The Context Pack (resume --pack)

memprove resume --pack                 # this Workspace's Task, budget 9000
memprove resume --pack --budget 3000
memprove resume --pack --query "truncated header"
memprove resume --pack --claims 3,5-7  # the full lines of those claims, no budget
memprove resume --text                 # one line per statement, for a person

The Pack is a fixed set of lines, in this order: Pack: (format, capsule_digest, ledger sequence, probed HEAD, budget), Ontology: (the version of the verdict rules), Query: (only with --query: the phrase and the matches by rank), Task「<title>」 (<state>): <verdict> with the reason codes and the counts of active, superseded and withdrawn statements, Integrity: (every record of the Task recomputes, or which one does not), Sessions: (how many in all, ended normally, ended abandoned, not ended, and the last part counting the Sessions that declared the Pack they continue from). Then a Session line for every Session that has not ended, with its start, its last write, its agent (; agent「<name>」(<role>[, dispatched by Session <j>])), continues from the Pack at ledger sequence N (self-reported, digest sha256:<first 12 hex>) when it declared one, and the takeover hint when it is not yours. The Session lines list every Session that has not ended, the latest Session that ended, and the latest three that ended abandoned (ended abandoned); Every other Session is only counted on the summary line. Then Last passed: (the latest passed test run, its commit and the arguments that record it again), Can:/Cannot: lines for completing the Task, beginning a Session and recording a fact on the latest run (with the Problem code a write would get), Base: lines (each commit facts are bound to and the files changed since), the claim lines (#n verified: 「…」 (evidence bound to commit …; from Session k), #n fact stale: …, #n open item: …, #n decision: …, #n next action: …), Rerun: (the command to run again on the current HEAD and the facts it would confirm) or No rerun needed:, Gaps:, Omitted: and Digest:.

• The budget counts UTF-16 code units of the whole output. The default is 9000; --budget takes a whole number from 1000 through 1000000. When the Pack is longer, it leaves things out in steps that stop as soon as it fits: level 1 writes only the number of changed files on each base line; level 2 shortens each body longer than 120 characters to its first 120 and says how many were left out; level 3 omits whole bodies; level 4 folds the verified active facts into one Folded: line with their numbers; level 5 stops listing claims one by one, facts first, then Decisions, then open items, and counts them on a Not listed: line. The next action is never shortened and the first, integrity, session, Can:/Cannot: and Omitted: lines are never left out. The Omitted: line says the level and how many bases, bodies and claims were affected. When even level 5 does not fit, nothing is printed and the command exits 1, and standard error names the least budget that fits.
• The last line is Digest: sha256:<pack_digest>, the SHA-256 of the UTF-8 bytes of every line before it. The Pack holds no time of its own, so the same ledger, probe and budget give the same bytes.
• Every text an agent wrote is between 「」, with control and format characters, line separators and the quote-like characters escaped, so a statement cannot forge a Pack line. Put the conclusion in the first sentence of a statement’s text: the Pack may keep only its first 120 characters.
• --pack --claims <NUMBERS> prints the full Pack lines of the claims named, superseded and withdrawn ones too, after the base lines they name, with no budget. --text --claims <NUMBERS> prints only the resume --text --all lines of the claims named, in full, superseded and withdrawn ones too. Both take numbers and ranges separated by commas (3,5-7).
• --query <TEXT> (1 to 256 characters) selects the statements that match the phrase, best first (an in-memory FTS5 index: terms of three or more characters by substring, shorter ones as whole words; the kind words fact, open_item, decision and next_action match statements of that kind), with what supersedes them and the next action in force; the Query: line gives the matches by rank, and 0 matches leaves the statement section empty.
• resume --text is the same record for a person: one line per statement (#n verified: 「…」 (evidence bound to commit X, current HEAD Y; verdict: verified; from Session k)), the Rerun: line with the listed argv (the env NAME=value prefix included; A run with a different argv is a different command), and one last line for the statements that a later one superseded or withdrew. That line counts each kind and names their numbers; resume --text --all lists each of those statements on its own line instead.
• In a tampered Pack the first Integrity: line adds claim numbers may differ from before, no line says verified, and Cannot: lines give evidence_stale.

From docs/guide.md at 2625064cbf0a · Markdown