Telegram Import

2026.08.02.1

Real-fix for all eight remaining latent bugs (LB-2..LB-9), all tracked on the LOCAL telegram-import-latent-bugs issue-lifecycle model (never filed to the swamp.club Lab — see CLAUDE.md's anti-bypass rule). model.version and manifest.yaml both bump 2026.08.01.2 -> 2026.08.02.1. LB-0 and LB-1 (both already fixed) are untouched and stay green; every benign/golden/property assertion is byte-identical.

• LB-2 (note-path traversal via msg.id, MEDIUM): added assertSafeSlug, called as the first statement inside the note-create try — rejects a slug containing a / or a .. segment before it can reach the CLI create path= argument verbatim (the vaultRoot branch was already confined by resolveVaultPathSafe; only the CLI branch had no check of its own). A rejection lands in the existing catch: the note for that message is skipped and recorded in errors[], and the post resource is still written. Pinned with a non-vacuity test: a normal numeric-id slug still reaches create path= unchanged.
• LB-3 (YAML frontmatter injection, MEDIUM): added yamlDq, escaping backslash/quote/CR/LF/other control characters into a valid single-line YAML double-quoted scalar. Wraps channel, forwarded_from, and the note title — all three are interpolated straight from the Telegram export. An embedded " plus CR/LF can no longer break out of the frontmatter block to inject sibling YAML keys. Byte-identical to the previous bare interpolation for any input with no backslash/quote/control character (the common case; the golden-fixture and property suites, which never generate such input, are the guard).
• LB-4 (one malformed message aborted the whole import, MEDIUM): the entire per-message loop body is now wrapped in its own try/catch. A failure (e.g. noteSlug throwing on a non-string date) is recorded in errors[] as Skipped message (id …): … and the loop moves on to the next message, instead of rejecting the whole import call uncaught. writeResource for the result summary is still called only after the loop completes — unchanged — but now it is always reached. Pinned: totalMessages still counts all messages (the failure does not zero the import), and the messages before AND after the failing one both get their post resource written.
• LB-5 (unvalidated top-level export shape, MEDIUM): added ExportSchema (z.object({ name: z.string(), messages: z.array(z.unknown()) })), safeParse'd against the freshly-parsed JSON before data.name/ data.messages are trusted, throwing a clear Invalid Telegram export: … Error on failure. data itself is left as whatever JSON.parse returned (not replaced by the parsed/narrowed value) so every per-message field access downstream keeps its prior permissive typing — this is purely a fail-closed shape check, not a schema migration. A missing messages array now throws a clear Error instead of an opaque TypeError; a missing name now fails loudly instead of silently rendering the literal string "undefined" into every note's channel frontmatter line.
• LB-6 (find had no timeout, exit code never checked, LOW): find is now run under an AbortController with a 30s timeout (clearTimeout in a finally, so no timer is ever leaked past the call), and findOut.success/ .code are checked explicitly before reading stdout — a genuinely FAILING find now throws a distinct find failed (exit N): <stderr> instead of the generic No result.json found an empty-but-successful find produces (both messages stay pinned, one per outcome).
• LB-7 (text.substring(0, 500) could split a surrogate pair, LOW): post text is now truncated with Array.from(text).slice(0, 500).join("") — truncation is by Unicode CODE POINT, never cutting a surrogate pair in half. Byte-identical to the old substring behavior for any text with no astral-plane characters at the truncation boundary (the property suite's ASCII-only generator never exercises this, so it stays green unmodified).
• LB-8 (unbounded result.json size, LOW): added a Deno.stat size check against a new internal MAX_RESULT_JSON_BYTES (50 MB) constant, run BEFORE Deno.readTextFile/JSON.parse ever touch the file — an oversized result.json now throws a clear "too large to import" Error instead of being read into memory and parsed regardless of size.
• LB-9 (leading-dash zipPath positionally ambiguous, LOW): a zipPath starting with - is now normalized to a ./-relative form (./${zipPath}) before being placed in unzip's argv — Info-ZIP's unzip does not honor a -- end-of-options marker, so a real unzip binary would otherwise positionally misread a leading-dash path as a flag. The already-closed command-injecti

2026.08.01.2

Adds an optional headless vaultRoot filesystem backend to import, so the import can run with the Obsidian desktop app closed (swamp-workspace #57; mirrors the CLI/filesystem backend split done for @magistr/obsidian-vault in PR #56 — see that PR for the path-confinement rationale). The Obsidian CLI (getVaultPath + runObsidian("create", ...)) is kept as the fallback for when vaultRoot is not set. Cross-reference: swamp-workspace#57.

• Added the vaultRoot global argument. When set, the vault path resolves to it directly (skipping the obsidian vault ... info=path CLI call), and the note is written with a confined atomic write instead of runObsidian("create", ...).
• Added resolveVaultPath/resolveVaultPathSafe (realpath + symlink refusal
  • .. rejection) and the atomic-write helpers (writeAtomic, ensureParentDir, chmodQuietly), copied VERBATIM (same names/comments, per the approved plan's scope constraint against a shared cross-extension module — swamp bundles each extension independently) from obsidian-vault/extensions/models/obsidian_vault.ts (PR #56). The note write now resolves through resolveVaultPathSafe before every mkdir/write, closing folder-traversal and symlink-escape vectors on the new headless path (the CLI fallback is untouched and keeps whatever behavior it already had).
• Upgraded isPathContained/safeCopyMedia (LB-1's extractDir confinement) from lexical-only to realpath-aware: after the existing lexical containment check passes, every existing path segment between extractDir's realpath and the candidate is walked with Deno.lstat, refusing to follow a symlink — closing the "symlink created inside extractDir that points outside it" residual the 2026.08.01.1 CHANGELOG entry documented as accepted. A segment that does not exist yet is not an error, same as before.
• No npm:yaml dependency was added — this model emits brand-new hand-built frontmatter into notes it owns, it never round-trips existing frontmatter, so PR #56's yaml-Document rationale does not apply here. Every hand-built frontmatter string stays byte-for-byte identical to before this change.
• Dot-dir/.trash exclusion is N/A: import writes into a caller-named folder, it never walks the vault tree (covered by a covered-negative test in the adversarial suite).
• Real-world behavior note: writing the note directly via Deno.writeTextFile (through the new atomic-write helper) may not be byte-identical to what the real Obsidian CLI's create command would have produced on disk — the CLI has never been observed to differ in this repo's tests (it's always stubbed), but a real obsidian create call could in principle normalize a trailing newline differently than a raw Deno.writeTextFile. Not reproduced or fixed here, just flagged.
• Extended all five test suites (contract-fixture, methods, adversarial, coverage, property-invariant-flow) with vaultRoot coverage: a golden fs-backend run against fixtures/basic/result.json, a method-level test proving the CLI is never invoked when vaultRoot is set (Deno.Command stubbed to throw on "obsidian"), a backend-selection precedence branch matrix, path-confinement adversarial tests (.. traversal and symlinked folder segment refused, /var-vs-/private/var real-root containment), and a property test asserting exactly one note per message with frontmatter round-trip and no path escaping the vault's real root, for any synthetic set of messages with unique ids. telegram_import_test_helpers.ts gained two new StubConfig options (realMkdir, throwOnObsidian) so these new tests can let Deno.mkdir run for real against a real Deno.makeTempDir vault while keeping Deno.Command/copyFile stubbed. No new committed fixture files were needed.
• manifest.yaml/model version: 2026.08.01.1 -> 2026.08.01.2.

Release 2026.07.16.2 — align model versions with manifests

Maintenance release across the @magistr extensions. For most packages this carries no functional change: the only edit is the model's version: field, brought back in line with its manifest version so the published model type version and the package version no longer drift.

Functional changes in this release are limited to:

• anime-cron: normalizeTitle now strips a ": subtitle" suffix and a trailing parenthesized year before comparison, fixing dedup false-misses where the torrent title carries a subtitle or year that the AniList romaji does not.

• arckit: first publish. Standalone ArcKit port — a 12-phase architecture governance state machine with 65 bundled templates, driven by a bundled skill.

Also tracks three extensions (kaiten, observability-agent, music-library) that previously existed only as untracked working-tree directories, recovered from stashes.

Merge pull request #4 from umag/extensions/magistr-grade-a-workspace

extensions: stage 15 @magistr extensions as Grade A workspace dirs + wire CI