IMAP Extensions Working Group                                 M. Crispin
Internet Draft: IMAP THREAD                                 K. Murchison
Document: internet-drafts/draft-ietf-imapext-thread-09.txt    March internet-drafts/draft-ietf-imapext-thread-10.txt     June 2002


Status of this Memo

   This document is an Internet-Draft and is in full conformance with
   all provisions of Section 10 of RFC 2026.

   Internet-Drafts are working documents of the Internet Engineering
   Task Force (IETF), its areas, and its working groups.  Note that
   other groups may also distribute working documents as Internet-

   Internet-Drafts are draft documents valid for a maximum of six months
   and may be updated, replaced, or obsoleted by other documents at any
   time.  It is inappropriate to use Internet-Drafts as reference
   material or to cite them other than as "work in progress."

   The list of current Internet-Drafts can be accessed at

   To view the list Internet-Draft Shadow Directories, see

   A revised version of this draft document will be submitted to the RFC
   editor as a Proposed Standard for the Internet Community.  Discussion
   and suggestions for improvement are requested, and should be sent to
   ietf-imapext@IMC.ORG.  This document will expire before 21 August 20 December
   2002.  Distribution of this memo is unlimited.


   This document describes the server-based threading extension to the
   IMAP4rev1 protocol.  This extension provides substantial performance
   improvements for IMAP clients which offer threaded views.

   A server which supports this extension indicates this with one or
   more capability names consisting of "THREAD=" followed by a supported
   threading algorithm name as described in this document.  This
   provides for future upwards-compatible extensions.

Extracted Subject Text

   Threading uses a version of the subject which has specific subject
   artifacts of deployed Internet mail software removed.  Due to the
   complexity of these artifacts, the formal syntax for the subject
   extraction rules is ambiguous.  The following procedure is followed
   to determine the actual "base subject" which is used to thread:

        (1) Convert any RFC 2047 encoded-words in the subject to
        UTF-8.  Convert all tabs and continuations to space.
        Convert all multiple spaces to a single space.

        (2) Remove all trailing text of the subject that matches
        the subj-trailer ABNF, repeat until no more matches are

        (3) Remove all prefix text of the subject that matches the
        subj-leader ABNF.

        (4) If there is prefix text of the subject that matches the
        subj-blob ABNF, and removing that prefix leaves a non-empty
        subj-base, then remove the prefix text.

        (5) Repeat (3) and (4) until no matches remain.

           Note: it is possible to defer step (2) until step (6),
           but this requires checking for subj-trailer in step (4).

        (6) If the resulting text begins with the subj-fwd-hdr ABNF
        and ends with the subj-fwd-trl ABNF, remove the
        subj-fwd-hdr and subj-fwd-trl and repeat from step (2).

        (7) The resulting text is the "base subject" used in

   All servers and disconnected clients MUST use exactly this algorithm
   when threading.  Otherwise there is potential for a user to get
   inconsistent results based on whether they are running in connected
   or disconnected IMAP mode.

Sent Date

   As used in this document, the term "sent date" refers to the date and
   time from the Date: header, adjusted by time zone.  This differs from
   date-related criteria in SEARCH, which use just the date and not the
   time, nor adjusts by time zone.

Additional Commands

   This command is an extension to the IMAP4rev1 base protocol.

   The section header is intended to correspond with where it would be
   located in the main document if it was part of the base

6.3.THREAD.     THREAD Command

   Arguments:  threading algorithm
               charset specification
               searching criteria (one or more)

   Data:       untagged responses: THREAD

   Result:     OK - thread completed
               NO - thread error: can't thread that charset or
               BAD - command unknown or arguments invalid

      The THREAD command is a variant of SEARCH with threading semantics
      for the results.  Thread has two arguments before the searching
      criteria argument; a threading algorithm, and the searching
      charset.  Note that unlike SEARCH, the searching charset argument
      is mandatory.

      There is also a UID THREAD command which corresponds to THREAD the
      way that UID SEARCH corresponds to SEARCH.

      The THREAD command first searches the mailbox for messages that
      match the given searching criteria using the charset argument for
      the interpretation of strings in the searching criteria.  It then
      returns the matching messages in an untagged THREAD response,
      threaded according to the specified threading algorithm.

      Sorting is in ascending order.  Earlier dates sort before later
      dates; smaller sizes sort before larger sizes; and strings are
      sorted according to ascending values established by their
      collation algorithm (see under "Internationalization

      The defined threading algorithms are as follows:

         The ORDEREDSUBJECT threading algorithm is also referred to as
         "poor man's threading."  The searched messages are sorted by
         subject and then by the sent date.  The messages are then split
         into separate threads, with each thread containing messages
         with the same extracted subject text.  Finally, the threads are
         sorted by the sent date of the first message in the thread.

         Note that each message in a thread is a child (as opposed to a
         sibling) of the previous message.

         The REFERENCES threading algorithm is based on the algorithm
         written by Jamie Zawinski which was used in "Netscape Mail and
         News" versions 2.0 through 3.0.  For details, see

         This algorithm threads the searched messages by grouping them
         together in parent/child relationships based on which messages
         are replies to others.  The parent/child relationships are
         built using two methods: reconstructing a message's ancestry
         using the references contained within it; and checking the
         subject of a message to see if it is a reply to (or forward of)

            Note: "Message ID" in the following description refers to a
            normalized form of the msg-id in [RFC 2822].  The actual
            text in an RFC 2822 may use quoting, resulting in multiple
            ways of expressing the same Message ID.  Implementations of
            the REFERENCES threading algorithm MUST normalize any msg-id
            in order to avoid false non-matches due to differences in

            For example, the msg-id
            and the msg-id
            MUST be interpreted as being the same Message ID.

         The references used for reconstructing a message's ancestry are
         found using the following rules:

            If a message contains a [NEWS]-style References header line,
            then use the Message IDs in the References header line as
            the references.

            If a message does not contain a References header line, or
            the References header line does not contain any valid
            Message IDs, then use the first (if any) valid Message ID
            found in the In-Reply-To header line as the only reference
            (parent) for this message.

               Note: Although [RFC 2822] permits multiple Message IDs in
               the In-Reply-To header, in actual practice this
               discipline has not been followed.  For example,
               In-Reply-To headers have been observed with email
               addresses after the Message ID, and there are no good
               heuristics for software to determine the difference.
               This is not a problem with the References header however.

            If a message does not contain an In-Reply-To header line, or
            the In-Reply-To header line does not contain a valid Message
            ID, then the message does not have any references (NIL).

         A message is considered to be a reply or forward the subject
         extraction rules, applied to the original subject, remove any
         of the following: a subj-refwd, a "(fwd)" subj-trailer, or a
         subj-fwd-hdr and subj-fwd-trl.

         The REFERENCES algorithm is significantly more complex than
         ORDEREDSUBJECT and consists of six main steps.  These steps are
         outlined in detail below.

         (1) For each searched message:

            (A) Using the Message IDs in the message's references, link
            the corresponding messages (those whose Message-ID header
            line contains the given reference Message ID) together as
            parent/child.  Make the first reference the parent of the
            second (and the second a child of the first), the second the
            parent of the third (and the third a child of the second),
            etc.  The following rules govern the creation of these

               If a message does not contain a Message-ID header line,
               or the Message-ID header line does not contain a valid
               Message ID, then assign a unique Message ID to this

               If two or more messages have the same Message ID, assign
               a unique Message ID to each of the duplicates.

               If no message can be found with a given Message ID,
               create a dummy message with this ID.  Use this dummy
               message for all subsequent references to this ID.

               If a message already has a parent, don't change the
               existing link.  This is done because the References
               header line may have been truncated by a MUA.  As a
               result, there is no guarantee that the messages
               corresponding to adjacent Message IDs in the References
               header line are parent and child.

               Do not create a parent/child link if creating that link
               would introduce a loop.  For example, before making
               message A the parent of B, make sure that A is not a
               descendent of B.

                  Note: Message ID comparisons are case-sensitive.

            (B) Create a parent/child link between the last reference
            (or NIL if there are no references) and the current message.
            If the current message already has a parent, it is probably
            the result of a truncated References header line, so break
            the current parent/child link before creating the new
            correct one.  As in step 1.A, do not create the parent/child
            link if creating that link would introduce a loop.  Note
            that if this message has no references, that it will now
            have no parent.

               Note: The parent/child links created in steps 1.A and 1.B
               MUST be kept consistent with one another at ALL times.

         (2) Gather together all of the messages that have no parents
         and make them all children (siblings of one another) of a dummy
         parent (the "root").  These messages constitute the first
         (head) message of the threads created thus far.

         (3) Prune dummy messages from the thread tree.  Traverse each
         thread under the root, and for each message:

            If it is a dummy message with NO children, delete it.

            If it is a dummy message with children, delete it, but
            promote its children to the current level.  In other words,
            splice them in with the dummy's siblings.

            Do not promote the children if doing so would make them
            children of the root, unless there is only one child.

         (4) Sort the messages under the root (top-level siblings only)
         by sent date.  In the case of an exact match on sent date or if
         either of the Date: headers used in a comparison can not be
         parsed, use the order in which the messages appear in the
         mailbox (that is, by sequence number) to determine the order.
         In the case of a dummy message, sort its children by sent date
         and then use the first child for the top-level sort.

         (5) Gather together messages under the root that have the same
         extracted subject text.

            (A) Create a table for associating extracted subjects with

            (B) Populate the subject table with one message per
            extracted subject.  For each child of the root:

               (i) Find the subject of this thread by extracting the
               base subject from the current message, or its first child
               if the current message is a dummy.

               (ii) If the extracted subject is empty, skip this

               (iii) Lookup the message associated with this extracted
               subject in the table.

               (iv) If there is no message in the table with this
               subject, add the current message and the extracted
               subject to the subject table.

               Otherwise, replace the message in the table with the
               current message if the message in the table is not a
               dummy AND either of the following criteria are true:

                  The current message is a dummy, OR

                  The message in the table is a reply or forward (its
                  original subject contains a subj-refwd part and/or a
                  "(fwd)" subj-trailer) and the
                  current message is not.

            (C) Merge threads with the same subject.  For each child of
            the root:

               (i) Find the subject of this thread as in step 4.B.i 5.B.i

               (ii) If the extracted subject is empty, skip this

               (iii) Lookup the message associated with this extracted
               subject in the table.

               (iv) If the message in the table is the current message,
               skip this message.

               Otherwise, merge the current message with the one in the
               table using the following rules:

                  If both messages are dummies, append the current
                  message's children to the children of the message in
                  the table (the children of both messages become
                  siblings), and then delete the current message.

                  If the message in the table is a dummy and the current
                  message is not, make the current message a child of
                  the message in the table (a sibling of it's children).

                  If the current message is a reply or forward and the
                  message in the table is not, make the current message
                  a child of the message in the table (a sibling of it's

                  Otherwise, create a new dummy message and make both
                  the current message and the message in the table
                  children of the dummy.  Then replace the message in
                  the table with the dummy message.

                     Note: Subject comparisons are case-insensitive, as
                     described under "Internationalization

         (6) Traverse the messages under the root and sort each set of
         siblings by sent date.  Traverse the messages in such a way
         that the "youngest" set of siblings are sorted first, and the
         "oldest" set of siblings are sorted last (grandchildren are
         sorted before children, etc).  In the case of an exact match on
         sent date or if either of the Date: headers used in a
         comparison can not be parsed, use the order in which the
         messages appear in the mailbox (that is, by sequence number) to
         determine the order.  In the case of a dummy message (which can
         only occur with top-level siblings), use its first child for

               S: * THREAD (166)(167)(168)(169)(172)(170)(171)
                  (173)(174 175 176 178 181 180)(179)(177 183
                   182 188 184 185 186 187 189)(190)(191)(192)
                  (193)(194 195)(196 197 198)(199)(200 202)(201)
                  (203)(204)(205)(206 207)(208)
               S: A283 OK THREAD completed
               S: * THREAD
               S: A284 OK THREAD completed
               C: A285 THREAD REFERENCES UTF-8 SINCE 5-MAR-2000
               S: * THREAD (166)(167)(168)(169)(172)((170)(179))
                  ((177)(183)(182)(188 (184)(189))(185 186)(187))
                  (190)(191)(192)(193)((194)(195 196))(197 198)
                  (199)(200 202)(201)(203)(204)(205 206 207)(208)
               S: A285 OK THREAD completed

        Note: The line breaks in the first and third client
        responses are for editorial clarity and do not appear in
        real THREAD responses.

Additional Responses

   This response is an extension to the IMAP4rev1 base protocol.

   The section heading of this response is intended to correspond with
   where it would be located in the main document.

7.2.THREAD.     THREAD Response

   Data:       zero or more threads

      The THREAD response occurs as a result of a THREAD or UID THREAD
      command.  It contains zero or more threads.  A thread consists of
      a parenthesized list of thread members.

      Thread members consist of zero or more message numbers, delimited
      by spaces, indicating successive parent and child.  This continues
      until the thread splits into multiple sub-threads, at which point
      the thread nests into multiple sub-threads with the first member
      of each subthread being siblings at this level.  There is no limit
      to the nesting of threads.

      The messages numbers refer to those messages that match the search
      criteria.  For THREAD, these are message sequence numbers; for UID
      THREAD, these are unique identifiers.

   Example:    S: * THREAD (2)(3 6 (4 23)(44 7 96))

      The first thread consists only of message 2.  The second thread
      consists of the messages 3 (parent) and 6 (child), after which it
      splits into two subthreads; the first of which contains messages 4
      (child of 6, sibling of 44) and 23 (child of 4), and the second of
      which contains messages 44 (child of 6, sibling of 4), 7 (child of
      44), and 96 (child of 7).  Since some later messages are parents
      of earlier messages, the messages were probably moved from some
      other mailbox at different times.

      -- 2

      -- 3
         \-- 6
             |-- 4
             |   \-- 23
             \-- 44
                  \-- 7
                      \-- 96
   Example:    S: * THREAD ((3)(5))

      In this example, 3 and 5 are siblings of a parent which does not
      match the search criteria (and/or does not exist in the mailbox);
      however they are members of the same thread.

Formal Syntax of THREAD commands and Responses

   thread-data       = "THREAD" [SP 1*thread-list]

   thread-list       = "(" thread-members / thread-nested ")"

   thread-members    = nz-number *(SP nz-number) [SP thread-nested]

   thread-nested     = 2*thread-list

   thread            = ["UID" SP] "THREAD" SP thread-algorithm
                       SP search-charset 1*(SP search-key)

   thread-algorithm  = "ORDEREDSUBJECT" / "REFERENCES" / atom

   The following syntax describes subject extraction rules (2)-(6):

   subject         = *subj-leader [subj-middle] *subj-trailer

   subj-refwd      = ("re" / ("fw" ["d"])) *WSP [subj-blob] ":"

   subj-blob       = "[" *BLOBCHAR "]" *WSP

   subj-fwd        = subj-fwd-hdr subject subj-fwd-trl

   subj-fwd-hdr    = "[fwd:"

   subj-fwd-trl    = "]"

   subj-leader     = (*subj-blob subj-refwd) / WSP

   subj-middle     = *subj-blob (subj-base / subj-fwd)
                   ; last subj-blob is subj-base if subj-base would
                   ; otherwise be empty

   subj-trailer    = "(fwd)" / WSP

   subj-base       = NONWSP *([*WSP] NONWSP)
                   ; can be a subj-blob

   BLOBCHAR        = %x01-5a / %x5c / %x5e-7f
                   ; any CHAR except '[' and ']'

   NONWSP          = %x01-08 / %x0a-1f / %x21-7f
                   ; any CHAR other than WSP

Security Considerations

   Security issues are not discussed in this memo.

Internationalization Considerations

   By default, strings are threaded according to the "minimum sorting
   collation algorithm".  All implementations of THREAD MUST implement
   the minimum sorting collation algorithm.

   In the minimum sorting collation algorithm, the Basic Latin
   alphabetics (U+0041 to U+005A uppercase, U+0061 to U+007A lowercase)
   are sorted in a case-insensitive fashion; that is, "A" (U+0041) and
   "a" (U+0061) are treated as exact equals.  The characters U+005B to
   U+0060 are sorted after the Basic Latin alphabetics; for example,
   U+005E is sorted after U+005A and U+007A.  All other characters are
   sorted according to their octet values, as expressed in UTF-8.  No
   attempt is made to treat composed characters specially, or to do
   case-insensitive comparisons of composed characters.

        Note: this means, among other things, that the composed
        characters in the Latin-1 Supplement are not compared in
        what would be considered an ISO 8859-1 "case-insensitive"
        fashion.  Case comparison rules for characters with
        diacriticals differ between languages; the minimum sorting
        collation does not attempt to deal with this at all.  This
        is reserved for other sorting collations, which may be

   ;;;   *** ITEM FOR DISCUSSION ***
   Other sorting collations, and the ability to change the sorting
   collation, will be defined in a separate document dealing with IMAP

   It is anticipated that there will be a generic Unicode sorting
   collation, which will provide generic case-insensitivity for
   alphabetic scripts, specification of composed character handling, and
   language-specific sorting collations.  A server which implements
   non-default sorting collations will modify its sorting behavior
   according to the selected sorting collation.

   Non-English translations of "Re" or "Fw"/"Fwd" are not specified for
   removal in the extracted subject text process.  By specifying that
   only the English forms of the prefixes are used, it becomes a simple
   display time task to localize the prefix language for the user.  If,
   on the other hand, prefixes in multiple languages are permitted, the
   result is a geometrically complex, and ultimately unimplementable,
   task.  In order to improve the ability to support non-English display
   in Internet mail clients, only the English form of these prefixes
   should be transmitted in Internet mail messages.

A.      References

   [ABNF] Crocker, D., and Overell, P. "Augmented BNF for Syntax
   Specifications: ABNF", RFC 2234, November 1997.

   [NEWS] Horton, M., and Adams, R., "Standard for interchange of USENET
   messages", RFC-1036, AT&T Bell Laboratories and Center for Seismic
   Studies, December, 1987.

   [RFC-2822] Resnick, P. "Internet Message Format", RFC 2822, April

Author's Address

   Mark R. Crispin
   Networks and Distributed Computing
   University of Washington
   4545 15th Avenue NE
   Seattle, WA  98105-4527

   Phone: (206) 543-5762

   EMail: MRC@CAC.Washington.EDU

   Kenneth Murchison
   Oceana Matrix Ltd.
   21 Princeton Place
   Orchard Park, NY 14127

   Phone: (716) 662-8973 x26