fix(cli): a run pulled from a gateway can be forked, and says so when it cannot - #89
Merged
Merged
Conversation
… it cannot
`orca pull` never brings the workspace snapshots down. A gateway archive carries
manifest.json, events.jsonl, redactions.json and blobs/ — the allowlist both
sides implement — and never the `fs/` shadow store, so a pulled run's
`fs.snapshot` events name trees whose objects were never sent.
The fork path restored from those trees unconditionally and died inside git:
git read-tree failed (128): fatal: failed to unpack tree object 4ec8e609ea10…
which names neither the cause nor anything to do about it. That is the exact
path the gateway console prints under "fork it locally": `orca pull <run>`, then
`orca compare last --from N`.
`replay.ts` already guarded the neighbouring mistake — restoring from the WRONG
store — and said so in a comment. It could not guard against there being no
store at all, because nothing asked.
- `ShadowIndex.has(tree)` answers it without writing a byte, the same way
`gitlinks` already answers "would this restore refuse?" before it refuses.
- Missing tree now reports `fork.no_snapshot` with the tree, the store, the
cause, where the run was recorded, and the command that works — then fails
with one line, because `compare` renders each leg's error in one table cell
and anything past the first line is lost exactly where it is most read.
- `--no-fs` now covers the restore as well as the fork's own capture, so
"the conversation is what I am forking" is sayable. It was undocumented on
`replay`, absent from `compare`'s allowlist, and dropped from the argv
`compare` builds per fork — the silent no-op that list already warns about
twice, for two spellings of the same fork.
Verified against the live gateway, three rounds of record → push → pull →
compare: the default reports `fork.no_snapshot` and refuses, `--no-fs` passes.
Forking in the directory that still has the store restores the workspace as
before. A run whose checkpoint is git's empty tree needs none of this — the
empty tree resolves anywhere — so a conversation-only run pulls and forks with
no friction at all.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
![orcacode-review[bot]](https://avatars.githubusercontent.com/in/4254738?s=60&v=4){kind=small}
There was a problem hiding this comment.
🐳 OrcaCode Review
✅ No findings — nothing to flag in this PR. Great work!
OrcaCode Review — Route Smarter. Ship Safer. Spend Less.
Engine-reported: 764 calls · 81.9M tokens · 99% cached
❤️ Share · Install OrcaCode Review
Free on GitHub — the review runs on your own OrcaRouter key. If it helped, a shout-out goes a long way.
Share: X · Reddit · LinkedIn
Follow: X · Discord · LinkedIn · OrcaRouter
`prettier --check .` is the first step of the CI check job, and the argv array in the new test was written compact where prettier wants one element per line. Nothing else changed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
There was a problem hiding this comment.
🐳 OrcaCode Review
✅ No findings — nothing to flag in this PR. Great work!
OrcaCode Review — Route Smarter. Ship Safer. Spend Less.
Engine-reported: 521 calls · 45.1M tokens · 99% cached
❤️ Share · Install OrcaCode Review
Free on GitHub — the review runs on your own OrcaRouter key. If it helped, a shout-out goes a long way.
Share: X · Reddit · LinkedIn
Follow: X · Discord · LinkedIn · OrcaRouter
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Orca-Code-Review — push 2
✅ no blocking findings
The gateway console prints this under Compare from step → "fork it locally", with the
promise "fork from this step with the same files and conversation prefix":
Copy-pasted verbatim, the second command dies:
Why
orca pullnever brings the workspace snapshots down, and cannot. The archive allowlist ismanifest.json,events.jsonl,redactions.jsonandblobs/—PUSHED_TOP_LEVELinsync.ts, and the same set the gateway enforces. I checked the gateway's half by pushinghand-built archives one member at a time:
fs/objects/<ab>/<hex>"…/fs/objects/…" is not part of a trace archivefs/HEAD,fs/config,fs/hooks/…,fs/info/excludeshell-frames.jsonlSo a pulled run's
fs.snapshotevents name trees whose objects were never sent, and therestore had no way to know that.
replay.tsalready guarded the neighbouring mistake — restoring from the wrong store — andsays so in a comment right above the call. It could not guard against there being no store at
all, because nothing asked.
What this changes
ShadowIndex.has(tree)answers it without writing a byte, the waygitlinksalreadyanswers "would this restore refuse?" before it refuses.
fork.no_snapshot— the tree, the store, the cause, where the runwas recorded, and the command that works — then fails with one line. The detail goes in the
warning because
comparerenders each leg's error in one cell of a table, and anything pastthe first line is lost exactly where it is most likely to be read.
--no-fsnow covers the restore, not just the fork's own capture, so "the conversation iswhat I am forking" is sayable. It was undocumented on
replay, missing fromcompare'sallowlist, and dropped from the argv
comparebuilds per fork — the silent no-op that listalready warns about twice, for two spellings of the same fork.
Verification
Against the live gateway, three rounds of record → push → pull → compare, with a non-empty
workspace so the checkpoint names a real tree: the default reports
fork.no_snapshotandrefuses 3/3,
--no-fspasses 3/3. Forking in the directory that still has the store restoresnotes.mdandsrc/mod.jsinto the worktree exactly as before.A run whose checkpoint is git's empty tree needs none of this — the empty tree resolves in any
repository — so a conversation-only run pulls and forks with no friction, which is the common
case and now measurably so.
The new test deletes the run's
fs/store, which is what the wire leaves behind. Without thischange it fails with the original
failed to unpack tree object; with it, both branches pass.Full suite on Windows: 10 failed / 2414 passed / 40 skipped — the same 10 pre-existing platform
failures as
main, one more passing test.compare.test.ts > forks every model from the SAME parent runis one of those 10 and fails identically with this change stashed.Not in scope
orca push --fsstill packsfs/and the gateway still refuses it —--fshas been unable tocomplete since it was added in #35, and was unreachable behind the flag allowlist until #88 made
it callable. Worth settling separately: either the wire learns to carry snapshots, or
--fsgoes,or push translates the rejection. It is not what the console's "fork it locally" path needs.
🤖 Generated with Claude Code