Repository navigation
v0.9.0
Capshelf 0.9 makes every command that describes an mcp or subagents item
say which runtimes it covers. Either kind can carry a Claude source, a Codex
source, or both, and 0.8 printed one output path for all three cases, so a
Claude-only MCP server printed .mcp.json and said nothing about Codex. This
release also finishes moving add onto the pinned commit: fragments were the
last kind whose install target set came from the data repo's working tree while
apply took it from the lock. The lock format is unchanged — 0.8 and 0.9 read
each other's locks, and there is no migration; run capshelf update skills/capshelf once per project to pick up the refreshed bundled skill. Full
details in docs/whats-new-0.9.md.
Breaking changes
-
addinstalls every fragment target the pinned commit holds.It used to pick targets from the data repo's working tree while
applypicked
them from the lock, and the cleanliness gate cannot see a deletion that
removes a canonical path. Anmcpitem withcodex.tomldeleted in the
worktree was therefore accepted, pinned to a commit that still contained the
file, and installed with only.mcp.jsonwritten.A project that depended on that behavior will now get
.codex/config.toml
written where 0.8 wrote only.mcp.json. Commit the deletion in the data repo
if the target is genuinely retired. -
A canonical path committed as a directory or an empty tree is refused.
Both used to install, save a lock, and fail in a later
applywith a parser
error, leaving a broken lock already on disk. They are refused at pin time
now.
Enhancements
- Report runtime target coverage on
add,show,status, andshare: every candidate target marked present or absent, with the output path it feeds and the canonical source path a missing one reads, read at the pinned commit rather than the worktree. - Add
targetCoveragetoadd,show,share, and everystatus --jsonrow, carryingpresent: true | false | nullpluscoverageStatewhen unknown;sources,dst, and the subagenttargetsarray keep their existing shapes. - Cover target coverage in the bundled
capshelfskill, which guided agents through themcpand subagent workflows without mentioning the report those commands print.
Bug fixes
- Fire the Codex project-trust warning only when an
mcpitem's locked commit has a Codex source; it used to fire for everymcpitem, so a Claude-only item warned about the harness it does not reach and said nothing about the one it does not cover.codex-configitems keep the unconditional warning, and unknown coverage still warns — the gate removes a warning only on evidence. - Prove a pinned fragment source is readable before consent, so an unreadable object no longer plans as an empty contribution and locks as a covered target.
- Stop counting a committed symlink as a covered target; Git stores one as a
blobwith mode120000, so filtering on object type reported it as a present source for an item capshelf refuses to pin outright. - Report every output
rmreconciled instead of the first, so removing a two-sourcemcpitem names.codex/config.tomlas well as.mcp.json. - Fix
ls --hererenderingundefinedin the identity column for data rows, andadd --jsonandrevert --jsondropping the key entirely;ls --here --jsongainslockedSha, the namestatus --jsonalready uses. - Stop
showreporting(update available)for every item forever, including immediately after a cleanadd; it compared a content hash of the data repo's working tree against the locked pin digest, two schemes over two inputs that can never be equal.statusowns update detection and reported these items correctly all along. - Stop
statusexiting 1 for a whole report when one subagent's locked tree resolves but cannot be read. - Stop
addre-pinning after consent, so a data-repo commit landing in that window can no longer add a target the consent gate never saw. - Correct three user-facing strings:
--kindhelp omittedsubagents, thelockdescription promised an inspect command, and the "data repo not found" error named a default path capshelf never reads.