fix(cursor): migrator path-priority + ambiguity surface (closes #844 regression) (#877)

* fix(cursor): migrator path-priority + ambiguity surface (closes #844 regression)

The legacy-to-ACP migrator's `findLegacyChatStore()` walks
`~/.cursor/chats/<workspace-hash>/<cursorSessionId>/store.db` via
`readdirSync()` and returns the FIRST match. When the same cursor
session id exists in more than one workspace-hash drawer (operator
opened the session from a worktree, an old workspace clone, etc.)
the readdir order picks an arbitrary candidate. The migrator then
transplants alien content into the ACP target, deletes the source
drawer, and reports success - because the verify probe only checks
"loads cleanly", not "loaded the right content". Operator session
resurrects with no recall of its real history.

Four-part fix (all four must land together):

1. Path-priority discovery in `findLegacyChatStore(id, home, cwd?)`:
   - Optional 3rd arg = canonical workspace path (caller passes
     `session.metadata.path`).
   - Compute md5(cwd) and check that drawer FIRST.
   - Fall back to readdir scan only if the canonical drawer is empty.
   - If 2+ candidates remain after fallback, throw
     `AmbiguousLegacyStoreError` listing all of them
     (workspaceHash, sizeBytes, mtimeMs).
2. Ambiguity surface in `maybeAutoMigrateLegacyCursorSession`:
   - Catch `ambiguous_legacy_store` / `size_mismatch` refusals and
     promote `cursorMigrationState` from 'in_progress' to a new
     'ambiguous' state instead of silently clearing the banner.
     Operator sees an actionable web-banner.
3. Size sanity check before transplant:
   - Compare HAPI's known message count (new `MessageStore.countMessages`
     + `CursorLegacyMigratorDeps.getHapiMessageCount` dep) against
     the candidate `store.db`'s blob count. If message count > 100
     AND blob count < messageCount/4, refuse with `size_mismatch`.
   - Skipped when message count is 0 (brand-new session) or the dep
     is unwired (unit tests, CLI direct callers).
4. Diagnostic logging on every successful transplant:
   - `[migrator] transplanted` info log capturing cursorSessionId,
     picked workspaceHash, candidate count discovered, sourceBytes,
     sourceBlobCount, targetAcpPath, sourceRemoved, canonical-path
     md5. Future regressions of this bug shape are diagnosable from
     `journalctl -u hapi-hub` without blob-overlap forensics.

Tests added in `hub/src/cursor/cursorLegacyMigrator.test.ts`:
  - regression guard for single-drawer discovery
  - canonical-path wins over readdir order
  - ambiguity throws with all candidates listed (3-drawer + 2-drawer
    no-canonical-arg variants)
  - canonical-path resolves ambiguity cleanly
  - listLegacyChatStoreCandidates enumeration
  - workspaceHashFromPath shape
  - migrateOne happy path with canonical workspace + 3 sibling decoys
  - migrateOne refuses with ambiguous_legacy_store (3 drawers, no
    canonical match) and leaves all sources untouched
  - migrateOne proceeds when canonical path resolves
  - size_mismatch refuses tiny candidate when messageCount=6000
  - size_mismatch passes when candidate blob count meets the floor
  - size sanity skipped on messageCount=0, missing dep, throwing dep,
    boundary (messageCount=100)
  - countLegacyStoreBlobs returns counts / null on bad path
And in `hub/src/sync/syncEngineAutoMigrate.test.ts`:
  - cursorMigrationState promoted to 'ambiguous' on
    ambiguous_legacy_store / size_mismatch refusals.

Schema:
  - `shared/src/schemas.ts`: cursorMigrationState enum gains 'ambiguous'.
  - `shared/src/apiTypes.ts`: CursorMigrateRefusalReason gains
    'ambiguous_legacy_store' + 'size_mismatch'.

Real-world repro (operator's tooling session, 2026-06-09): three legacy
drawers contained one cursor session id - one with the real 21k-blob
history, two with stale 19/568-blob diagnostic snapshots. Migrator
silently transplanted the 568-blob alien content; resurrected session
had no memory of prior history. Manual rescue completed; this fix
prevents recurrence and surfaces the ambiguity to the operator instead.

* fix(cursor): address cold review on migrator path-priority fix

Self-review against the cold-PR rubric surfaces four polish items on
the previous commit; all four addressed in-loop before push.

- Major: `migrator:transplanted` candidate count was captured AFTER
  the source rm, so for the dominant single-candidate happy path the
  log reported `candidateCount=0, sourceRemoved=true`. Useless for
  diagnosing a future regression of the bug shape this PR is fixing.
  Snapshot candidates + source-side size + source-side blob count
  BEFORE any destructive step and use those for the log.
- Minor: `sourceBytes` and `sourceBlobCount` were read from the
  destination path (acpSessionDir/store.db). The cp guarantees they
  match, but the field names imply source-side measurement. Now they
  measure the source directly.
- Minor: `setCursorMigrationStateAmbiguous` silently returned false on
  cache miss / repeated version mismatch / write failure, letting the
  finally{} block clear the banner without any log. Now emits a
  warn-level log so the gap is diagnosable from journalctl.
- Minor: `findLegacyChatStore` is exported public API and used as a
  free function in unit tests. An out-of-band caller bypassing
  preflightSession could pass `..` or `/etc/passwd` and have the inner
  `join(chatsRoot, wsh, id, 'store.db')` resolve to an arbitrary on-
  disk path. The probe is read-only `statSync` so blast radius is
  small, but enforce the same CURSOR_SESSION_ID_RE at the function
  boundary as a defence-in-depth. New unit test locks the behaviour.

Hub test suite: 414 pass, 0 fail. Typecheck clean across cli/web/hub.

* fix(cursor): cold-review polish on migrator path-priority (tiann/hapi#873)

- Web `CursorMigrationBanner` now renders a "Manual review needed"
  state for `cursorMigrationState === 'ambiguous'` (Major #1: caller
  was promoting the metadata flag but no UI surfaced it).
- Pin the md5-fixture contract for `workspaceHashFromPath`: raw,
  no-normalization, trailing-slash-distinct hashes computed via
  `printf '%s' <path> | md5sum` (Major #2: prevents algorithm drift
  that would silently revert path-priority discovery to fallback).
- Snapshot full candidate set BEFORE the canonical fast-path resolves
  a single drawer so the `migrator:transplanted` log reports the
  decision-time count, not a post-rm undercount (Minor #1).
- Warn log when canonical-path drawer is missing but readdir hands
  back exactly one candidate - regression-equivalent behaviour, but
  the size mismatch warrants a journalctl trail (path-normalization
  corner case the maintainer can grep for).
- Boundary test: `messageCount = 101` (first value above the skip
  threshold) engages the size sanity check, pinning the cutoff
  contract (Nit).
- Schema docstring on `cursorMigrationState` enum spelling out the
  banner contract per value (Nit).
- syncEngine `getHapiMessageCount` warn-logs `countMessages` throws
  instead of silently downgrading to 0 (would chronically disable
  the floor).

Drafted with claude-4.6-sonnet-thinking via Cursor; reviewed and
tested by the operator. tiann/hapi#873.

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(cursor): correct log-search strings in ambiguous banner copy

The en/zh-CN locale strings told users to grep for
'migrator:ambiguous_legacy_store' and 'migrator:size_mismatch'
but the hub emits '[migrator] ambiguous legacy store; refusing
transplant' and '[migrator] size sanity check refused transplant'.

Fix both locale files to quote the actual log prefix so the
journalctl grep the operator is directed to actually hits.

Addresses tiann/hapi#877 bot finding (Minor).

Co-authored-by: Cursor <cursoragent@cursor.com>

* fix(cursor): address #877 bot Minor findings (trim + boundary guard)

- Remove .trim() from canonical path before hashing: Cursor hashes
  raw workspace-path bytes; trimming a POSIX path with leading/
  trailing spaces would hash to the wrong drawer, causing a false
  canonical miss and potential ambiguity refusal.

- Add CURSOR_SESSION_ID_RE guard to listLegacyChatStoreCandidates:
  the function was exported without the same traversal-ID boundary
  check present in findLegacyChatStore. A future direct caller
  bypassing findLegacyChatStore could stat paths outside the intended
  <wsh>/<cursorSessionId>/store.db shape.

- Move CURSOR_SESSION_ID_RE declaration above both functions that
  reference it so there is no temporal-dead-zone hazard.

Addresses tiann/hapi#877 bot review Minor findings.

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
HeavyGee
2026-06-11 09:11:58 +08:00
committed by GitHub
co-authored by Cursor
parent 434cd9021d
commit 3e2e48222a
12 changed files with 932 additions and 43 deletions
+2
View File
@@ -179,6 +179,8 @@ export type CursorMigrateRefusalReason =
| 'session_resumed_during_migrate'
| 'legacy_store_modified_during_migrate'
| 'cross_host_session'
| 'ambiguous_legacy_store'
| 'size_mismatch'
| 'internal_error'
export const UploadFileRequestSchema = z.object({
+7 -1
View File
@@ -39,7 +39,13 @@ export const MetadataSchema = z.object({
opencodeSessionId: z.string().optional(),
cursorSessionId: z.string().optional(),
cursorSessionProtocol: z.enum(['acp', 'stream-json']).optional(),
cursorMigrationState: z.enum(['in_progress']).optional(),
// Drives the web `CursorMigrationBanner`:
// 'in_progress' = legacy-to-ACP transplant currently running; banner shows spinner + "Upgrading..."
// 'ambiguous' = migrator refused to transplant (ambiguous source drawer OR size mismatch);
// banner switches to "Manual review needed" until the operator resolves on disk.
// undefined = no migration in flight; banner hidden.
// tiann/hapi#873.
cursorMigrationState: z.enum(['in_progress', 'ambiguous']).optional(),
kimiSessionId: z.string().optional(),
tools: z.array(z.string()).optional(),
slashCommands: z.array(z.string()).optional(),