Repository navigation
feat(store): delete objects and remember when they were stored - #36
Open
PiquelChips wants to merge 1 commit into
Open
PiquelChips wants to merge 1 commit into
PiquelChips wants to merge 1 commit into
Conversation
PiquelChips
added
enhancement
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 29 minutes. View limit details
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. |
PiquelChips
force-pushed
the
feat/collection
branch
2 times, most recently
from
November 4, 2026 18:33
eaf10c5 to
dac8f33
Compare
PiquelChips
force-pushed
the
feat/collection
branch
8 times, most recently
from
November 4, 2026 22:26
59d3a74 to
df5ed59
Compare
PiquelChips
force-pushed
the
feat/collection
branch
from
November 8, 2026 11:34
df5ed59 to
d20b36a
Compare
PiquelChips
force-pushed
the
feat/collection
branch
2 times, most recently
from
November 8, 2026 13:57
0ebf8db to
3e79923
Compare
PiquelChips
force-pushed
the
feat/collection
branch
from
November 8, 2026 14:38
3e79923 to
ab59a96
Compare
PiquelChips
force-pushed
the
feat/collection
branch
from
November 8, 2026 16:36
ab59a96 to
e414669
Compare
PiquelChips
added a commit
that referenced
this pull request
Nov 8, 2026
Nothing could attach claims to revisions: CI results, reviews, a recorded abandonment. Nor could a result be carried over to a new revision whose inputs hadn't changed. **Commands:** - `origin annotate` issues a statement about a revision. It supersedes the user's previous statement of the same kind about that revision. Its details come inline with `-m "<text>"`, such as a review's message, or from a file with `--payload <file>`, where `-` reads standard input so a log can be piped in. `--withdraw` takes the user's statement back: it supersedes it with one that has no verdict, refused when there's nothing to take back, and `origin statements` shows it as withdrawn. `--on <statement>` makes another statement the subject instead of a revision, such as a `flaky` mark on a failed check, and `origin statements` lists each statement's own statements indented under it. - `origin statements` lists the statements about a revision. - `origin promote` carries promotable results over from records with the same input digest. Each one becomes a new statement whose `promoted_from` points at the original. **Rules:** - **The input digest** hashes what a snapshot stores at each declared path, plus the check's definition and toolchain. It costs one lookup per path. - **A result that ran there wins.** A promotion stops counting once the record has a statement of that kind from the original's issuer that isn't a promotion, a result or a withdrawal, whichever came first. `Statement::drop_replaced_promotions` holds the rule, so the server and plugins apply it alike. - **Promotions are checked against their original.** `Store::claim` only accepts one whose kind, verdict, inputs and payload match the original's, about a record with the same input digest. A promotion counts as its original's issuer. - **Supersession** only counts within one issuer. - **Reserved kinds have authors.** - `core.abandon`: only the change's creator, the author of its first revisions, or an admin from #6. - `core.conflict`: a node, or the revision's author. - `core.divergent`: a node. - Reserved kinds never promote. Deleting objects (#36) and payload retention with `origin gc` (#37) used to be part of this PR. They're now their own. **Benchmarks:** an input digest costs about 40 µs. Searching 20k statements for promotions takes 1.2 ms with the digest cache, or 41 ms without it. **Review fixes:** `@` names the working copy's own revision, a prefix matching both a change and another revision is ambiguous, `statements` lists a promotion whose original is missing, `new` on a stale revision flags the old change's divergence, and `abandon` checks authorization before writing anything. Proved with focused tests that fail without their fix, and `just` (332 tests). Reviewed with review-loop: 2 rounds with 2× GPT-6.1 Sol and Claude Opus 5.5. Verdict: approved **Withdrawals:** Reviewed with review-loop: 3 rounds with 2× GPT-6.1 Sol and Claude Opus 5.5. Verdict: approved **Statements about statements:** Reviewed with review-loop: 1 round with 2× GPT-6.1 Sol and Claude Opus 5.5. Verdict: approved > [!NOTE] > Written by Claude Opus 5.5 in Claude Code, reviewed by GPT-6.1 Sol and Claude Opus 5.5 over three review rounds (via T3 Code). 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- This is an auto-generated comment: release notes by coderabbit.ai --> <!-- coderabbit-ai-generated:v1 --> ## Annotation model and authorization - Added the public `annotations` module to issue, query, and promote statements. - Added `Statement.promoted_from` to identify the source statement for a promotion. Added checks for reserved kinds and statement promotability. - Added issuer-scoped supersession. A statement can supersede an earlier statement only from the same issuer. - Added authorization rules for reserved kinds: - `core.abandon` on a revision is limited to the creator of its change, based on the authors of its root revisions. - `core.conflict` can be issued by a node, or by the author of the revision it describes. A node can also issue it about a non-revision record. - `core.divergent` can be issued only by a node. - Reserved kinds cannot be promoted. - Added divergence flagging when a change has multiple heads. The generated statement uses a deterministic head and includes the sorted head IDs. - Added payload storage and retrieval support through `Store::try_blob`. ## Promotion and input digests - Added promotion discovery and validation. A promotion must reference an existing promotable statement and match its kind, verdict, inputs, and payload. The original and target records must have snapshots with matching digests for the declared input paths. - Added input-digest computation over sorted, unique input paths. The digest distinguishes missing entries from present entries and propagates lookup errors. - Added index support for promotion groups, statement supersession, and cached input digests. The index log format is now version 2. - Added benchmarks for input-digest calculation and promotion searches, including cached and uncached digest cases. ## CLI and workspace - Added `origin annotate`, `origin statements`, and `origin promote`. - `origin annotate` accepts a revision, kind, verdict, optional inputs, and optional payload. It can supersede the user’s current statement of the same kind for that revision. - `origin statements` lists statements and reports verdicts, supersession, promotion sources, inputs, and payload summaries. - `origin promote` writes eligible promotions for a revision. - Added revision-prefix lookup and updated the README usage examples. - Workspace saves now flag divergence. Abandonment listings now include only authorized `core.abandon` statements. ## Shared types and tests - Added `Kind::is_reserved`, `Statement::is_promotable`, `Statement::promotion_group`, and input specification and snapshot digest methods. - Added CLI integration tests for divergence, abandonment authorization, annotation, supersession, payloads, external inputs, and promotion behavior. - Updated statement fixtures and encoded-object expectations for `promoted_from`. <!-- end of auto-generated comment: release notes by coderabbit.ai -->
PiquelChips
force-pushed
the
feat/collection
branch
2 times, most recently
from
November 10, 2026 17:22
386905d to
eb80ceb
Compare
Nothing could delete an object, and nothing knew how long one had been stored. `Store::remove` deletes objects durably, loose and packed: each pack holding one is rewritten into a single new pack without them (`Store::remove_packed`, which shares `replace_packs` with compaction). `Store::stored_at` says when this node stored an object: its loose file's modification time, which pack indexes now record for each object so packing never resets it. Deleting must never race a writer that found an object already stored and skipped writing it. Such writers hold the store's objects (`Store::hold_objects`, a shared lock) from finding the object to writing what refers to it, as issuing a statement and snapshotting the working copy do. A collector takes the lock exclusively (`Store::lock_for_collection`), and a batch rereads the pack manifest before it trusts a pack to hold an object (`Batch::packed`), since a collection may have rewritten that pack without it. `Store::reach` finds every object some trees refer to, which a collector keeps.
PiquelChips
force-pushed
the
feat/collection
branch
from
November 13, 2026 22:25
eb80ceb to
cad0322
Compare
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
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.
Nothing could delete an object from the store, and nothing knew how long one had been stored. Payload retention, in the next PR, needs both. Deleting must also never race a writer that reuses an object it found already stored.
Deleting:
Store::removedeletes objects durably, loose and packed. Each pack holding one is rewritten into a single new pack without them (remove_packed, which sharesreplace_packswith compaction).Store::reachfinds every object some trees refer to, which is what a collector must keep.When an object was stored:
Store::stored_atis when this node first wrote the object: its loose file's modification time.Collection never races a writer:
Store::hold_objects, a shared lock) until it has written what refers to that object. Issuing a statement and snapshotting the working copy both do.Store::lock_for_collection).Batch::packed), since a collection may have rewritten that pack without it.This used to be part of #12. The tests that run a real collector against these locks come with the collector, in the next PR.
Review fixes: promotions, in
promoteand in the queue, hold the shared collection guard until they are published, so collection never deletes a payload a new promotion references. Proved with focused tests that fail without their fix, andjust(358 tests).Reviewed with review-loop: 3 rounds with 2× GPT-6.1 Sol and Claude Opus 5.5. Verdict: approved
Note
Written by Claude Opus 5.5 in Claude Code (via T3 Code)
🤖 Generated with Claude Code