisor state equal across arms (CLAUDE_CODE_DISABLE_ADVISOR_TOOL).
- New receipt evidence/receipts/claude-advisor-usage-scan-20260928.json:
  count-only scan of this host's transcripts (local observation), with its
  script, hash, synthetic-fixture control and scan time.
- Brief contract: writers must not delete, skip, weaken, rewrite or
  special-case a test to pass it; report a wrong test with its failing output
  (platform prompting best practices, "Avoid focusing on passing tests").
- Adopt step 3: CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS for claude -p runs
  (env-vars docs; Claude Code 2.1.182 or later).
- SOTA references section pointing to the 2026-09-27 review record.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…w section, dated pointers

docs/decisions/2026-09-28-community-sweep.md merges both syntheses of the Opus 5.5
Claude Code practice review: applied items, amendments to M1-M3, M6-M9 and M14,
keep-but-compare M15-M52, rejections R41-R79, consensus maps, closures, the 22
community pins with full SHAs, primary sources, not-covered lists, deferred,
handed-off and host-scope items, and overturn conditions.

docs/community-native-practice.md gains the 2026-09-27 review table, a dated
checkpoint-limits correction to the Recovery row and a dated claude-hud note.
The 2026-09-24 sweep and the max-effort record gain dated pointers.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…agents skill

The section now asks for each modified file's branch or worktree rather than the
file list, which the 2.1.283 built-in summarizer prompt already requests
("examined, modified, or created"), and names each open claim's provenance. The
cleanup addendum records the check, and MI-7 carries the new CLAUDE.md size.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…er (#447), CT-7 cache reason scoped

MI-7 now gives each SubagentStart carrier block measured at f508ffb, notes that native acceptance covers Agent-tool children only, re-measures MEMORY.md, drops the unsourced imported-files clause and re-measures after AN-15 (review findings F1, F5). CT-7 keeps the natural-break endorsement and ties the cache-lifetime point only to compacting before stepping away (F8).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

… packet; compaction reasons separated; child-model roles

Fast mode bills usage credits on subscriptions rather than drawing only on them; the dev-container sentence becomes a pointer; plan mode's unenforced block covers an attempted edit (review finding F5). The cache-lifetime reason now attaches only to compacting before stepping away, with the Opus 5.5 blog cited (F8). The child-model parenthetical names Opus for every shipped role but source-scout and dates only the isolated-builder and stack-verifier move (F10).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

… their packet items

The gateway caveat drops the proxy-forwarding cause, the request-failure claim and the stays-off claim, and names tool_reference through the proxy-arm control. The cache-counter row drops the field list, the query_source values and the OTEL_LOG_TOOL_DETAILS clause (review finding F5).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…s the unverified quote

The brief-contract clause now says what the writer must not do and what the writer reports, and links the prompting best-practices section without a quote the item packet does not carry (review findings F10, F5).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

… complete inventory, audit run conditions

The record notes the rebase onto f508ffb, hands AN-15 off to the token-lanes owner with its two carrier lines, and adds a carrier status line (review finding F1). The 2026-09-24 tests and adoption rules still apply except M7's automatic lowering and M9's reset wait, and M1 and M6 quote their standing rules (F2, F3). A10 is handed off, terminal_bell is host scope and AN-03 names the practice page (F4). The -p JSON rows say the array shape comes from the binary read and the fixture (F6). The PreCompact audit carries its scratch-state conditions in the record and the cleanup addendum (F7). M47 adds only the missing mode flag (F9). The CLAUDE.md timing note names the omitClaudeMd roles instead of every non-blind child.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

The carrier sentence names the child's start context as the place the role-matched block is added (review finding F1 wording).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…E_TOOL_SEARCH behaviour

The overturn condition names a gateway shown to forward tool_reference blocks, matching the trimmed gateway caveat and the proxy-arm control (review finding F5).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…references

catalogs/foundation/practice-references.json lists the 22 community repositories the
sweep surveyed and the three upstream Anthropic repositories its syntheses cited as
primary sources (anthropics/claude-code CHANGELOG.md, anthropics/claude-code-action
v1.0.235, anthropics/claude-agent-sdk-python), each with its full-SHA pin, the pin's UTC
committer date, stars at check (discovery metadata only), the dimensions it informed and
the wave's decision record path.

Sources: pins from the sweep's reference list, each resolved with
`gh api repos/{repository}/commits/{pin}` on 2026-09-28 (all 25 resolve); used_for texts
verbatim from the two syntheses' reference lists; dimension topics verbatim from their
consensus maps; claude-code-action tags v1.0.235 and v1.0.233 read from its tag list.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

… weekly workflow

tools/sota-convergence/practice_references.py reports, per catalog repository through gh:
archived; default-branch commits since the pin and the head SHA (GitHub REST compare,
github/rest-api-description@c6721f32 commit-comparison schema); stale per the OpenSSF
Scorecard Maintained check, archived or no default-branch commit in the last 90 days
(ossf/scorecard@8788fc28 checks/evaluation/maintained.go:31,88-89,
probes/hasRecentCommits/impl.go:52-57); renamed (repos/get 301) or missing. It imports
fetch_repository, gh_api and is_expected_missing from github_freshness.py and
unique_json from scripts/catalog_decisions.py. Exit 0 for any well-formed catalog;
2 only for a malformed catalog or an invalid argument, and then no report is written.

.github/workflows/practice-references-freshness.yml runs it weekly (Thursday 06:41 UTC,
off the hour and away from the Monday lanes' shared GITHUB_TOKEN budget) and on dispatch,
with catalog-freshness.yml's conventions and exact action pins; it writes the Markdown
table to the job summary and uploads the JSON report. The job fails only when the
checker's offline tests fail or the catalog is malformed.

tests/test_practice_references.py covers the fresh, drifted, stale-by-90-days, archived,
renamed, missing, unresolved-pin and malformed-catalog cases offline, including observe()
against a fake gh. tests/test_workflow_security_coverage.py lists the new workflow.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…cted security-review pin

anthropics/claude-code gains the main synthesis's use of the same pin, verbatim from the
sweep's references_main: "version facts 2.1.32 to 2.1.283 (prompt-audit, Opus 5.5
default, effort defaults, cacheTtl, summarize)" (ref
https://github.com/anthropics/claude-code/blob/7779afb12e36/CHANGELOG.md). The earlier
extraction matched only the owner/name@sha form and skipped this blob URL.

anthropics/claude-code-security-review@0c6a49f joins as a
primary reference: the supplement synthesis cites it by pin, used for "Rejected:
README.md:45 warns it is not hardened against prompt injection.", and its rejection rests
partly on staleness ("Stale: the last commit is 2026-02-11"), so the weekly lane now
watches it. Read 2026-09-28 with gh: committer date 2026-02-11T18:01:23Z (pin_date
2026-02-11), stargazers_count 6276, the default-branch head equals the pin, and
README.md:45 at the pin says the action "is not hardened against prompt injection
attacks and should only be used to review trusted PRs".

The scope sentence now says four Anthropic repositories and names the one pinned source
left out on purpose: git/git at v2.43.0 is a git source tag, not a Claude Code practice
reference. The committed-catalog test asserts 26/22/4 and both claude-code dimensions.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…must exist, decision note

Classification. is_expected_missing() tested "404", "not found" and "429" anywhere in the
error text, and gh_api() returns a timeout as str(subprocess.TimeoutExpired), which quotes
the command's path with both compared SHAs. A 60 s compare timeout whose SHA contained
"404" was reported pin_unresolved, a "No commit found for SHA: ... (HTTP 422)" whose SHA
contained "429" was reported fetch_error, and a repository timeout whose slug contained
"404" was reported missing. gh_http_status() now reads the status only from gh's own error
line, "gh: <message> (HTTP <status>)" or "gh: HTTP <status>"
(cli/cli@0cf1092, tag v2.101.0, pkg/cmd/api/api.go:684-685
and :553-557). missing is a 404 on repos/{slug}; pin_unresolved is a 404 or 422 on the
comparison (github/rest-api-description@c6721f32 api.github.com.json:55474-55483 document
404 and 422 beside 500 and 503). Everything else, timeouts included, is fetch_error.
The regression tests failed against the previous code and pass now. Their gh strings are
the ones gh 2.101.0 returned on 2026-09-28: "gh: Not Found (HTTP 404)" for a missing
repository and for an unknown SHA on either side of a comparison, and "gh: No commit found
for SHA: <sha> (HTTP 422)" for commits/<unknown sha>.

Records. record_issues() requires each entry's record to be an existing file confined to
the repository root: scripts/catalog_decisions.py's safe_file() (no symlink, no escape),
the rule scripts/landscape.py's single_lane_decision_path_issue() applies to a
single-lane record. main() applies it under --root, which defaults to the checkout. A
missing record is a malformed catalog: exit 2 and no report. validate_catalog() stays
pure, and the committed-catalog unit test stays format-only, because the wave's record,
docs/decisions/2026-09-28-community-sweep.md, lands with the practice unit.

Decision note. The module docstring now records the 2026-09-28 choice, the alternatives
compared and the overturn condition. The alternatives are running the OpenSSF Scorecard
Maintained check itself (ossf/scorecard@8788fc28 README.md:525-528, :387, :181-184;
docs/checks.md:404-408), adding the repositories to the Monday catalog-freshness lane
(fetch_repository() compares no pin) and having no scheduled check.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…d-records test

The practice record and the reference catalog land together, so the committed
catalog's records now exist and are asserted (load_catalog with root). The
github-automation and foundation catalog indexes name the new weekly
report-only lane.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

… drift, handoffs; RTK five exclusions

The approved host-scope items (AN-17, AN-18 and AN-19 host halves, AN-26,
A09) and the install-audit gaps are recorded as applied, each with its
read-back. The jCodeMunch user-scope drift against the 2026-09-25 addendum is
recorded with its two options, owner decision pending. The SOTA-sources line
is folded into #444, AN-13 stays open, and AN-15 lands as its own change.
token-practice.md now counts the recipe's five exclude_commands entries.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

A repository, pin or record ending in a newline passed validation, because
Python's "$" also matches just before a trailing newline (Codex review).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…t; consistency fixes

The host-scope items were stated as applied and read back without retained
evidence (Codex review). A value-free read-back receipt now records them on
this host: the deny set and two bypass-mode claude -p probes (git clean -fn
denied; a force-with-lease dry run allowed), the ConfigChange log, the
status-line export with a negative control on claude-hud 0.8.0's diff step,
the bell, Headroom's no-egress env, agent-browser and QMD. Its live -p runs
also returned the array shape. M31 no longer lists open-code-review, which
M42 excludes, and the notifications row records A09 as applied.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…-26 and verbose checked

The Codex re-check found the host claims still wider than the receipt: the
ConfigChange behaviour and the install actions were stated but not observed,
AN-26 and verbose were unchecked, and refreshInterval was kept as a raw value.
The receipt is re-recorded with booleans for refreshInterval, verbose and the
two user CLAUDE.md phrases, object-only log-line parsing and Headroom's scope,
and the record now separates the operator's account from what is read back.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…ceipts

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>