Keep an agent-readable docs/ folder in sync with OneDrive, SharePoint, Outlook and Teams by processing only what changed.
A since-token per source says what changed. A manifest with three hashes decides, before a single file byte is read,
whether that change costs anything. Git carries the result, so "what changed since Tuesday" is git log.
Note
Status: implemented. The design's build order (§10, weeks
1–3) is built as src/agentsync/, the command-line tool agentsync. It has the local walk and
inbox arms, the SQLite manifest with the H0/H1/H2 classifier, converters with a write-once cache, docs/mirror/ with
one git commit per cycle, the Graph drive, mail and Teams arms, the curation map with its STALE lint, and a
single-writer LaunchAgent. It was built and measured on this Mac, including runs against a live OneDrive File
Provider mount: a no-op cycle over the live OneDrive source fell from 12.9 s to 0.51 s in commit 209bc80.
- Works with zero IT involvement: a
localsource over the OneDrive sync folder, plus a manualinbox(day 1). Install needs no admin rights. - Needs the tenant: the Graph arms (libraries you do not sync, Outlook folders, Teams channels and chats)
need an Entra app registration and admin consent from IT (
docs/deploy/it-request.md). Five probes can still only be measured on the target tenant (what is measured). After sign-in,scripts/tenant-probes.shruns probes 1, 2, 4b and 9, and probe 8 is run by hand. - Rollout: the dated rollout names an owner and a proof command for each step. Its readiness table lists what is still open.
CORRECTED (2026-09-29): this note said "Status: a design with measured probes. There is no implementation yet. The build order is §10 of the design. It starts once a target corporate tenant is chosen, because five of the probes can only be measured there." The hero line said "a design with measured probes, not yet an implementation". The implementation landed on 2026-09-29, and it did not wait for the tenant. The arms that need no tenant run today, and the Graph arms wait only on the IT request.
Copy this block into Claude Code, GitHub Copilot CLI or any coding agent that can run shell commands on the Mac. It installs, asks you which folders to sync, waits while you click Allow on one macOS prompt, verifies every step, and writes the IT request as a draft file that it never sends. You fill in the IT contact and send it yourself. Doing it by hand instead: Install.
Set up agentsync on this Mac. agentsync keeps a local, agent-readable git repo (~/agent-context/docs) in sync with
the OneDrive and SharePoint folders this Mac syncs. Source: https://github.com/renchris/agent-context-sync
(docs/deploy/README.md there explains every step). Run each command yourself and show me its output.
Rules: no sudo; never push, upload or email anything; do not edit my shell profile; do not change Keychain, MDM,
System Settings or privacy (TCC) settings; do not delete, reset or stash anything; do not open or read the files
inside my OneDrive folders (listing folder names is fine). If a command fails and this prompt does not say what to
do, show me the output and stop. If your tool refuses to run a command, show it to me and I will run it myself.
1. Preflight. Run `sw_vers` and `xcode-select -p`. If xcode-select prints no path, stop and tell me: "Install the
Xcode Command Line Tools with `xcode-select --install`, or request them from IT through Self Service if that asks
for an admin password; then paste this prompt again." Run
`git ls-remote https://github.com/renchris/agent-context-sync.git HEAD`; if it fails, show me the error (a
corporate proxy may need HTTPS_PROXY set) and stop.
2. Get the code. If ~/src/agent-context-sync exists, run `git -C ~/src/agent-context-sync pull --ff-only`;
otherwise run `git clone https://github.com/renchris/agent-context-sync.git ~/src/agent-context-sync`. Then run
`~/src/agent-context-sync/scripts/install.sh --help | grep -c -- --source-local`; if it prints 0, stop and tell
me the published installer is older than this prompt.
3. Choose what to sync. Tell me that macOS may ask once whether this terminal app can access files managed by
OneDrive, and that I should click Allow. Run `ls -d ~/Library/CloudStorage/*/*/ ~/Library/CloudStorage/*/*/*/`.
If it prints "no matches found" or "No such file or directory", stop: OneDrive is not signed in on this Mac.
Show me the list and ask which folders to sync, suggesting project folders rather than a whole library. Online-
only files in them are downloaded up to 1 GiB per folder per sync, so large folders take several syncs.
4. Install (no admin rights). Tell me: "macOS may now ask whether agentsync-launcher may access files managed by
OneDrive. Click Allow." Run `~/src/agent-context-sync/scripts/install.sh --source-local "<folder>"`, one
--source-local per folder I chose, each a full path starting with $HOME/Library/CloudStorage/. It is safe to
re-run. If it exits non-zero, show me the output and stop. Ignore its "next:" and PATH hints: this prompt
covers the next steps, and every command below uses the full path ~/.local/bin/agentsync.
5. Check. Run `~/.local/bin/agentsync doctor`. These [warn] lines are expected and need nothing now:
launcher.signature and launcher.requirement (ad-hoc signature), launchd.poll and launchd.reconcile (not installed
until step 7). If a tcc line says TCC_PENDING, ask me to click Allow, then re-run doctor. For any other [FAIL],
apply its fix only if it is an agentsync or install.sh command, then re-run doctor; show me anything that remains.
6. First sync by hand. Run `~/.local/bin/agentsync sync --once`, then `~/.local/bin/agentsync status` and
`git -C ~/agent-context/docs log --oneline -3`. It worked if sync exits 0, status shows "baseline complete" and a
"last success" time for every folder, and the log shows a "sync:" commit. Otherwise show me and stop.
7. Background sync. Tell me: "macOS may ask again whether agentsync-launcher may access files managed by OneDrive.
Click Allow." Run `~/src/agent-context-sync/scripts/install.sh --confirm-install-agent`; if it exits non-zero,
show me the output and stop. Wait until I say I clicked Allow or that no prompt appeared. Then run
`launchctl kickstart -k gui/$(id -u)/com.agentsync.poll`, wait 60 seconds, and run
`launchctl print gui/$(id -u)/com.agentsync.poll | grep -E "runs|last exit code"` and
`~/.local/bin/agentsync status`. Done means "last exit code = 0" and no TCC_PENDING line in status. If the exit
code is 79 or status shows TCC_PENDING, ask me to click Allow (or to turn on agentsync-launcher under System
Settings > Privacy & Security > Files and Folders), then repeat the kickstart check.
8. IT request, for Outlook, Teams and SharePoint sites this Mac does not sync. Read
~/src/agent-context-sync/docs/deploy/it-request.md and write an email draft from it to
~/agent-context/it-request-draft.md (not inside the repo, which you must not change). Fill <name> with the output
of `id -F`, <serial> from `system_profiler SPHardwareDataType | grep "Serial Number"`, and <org> from the name
after "OneDrive-" in ~/Library/CloudStorage. Leave <UPN>, <team> and the IT contact as placeholders and list them
at the top of the draft for me to fill in. Name ~/src/agent-context-sync/docs/deploy/entra-app.json as the
attachment. Do not send anything.
9. Finish with five lines: the folders synced (full paths); the docs repo path and its latest commit; the
background sync result from step 7 (last exit code); the doctor result (how many FAIL and warn lines); and
~/agent-context/it-request-draft.md with the placeholders I still have to fill.
Coding agents answer best from a folder of markdown they can read and grep. A company's knowledge lives somewhere else: tens of thousands of Office files, PDFs, mail and chat in Microsoft 365, changing in place under the same name, and mostly online-only on the laptop. Re-reading all of it on every sync means hours of downloads. The design gets the diff in four steps, and each step exists so that the next one can skip work:
- Ask each source what changed, and make expiry cheap. A Graph delta link, an FSEvents event id or a metadata walk. When a token expires, the fallback compares metadata and reads zero file bytes.
- Decide with three hashes before reading a byte. H0 decides whether to read, H1 whether to convert, H2 whether anything downstream changed. Office changes bytes on every no-op save, and H2 is what makes that save free.
- Read bytes only on purpose. An online-only file is a placeholder. The pipeline refuses to download it by default, and one budgeted step opts in.
- Publish a pure function of the source, in git. Converted pages are machine-generated and diffable, and each curated page records which version of each source it was written from.
Interactive Diagram
flowchart TD
GD["<b>L0</b> · OneDrive · SharePoint<br/>Outlook · Teams"]
FP["<b>L0</b> · local sync-client folder<br/>(File Provider)"]
MD["<b>L0</b> · manual drop"]
A["<b>L1 · Arm A · Graph delta</b><br/>stored deltaLink<br/>1 RU per poll"]
B["<b>L1 · Arm B · local walk</b><br/>FSEvents event id,<br/>getattrlistbulk walk"]
C["<b>L1 · Arm C · inbox</b><br/>same walk,<br/>max(created, modified)"]
M["<b>L2 · Manifest</b><br/>identity = stable id, never path<br/>H0 stat tuple, then H1, then H2"]
U(["<b>unchanged</b><br/>zero bytes read"])
X["<b>materialise</b><br/>budgeted download"]
V["<b>L3 · Convert</b><br/>write-once cache<br/>keyed on converter + content"]
E(["<b>early cutoff</b><br/>nothing downstream moves"])
MI[("<b>docs/mirror/</b><br/>pure function of source")]
T["<b>L4 · Curate</b><br/>docs/topics/ pins<br/>at_rendered_sha256"]
S[("<b>L5 · Surface</b><br/>docs/ in git · INDEX.md<br/>what changed = git log")]
O["<b>L6 · Operate</b><br/>one writer · flock · heartbeat<br/>hourly full reconcile"]
GD --> A
FP --> B
MD --> C
A --> M
B --> M
C --> M
M -->|"hash or stat tuple equal"| U
M -->|"maybe changed"| X
X --> V
V -->|"H2 output hash equal"| E
V -->|"H2 differs"| MI
MI --> T
T --> S
O -.->|"re-enumerates metadata"| M
classDef neutral fill:#161b22,stroke:#3d444d,color:#e6edf3
classDef amber fill:#2d2111,stroke:#f0a33a,color:#e6edf3
classDef green fill:#0f2a19,stroke:#3fb950,color:#e6edf3
class GD,FP,MD,A,B,C,M,MI,T,S,O neutral
class X,V amber
class U,E green
Colour means cost: amber steps read file bytes, green steps are free exits, everything else touches metadata only.
Every source already keeps a change log of its own. The pipeline stores one since-token per source and asks for the changes since that token, instead of looking at the files:
| Source | Since-token | Cost of asking |
|---|---|---|
| OneDrive, SharePoint, Outlook folders, Teams channels | Microsoft Graph deltaLink |
1 resource unit per poll: a drive polled every 60 s uses 0.12 % of the smallest per-app daily budget (Microsoft's published defaults). CORRECTED (2026-09-29): the resource-unit figure holds for OneDrive and SharePoint drives only. Outlook and Teams are throttled under their own Graph service limits, not resource units (Outlook: 10,000 requests per 10 min and 4 concurrent requests per app per mailbox), and neither budget is modelled yet. Under auth rung (ii) the per-app bucket belongs to the first-party Microsoft Graph PowerShell app and is shared with every user of that app in the tenant |
| A folder synced by the OneDrive client | FSEvents event id, backed by a getattrlistbulk metadata walk |
0.13–0.31 s per 100,000 files on local APFS (C11); a 2,000-file walk costs the same on OneDrive's File Provider as on local disk (C14 §3). CORRECTED (2026-09-29): that holds for directories the provider had already listed; the first walk of a never-listed directory is a provider round trip whose cost has no receipt |
| A manual drop folder | the same walk | the same |
Every token expires by design: a 410 from Graph, a wrapped FSEvents journal, a changed volume UUID. So the part that carries the load is the fallback, not the token: list the metadata again, diff it against the manifest, and read no file bytes. Deletions are trusted only inside a listing that completed.
Interactive Diagram
flowchart TD
ENTRY(["<b>new source</b><br/>or the hourly schedule"])
SYNC["<b>synced</b><br/>token stored outside git"]
POLL["<b>poll with the token</b><br/>Graph delta: 1 RU<br/>FSEvents: resume from event id"]
COMMIT(["<b>commit docs/</b><br/>then advance cursor"])
WAIT(["<b>back off</b><br/>Retry-After"])
EXP["<b>token expired</b><br/>by design"]
BAD["<b>cursor corrupt</b><br/>alarm, drop it"]
FULL["<b>full enumeration</b><br/>metadata only, zero file bytes<br/>diff the listing against the manifest<br/>deletes allowed only once it completes"]
ENTRY --> FULL
FULL -.->|"last page: store token"| SYNC
SYNC --> POLL
POLL -->|"200"| COMMIT
POLL -->|"429 / 503"| WAIT
POLL -->|"410 Gone<br/>journal wrap<br/>UUID change"| EXP
POLL -->|"400<br/>malformed"| BAD
EXP --> FULL
BAD --> FULL
classDef neutral fill:#161b22,stroke:#3d444d,color:#e6edf3
classDef red fill:#2d1417,stroke:#ff7b72,color:#e6edf3
class ENTRY,SYNC,POLL,COMMIT,WAIT,FULL neutral
class EXP,BAD red
A symlink into the OneDrive folder cannot play this role. It carries no since-token, and the agent's own tools do not
see through it: in a three-file fixture, ripgrep, BSD grep -r and -R, find, and Claude Code's Grep and Glob each
found 1 of 3 files, all with exit 0, and git stores the link as a single blob
(C1). The local walk, by contrast, gets a change signal from
the kernel: ATTR_CMN_GEN_COUNT, returned for 2,000 of 2,000 downloaded File Provider items and for the one placeholder
tested (C14 §3). CORRECTED (2026-09-29): this sentence said
"2,000 of 2,000 File Provider items, placeholders included" and called the counter "a real change signal". The 2,000
were all downloaded (dataless=0), only one placeholder was tested, the counter also moves on evict/download churn with
no content change, and whether it moves on an in-place edit on File Provider is unmeasured. A moved counter means "hash
to confirm", not "edited".
Identity is the stable id (Graph driveItem.id, or the file id locally), never the path, so a folder rename is one
record rather than a subtree of deletes and creates. Each observed object then passes up to three hashes, and every
decision before the byte read costs nothing:
| Hash | Computed from | Decides |
|---|---|---|
| H0 | the provider's quickXorHash, else (fileid, gen_count) and the stat tuple |
whether to read the bytes at all |
| H1 | a canonical content hash: sorted (part, bytes) with volatile parts removed |
whether to convert |
| H2 | the converter's output | whether anything downstream changed |
Interactive Diagram
flowchart TD
O["<b>observed object</b><br/>keyed by source_id, never by path"]
D{"dataless<br/>placeholder?"}
DL(["<b>DATALESS</b><br/>skip, never open"])
H0{"<b>H0 · zero bytes</b><br/>provider hash, else<br/>(fileid, gen_count) + stat"}
U(["<b>UNCHANGED</b><br/>zero bytes read"])
R["<b>read bytes</b><br/>budgeted materialise"]
H1{"<b>H1</b> · canonical<br/>content hash"}
T(["<b>TOUCHED, NOT CHANGED</b><br/>no-op save: stop"])
CV["<b>convert</b><br/>write-once cache"]
H2{"<b>H2</b> · hash of<br/>converter output"}
EC(["<b>early cutoff</b><br/>no commit"])
G[("<b>commit to docs/</b><br/>dependents flagged STALE")]
O --> D
D -->|"yes"| DL
D -->|"no"| H0
H0 -->|"equal"| U
H0 -->|"moved, or new id"| R
R --> H1
H1 -->|"equal"| T
H1 -->|"differs"| CV
CV --> H2
H2 -->|"equal"| EC
H2 -->|"differs"| G
classDef neutral fill:#161b22,stroke:#3d444d,color:#e6edf3
classDef amber fill:#2d2111,stroke:#f0a33a,color:#e6edf3
classDef green fill:#0f2a19,stroke:#3fb950,color:#e6edf3
class O,D,DL,H0,G neutral
class R,H1,CV,H2 amber
class U,T,EC green
H2 is not an optimisation. Real Office apps never re-save a file byte for byte, even with no edit (C12 §6):
| Format | What changed on a no-op save, outside docProps/ |
Converter output (H2) |
|---|---|---|
.pptx |
nothing | identical |
.xlsx |
xl/workbook.xml: a new documentId GUID on every save |
identical |
.docx |
word/document.xml and word/settings.xml: new revision-save ids (w:rsid*) |
identical |
A LibreOffice round trip changed the styles part and every sheet. Without H2, each of those saves would become a commit
that says nothing. CORRECTED (2026-09-29): "identical" was measured on the converter's body output only (pandoc
gfm, an openpyxl values-and-formulas dump, slide shape text), and on one generated fixture per format, not on a
full mirror page. The design's mirror frontmatter also carries source_etag, source_modified, source_version and
content_sha256, which change on every no-op save, so the page as specified would still change; that is a design gap,
and the implementation (docs/plans/implementation.md) resolves it.
The bookkeeping is cheap: a 200,000-row manifest is 31 MB of JSON and loads in 0.12 s
(design §4.1). CORRECTED (2026-09-29): that figure
was measured on a four-field row (path, size, mtime, sha256; 155 B/row,
D-manifest-buildgraph), not on the design's ~25-field source row,
which will be several times larger (estimated from the field count, not measured).
On macOS, OneDrive keeps most files online-only: a placeholder with the full size, zero allocated blocks and the
SF_DATALESS flag. Whether reading one downloads it depends on the caller's materialization policy, and the default
differs by context. This is a real run of the probes against a OneDrive for Business folder:
launchd-run.sh submits readfp as a real launchd job, which reports its default policy: off. readfp off applies
that policy to the placeholder, and the read fails with EDEADLK. The login shell's default is on, so the same read
downloads 2 MB. The direct read from a launchd job fails the same way, errno 11
(C14 §2). Recorded headlessly with VHS
(scripts/record-demo.sh).
Two consequences shape the design:
- A launchd job is fail-closed for free. The walk and hash stages inherit the refusing policy, and one budgeted
materialise()step opts in withsetiopolicy_np(…_ON)or the job'sMaterializeDatalessFileskey. CORRECTED (2026-09-29): two limits. (1)MaterializeDatalessFilesapplies to the whole job (C14 §2), so setting it on a job that also walks and hashes would switch those stages to downloading; inside that job,setiopolicy_np(…_ON)inmaterialise()is the only route that keeps them fail-closed. (2) "Fail-closed" covers the materialization policy, not file access: on 2026-09-24 a freshly built, unapprovedreadfpunder launchd blocked insideopen()on a File Provider path for more than 20 s instead of returningEDEADLK(probes/README). That is still unexplained, so a new binary may hang rather than fail. Every CloudStorage call needs a per-call timeout, and the launchd access matrix (unapproved vs approved binaries) is an open probe. - A stray walk from a login shell is a download.
rgor a hash pass over the sync folder would pull down the whole library, and macOS evicts it again under disk pressure. That is why phase 1 reads no bytes, and why a placeholder is recorded asdataless, a state distinct from both changed and unchanged.
docs/mirror/ is generated one to one from the sources and never hand-edited. Its frontmatter holds content-derived
fields only (no converted_at), so an unchanged source re-renders to identical bytes. docs/topics/ is written by the
agent, and each page pins the exact version (at_rendered_sha256) of every mirror page it cites. In the design, the two
questions an agent asks are each one command:
git log --since=2026-09-01 --stat -- docs/mirror # what changed upstream
sh refresh-queue.sh docs/DEPENDS.tsv # which curated pages are now stale, and whyThe refresh queue is one awk pass over a generated DEPENDS.tsv: 0.08–0.12 s over 1,600 rows
(C11 §3). It reports STALE, SOURCE-DELETED and SOURCE-UNREADABLE
separately because each needs a different action (design §4.5).
Freshness never rides on file times: git clone resets every mtime, and Claude Code's Glob orders its results by mtime
(G §4 measures the clone reset;
C7 (a) the mtime ordering). CORRECTED (2026-09-29): this cited the
retrieval review, which contains neither measurement.
The probes in design §9 each decide one part of the architecture. Four ran on a Microsoft 365 Business Standard tenant on macOS 15.7.9 (2026-09-23):
| # | Probe | Result |
|---|---|---|
| 3 | Which FSEvents fire when a file is edited on the web | A downloaded file raises Modified on the CloudStorage path within about 20 s. A new folder raises only its directory event, so a new directory needs a walk (C14 §4). CORRECTED (2026-09-29): case (ii), an online-only file, was not measured |
| 4 | Reading a placeholder, by context | Login shell downloads; launchd job fails EDEADLK, errno 11 (C14 §2) |
| 5 | Walk cost and GEN_COUNT on File Provider |
GEN_COUNT on 2,000 of 2,000 items; walk cost equal to local APFS (C14 §3). CORRECTED (2026-09-29): the 2,000 were downloaded items in already-listed directories, plus one placeholder; cold (never-listed) walk cost and GEN_COUNT on an in-place edit are unmeasured |
| 6 | Office no-op re-save | Never byte-stable; H2 identical for .pptx, .xlsx and .docx (C12 §6). CORRECTED (2026-09-29): measured on generated fixtures, one per format, not on real tenant files, and on the converter body only |
Five depend on the tenant itself, so they wait for the target one: whether SharePoint libraries return quickXorHash
in delta (1), whether a non-admin may consent to Files.Read.All and Sites.Read.All (2), whether two downloads of a
sensitivity-labelled file are byte-identical (4b), whether the spreadsheet converter reads a real Excel-saved workbook
correctly (8), and how long an idle delta token lives (9).
CORRECTED (2026-09-29): §9 has ten probes (1–9 plus 4b), not nine. Four are measured (3 and 6 only in part, as
marked above), five are unmeasured (1, 2, 4b, 8, 9), and one is partly measured: probe 7, converting a 50-file sample
twice with each converter, has pandoc ×2 on one .docx (C13) and
MarkItDown and PyMuPDF4LLM in the E report. The 2026-09-29 readiness audit adds
open probes, listed at the end of
design §9: GEN_COUNT on an
in-place edit on File Provider, the launchd access matrix, device-code sign-in under Conditional Access, Teams channel
delta, and shared-mailbox delta.
agentsync runs as you on a Mac and needs no admin rights. You need the Xcode Command Line Tools first
(xcode-select -p prints a path). They provide git and the compiler that builds the launcher. To have a coding
agent do all of this for you, paste the block in Set up on a new Mac: one prompt.
git clone https://github.com/renchris/agent-context-sync.git ~/src/agent-context-sync
ls -d ~/Library/CloudStorage/*/*/ # the OneDrive and SharePoint folders this Mac syncs
# uv, agentsync, the signed launcher, sources.toml with one live source per folder; ends with a NEXT: line
~/src/agent-context-sync/scripts/install.sh --source-local "$HOME/Library/CloudStorage/OneDrive-Contoso/Projects"
agentsync sync --once # a first cycle by hand, then start background sync:
~/src/agent-context-sync/scripts/install.sh --confirm-install-agentThe installer is safe to re-run and never prompts, and --dry-run prints every step first. Repeat
--source-local for each folder. On a Mac that already has sources.toml, it adds only the folders not yet in it
(agentsync add-source FOLDER does the same for one folder). --confirm-install-agent
installs two LaunchAgents: a poll every 5 minutes and an hourly reconcile. On the first background run, macOS asks
once for permission for the launcher to read OneDrive files. agentsync sync --once and agentsync status check a
cycle by hand. docs/deploy/README.md covers the rest: every installed path, the one-time
Allow click, the exit codes, what needs IT, and agentsync offboard.
CORRECTED (2026-09-29): three exceptions. The H2 comparison in
C12 §6 is not computed by any tracked script (see
probes/README). The 0.044 s first-walk figure in probes/README.md has no
receipt. And the design's "six-verdict fixture" re-run of the refresh queue has no receipt; C11 §3 records a 1,600-row
all-fresh run and a two-row fixture.
| Path | What it holds |
|---|---|
docs/design/agent-context-sync.md |
The full design: layers L0–L6, seven invariants, 25 ranked failure modes with the element that closes each, rejected alternatives, and the build order |
docs/design/receipts/ |
14 research-axis reports, 15 verifier reports and 4 review lenses, with the commands behind each number. CORRECTED (2026-09-29): this said 14 verifier reports; C15 (corporate controls: sign-in, Graph scopes, TCC, labels, TLS, PDF route, retention) is the fifteenth |
probes/ |
The macOS probes in C and Swift, plus the Office re-save script. make -C probes, then probes/README.md |
docs/diagrams/ |
Mermaid sources for the diagrams. npm run diagrams re-renders them, and CI fails when a render is stale |
src/agentsync/, tests/ |
The implementation and its test suite (uv run pytest -q). The last measured run is recorded in the dated rollout |
docs/deploy/, scripts/install.sh |
The corporate pack: the no-admin install, the IT request, the Entra app manifest, the PPPC profile, data governance and the tenant probes |
docs/plans/implementation.md |
The implementation plan, the dated rollout, and the readiness table that says what is closed and who owns what is still open |
The receipts were published with account names, tenant ids, file names and local paths replaced by placeholders
(Contoso, user@example.com, ~/). Every measurement is unchanged.