From a39ec9412e79a8595d41597fd453dad0cceeb415 Mon Sep 17 00:00:00 2001 From: Will Kahn-Greene Date: Fri, 25 Sep 2026 08:50:28 -0400 Subject: [PATCH 1/3] docs: the upload-through-the-root plan Answers #186: open an attachment through the os.Root that checked it, and verify the upload sends the bytes whose checksum it records. --- _plans/052_upload-through-the-root.md | 81 +++++++++++++++++++++++++++ 1 file changed, 81 insertions(+) create mode 100644 _plans/052_upload-through-the-root.md diff --git a/_plans/052_upload-through-the-root.md b/_plans/052_upload-through-the-root.md new file mode 100644 index 0000000..a530d11 --- /dev/null +++ b/_plans/052_upload-through-the-root.md @@ -0,0 +1,81 @@ +# 052: upload an attachment through the root that checked it + +Answers #186. The converter checks an image through the documentation root's +`os.Root`, but the client computes its checksum and uploads it through a plain +path. So the check and the two reads are three separate lookups, and only the +first is bounded. A directory on the path replaced by a symbolic link between +the check and the reads makes markfluence upload a file from outside the root, +which **S2** (`no-read-outside-root`) rules out. It is a race, not a static +hole: no layout of files on disk triggers it by itself. + +The same shape has a second, milder consequence, which this plan fixes too: +the checksum and the upload are two opens, so the bytes uploaded can differ +from the bytes whose checksum is recorded in the attachment's comment. + +## Decisions + +**D1. A `LocalAttachment` carries the root, and opens through it.** It gains +`Root *os.Root` (tagged `json:"-"`, so `check --show-html --json` and the +schema are unchanged) and an `Open() (*os.File, error)` method: + +- with a `Root`, it opens `Source`, the root-relative path the converter + checked, through `Root.Open`, which refuses an escape through a symbolic + link anywhere on the path; +- without one, it opens `Path` with `os.Open`. + +`Path` stays: it is what error messages and `check --show-html` show. + +**D2. The converter sets `Root`.** `images.go` already holds `r.root.FS`; it +sets it on every attachment it records. + +**D3. `attachment-upload` sets no `Root`.** Its files are named on the command +line by the person running it, not by a reference in a Markdown file, so S2 +does not govern them, and a file given with `--name` need not be under any +root at all. It still gets D4, which removes its own check-then-open race +(`os.Stat` and a later open). + +**D4. Whatever is opened must be a regular file, checked on the handle.** +`Open` calls `Stat` on the opened file and refuses anything that is not +regular. The converter's `Lstat` and `attachment-upload`'s `os.Stat` checked a +path; this checks the thing actually read, so a FIFO swapped in cannot hang +the upload and a directory cannot be "uploaded". + +**D5. The upload verifies the checksum it recorded.** `planAttachments` +computes the checksum through `Open`; `uploadAttachment` hashes the bytes as +it copies them into the request, and before sending compares the result with +the planned checksum. On a mismatch it sends nothing and fails with an error +naming the file, `changed while publishing`. The upload already reads the whole +file into memory, so this costs one hash and no extra read. The error is local +(not a `requestError`), so `--json` reports it as a local failure, not +`NETWORK`. + +## Files + +| file | change | +|---|---| +| `internal/attachref/attachref.go` | `Root` field, `Open` method | +| `internal/convert/images.go` | set `Root` | +| `internal/client/client.go` | `planAttachments` and `uploadAttachment` open through `Open`; `uploadAttachment` takes the planned checksum and verifies it; `fileChecksum` takes the attachment | +| `cmd/attachmentupload/attachmentupload.go` | nothing required; the `os.Stat` check stays for its early, friendly error | +| `CLAUDE.md` | `internal/attachref` has no bullet; the client bullet notes that attachments open through the root | + +## Tests + +- `attachref`: `Open` with a root opens the file; refuses an escape through a + symlinked directory (built after the root is opened, which is the race); + refuses a non-regular file (a FIFO, a directory); without a root opens + `Path`. +- `client`: `SyncAttachments` with an attachment whose root-relative path + escapes through a symlinked directory fails and uploads nothing; a file + whose content changes between plan and upload fails with `changed while + publishing` and uploads nothing (a test hook between the two, or a plan + built by hand with a wrong checksum). +- `convert`: an image attachment carries the root. + +## Not in scope + +- **Holding one file handle from plan to upload.** It would make D5 + unnecessary, but a page with many images would hold many descriptors open + across network calls. Reopening through the root and verifying the checksum + gives the same guarantee. +- **Bounding `attachment-upload` by a root** (D3). From 3e76e446e1c9bf7f69573febb26febff96b8f83b Mon Sep 17 00:00:00 2001 From: Will Kahn-Greene Date: Fri, 25 Sep 2026 08:53:06 -0400 Subject: [PATCH 2/3] fix(client): upload an attachment through the root that checked it Closes #186. The converter checks an image through the documentation root's os.Root, but the checksum and the upload opened it by absolute path, so a directory on that path replaced by a symbolic link in between could make markfluence upload a file from outside the root (S2). LocalAttachment now carries the root (json:"-") and an Open method that reads Source through it; the converter sets it. attachment-upload's files have no root, since the person named them on the command line. Either way the open is non-blocking and the handle must be a regular file. The upload also hashes what it sends and refuses, sending nothing, when the bytes no longer match the checksum taken while planning: the comment would otherwise misdescribe the attachment's content. --- CLAUDE.md | 3 +- internal/attachref/attachref.go | 46 ++++++++++ internal/attachref/attachref_test.go | 114 +++++++++++++++++++++++++ internal/client/client.go | 39 ++++++--- internal/client/client_test.go | 56 +++++++++++- internal/convert/images.go | 3 + internal/convert/namecollision_test.go | 12 +++ 7 files changed, 257 insertions(+), 16 deletions(-) create mode 100644 internal/attachref/attachref_test.go diff --git a/CLAUDE.md b/CLAUDE.md index 297319a..464263d 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -73,10 +73,11 @@ Module `github.com/mozilla/markfluence` (`go 1.25`). `main.go` is a shim to `cmd - `internal/actionlog` — the append-only record of what markfluence published, one log per project root at `/.markfluence/log.jsonl` (#149). `create`, `update` and `export` each append a line as a page completes, carrying the page version they left behind and `Sum`'s hash of what the body `PUT` sent; the last successful line for a file is that copy's **merge base**. It exists because nothing else can tell "the page differs because I have edits" from "the page differs because somebody published first" — that needs what *this copy* was derived from, which is a per-copy fact no page-side state can hold. Three things about it are load-bearing. **It is not committed**, and the reason is structural: a shared repository is itself a declaration that the repository is the source of truth, which is the arrangement where `update --force` is the answer and no base is consulted — so the log serves a local copy with the source of truth in Confluence, where per-checkout state is the right shape. The directory ignores itself (a planted `.gitignore` holding `*`, never overwritten) rather than editing a `.gitignore` markfluence does not own. **Nothing in it may fail a command**: a missing, unreadable, corrupt or half-written log degrades the check that reads it and never the run, which is why `read` skips a line it cannot parse instead of erroring and why a failed `Append` is a warning at the call site — the page is published by then, so failing would report that it was not. And **`Sum` covers exactly what the body `PUT` sends**, the resolved title and the rendered body: the title because `convert.ConfluencePage` carries none (a render-only hash would skip the publish of a file whose only change was its title), and *not* page width, labels or attachments, each of which has its own pass that runs whether or not the body is republished — folding them in would bump the page version for a change that never touched the body. Its parts are length-prefixed, which is part of the persisted format: changing the framing invalidates every recorded base. A root with `File == ""` gets no log at all (`For` returns nil, and every method is nil-safe), matching #139's rule that a root with no project file refuses rather than creating one. `Cache` hands out one `Log` per root so a batch reads each log once, and a batch spanning roots writes to several. `update` reads it for **two orthogonal checks, both only when `--force` is absent** (`--force` means always PUT, and no logic may suppress the request): the logged **`page_version`** against the live one decides **divergence** and *refuses* the file with `CodeConflict` — deliberately not qualified by the sha, since the case where you have no local edits is the worse one, publishing their work away with the bytes they started from — and the logged **`publish_sha256`** against this run's decides **idempotence** and skips the *body* `PUT` alone, leaving the attachment, width and label passes to run. An entry naming a different `page_id` is discarded rather than compared (a retarget would otherwise read as "the page moved 40 versions"), and the two fields degrade independently, so an `export` line written before its sha pass still refuses a moved page. A body-unchanged skip **records a line too**, which is load-bearing: once the sha does the skipping most runs skip, and a publish-only log would never keep a base current in a tree that is already published. Principles: **S8** (`no-overwrite-of-a-moved-page`), decided by the logged page version, and **L4** (`publish-is-idempotent`), decided by the publish sha rather than by mtime; both accept that a file with no base, or a root with no project file, gets neither. - `internal/pageslug` — `Slug`/`For`/`Filename`: a title to a filename-safe slug. A package rather than a helper because `export`, `read` and `attachment-download` all place attachments under a page's own directory and must agree. It lowercases (so case-variant titles collide and can be caught) and drops `/` (so no title can inject a path separator); it is lossy, and no readable slug can avoid being, so the caller decides what a collision means. Known limit: NFD and NFC spellings of one title are different Go strings but one filename on APFS, so that pair is not disambiguated. - `internal/pagedoc` — a fetched page as a Markdown document: `Render` (frontmatter + converted body), `Frontmatter` (which since #168 also emits `page_status`, widening `RenderFrontmatter`'s positional signature and adding an unconditional `GET /content/{id}/state` per page — a third best-effort read beside the width and label ones, so `export --space --depth all` pays it for every page in the space), and the lookups the converter can't do for itself — `Sources`/`SourcesFrom` (attachment name → recorded source path) and `PageLinks` (the page an `` points at → its URL). **One conversion, parameterized by a `Placement`**: where the page's file sits, where its unrecorded attachments go, what `parent:` says, and the attachment listing the caller already has. `read`, `export` and `attachment-download` all go through `Options`/`AttachmentDirFor` rather than assembling their own, so they cannot drift by accident — only by argument. For a page at the top level of what is being written, which is what `read` prints and what a single-page export writes, `read` and `export` are byte-identical; deeper in a tree they differ in exactly the position-dependent parts (a sourced attachment's `../` prefix, a `-` suffix a sibling forced, and `parent:`), because `read` has no tree to be positioned in. It needs a client (page width, attachment list, title lookups), which is why it isn't in `internal/convert` — that package is deliberately client-free, and it's why `StorageToMarkdown` takes those maps rather than fetching them. Every one of them is best-effort in the same shape: no references in the body means no request at all, and a lookup that fails is omitted rather than fatal (an omitted page link renders as raw storage, not as a link with no destination). It also owns **`UserCache`**, the user-name lookup both mention directions share (#91): a per-run, **cross-page** cache, threaded in from the caller the way `project.Cache`/`linkindex.Cache` are, because the obvious structure is wrong — `PageLinks` builds its space-id map per page and `Options` is built per page, so a user map written that way would re-resolve the same twelve people on every page of a 200-page export. It **remembers misses**, or a page mentioning deactivated people costs a request each, every page, to learn the same failures. Not persisted to disk, and the reason is **L2**: output must depend only on the files on disk, not on what a cache happens to hold. `MentionWarnings` is the forward direction's use of it, and sharing the cache is what makes that warning affordable — publishing needs no display names at all, so that lookup exists purely to report an id that names nobody. `PageLinks` resolves a space id **once per space key**, not once per link, and refuses to search site-wide when it can't scope a title to a space — a same-titled page in the wrong space is a wrong answer, which is worse than the passthrough a miss produces. +- `internal/attachref` — `LocalAttachment`, the local file to upload as one page's attachment, shared by `internal/convert` (which finds them) and `internal/client` (which uploads them) so the converter stays client-free. Its `Open` is the only way an upload reads the file (#186, `_plans/052`): with a `Root` — the documentation root's `os.Root`, which the converter sets and which is `json:"-"` so `check --show-html --json` is unchanged — it opens `Source` *through the root that checked it*, because the converter's `Lstat`, the checksum and the upload were otherwise three lookups of which only the first was bounded, and a directory swapped for a symlink in between made markfluence upload a file from outside the root (S2). Without a `Root` it opens `Path`, which is `attachment-upload`'s case: a file named on the command line is not a Markdown reference, and `--name` lets it live anywhere. Either way the open is non-blocking and the *handle* must be a regular file, so a FIFO swapped in is refused rather than hanging the upload. - `internal/attachfile` — `Resolve` (where an attachment goes under a destination root, **including the traversal clamp**) and `Write` (download it there, honoring force/dry-run). Shared by `attachment-download` and `export`; the clamp must never exist in two copies. - `internal/pagetree` — `Walk`, the traversal of pages *and folders* under a node, plus `WalkSpace` (the same traversal seeded from a space's root pages, via `client.ListSpaceRootPages`) and `AllDepths`. Both go through one `walker`, so the depth rule and the visited guard exist in a single copy. It is a package rather than command-local because listing a subtree and exporting one (#59) need the identical walk, and its rules must not exist in two copies: siblings arrive from two requests (`/child/page`, `/child/folder`) and are **merged by `extensions.position`**, or the output loses the order Confluence displays; a folder **counts as a level** like a page, which is only reasonable because folders are reported rather than silently traversed; and the walk descends folders even when only pages matter, since a folder may hold the only pages in a subtree. `nodeURL` uses `SiteURL()` — a v1 child row carries `webui` but no `base`. A visited set guards the unbounded case. - `internal/pageref` — `Resolve`, the single page-argument resolver: a numeric id, a Confluence page **or folder** URL (`pagePathRE` matches both `/pages/` and `/folder/`, since `children` takes a folder and a folder URL is what a browser hands you — the id is all it returns, so a command that can only use a page reports its own not-found), or a `.md` file that **declares** a `page_id` — in its own frontmatter or in a `pages:` entry for it (#139), resolved through `pagemeta` after discovering the root from the file's own directory (stat'd first, so `123.md` is a file). Every command taking a page uses it, which is why the manifest lookup lives here rather than at the seven call sites: without it the page argument meant one thing to `update` and another to `page-info`/`read`/`children`/`export`/`attachment-*`, so a file `update` could publish could not be named to any of them. A **disagreement** between the two locations is fatal here — it is the question being asked — while a **malformed** `markfluence.yaml` is not: a project file this resolver never consults must not make `page-info 123` fail, and the commands that bound reads by the root report it themselves. `message.go` also owns the wording for the two ways a *frontmatter* `page_id` is wrong — `NotFoundMessage` (caller supplies the remedy, which differs per command) and `NotNumericMessage` — because `create` and `update` report both and `check` reports the non-numeric one, and a reader should recognize the same problem across all of them. They return strings, not errors: `create` wraps the text in its typed `pageIDFailure` (which also carries the `--json` fields), the others want a plain error. Anything checking a `page_id` before a request uses `IsDigits`, since the API answers a non-numeric id with a 400 whose body says nothing useful. -- `internal/client` — `ConfluenceClient` over `net/http` with basic auth. Built from a `Config` (site URL, cloud ID, username, token) via `New`; it carries **two bases**: `BaseURL()` is where requests go (the gateway when a cloud ID is set) and `SiteURL()` is always the site. Anything a reader sees uses `SiteURL()` — printed page URLs and, critically, the `baseURL` handed to `convert.MdToConfluence`, since rewritten links are published *into* the page. Pages are Confluence **v2**; attachment writes and the user lookup are **v1** (`/wiki/rest/api/...`). A **folder** — the Cloud content type that can parent a page — has its own v2 route, `GetFolderOrNil` against `/wiki/api/v2/folders/{id}`, because every v2 *page* route answers a folder id with 404; enumerating children, if it is ever added, must be v1, since v2 cannot list inside a folder at all and its page-children route silently omits folders ([docs/confluence/folders.md](docs/confluence/folders.md)). Page **status** — the title lozenge — is v1 only and lives in `state.go` (`PageState`/`AvailableStates`/`SetPageState`, plus `StateVocabulary`, which keeps the space's statuses and the caller's own custom ones in separate fields because only the first are valid for a file); v2 carries no state field on a page in any form and there is no expansion that adds one, so a lozenge is one extra request per page, always. `AvailableStates` must be asked about the page the status is going on — its answer varies by caller *and* page, and it needs edit permission on that page. There is deliberately **no `ClearPageState`**: the `DELETE` route exists, but nothing can reach it until a clearing spelling does, and an unused write method is a loaded gun. Typed `HTTPError`, per-attempt context timeouts, centralized retry/backoff in `send`. `HTTPError.Error()` appends a **hint** for the three auth failures whose status misleads, matched on the response *body* rather than deduced from the status and always **appended** to it, never replacing it. The one that matters: **a rejected credential is a 404 on every v2 route**, so a revoked token used to make `read` answer `page ... not found` about a page that exists. `RejectedCredential` tells it apart by the fact that every genuine v2 404 *names* what it could not find and the auth one does not, `notFound` gates the three `…OrNil` helpers on it so they stop reading it as "absent", and `jsonout.CodeFor` checks it before the status switch so `--json` reports `AUTH` rather than `NOT_FOUND`. **Two error types on the request path, and one predicate for them**: an `*HTTPError` once a response has a status, an unexported `requestError` when there is none (a transport failure, a request that would not build, a body that would not decode), and `FromRequest` answers whether an error is either. That is what lets a caller tell a server failure from a local one — `jsonout.CodeOr(err, fallback)` is the whole point of it, since `CodeFor` alone reports every non-`HTTPError` as `NETWORK` and so turns `no title given` into a network problem (#133). The rule is deliberately scoped to the request: `DownloadAttachment` writing to the caller's writer, `uploadAttachment` opening the caller's file, and `Resolve` reading the environment stay untyped, because tagging them would misreport an unreadable file as a network failure. The wrapper carries no message of its own, so `Error()` is the inner text verbatim and nothing a reader sees changed. A 403 that is not one of the two measured credential phrasings gets no hint, because that is what a genuine permission denial looks like ([docs/confluence/api.md](docs/confluence/api.md#scopes)). **Retry rules**: 429 for any method; 502/503/504 for idempotent methods; **any other 5xx only when the response carries `Retry-After`** — that is how a 500 becomes retryable, and it is why `parseRetryAfter` reports the header's *presence* apart from its delay (`Retry-After: 0` means "retry now", not "no header"). The exponential delay is jittered, a server-supplied `Retry-After` never is. Decisions go to a package-level hook (`SetRetryLogger`, set once in `root.go` beside `ui.SetDebug`) and fire whichever way they went, because `internal/client` prints nothing and a silent twelve-minute retry storm is otherwise indistinguishable from a hang. **A versioned PUT is not as idempotent as its method**: `SetContentProperty` retry-once on top (recovers a lost create-POST response) and `UpdatePage`'s `updateLanded` both exist for the same reason — a write whose response was lost gets re-sent, and the re-sent version is refused. `updateLanded` requires version *and* title *and* body to match what was sent, since a concurrent edit could have produced the version alone and claiming success over someone else's content is worse than a false failure ([docs/confluence/api.md](docs/confluence/api.md)). `SyncAttachments` (skip/update by a SHA-256 recorded in the attachment's comment, alongside the source path so `read` recovers image paths exactly; only the current comment form is parsed — an attachment stamped by a markfluence predating a comment-format change reads as unmanaged and is re-uploaded once, the same as any hand-uploaded file — except that a *recorded path disagreeing with the local source* is an update even when the checksum matches, so a mangled path repairs itself instead of surviving every later publish; a comment with no source recorded at all is not a disagreement. Every text part of the upload form must go through `writeTextField`, never `multipart.Writer.WriteField`, which emits no charset and gets decoded as Latin-1), `_links.next` pagination. **Four pagination schemes, and picking the wrong one truncates silently.** v1 *child/attachment* collections page through the generic `listV1` helper by `start`/`limit` offset, never `_links.next` (absent when the results fit one page, so it cannot terminate a loop); `ListAttachments`, `ListChildPages`, and `ListChildFolders` all go through it. v2 collections page through `listV2` by the cursor in `_links.next` (whose loop is `walkV2`, shared rather than copied so a counting caller can stream — a second implementation of v2 paging is how one of them comes to terminate on a short page), which is a `/wiki`-prefixed absolute path `resolveNext` handles unchanged; `ListContentProperties` and `SearchPagesByTitle` share it. **`/wiki/rest/api/search` is neither**: it ignores `start` outright, its `next` is context-relative so it needs the `/wiki` prefix `resolveNext` does not add, a short page does *not* mean the end, and `totalSize` can be nonzero against an empty `results` — so `searchCQL` terminates only on a missing `next` and nothing may branch on `totalSize` ([docs/confluence/search.md](docs/confluence/search.md)). `searchCQLBounded` adds a row bound under it (`SearchCQL` is that call with no bound, which is why `find` is unaffected): it asks for `max+1` and reports the surplus as `more`, since `totalSize` cannot supply a count. **`/wiki/rest/api/space` is a fifth, and the one that punishes the obvious choice**: it pages by `start`/`limit` offset exactly as the child collections do, and **a short page is not the end** — asked for 250 from `start=0` it answered 200, and `start=200` then answered 250 more, against 525 spaces. `listV1` stops on that short page, so `WalkSpaceOperations` has its own loop terminating on an **empty** page, advancing by rows *returned* rather than by the limit asked for, bounded by `maxSpacePages` since an empty page is the only end signal offset paging has here. The first version of the probe that found this trusted the short page and reported 200 spaces with total confidence ([users.md](docs/confluence/users.md)). **`/wiki/rest/api/search/user` is a fourth**, and the one that looks most like an existing scheme while not being it: it pages by `start`/`limit` offset exactly as `listV1` does, so `listV1` is the obvious home for it and is a trap — the route **caps a page at 100 rows while echoing back whatever limit was requested** (101, 250 and 500 all answer 100), and `listV1` asks for `v1PageSize = 250` and reads a short page as the end of the collection, so it would truncate every result set past 100 with no error at all. `SearchUsers` lives in its own `users.go` with `userPageSize = 100` and the measurement beside it for that reason, and `TestPageCapDoesNotTruncate` is the regression. It also carries `maxUserPages`, `searchCQLBounded`'s guard for the same hazard reached a different way: a short page is the *only* end signal offset paging here has, so a server that clamped `start` — or ignored it the way `/wiki/rest/api/search` ignores it outright — would return a full page forever and an unbounded walk would collect rows until it ran out of memory. Its `totalSize` is a *third* kind of wrong: not absent like v1's and not an estimate like `/search`'s, but the row count of the page just fetched, so `limit=3` answers 3 and `limit=500` answers 100 against 304 real matches. `user.go` holds the two identity routes (`CurrentUser`, `UserInfo` — both `read:confluence-user`, both seeing a deactivated account the directory cannot) and `WalkSpaceOperations`; `space.go` holds `GetSpace` (one v1 request answering identity, the caller's own operations, description, labels and the homepage *with its title*), `SpaceStateSettings` (space-admin only, so a 403 that is not a rejected credential is `(nil, nil)` rather than an error) and `WalkSpacePages`. Both space routes decode the space `id` as a `json.Number`: v1 reports it as a **number** where every v2 route reports a string, and `homepage.id` in the same response is a string. Full text goes through `SearchText`/`SearchRawCQL`, which return the cleaned `SearchMatch` the way `FindByTitle` returns `TitleMatch` — and **every field of a match comes from the row's `content` object**, because the row-level `title` is HTML-escaped *and* wrapped in `@@@hl@@@` markers where `content.title` is neither. The `excerpt` exists only at row level, so `cleanExcerpt` strips those markers, unescapes once, and collapses to one line — in the client, so the human and `--json` paths cannot disagree about it. `excerpt=highlight` is passed explicitly and **re-attached when following the cursor** (the `next` link carries `cql` and `limit` but not `excerpt`, and `doJSON` appends params with a bare `?`); an unrecognized value there yields an empty excerpt with a 200, so a rename by Atlassian degrades to no excerpts rather than an error. A row with no `content` object is skipped and **counted** — `type = space` answers with hundreds of them, and a silent skip would report a successful empty result. A bare v1 child row already carries `webui`, `status`, and `extensions.position`, so child listing needs no `expand`. `DownloadAttachment` goes through `send` (inheriting retry/backoff) against `_links.download`; **never** add a `CheckRedirect` that forwards headers — it would leak site credentials to Atlassian's media host, which neither needs nor wants them. `credentials.go` holds the credentials file's own functions — `CredentialsPath`, `ReadCredentials` (the rules `Resolve` uses, minus the permission warning, for a caller about to rewrite the file; one read returning a `CredentialsFile` of values, the lines a rewrite drops, and the mode), `WriteCredentials` (temp-file-and-rename at `0600`, writing through a symbolic link, quoting a value exactly when `readDotenv` would not read it back unchanged, then reading the result back and comparing), `DisplayPath`, and `FetchCloudID`, the unauthenticated `tenant_info` request, which bypasses `send` because `send` always sets basic auth. The setting names (`URLVar`/`UsernameVar`/`TokenVar`/`CloudIDVar`), `CredentialsDoc` and `LooseMode` are exported so `credentials-init` copies none of them. `HTTPError.ScopeMismatch`/`SiteRejectedAuth`, beside `RejectedCredential`, are the shapes `hint` matches, exported for `credentials-init`. `config.go` holds `Resolve` and the env-file reader (`loadDotenv`, which warns, over `readDotenv`, which does not), plus the **permission warning** (#136): a *regular* credentials file or `--env-file` reachable by anyone but its owner (`mode.Perm()&0o077`; a pipe from `--env-file <(pass show …)` reports 0440 and no chmod can fix it) *and* containing `CONFLUENCE_TOKEN` earns a warning naming the file, its mode, and the `chmod`. Both halves matter — a file holding only the URL and username leaks nothing, and a warning that fires on a file with no secret in it is how one becomes something people scroll past. It stats rather than lstats (a link's own `0777` would cry wolf over a `0600` target), lives in `loadDotenv` because that is the one function both the credentials file and `--env-file` pass through, and reaches the reader through `SetSecurityWarner` for the same reason `SetRetryLogger` exists — wired to `cmd/root.go`'s `reportSecurityWarning`, which prints it (human mode) *and* records it via `jsonout.AddWarning`, since stderr under `--json` is a schema-validated document with no room for a stray line. A group/world-*writable* file with no token in it is knowingly **not** covered: the same-source rule means a URL rewritten there can no longer be paired with a token from somewhere else, and the cloud ID follows the URL. Why each of these is shaped this way, with the evidence: [docs/confluence/api.md](docs/confluence/api.md) and [attachments.md](docs/confluence/attachments.md). +- `internal/client` — `ConfluenceClient` over `net/http` with basic auth. Built from a `Config` (site URL, cloud ID, username, token) via `New`; it carries **two bases**: `BaseURL()` is where requests go (the gateway when a cloud ID is set) and `SiteURL()` is always the site. Anything a reader sees uses `SiteURL()` — printed page URLs and, critically, the `baseURL` handed to `convert.MdToConfluence`, since rewritten links are published *into* the page. Pages are Confluence **v2**; attachment writes and the user lookup are **v1** (`/wiki/rest/api/...`). A **folder** — the Cloud content type that can parent a page — has its own v2 route, `GetFolderOrNil` against `/wiki/api/v2/folders/{id}`, because every v2 *page* route answers a folder id with 404; enumerating children, if it is ever added, must be v1, since v2 cannot list inside a folder at all and its page-children route silently omits folders ([docs/confluence/folders.md](docs/confluence/folders.md)). Page **status** — the title lozenge — is v1 only and lives in `state.go` (`PageState`/`AvailableStates`/`SetPageState`, plus `StateVocabulary`, which keeps the space's statuses and the caller's own custom ones in separate fields because only the first are valid for a file); v2 carries no state field on a page in any form and there is no expansion that adds one, so a lozenge is one extra request per page, always. `AvailableStates` must be asked about the page the status is going on — its answer varies by caller *and* page, and it needs edit permission on that page. There is deliberately **no `ClearPageState`**: the `DELETE` route exists, but nothing can reach it until a clearing spelling does, and an unused write method is a loaded gun. Typed `HTTPError`, per-attempt context timeouts, centralized retry/backoff in `send`. `HTTPError.Error()` appends a **hint** for the three auth failures whose status misleads, matched on the response *body* rather than deduced from the status and always **appended** to it, never replacing it. The one that matters: **a rejected credential is a 404 on every v2 route**, so a revoked token used to make `read` answer `page ... not found` about a page that exists. `RejectedCredential` tells it apart by the fact that every genuine v2 404 *names* what it could not find and the auth one does not, `notFound` gates the three `…OrNil` helpers on it so they stop reading it as "absent", and `jsonout.CodeFor` checks it before the status switch so `--json` reports `AUTH` rather than `NOT_FOUND`. **Two error types on the request path, and one predicate for them**: an `*HTTPError` once a response has a status, an unexported `requestError` when there is none (a transport failure, a request that would not build, a body that would not decode), and `FromRequest` answers whether an error is either. That is what lets a caller tell a server failure from a local one — `jsonout.CodeOr(err, fallback)` is the whole point of it, since `CodeFor` alone reports every non-`HTTPError` as `NETWORK` and so turns `no title given` into a network problem (#133). The rule is deliberately scoped to the request: `DownloadAttachment` writing to the caller's writer, `uploadAttachment` opening the caller's file, and `Resolve` reading the environment stay untyped, because tagging them would misreport an unreadable file as a network failure. The wrapper carries no message of its own, so `Error()` is the inner text verbatim and nothing a reader sees changed. A 403 that is not one of the two measured credential phrasings gets no hint, because that is what a genuine permission denial looks like ([docs/confluence/api.md](docs/confluence/api.md#scopes)). **Retry rules**: 429 for any method; 502/503/504 for idempotent methods; **any other 5xx only when the response carries `Retry-After`** — that is how a 500 becomes retryable, and it is why `parseRetryAfter` reports the header's *presence* apart from its delay (`Retry-After: 0` means "retry now", not "no header"). The exponential delay is jittered, a server-supplied `Retry-After` never is. Decisions go to a package-level hook (`SetRetryLogger`, set once in `root.go` beside `ui.SetDebug`) and fire whichever way they went, because `internal/client` prints nothing and a silent twelve-minute retry storm is otherwise indistinguishable from a hang. **A versioned PUT is not as idempotent as its method**: `SetContentProperty` retry-once on top (recovers a lost create-POST response) and `UpdatePage`'s `updateLanded` both exist for the same reason — a write whose response was lost gets re-sent, and the re-sent version is refused. `updateLanded` requires version *and* title *and* body to match what was sent, since a concurrent edit could have produced the version alone and claiming success over someone else's content is worse than a false failure ([docs/confluence/api.md](docs/confluence/api.md)). `SyncAttachments` (skip/update by a SHA-256 recorded in the attachment's comment, alongside the source path so `read` recovers image paths exactly; only the current comment form is parsed — an attachment stamped by a markfluence predating a comment-format change reads as unmanaged and is re-uploaded once, the same as any hand-uploaded file — except that a *recorded path disagreeing with the local source* is an update even when the checksum matches, so a mangled path repairs itself instead of surviving every later publish; a comment with no source recorded at all is not a disagreement. Every file is read through `LocalAttachment.Open`, twice — once for the checksum, once for the upload — and `uploadAttachment` hashes what it copies and sends nothing if it differs from the planned checksum ("changed while publishing"), since a comment misdescribing its content reads as up to date on the next publish. Every text part of the upload form must go through `writeTextField`, never `multipart.Writer.WriteField`, which emits no charset and gets decoded as Latin-1), `_links.next` pagination. **Four pagination schemes, and picking the wrong one truncates silently.** v1 *child/attachment* collections page through the generic `listV1` helper by `start`/`limit` offset, never `_links.next` (absent when the results fit one page, so it cannot terminate a loop); `ListAttachments`, `ListChildPages`, and `ListChildFolders` all go through it. v2 collections page through `listV2` by the cursor in `_links.next` (whose loop is `walkV2`, shared rather than copied so a counting caller can stream — a second implementation of v2 paging is how one of them comes to terminate on a short page), which is a `/wiki`-prefixed absolute path `resolveNext` handles unchanged; `ListContentProperties` and `SearchPagesByTitle` share it. **`/wiki/rest/api/search` is neither**: it ignores `start` outright, its `next` is context-relative so it needs the `/wiki` prefix `resolveNext` does not add, a short page does *not* mean the end, and `totalSize` can be nonzero against an empty `results` — so `searchCQL` terminates only on a missing `next` and nothing may branch on `totalSize` ([docs/confluence/search.md](docs/confluence/search.md)). `searchCQLBounded` adds a row bound under it (`SearchCQL` is that call with no bound, which is why `find` is unaffected): it asks for `max+1` and reports the surplus as `more`, since `totalSize` cannot supply a count. **`/wiki/rest/api/space` is a fifth, and the one that punishes the obvious choice**: it pages by `start`/`limit` offset exactly as the child collections do, and **a short page is not the end** — asked for 250 from `start=0` it answered 200, and `start=200` then answered 250 more, against 525 spaces. `listV1` stops on that short page, so `WalkSpaceOperations` has its own loop terminating on an **empty** page, advancing by rows *returned* rather than by the limit asked for, bounded by `maxSpacePages` since an empty page is the only end signal offset paging has here. The first version of the probe that found this trusted the short page and reported 200 spaces with total confidence ([users.md](docs/confluence/users.md)). **`/wiki/rest/api/search/user` is a fourth**, and the one that looks most like an existing scheme while not being it: it pages by `start`/`limit` offset exactly as `listV1` does, so `listV1` is the obvious home for it and is a trap — the route **caps a page at 100 rows while echoing back whatever limit was requested** (101, 250 and 500 all answer 100), and `listV1` asks for `v1PageSize = 250` and reads a short page as the end of the collection, so it would truncate every result set past 100 with no error at all. `SearchUsers` lives in its own `users.go` with `userPageSize = 100` and the measurement beside it for that reason, and `TestPageCapDoesNotTruncate` is the regression. It also carries `maxUserPages`, `searchCQLBounded`'s guard for the same hazard reached a different way: a short page is the *only* end signal offset paging here has, so a server that clamped `start` — or ignored it the way `/wiki/rest/api/search` ignores it outright — would return a full page forever and an unbounded walk would collect rows until it ran out of memory. Its `totalSize` is a *third* kind of wrong: not absent like v1's and not an estimate like `/search`'s, but the row count of the page just fetched, so `limit=3` answers 3 and `limit=500` answers 100 against 304 real matches. `user.go` holds the two identity routes (`CurrentUser`, `UserInfo` — both `read:confluence-user`, both seeing a deactivated account the directory cannot) and `WalkSpaceOperations`; `space.go` holds `GetSpace` (one v1 request answering identity, the caller's own operations, description, labels and the homepage *with its title*), `SpaceStateSettings` (space-admin only, so a 403 that is not a rejected credential is `(nil, nil)` rather than an error) and `WalkSpacePages`. Both space routes decode the space `id` as a `json.Number`: v1 reports it as a **number** where every v2 route reports a string, and `homepage.id` in the same response is a string. Full text goes through `SearchText`/`SearchRawCQL`, which return the cleaned `SearchMatch` the way `FindByTitle` returns `TitleMatch` — and **every field of a match comes from the row's `content` object**, because the row-level `title` is HTML-escaped *and* wrapped in `@@@hl@@@` markers where `content.title` is neither. The `excerpt` exists only at row level, so `cleanExcerpt` strips those markers, unescapes once, and collapses to one line — in the client, so the human and `--json` paths cannot disagree about it. `excerpt=highlight` is passed explicitly and **re-attached when following the cursor** (the `next` link carries `cql` and `limit` but not `excerpt`, and `doJSON` appends params with a bare `?`); an unrecognized value there yields an empty excerpt with a 200, so a rename by Atlassian degrades to no excerpts rather than an error. A row with no `content` object is skipped and **counted** — `type = space` answers with hundreds of them, and a silent skip would report a successful empty result. A bare v1 child row already carries `webui`, `status`, and `extensions.position`, so child listing needs no `expand`. `DownloadAttachment` goes through `send` (inheriting retry/backoff) against `_links.download`; **never** add a `CheckRedirect` that forwards headers — it would leak site credentials to Atlassian's media host, which neither needs nor wants them. `credentials.go` holds the credentials file's own functions — `CredentialsPath`, `ReadCredentials` (the rules `Resolve` uses, minus the permission warning, for a caller about to rewrite the file; one read returning a `CredentialsFile` of values, the lines a rewrite drops, and the mode), `WriteCredentials` (temp-file-and-rename at `0600`, writing through a symbolic link, quoting a value exactly when `readDotenv` would not read it back unchanged, then reading the result back and comparing), `DisplayPath`, and `FetchCloudID`, the unauthenticated `tenant_info` request, which bypasses `send` because `send` always sets basic auth. The setting names (`URLVar`/`UsernameVar`/`TokenVar`/`CloudIDVar`), `CredentialsDoc` and `LooseMode` are exported so `credentials-init` copies none of them. `HTTPError.ScopeMismatch`/`SiteRejectedAuth`, beside `RejectedCredential`, are the shapes `hint` matches, exported for `credentials-init`. `config.go` holds `Resolve` and the env-file reader (`loadDotenv`, which warns, over `readDotenv`, which does not), plus the **permission warning** (#136): a *regular* credentials file or `--env-file` reachable by anyone but its owner (`mode.Perm()&0o077`; a pipe from `--env-file <(pass show …)` reports 0440 and no chmod can fix it) *and* containing `CONFLUENCE_TOKEN` earns a warning naming the file, its mode, and the `chmod`. Both halves matter — a file holding only the URL and username leaks nothing, and a warning that fires on a file with no secret in it is how one becomes something people scroll past. It stats rather than lstats (a link's own `0777` would cry wolf over a `0600` target), lives in `loadDotenv` because that is the one function both the credentials file and `--env-file` pass through, and reaches the reader through `SetSecurityWarner` for the same reason `SetRetryLogger` exists — wired to `cmd/root.go`'s `reportSecurityWarning`, which prints it (human mode) *and* records it via `jsonout.AddWarning`, since stderr under `--json` is a schema-validated document with no room for a stray line. A group/world-*writable* file with no token in it is knowingly **not** covered: the same-source rule means a URL rewritten there can no longer be paired with a token from somewhere else, and the cloud ID follows the URL. Why each of these is shaped this way, with the evidence: [docs/confluence/api.md](docs/confluence/api.md) and [attachments.md](docs/confluence/attachments.md). - `internal/convert` — the converter (the crux). `MdToConfluence(md *frontmatter.MarkdownFile, root *project.Root, index *linkindex.Index, baseURL, spaceKey string) (*ConfluencePage, error)`. `root` bounds which images and parent references may be read (S1/S2) and is what an image's recorded `Source` is relative to; `index` is the tree-wide link/anchor index for `root` (`internal/linkindex.Build`), built once and shared across every file converted under it rather than rebuilt per conversion — both are discovered/built by the caller (`internal/project`/`internal/linkindex`), which is why this package stays client-free. It parses with goldmark (GFM) and renders through a custom `storageRenderer` registered at priority 100 (below the default HTML=1000 and table=500 renderers) that emits Confluence storage format. `shield.go` renames raw `ac:`/`ri:` tags to colon-free sentinels around the goldmark step so pasted storage passes through; `callouts.go` is an AST transformer + blockquote renderer for GitHub alerts; `mention.go` owns the user-mention mapping in both directions (#91): a mention is 80% of all `` usage, and it converts to `[@Display Name](https://home.atlassian.com/people/{accountId})`. Three things decide its shape, each measured rather than reasoned. **The URL is Atlassian Home, not the site** — Confluence's own renderer still emits `{site}/wiki/people/{id}`, which no longer resolves usefully in a browser, so a mention in Markdown names *no site*, needs nothing from configuration, and is therefore recognisable by `check` with no client at all. **Matching is on the path, ignoring host and query**, because several spellings of one target circulate (the Home URL, the modal's `?cloudId=` copy, the `/o/{orgId}` redirect, both Confluence forms, root-relative) and none of `cloudId`/`ref`/the org segment identifies the person — only the id does, and `ri:user` stores nothing else. **The `@` on the link text is the marker**, load-bearing rather than decoration: the URL cannot tell "mention this person" from "link to their profile", so without it anyone writing the second would silently get the first. The account id is **not** pattern-validated (two shapes are live on one instance, so a pattern tight enough for one rejects the other) and `ri:local-id` is never emitted (a mention carrying only the id resolves to the same person, verified via ADF). `MentionMarkdown` is the shared builder for a mention's whole Markdown line, exported because `user-find` (#143) prints exactly it and a second copy there would be a second place to get the `@` marker and name escaping right. `ConfluencePage.Mentions` reports the ids the *forward* direction emitted so the caller can warn about one that names nobody — the `Attachments` arrangement, and necessary because Confluence accepts any id and renders `@Unlicensed user` rather than failing, and the profile URL 200s either way. An unresolvable mention still renders as a link, `[@Unlicensed user](…)` — that wording mirrors Confluence because the only ids reaching it are the ones the page labels that way: a **deactivated account resolves normally** and keeps its name (measured across every mention on a real page — 18 of them, six departed, all 200, returning e.g. `Mark Reid (Deactivated)`), so a departed colleague never takes that branch. Name resolution is `pagedoc.UserCache`, a per-run cross-page cache, and the tri-state is the part to preserve: `client.LookupUser` separates a name from `ErrNoSuchUser` from an unaskable question, `StorageOptions.UserNames` carries that as name / `""` / absent, and only a **confirmed** absence renders the placeholder. Flattening those would write a fabricated name over a real one the moment a VPN dropped mid-export, across a whole tree, into a file that then looks authoritative — which is also why the cache remembers a 404 but not a timeout (one is an answer, the other is not) and why `MentionWarnings` warns only about a confirmed absence; `aclink.go` is the *inverse* direction's one element with enough shape to need its own file — ``, which the editor writes for every internal link and `MdToConfluence` never emits, so nothing in the regression suite covers it. One rule decides its whole mapping: **convert when the Markdown republishes to a link resolving to the same target, pass the storage through when it would not** — so a page link and a space link convert, while a mention and a space link convert, and an attachment link (only images are uploaded, so a relative href would be dead) and an unresolvable target stay raw, which the shield republishes byte-identical. A page target is a **title, never an id**, so `PageLinkTargets` reports what needs resolving and `StorageOptions.PageLinks` carries the answers back. An `ac:anchor` is **percent-encoded** where `confluenceSlug` output is not: decode it before matching a heading, leave it encoded inside a URL. A same-page anchor recovers its heading from the document rather than inverting the slug, which is impossible — `confluenceSlug` turns both a space and a hyphen into `-`. The survey the mapping rests on, and the `xml.HTMLAutoClose` trap that made `` crash the parser outright (#88), are in [docs/confluence/links-and-anchors.md](docs/confluence/links-and-anchors.md); Plain text used as a Markdown link's text goes through `escapeLinkText` (`\`, `[`, `]`), applied to the *raw* sources only — a page title, a space key, an anchor, a display name — and via `inlineTextForLink` to a body whose every descendant is a text node. Never to already-rendered output: an `ac:link-body` holding markup has been converted to Markdown already, and escaping it yields a literal `\*\*bold\*\*`. Both directions are tested, because a fix at either extreme passes one and fails the other. `attachname.go` owns the source-path→attachment-name mapping, which is now the path's **base name** and nothing else (#59/`_plans/029`): the name is the attachment's identity, so an encoded path moved the name every time the file moved and orphaned the old attachment, and the path is recorded in the comment anyway. The mapping is therefore lossy, and what the bijection used to buy is an explicit refusal — two assets in one document whose base names agree return a typed `NameCollisionError` from `MdToConfluence`, which is a *failure* and not a `Broken` entry, since nothing blocks a publish on `Broken`. `check` catches that error and reports it as `Broken` anyway, because there it is a document defect like a dead link rather than a converter failure. A stored name is never interpreted in the other direction either: `sourceFor` reads the recorded path or uses the name verbatim. What names Confluence accepts is in [docs/confluence/attachments.md](docs/confluence/attachments.md); `destination.go` owns the **other** codec, destination↔path (`decodeDestination`/`encodeDestination`), shared by images *and* doc links — a Markdown destination is a URL, so decode inbound (**before** `withinRoot`, or an encoded `..%2F` slips the clamp) and encode outbound in `storage_to_md.go` (or `export` emits Markdown that no longer parses, and `sourceFor`'s absolute-path refusal is undone by the next read); an undecodable destination is a literal `%` in a filename, not an error; the reasoning is in [docs/confluence/links-and-anchors.md](docs/confluence/links-and-anchors.md); `images.go` (resolution stays page-relative like GitHub; the documentation root — cwd — bounds what may be published, and an image above it is `IMAGE BROKEN`), `links.go` (GitHub/Confluence slugs, doc-link + anchor rewriting against `internal/linkindex`'s tree-wide index; `resolveDocKey` resolves a destination to the index's root-relative key and reports `escapes` — a purely lexical check on the *query* side, since the index itself needs no clamp: an escaping key can never be in it, built by walking downward from root). A doc-link target is one of four severities, #42: missing entirely or escaping root is **Broken** (`LINK BROKEN: … (not found|outside the documentation root)`) and replaces the whole `` element — tags and visible text alike — with that literal message, matching `images.go`'s precedent for a missing image (`renderLink` needs a small per-node flag, `linkBrokenText`, since goldmark still invokes a container node's renderer on the matching leaving call regardless of `WalkSkipChildren` on entering, and there is no `` to write in the broken case); existing on disk with no `page_id` yet is unchanged — a **warning**, the normal state of an unpublished tree; a `#fragment` matching no heading on an otherwise-resolving target also **warns**, gated on `linkindex.Index.FileExists` so a missing/escaping target isn't double-reported. `tables.go` (the `` tag, stamped with `data-layout="align-start"` so tables auto-size and left-align — this must stay if column widths are ever emitted, or a `` silently induces a layout; plus cells: an AST transformer consumes a leading `` comment in a cell and `renderTableCell` emits it as `data-highlight-colour`; `storage_to_md.go`'s `cellTexts` reverses this, reading `data-highlight-colour` back into a `bg:` marker (`cellBGNames`, the reverse of `tables.go`'s swatch map — a hex outside the 21 swatches round-trips as the literal hex, and where two names share a hex the British spelling wins, matching Confluence's own `-colour`), and a column's GFM alignment becomes a `

` wrapper **inside** the cell — never the `align` attribute the GFM renderer would emit, which is the one form Confluence discards. Only center and right are emitted: Confluence has no explicit left, so `:---` publishes bare and `read` recovers it as `---`. Since alignment is per-paragraph there and per-column in GFM, `columnSeparators` in `storage_to_md.go` takes each column's most common declared alignment (ties to the first seen) and drops the rest. Rows still fall through to the GFM renderer. A multi-line cell is one `

` per line — Enter in the editor starts a new `

`, it does not insert a `
` — so `renderCellLines` in `storage_to_md.go` joins sibling `

` children with a literal `
` rather than nothing: a GFM table row is exactly one physical line, so a real newline isn't an option, and the same substitution catches a bare mid-line `
` (Shift+Enter) that would otherwise render as the two-space hard break valid in ordinary block content but not inside a table row. A `

` tag, stamped with `data-layout="align-start"` so tables auto-size and left-align — this must stay if column widths are ever emitted, or a `` silently induces a layout; plus cells: an AST transformer consumes a leading `` comment in a cell and `renderTableCell` emits it as `data-highlight-colour`; `storage_to_md.go`'s `cellTexts` reverses this, reading `data-highlight-colour` back into a `bg:` marker (`cellBGNames`, the reverse of `tables.go`'s swatch map — a hex outside the 21 swatches round-trips as the literal hex, and where two names share a hex the British spelling wins, matching Confluence's own `-colour`), and a column's GFM alignment becomes a `

` wrapper **inside** the cell — never the `align` attribute the GFM renderer would emit, which is the one form Confluence discards. Only center and right are emitted: Confluence has no explicit left, so `:---` publishes bare and `read` recovers it as `---`. Since alignment is per-paragraph there and per-column in GFM, `columnSeparators` in `storage_to_md.go` takes each column's most common declared alignment (ties to the first seen) and drops the rest. Rows still fall through to the GFM renderer. A multi-line cell is one `

` per line — Enter in the editor starts a new `

`, it does not insert a `
` — so `renderCellLines` in `storage_to_md.go` joins sibling `

` children with a literal `
` rather than nothing: a GFM table row is exactly one physical line, so a real newline isn't an option, and the same substitution catches a bare mid-line `
` (Shift+Enter) that would otherwise render as the two-space hard break valid in ordinary block content but not inside a table row. A `