Local pages: a command-center page for lanes, PR waves, tokens, skills, manifests and owner items - #1012

Scope

• What this PR changes, in one or two sentences: adds command-center.html to the local pages, a nine-section view of sessions and lanes, orchestration, PR waves, the resolution landscape, SOTA skill invoke rates, tokens saved, the research queue, compounding manifests and owner items. It reads existing producers only, sixteen files through a new exact command_center role map in source_policy.json, and ships reviewed user units for the skill-usage reports plus token_tools_now.py, which runs each token tool's own report command. A missing input renders as a stated gap instead of stopping the refresh.
• Base commit: 900835c60ce9890cd704df4e98ded0b8aa91977d (main after Measure daily native skill use and test the security-audit trigger cue #970, which merged at 03:14:22Z on top of Add a local Claude review worker for essential pull request heads #953, Bound paper sell quote age independently of entry freshness #1015, docs: consolidate research convergence and close practice workstreams #1011, tools(landing): move the CC's landing scripts into the repository, with offline tests #975 and Regenerate the final-stack manifest after #1018 changed one of its sources #1025). The branch was rebased onto it at 03:23:42Z and replayed at 03:32:23Z (reflog). Measure daily native skill use and test the security-audit trigger cue #970 inserted a README paragraph directly above a sentence this PR changes, and both are kept verbatim. Every other change of the rebase is registry-only: the manifests/evidence.json entries for source_policy.json and tools/skill-usage/README.md, and Measure daily native skill use and test the security-audit trigger cue #970's new entry for tools/local-pages/README.md, take the measured value of each commit's file, and every commit's entries match its files (git range-diff against 51f85b5: commits 1-7 ! in those entries, and in commit 1 also the README context). Commits: 380dc3f, the page; 71b3338, the full landscape with the first Claude read's B1-B3; 7de947f, the idle-class units; 28ee5e1, the round-2 fixes; bc8ee15, the .target correction; 99777e2, the last wave's commit (Astra's first read of a737ef9, and the final Claude delta's wording); 47f064c, the Claude skill columns relabelled as calls, not sessions (Codex P2 at a737ef9, PRRT_kwDOUg_LrM6rSAfK); 3d7b1e7, the terminal commit under rule 4 (the 51f85b5 micro's B1, B2 and C, and Astra's two P1s at 51f85b5); 38302cd, the final commit under the CC's ruling on Astra's three P1s at 3d7b1e7; 66e8281, the fix commit under the CC's ruling on the GPT fix3 verdict at 38302cd and the Claude micro at 38302cd (below). Before this rebase the first seven were 33c9435, 40a11df, 3256b1d, 3614b78, 0c18ae0, d5bf784 and 51f85b5, on 791eab0.
• Lane: lane:foundation
• Owned paths touched: tools/local-pages/{command_center_view.py (new), token_tools_now.py (new), build_pages.py, sanitization.py, source_policy.py, source_policy.json, architecture_builder.py, architecture_view.py, assets/site.css, README.md, design-plan.md}, tools/skill-usage/{render_units.py (new), systemd/*.in and claude-lanes.jq (new), README.md}, manifests/evidence.json (three rehashed entries: source_policy.json, tools/skill-usage/README.md and tools/local-pages/README.md), tests/test_local_pages_command_center.py (new), tests/test_local_pages_token_tools.py (new), tests/test_skill_usage_units.py (new), tests/test_local_pages_sanitization.py, tests/test_local_pages.py, tests/test_local_pages_adoption_roles_boundary.py, tests/test_local_pages_architecture_read_boundary.py

At a scratch build of 38302cd (2026-10-12T06:57:31Z, tool clock; the same state root, rc 0), the nine sections read the same statuses as in this table: wired for landscape, manifests, orchestration, owner items and research, partial for PR waves, sessions, skills and tokens. At scratch builds of 66e8281 and of 38302cd from the same state root (generated 2026-10-12T13:21:55Z and 11:21:00Z, tool clock; rc 0 each) the nine sections read the same statuses at both trees: Sessions reads wired as well, as the state has drifted since 06:57Z, and PR waves, skills and tokens read partial. The per-row reasons above are from the earliest build.

Round 2, from the CC's adjudication of the Claude micro at 50977b0, adds one commit with six fixes, each with a test that fails at 2dfb3c8:

• An hcom roster that Fleet tracking could not read renders its gap with fleet_data's status and reason, so Sessions reads partial instead of wired (Codex :156).
• token_tools_now.py refuses a destination whose parent or any ancestor reaches the repository through a symlink (Codex :160).
• The count strip's saved-tokens figure is RTK's own, or UNKNOWN; another tool's report never stands in for it (Codex :1144).
• token_tools_now.py sanitizes dict keys as well as values, so its docstring's claim holds; the code was fixed rather than the claim narrowed.
• An encoded at sign inside a token over 4,096 characters is masked again, as b209afe masked %40 and @ there.
• A systemd template-unit name such as [email protected] is no longer read as an address. Since the round-1 mask, 9 such paths on 3 architecture layer pages had rendered as "[email omitted]"; the last part may not be one of the exempt unit types. Now 0 cells are masked and the 9 names read.

The final correction, from the CC's adjudication of the Claude micro at bdc0978, is one more commit. target is a delegated top-level domain as well as a systemd unit type, so it leaves the exempt list: an address there is masked again in its plain, %40 and @ forms. The other ten of systemd.unit(5)'s eleven types stay exempt. A class-guard test intersects the exempt list with the installed public suffix list and fails on any overlap, and the README and sanitizer name exactly the exempt types and the forms matched in place in long tokens. Both new tests fail at bdc0978.

Astra's first read of a737ef9 asked for changes: one P1 and two P2s, each reproduced through the real composer. They join the last wave's commit:

• P1: the shared email mask now takes the whole local part RFC 5322 §3.4.1 allows. That is a quoted string such as "reviewer"@example.org, whose quotes may be percent- or entity-encoded, or a run of atext. It reads the whole text with each token decoded, so a local part a quote, space or apostrophe splits is found whole, and it is anchored at the at sign. Corrections at later commits: the anchor alone did not make the cost linear, because the walk back over atext had no bound (the 51f85b5 micro's B1), so the terminal commit bounds it and replaces the first-letter-or-digit rule with any nonempty atext run (Astra's P1 1 at 51f85b5); and that atext set had left out the slash and always skipped leading backticks, so the final commit takes all 81 characters and sets aside only the three contexts listed below (Astra's P1 1 at 3d7b1e7). The pages and token-tools-now.json share it.
• P2: an in_review row that is not an object is rejected with a stated gap naming the count; PR waves reads partial, and the strip counts accepted rows only.
• P2: the number of pinned skills with no 7-day call counts only skills with a measured 7-day column: null and UNKNOWN when none is measured, and "N of M measured" when some are.
• The regressions use Astra's probe inputs and check the decoded page, receipt and token report for unique planted local-part and domain text, with no address grammar of their own. The 9 new or changed tests fail at a737ef9 (119 failures, counting subtests). Astra's own three probes fail there and pass at d5bf784. On 283,448 unique strings from the live inputs, catalogs, manifests and repository Markdown, the mask's output equals a737ef9's, and a full live build takes 43.2 s against 41.3 s.

Codex at a737ef9 (P2) found the Claude skill columns headed as sessions where child-usage.mjs counts Skill calls: it bumps skill_calls once per invocation (examples/claude-native/workflows/child-usage.mjs:3101) and sums them per child (:3158). 51f85b5 heads them "Claude calls in main sessions, 7 d" and "Claude calls in subagents, 7 d" in both tables, with the producer label and the README; the pinning header assertion fails at d5bf784.

The CC's rule-4 ruling on the 51f85b5 micro allowed one terminal commit, 3d7b1e7, which also folds in Astra's delta read of 51f85b5 (two P1s). Each item has a test that fails at 51f85b5:

• B1, cost: from each at sign, the walk back over atext stops after 256 characters (RFC 5321 §4.5.3.1.1 limits a local part to 64 octets) and never passes the address before it, so the cost is linear in the text's length. The three adversarial 50,000-character forms (%40 and &#64 written in atext, and apostrophe-joined %252540 tokens) take 0.034, 0.025 and 0.024 s, against 11.9, 10.8 and 5.0 s at 51f85b5, run back to back. A local part over 256 characters keeps its leading part visible.
• B2: overlapping or touching matches merge from their earliest start, so "w1 [email protected]"@d2.example.org is omitted whole.
• C, ruled a P1: the JSON inputs are parsed before masking, so a \" reaches the mask only inside a value that itself holds serialized JSON, or in the TSV queue. The mask now reads \" as a quote there, so \"w1 w2\"@d.example.org masks as its decoded form does. The fail-closed residue guard follows the mask: a value that still holds an at sign, raw or encoded (%40, %2540, @, @, \u0040), followed by a domain, directly after a word character or after a quote with only whitespace between, is dropped whole and reads [value omitted: an at sign beside a word remained after masking]. The command-center page states the count in its header and names the sections partial; other pages show the placeholder in place. No value is dropped on the live pages.
• Astra's P1 1: an empty quoted local part and a punctuation-only one, each on the target domain, are valid (CPython email._parseaddr, cited below) and were published. The mask now takes any nonempty atext run and quoted strings of 0 to 256 characters. Leading backticks were skipped as a Markdown code span instead of the first-letter-or-digit rule, so `@AGENTS.md` stayed; the final commit replaces that with a positively identified code span (below). Her inputs are regressions in plain, percent and entity forms, as keys and values.
• Astra's P1 2 (2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161): sources.html published the SHA-256 of private raw bytes: the seven command-center state inputs, cc-now.json and adoption-now.json. A digest of raw bytes would confirm a guessed masked value. The receipt marks those 9 inputs private and keeps their digests and byte counts. The composer's source notes and custody rows showed "private state input; this page withholds its digest, which the nonserved receipt keeps" (since the final commit they read "private, not published (host state)"), and no composer page shows their sizes. Repository inputs keep their SHA-256. The regression checks every page the composer builds, and the receipt outside the served root. Her read of 3d7b1e7 found the same class on the new page's footer and on the other state-root rows of sources.html, which the final commit withholds (below). The fix commit extends the rule to every composer page and to the architecture pages (below).

The final Claude delta at a737ef9 read PASS_WITH_P3. One claim in the correction's own lines was false at head, and the same commit qualifies it: the sanitizer docstring and the README now say an instance name that itself contains a dot still masks ([email protected] becomes [email omitted].service; Follow-ups, the bdc0978 micro's (a)5), and a test pins that behaviour. The same commit corrects the page's pending label: #1008 merged at 21:34:23Z without final-stack.json, so the label, from one constant, the README and the two pinning assertions read "pending #1007 (#1008 landed 2026-10-11 21:34:23Z)". #1007 has since landed final-stack.json, so the page now reads its terminal states and shows that label only for a checkout without the file.

The #1012 Claude read's B1-B3 are folded into this head. B1: the email mask lives in the shared sanitizer and matches each token after its decoding, so @, an encoded dot (%2E or .), %2540 and a semicolon-less &#64 are omitted on every page; the old page-local regex left all five readable. B2: each section's label follows what it renders, and every partial section names its reason. B3: each section and the count strip render in their own guard; a scalar codex_top_lanes_M and an integer above 1e308 degrade only their own section.

The CC's ruling on Astra's read of 3d7b1e7 (CHANGES_REQUESTED, three P1s; verdict sha256 e945d18d) allowed one more commit, 38302cd. Each item has a test that fails at 3d7b1e7:

• P1 1, valid addresses still published: the atext set had no slash and leading backticks were always skipped, so a lone slash, x/ and a lone backtick before an at sign stayed readable on command-center.html and as token-tools-now.json keys and values, also as / and ` (the percent forms were already masked at 3d7b1e7: %, the hex digits and F are atext). The walk now takes all 81 RFC 5322 atext characters (CPython's getatom accepts each). Exactly three contexts are set aside, each positively identified, and the sanitizer docstring and README list them (<at> stands for the at sign here, so this body writes no address): (1) inside a URL (a scheme, ://, and every character up to a space, an angle bracket or a quote) the slash ends the local part, so https://medium.com/<at>handle.name is a path, not an address, while an address after a path segment masks from its own first character; (2) an at sign right after a slash, with a name after it that a slash or an opening parenthesis follows, starts a path segment or a call (run_stream_events/<at>agent.tool/message_history, <at>DBOS.workflow()/<at>DBOS.step()), which keeps the existing run_events/…/history control, one string of a repository catalog (run_stream_events/… in catalogs/sota-convergence/manifest-20260923.json) and the DBOS pair in evidence/artifacts/sdk-runtime-coverage-20260922/durable-runtimes.json; the slash must be the character right before the at sign (RFC 5322 also allows a local part that ends in a slash, a cost Follow-ups records), so a natural address path segment masks whole with its directory prefix, gcloud's per-account legacy_credentials/<account address>/adc.json and x/<address>/y (a control and a pinned outcome in test_the_path_segment_exemption_needs_a_slash_right_before_the_at_sign, and two pairs in the composer test); (3) a backtick string that a backtick string of equal length closes after the at sign opens a Markdown code span (CommonMark 0.31.2) when it does not follow a letter or digit, a condition of this mask that CommonMark does not set and that masks more, so `@AGENTS.md` stays and a code span around an address keeps its backticks. Everything else is part of the local part: /<at>d, x/<at>d and a backtick that nothing closes are masked whole, and so is a host path with no scheme and nothing after it (medium.com/<at>handle.name). This corrects the lines above that said the slash is left out and that leading backticks open a code span, and the README and docstring claims at README.md:282 and sanitization.py:19-21 at 3d7b1e7.
• P1 2 (2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161), a digest of private bytes still published: the new page's footer served the SHA-256 of the native builder's render of the state root, whose raw bytes held a private address (Astra planted one in the gate owner and recomputed the digest), and sources.html served the raw digests of the gate receipt, the gap board, the roadmap inputs and 70 readiness receipts. The per-class policy of the 2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161 Follow-up now holds on the two pages: the composer marks an input private by provenance, when its path is under the state root (in the capture itself, so no name list can miss one), and the role projection's registry files the same way. The Command center and Sources pages print "private, not published (host state)" for each of them and for the native manifest in the footer; the nonserved receipt keeps every digest and private flag; repository inputs keep their SHA-256. The regression plants a run-time address in the gate owner, the gap owner, the roadmap index and the research queue, rebuilds the raw native manifest to show that the planted value is in it and that its digest is the receipt's, and checks that every digest on the two pages is that of a repository input and none is of host state. Scope per the ruling at the time: the two pages; the fix commit extends it to every page (below).
• P1 3, an address in a comment: two comments in sanitization.py and one test comment wrote out a complete address at a delegated domain (one of the two is a host path that the full atext grammar also reads as an address). They are prose now, and a test scans the local-pages and skill-usage sources and tests for any address that is not a no-reply form or on a name RFC 2606 and RFC 6761 reserve, with the sanitizer's own atext grammar and the systemd unit-type and code-span exemptions. The search the CC asked for over 51f85b5b..38302cd2, this body and the commit messages finds two more forms and no address: a string literal in tools/claude-review-worker/probes.py (from Add a local Claude review worker for essential pull request heads #953, outside this PR) where the escape of a newline precedes an AGENTS import, which the scanner's grammar reads as a one-character local part, and the pushed message of 3d7b1e7, which words an AGENTS import in apostrophes (history is not rewritten, and the squash body words it without the form). Everything else is a no-reply trailer or an example on a reserved name.

Astra's three P3s at 3d7b1e7 are in Follow-ups.

The CC's ruling on the GPT fix3 verdict at 38302cd (fix3-pr1012-38302cd2a/VERDICT.md, CHANGES_REQUESTED, two P1s) and the Claude micro at 38302cd (pr1012-claude-micro-38302cd2.md) allowed one fix commit, 66e8281: every page and output follows the privacy rule, the composer pages included (CC correction #186); any fallback that publishes undecoded text unchecked closes by omitting it fail-closed; the performance controls stay. Each behaviour change has a test that fails at 38302cd; the citation and wording items are text only:

• The verdict's BLOCKER 1 (2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161): no page carries the native readiness manifest digest. Every composer page and every Architecture page prints "private, not published (host state)" in its footer, as the Readiness heading does, and the build's run summary for the unit journal names the nonserved receipt that keeps the digest. Before publication each page passes sanitization.withhold_digests with the SHA-256 of every record the refresh read from the state root, so no page carries such a digest or a prefix of one 7 or more hex characters long, content-addressed names included (receipt-[digest prefix withheld].json; the micro's (d)3). The all-outputs regression rejects every 7-character prefix of the native manifest digest and of each private input digest on every output file.
• The micro's (c), the verdict's exact fix for BLOCKER 1: the Architecture pages follow the same provenance rule. The custody table, the host-receipt table, the layer headers and every other cell withhold the digest of each record read from the state root (the adoption snapshot and its content-addressed name, the retained refresh projection, the automation projection, the host-receipt index and the receipts it declares); the architecture receipt keeps them, and a detail page is named by the digest of the bytes it publishes.
• The verdict's BLOCKER 2: no token over 4,096 characters is published unchecked. The in-place mask still runs first without decoding (the cost bound). The token is then decoded with the same two rounds: one that an address touches once decoded, or whose decoded text turns the character beside an in-place mask into part of the address, is omitted whole as [email omitted], as a shorter token is; one whose decoded text, or the text it encodes as hex or base64, holds a home path, a session id, a task handle, an account URL or the home name reads [long token omitted: it decodes to private text], and the command-center page counts those tokens and names their sections partial. The regressions cover the verdict's four entity forms with a pinned-decoder control (CPython html.unescape and email._parseaddr at cbc944f4), partly entity-encoded and double-encoded local parts, a double-encoded at sign and dot, and the other private classes, at the sanitizer, the real composer and the token report's keys and values.
• The micro's (b), Codex's five threads at 38302cd: a run pin is that of the run the same-layer final-stack winner names in run, and a winner without one shows its run_reason (catalogs/landscape/final-stack.json#/definitions, winners[].run); a skill the Claude lanes report measured keeps those counts when the host report has no row for it; a current daily projection makes an absent or stale hourly report it replaces read "not used" instead of a gap; a read queue without a parseable updated_utc leaves Orchestration partial; each Adoption token-efficiency row is labelled by its case-folded name, RTK included. Each thread carries a reply that names 66e8281 and its regression, and all five are resolved.
• The verdict's follow-up 4: the two line citations of skill_usage.py in comments read :603, :615 and :652 at 900835c. The micro's (d)1 and (d)2: the docstring, comments and README say that RFC 5322 allows a local part ending in a slash, and that the code-span opener's letter-or-digit condition is this mask's own, stricter than CommonMark.
• The verdict's review_probes.py gives FAILED (failures=13) on an export of 38302cd and OK at 66e8281.

SOTA sources

• frontend-design, anthropics/skills at 33375500bcea98d610eb30ce10ac4e59b89c390d: https://github.com/anthropics/skills/tree/33375500bcea98d610eb30ce10ac4e59b89c390d/skills/frontend-design, SKILL.md SHA-256 d91970639e9f5c37682ac7ab60094d35f1c7c1f38d731bd56396563aee10c1d3 (matches adoption/skills/manifest.json). Used for the plan and critique: a compact count strip that doubles as the in-page index, one emphasized element (the PR register's next-gate column), no card kit; recorded in design-plan.md, "Command center".
• dataviz, the skill bundled with Claude Code 2.1.296: references/choosing-a-form.md (SHA-256 fbd8d8fb385670569c16cdcfe3d0e05281d0a33f892ef09a54b5c5780dc548e2), references/marks-and-anatomy.md (2d6a992d…), references/anti-patterns.md (8526a6c6…), references/interaction.md (2050b65e…), references/palette.md (96416245…), scripts/validate_palette.js (d6bcfe60…). Method: counts are figures, not charts; token use by lane is the only chart (one series, slot-1 hue, bars beside the numbers in the same table); flags are status pills that always carry their words. validate_palette.js "#2a78d6" --mode light --surface "#ffffff" and "#3987e5" --mode dark --surface "#17283b" print ALL CHECKS PASS (4.42:1 and 4.11:1).
• CPython at the v3.13.16 mirror (python/cpython at cbc944f4bc59639a444dd971c737788ba2283a91, the installed python3 version) for the B1 decoders the shared sanitizer already calls: Lib/html/__init__.py:118 (&(#[0-9]+;?, the semicolon is optional) and :122 (unescape), Lib/html/entities.py:567 ('commat;': '@') and :1736 ('period;': '.'), Lib/urllib/parse.py:694 (unquote).
• For the terminal commit: RFC 5321 §4.5.3.1.1 (https://www.rfc-editor.org/rfc/rfc5321#section-4.5.3.1.1), which limits a local part to 64 octets, for the 256-character bound; RFC 5322 §3.2.3, §3.2.4 and §3.4.1 (https://www.rfc-editor.org/rfc/rfc5322) for atext, quoted strings and the addr-spec; and CPython at the same mirror, Lib/email/_parseaddr.py getaddrspec (:365-385), getatom (:475-493), getquote (:463) and getdelimited (:426-461, where a quote right after the opening one ends the empty string at :449), which accept ""@d and !!!@d.
• For the final commit: RFC 5322 section 3.2.3 (https://www.rfc-editor.org/rfc/rfc5322#section-3.2.3), whose atext lists the slash and the backtick; CPython at the same mirror, Lib/email/_parseaddr.py getatom (:475-493), which Astra's probe ran for all 81 characters; the CommonMark specification 0.31.2 (commonmark/commonmark-spec, tag 0.31.2 = 9103e341a973013013bb1a80e13567007c5cef6f, spec.txt blob f1fab281e98b6006a62afcc59ed910ae4bc6741f, 205,025 bytes, fetched through the GitHub contents API at that commit), :5865-5872 (a backtick string is a run of backticks neither preceded nor followed by one; a code span begins with one and ends with a string of equal length) and :6072-6073 (a string that no string of equal length closes is literal backticks); and RFC 3986 section 3.1 (https://www.rfc-editor.org/rfc/rfc3986#section-3.1), scheme = ALPHA *( ALPHA / DIGIT / "+" / "-" / "." ), for the URL scan. The digest class rule has no new upstream: it is the CC's 2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161 ruling, Astra's executed reproduction (review_probes.py and review_static.py in her packet fix2-pr1012-3d7b1e7a9) and the code at 3d7b1e7 (build_pages.py:200, :365, :432, :588, :613); the hashing citation she lists under P3 stays open (Follow-ups).
• For the fix commit: CPython at the same mirror (cbc944f4, v3.13.16; the installed Lib/html/__init__.py and Lib/email/_parseaddr.py are byte-identical to it), Lib/html/__init__.py:118 and :122 and Lib/urllib/parse.py:694 for the two decoding rounds a long token now passes, Lib/email/_parseaddr.py getaddrspec (:365-385) for the pinned-parser control, and Lib/hashlib.py:33 (hexdigest(), hexadecimal digits only) and :58 (sha256 always supported) for the SHA-256 hex digests and prefixes the pages withhold; catalogs/landscape/final-stack.json at 900835c, #/definitions term winners[].run ("The runs[] id whose repository is the winner's ..., or null with run_reason ..."), for the run-pin rule. The review inputs are the GPT fix3 verdict and its review_probes.py (packet fix3-pr1012-38302cd2a) and the Claude micro at 38302cd; the digest rule itself remains the CC's 2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161 ruling.
• Token tools at their installed versions, commands and flags read from each --help on 2026-10-11: RTK 0.51.0 rtk gain --format json; Headroom 0.40.0 headroom agent-savings --check-perf --hours 24 --format json (exits 1 below the profile target) and headroom output-savings (its help: counterfactual unless an A/B holdout is set, HEADROOM_OUTPUT_HOLDOUT>0), run with HEADROOM_UPDATE_CHECK=off, HEADROOM_OFFLINE=1 and DO_NOT_TRACK=1; context-mode 1.0.169 doctor through the plugin's cli.bundle.mjs, whose savings are the MCP tool ctx_stats (bin/statusline.mjs:236-242 prints a fixed headline when no session exists, so it is not used).
• tools/skill-usage/skill_usage.py at 900835c for the skill semantics the page shows: codex.enabled is the manifest's codex_enabled flag while Codex reads are counted for every manifest skill (:603, :652), and Claude reads "not-measured" when /skill-doctor did not run (:615), so the page keeps those counts and leaves install state unknown.
• Measure daily native skill use and test the security-audit trigger cue #970's tools/skill-usage/daily_skill_usage.py, landed on main at 900835c with the same blob (02ddea79) as the branch pin d38ab63b763819eaeacc7d3ccb0cb5d94c73430e the page was written against: project_reports (:36) returns skill-invoke-rate-daily/1 with host rows, lanes and measured (:88-95), written to coordination/command-center/skills/skill-invoke-rate-latest.json (adoption/templates/systemd/skill-invoke-rate.service:13). The page reads only that output file under its grant and imports no Measure daily native skill use and test the security-audit trigger cue #970 code.
• Round 2 sources: fleet_data.py _tracking_hcom (lines 638-675 at 2dfb3c8) for the UNKNOWN roster shape the test builds; tools/skill-usage/render_units.py _external_path (lines 56-61) for the path-and-resolve() check destination() reuses; CPython at the v3.13.16 mirror (cbc944f4bc59), Lib/pathlib/_local.py:664-670, where Path.resolve(strict=False) is os.path.realpath and resolves every existing symlink; sanitization.py at 2dfb3c8, whose sanitize() maps dict values only (:115-122) and whose _encoded_token leaves tokens over 4,096 characters undecoded (:69); b209afe's page-local mask (command_center_view.py:38), which matched %40 and @ in place; systemd.unit(5) at systemd 259 (DESCRIPTION: service, socket, device, mount, automount, swap, target, path, timer, slice and scope) for the unit types a template-unit name ends in; and, for the final correction, the installed public suffix list (Debian publicsuffix 20260129.1928-1, /usr/share/publicsuffix/public_suffix_list.dat:10339-10341: "target : Target Domain Holdings, LLC", citing https://www.iana.org/domains/root/db/target.html), which the class guard reads.
• systemd 259 (259.5-0ubuntu3.4, the installed manager) manuals for the rendered units: systemd.syntax(7), systemd.unit(5) (%%), systemd.service(5) (the : ExecStart prefix), systemd.exec(5) (Nice= -20 to 19; IOSchedulingClass= realtime, best-effort or idle), systemd.resource-control(5) (CPUWeight=idle sets the cgroup v2 cpu.idle attribute), systemd.timer(5) and systemd-analyze calendar. The unit tests round-trip every ExecStart through those quoting rules and run systemd-analyze --user verify.
• jq 1.8.1 manual (the installed jq) for tools/skill-usage/systemd/claude-lanes.jq.
• Pattern reference on an unlanded branch, not a main pin: tools/starred-maintenance/render_units.py and its systemd/*.in templates on the Converge 373 public stars and schedule weekly review intake #978 branch codex/ns2604-starred-convergence-20261010 at 2b5c9abe0d7b0f4cf56d8502f887261a0c8cc851.
• Producers used unchanged at the base commit: tools/skill-usage/skill_usage.py, examples/claude-native/workflows/child-usage.mjs --lanes-sweep, the co-op hourly_usage_refresh.sh outputs (token-window-now.json, adoption-now.json), and source_policy.py ArchitectureReads, which CommandCenterReads extends.
• The meter element: WHATWG HTML Living Standard, https://html.spec.whatwg.org/multipage/form-elements.html#the-meter-element
• Test-time names scan follows the LANDING_PROHIBITED_NAMES convention from tools(landing): move the CC's landing scripts into the repository, with offline tests #975 (af9d19011, tests/test_landing_scripts.py).

Evidence-class table

Local commands run

$ TMPDIR=<scratch>/tmp python3 -E -s -B -m unittest <every tests/test_local_pages*.py, tests/test_skill_usage*.py and tests/test_north_star*.py module>
Ran 825 tests; OK (skipped=8)   # at 66e82815; 812 at 38302cd2
$ python3 -m unittest <the 15 new or changed tests>   # on an export of 38302cd2 with the five changed test files copied in
FAILED (failures=54, errors=4)   (rc 1); at 66e82815: OK (rc 0)
$ REVIEW_TREE=<tree> python3 -B -I review_probes.py   # the GPT fix3 packet's probes, copied out of the packet
FAILED (failures=13) at 38302cd2; OK at 66e82815
$ python3 scripts/validate.py
{"components": 70, "hashed_files": 11738, "profiles": 4, "receipts": 239, "status": "passed"}   (rc 0)
$ python3 scripts/evidence_manifest.py --check
{"files": 11738, "status": "passed"}   (rc 0)
$ gitleaks git --log-opts="900835c6..HEAD" --redact --no-banner .
10 commits scanned; no leaks found   (rc 0)   # gitleaks 8.30.1
$ git merge-tree --write-tree origin/main HEAD   # origin/main 429f0421
bd60388f893a5d09e67c07fac4a19ccebf93f34e   (rc 0)
$ python3 tools/local-pages/token_tools_now.py --out <scratch>/token-tools-now.json --context-mode-cli <plugin cache>/context-mode/1.0.169/cli.bundle.mjs   (rc 0)
$ python3 tools/skill-usage/render_units.py --repo <clone> --skills-dir <state>/coordination/command-center/skills --codex-root <home>/.codex/sessions --claude-root <home>/.claude/projects --output-dir <review dir>
$ systemd-analyze --user verify <review dir>/*   (rc 0)

Decision record

None: presentation, a reporter that runs existing tools' own commands, and units that schedule existing producers. The section-to-source map, the command_center roles, the optional cc-now.json projections, staleness limits, flags and the not-wired list are in tools/local-pages/README.md; unit cost and cadence are in tools/skill-usage/README.md.

Host evidence

Not applicable: no file under evidence/hosts/ changes.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a
  version comment (no floating tags). No workflow changes.
• New/changed workflows declare top-level permissions: {} and grant each job
  only what it needs (contents: read, or an explicitly justified addition). No workflow changes.
• No secrets are printed, logged or committed; no new required secret was
  added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted,
  moved or overwritten).

Follow-ups

• The squash body at landing carries the micro's s1-s4 edits (the owner-free wording of the landscape entry, in-place notes on the corrected claims, the percent-form correction and the complete 2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161 list) and this commit's message as its last entry.

• The residue guard's whitespace form covers a quoted local part only: an unquoted w @d and a zero-width-space at-sign form pass unchanged (0 live strings on the 43 pages). Follow-up for this lane: extend the guard to those forms, for example with the normalize-then-scan approach Scan the pull request description for home paths and credentials #991 adopts for PR bodies.

• C's JSON escaping cites no source: CPython Lib/json/encoder.py:21-23 at cbc944f4.

• The build pair "43.8 s against 44.6 s" is one run's record; a later reader's pairs had the head 4-9 s slower, so no direction is claimed.

• (a)5 from the 50977b0 micro: the last wave's claim that the at-sign anchor made the cost linear was false at d5bf784 and 51f85b5. The walk back over atext had no bound, so one 50,000-character a%40b.cc run took 11.9 s at 51f85b5 (the 51f85b5 micro's B1). The terminal commit bounds the walk at 256 characters: 0.034 s.

• (a)6: read_command_center_inputs catches (OSError, ValueError, UnicodeError) (build_pages.py:354), so a deeply nested valid JSON input (RecursionError) stops the refresh; add RecursionError to that except. No producer writes such nesting.

• (a)7: "in use" also counts Codex name mentions (SKILL_COLUMNS[1:]); 0 of the 29 live skills are in use by mentions alone.

• (a)8, with the shared identity filter (Codex :103): with HOME=/root, the shared sanitizer's identity filter masks the word "root" in generated labels. The root cause is that filter, which masks a home name such as "root" in every label on every page; it is outside this PR.

• (a)9 (Codex :146): tests/test_skill_usage_units.py skips verification only on bus errors, not on "Failed to initialize manager" where no user manager exists. Not reproduced here (P3).

• (a)11, carried from the first read: landings objects are counted but not listed; summary focus, now also .cc-more summary; print min-width, now also .cc-skill-matrix (64rem) and .cc-layers (72rem).

• (a)12: a wrong-typed optional sub-input takes down its whole section (Fleet tracking.lanes in Skills, Adoption orchestration in Orchestration). fleet_data emits lists, so no producer reaches it.

• From the bdc0978 micro:

  • (a)4: for fleet_data's tracking-level exception shape (fleet_data.py:1029-1033 at 3d7b1e7), the hcom gap reads "(not reported)" without that shape's UNKNOWN status and reason; a fallback to tracking.status and tracking.reason would carry them. Sessions still reads partial.
  • (a)5: a dotted template-unit instance such as [email protected] still masks as "[email omitted].service", because the pattern backtracks to the .conf label. 0 such names render. The README and docstring now say so, and a test pins it.
  • (a)6: a destination that is a symlink loop passes destination(), because non-strict realpath leaves the loop unresolved, so the tools run (about 8 s) before publish fails with rc 2. Nothing is written.
  • (a)7: the hcom test's UNREPORTED fixture has no hcom object, while build_pages's real adapter-failure tracking carries one ("(UNKNOWN): Fleet adapter unavailable").
  • (a)8, wording: a malformed reported roster (agents not a list) gaps as "(reported)"; _names_rtk matches "rtk" inside any word; a roster reported with rejected rows reads wired without its rejection reason, which fleet.html carries.
  • (b): the hostname appears on 4 architecture layer pages (24 occurrences: restic backup and prune unit names and host-receipt labels), unchanged by this PR; P3 under the owner's hostname ruling.
  • Its (a)3, the README's claim about forms in long tokens, is fixed in the final correction.

• context-mode 1.0.169 doctor requests https://registry.npmjs.org/context-mode/latest on every run, so each token_tools_now.py run makes one external request; the README's offline sentence covers Headroom only.

• From the first read: backslash (Windows-style) home paths for other accounts pass the shared sanitizer. Generate the canonical final-stack manifest (W4 PR 2) #1007 has landed, and the page reads its final-stack.json under the grant already in place.

• From Astra's first read of a737ef9 (follow-ups, carried items): the context-mode doctor's external version lookup; the tracking-level hcom reason (the bdc0978 micro's (a)4); the RecursionError and unwritable-destination boundaries ((a)6); the rendering and wording items ((a)7, (a)8, landings objects, dotted unit instances, focus and print); hostname hygiene (P3 under the standing ruling); and the citation captures she lists as unverifiable from her archive.

• From the 51f85b5 micro and the CC's rule-4 ruling (P3):

  • obsolete-syntax local parts: "w1".w2@d still leaves "w1" beside the placeholder, and w1."w2"@d leaves w1.; 0 corpus strings hold either;
  • the over-masked JSON-ish list item: ["a", "@AGENTS.md"] becomes ["a[email omitted]"]; privacy-safe, and 0 placeholders render on the live pages;
  • a quoted local part over 256 characters, and whitespace between a quoted local part and the at sign: the residue guard now drops both whole, and a test records both;
  • measured at 3d7b1e7 and corrected by the final commit: a local part made only of backticks stayed readable; it now follows the code-span rule, so an unclosed backtick is part of the local part and masks. The apostrophe-quoted AGENTS-import strings (no backtick) mask at both 3d7b1e7 and 38302cd, because the apostrophe is atext.

• From Astra's delta read of 51f85b5 (P3, her line numbers at 51f85b5):

  • the _skills return annotation (command_center_view.py:689) still declares int | None where :808 returns the summary dictionary;
  • the recovery and interpretation limits already carried above: RecursionError before the panel guards (build_pages.py:349), collection before some destination failures (token_tools_now.py:210), the tracking-level hcom reason (command_center_view.py:154), fewer landing representations listed than the strip accepts (:313), mentions as use (:715) and dotted unit names (sanitization.py:38);
  • evidence retention: the malformed-row and measured-skill corrections need a pinned source locator for their reporting convention;
  • hostname hygiene: 26 occurrences on 24 lines in unchanged parts of manifests/evidence.json and tools/local-pages/source_policy.json, and 0 account-name hits (P3 under the hostname ruling);
  • the missing captures and unexecuted host boundaries she inventories.

• 2026-09-24 community sweep: effort-guard SessionStart notice, working plugin-install commands, dated dispositions #161: the digest policy per input class is that repo-pinned integrity digests whose preimages are published stay, and digests of privately rendered state are withheld. The final commit applied it by provenance to command-center.html and sources.html; 66e8281 applies it to every page the refresh publishes, the composer's and the Architecture builder's, and to the build's run summary. The digests stay in the nonserved receipts and in the manifests/evidence.json or catalogs/north-star/readiness.json rows that already publish some of them. Not covered: a hex run longer than 64 characters or a digest inside a longer hex run is not read as a digest (0 such runs on the 45 live outputs); and a private digest is withheld only when the refresh records it, so a page that quotes a digest it did not read keeps it. The opposite cost: any run of 7 to 64 hex characters that equals the start of a private digest reads as withheld, also a number (the live outputs hold 1,347 distinct runs, 13 of them 7 characters long, and all 659 withheld prefixes are content-addressed names).

• Known limits of the final commit's contexts (measured, P3): the path-segment exemption, which the CC accepted on this condition, fires only when the character right before the at sign is a slash and the character right after the domain is a slash or an opening parenthesis. Its cost is that an address whose local part ends in a slash, with that continuation, stays readable (RFC 5322 allows such a local part; none is known in use here); the opposite cost is that an address inside a path masks together with its directory prefix outside a URL (legacy_credentials/<account address>/adc.json reads [email omitted]/adc.json, and ${USER_HOME}/.config/gcloud/legacy_credentials/<account address>/adc.json reads the same, because the placeholder and the directories are atext in front of the address), while inside a URL the slash ends the local part and the prefix stays; a scheme-less host path with nothing after it masks; two addresses joined by a slash merge into one placeholder; package URLs of the form pkg:pypi/NAME plus a version after an at sign (2 corpus strings, none rendered) mask from the ecosystem segment. Tokens over 4,096 characters are closed by 66e8281 (the GPT fix3 verdict's BLOCKER 2): each is decoded and checked after the in-place mask, and one whose decoded text still shows an address or another private value is omitted whole, so the long token's own prefix goes with it. A fully percent-encoded URL with an encoded at sign (https%3A%2F%2Fmedium.com%2F%40…) masks from its scheme, at 3d7b1e7 too: the decoded pass leaves it, then the raw pass reads %2F as atext (0 corpus strings, privacy-safe). A non-ASCII symbol before an at sign (RFC 6531) stays readable, while a letter or digit is dropped by the residue guard, as before. The residue guard is unchanged: the mask now classifies every atext predecessor itself, and the three contexts cannot be told apart by a predecessor-only pattern.

• From Astra's read of 3d7b1e7 (P3, her line numbers there): (4) source and capture evidence is incomplete (pr-body.md:53-:65 and :74-:88, tools/skill-usage/render_units.py:5, build_pages.py:366): the frontend-design commit mirror, the dataviz files, the RFC captures, the token tools' version and help captures, Converge 373 public stars and schedule weekly review intake #978's historical renderer and the live-page, browser and producer measurements are not retained in her archive, and the digest correction lacks a pinned hashing or publication citation; (5) adjacent recovery and interpretation limits (build_pages.py:349, :629, :664, token_tools_now.py:210, command_center_view.py:154, :313, :715, render_units.py:134): pre-render JSON recursion before the panel guards, destination failures after collection, the tracking-level hcom reason, fewer landing representations than the strip accepts, mentions as use, and per-file rather than whole-directory publication; (6) historical measurement and hostname hygiene: the terminal commit's older figures (pr-body.md:89) and 26 hostname occurrences on 24 lines in unchanged parts of manifests/evidence.json and tools/local-pages/source_policy.json, with 0 account-name hits.

• From the GPT fix3 verdict at 38302cd (P3): (3) source and capture gaps (its SOURCE-AUDIT.md): UNMIRRORED CommonMark, RFC and dataviz material, installed-tool help and live-page captures; the pinned hashing citation is now under SOTA sources (Lib/hashlib.py at cbc944f4); (5) the adjacent recovery and interpretation boundaries carried above; (6) hostname hygiene (26 occurrences on 24 unchanged lines) and the body and squash-copy capture identities it could not authenticate.

• From the Claude micro at 38302cd (P3): (d)4, the squash body's final entry gives one build pair without the no-direction note; (d)5, no planted live-state build this round, since the source policy refuses the override route (the composer route is covered by the fixture tests through refresh()); (d)6, the body's dated status snapshot, now with the 66e8281 build beside it.

• Research queue (rule 8, from the CC's rule-4 ruling): one maintained, cited email detector for publication scans and page masking, shared with Refuse email addresses in the publication scan #1002. Candidates: the WHATWG HTML "valid email address" grammar, CPython email._parseaddr as a validator, Presidio's EmailRecognizer and the email-validator package. The chosen detector replaces both hand-rolled implementations.

• The CC installs the two rendered unit pairs after landing (re-rendered against the live clone); until then the hourly reports read stale and the daily Codex-lanes report is absent. The first daily run's duration should be read from the journal, because its three-hour limit is a ceiling, not a measurement.

• Measure daily native skill use and test the security-audit trigger cue #970 has landed (900835c). Once the CC deploys its daily timer, the page uses its skill-invoke-rate-latest.json while the file is under 30 hours old; the grant is in place.

• tools/token-report: its manifest is outside the page roots, and its README keeps raw reports private; wiring needs a separately reviewed sanitized receipt.

• Fold the legacy cc_status_page.py panels into this page (CC follow-up).

• Firefox rendering, including the ::-moz-meter-bar style, is untested.

• Outside this PR: cc-usage-refresh.service, which produces token-window-now.json and adoption-now.json, failed at 18:01:49Z on the CC's northstar_page.py:1461 (TypeError on dict rows in cc_agents.running).

🤖 Generated with Claude Code