Repository navigation
Releases
v1.0.0-pre.19
Nineteenth dogfood pre-release. It fixes data-loss bugs in the replay of uncommitted trunk edits and in maw ws recover --restore-file; the worst, found by the codex sweep, can delete files outside the repository and is in pre.18 and earlier. It also makes uncommitted trunk edits survive more merge and crash paths, and restarts the SG1 soak on a stricter harness. Upgrading is recommended. One behaviour change: undo is now a reserved workspace name.
Data loss fixed
- Replaying uncommitted trunk edits no longer deletes or writes files outside the repository (bn-3u5a4, found by the codex sweep). Suppose a merge replaced a tracked directory
dwith a symlink to a directory outside the repo, and your uncommitted state deletedd/x. The replay then deletedxin the symlink's target. Writes could follow a symlinked parent the same way. Every replay path is now checked component by component: a symlinked parent makes the replay refuse, and your state stays pinned and reported. This bug is in pre.18 and earlier. maw ws recover --restore-file <dir>no longer overwrites an ignored file in the destination without--force(bn-3u5a4). Status omits ignored files, so the check missed them. With--force, it no longer deletes an uncommitted symlink nested inside a directory in the way.- A crash inside the snapshot-failed fallback no longer leaves your edits only in a recovery ref (bn-1eg2u). The fallback now writes a checkout intent before its forced checkout, so
maw ws merge --recoverreplays your edits.
Uncommitted trunk edits across merges
- The snapshot-failed fallback now merges your edits instead of keeping only the merged version (bn-2ds48). It replays your pinned state through the same path a crash recovery uses. When you and the merge edit different hunks of one file, both survive; the same hunk gets conflict markers plus a report. Before, only a crash recovery merged them.
- An uncommitted file↔directory swap on a path the merge didn't touch stays in place (bn-ihi4h). Before, it was always replaced by the merged version and reported. The replay also applies deletions before additions now, and removes a directory only if it holds nothing but empty directories.
- Directory conflicts have a one-step restore (bn-1eg2u).
maw ws recover --ref <pin> --restore-file <dir>restores a whole directory, untracked files included. It removes the merged file or directory in its way only if that content is committed and clean, unless you pass--force. Conflict messages now say "replaced by file p" instead of "deleted". - A merge that only flips +x on a file you replaced with a symlink is reported as a type conflict (bn-ihi4h). The printed restore command puts your symlink back.
- The replay's filtered snapshot is pinned at
refs/manifold/replay/<ws>while the replay runs (bn-1eg2u), so a concurrentgit gc/prune can no longer remove it mid-merge. - Every restore command printed after a failed replay runs as printed (bn-7xuvm, bn-3u5a4). The broken
git stash apply <oid>hint is gone; maw prints one command per path, and one per file↔directory swap.
Recovery snapshots and workspace names
maw gc --recovery-snapshotskeeps themaw undopin that a redo still needs (bn-43x5k).refs/manifold/recovery/undo/*pins are no longer treated as a destroyed workspace namedundo. The pin the next redo re-applies is protected like a live workspace's pin: only--include-liveconsiders it, dropping it needs--force, and the refusal listing marks itUNDO. If the op log can't be read, every undo pin is protected.--include-live --forceno longer leaves a destroy record that points at a deleted ref (bn-43x5k). When a workspace name was reused, the old record is removed in the same pass.undois a reserved workspace name (bn-asqh7).maw ws create undoandmaw ws recover … --to undorefuse and suggest another name.maw doctorwarns about an existingundoworkspace, and every command in its rename hint runs as printed.
Agent guidance
- The examples in
maw --help,maw ws merge --help,maw tldrandmaw agents showrun non-interactively as written (bn-39nsp). They use--from main, pass--messageon every landing merge, resolve conflicts withmaw ws resolve --list/--keep, and say that an orchestrator (e.g. edict) may run the merge and that the lead pushes. A test runs them with stdin closed. - Suggested merge commands in hints and JSON fields (
recommended_command,recommended_action) include--message "<msg>"(bn-hfge7). Replace the placeholder with a real message.
Verification (the v1.0 gate)
- The SG1 soak's checks judge more and are stricter (bn-36chi, bn-ihi4h, bn-1axaz).
- A pre.18 soak seed showed that recoveries were settled before the displacement check ran, so a displaced path during a recovery could go unjudged. Recoveries are now judged first.
- Edits both sides made to one file are checked against a
git merge-filereference. A merged side kept silently where the reference conflicts is now a failure. - Also checked now: the exec bit on newly added paths, one-sided file↔directory changes, tainted recoveries, and a report that names the path together with its recovery command.
- Replacing an untouched swap is now a failure, even when reported.
- New crash windows cover the fallback and replay.
slot.shstops a slot whose trunk evidence falls below a minimum.- The campaign restarts on this release.
just checknow runs thefailpoints-feature test suites, and CI's verify workflow runs them too (bn-36chi). Until now, no gate ran the fallback and crash tests behind that feature.just checkalso runsmaw-scenario's tests, where the soak's seed-stream checksums are pinned.notes/v1.0-acceptance.mdlists each v1.0 gate as a yes/no check with its owner, status and evidence (bn-l6o7).
v1.0.0-pre.18
Eighteenth dogfood pre-release. It fixes data-loss bugs found by running the deterministic-simulation tests on the consolidated layout for the first time and by widening the SG1 soak to cover uncommitted trunk edits across merges. Several of these bugs shipped in pre.17 and earlier. Upgrading is strongly recommended. One behaviour change affects scripts: maw gc --recovery-snapshots --older-than 0 now needs --force.
Data loss fixed
- Replacing a tracked trunk directory with a symlink no longer wipes every uncommitted trunk edit (bn-1dlkd). gix status aborted on the file-type change, so the merge's snapshot failed, its fallback captured nothing, and the forced checkout overwrote the worktree while
maw ws mergeprinted[OK]and left no recovery ref. Status now falls back togit statuswhen gix fails; the merge refuses to touch a worktree whose uncommitted state it could not capture completely (the journal is kept, somaw ws merge --recoverfinishes later); and the directory/symlink change is reported as a conflict with your version pinned. - Checkout no longer deletes files through a symlink (bn-1dlkd). Its stale-file cleanup followed symlinks, so it could delete files outside the tracked tree through a symlink you created, and it never removed a tracked symlink to a directory.
- FF-absorb no longer detaches the consolidated repo root (bn-1jfui). When the merge then failed after the absorb, your next
git commitin the root landed offmain, and the next merge orphaned it. - Untracked files with non-UTF-8 names are no longer deleted by a merge (bn-31ijd, found by the codex sweep). Status silently skipped such paths. maw now refuses rather than omits: a merge stops with its journal kept, and
--recoverworks after the file is renamed. Full non-UTF-8 support is tracked separately. - Plain
maw gcno longer deletes the default workspace's refs on the consolidated layout (bn-1m3lx). It treatedrefs/manifold/{head,epoch/ws,ws}/defaultas dangling, which wiped the default workspace's op log (maw undothen reported nothing to undo) whilemaw doctorreported OK.
Correctness
- A merged change that clears an executable bit (755 to 644) is applied (bn-2nnuz). Checkout only ever added +x, so every checkout path — the trunk update after a merge,
ws sync, sibling auto-rebase, advance and recover — left the file executable, and the next commit reverted the merged mode. The snapshot-failed fallback also keeps your own mode change. - Printed restore commands quote filenames (bn-31ijd).
Recovery snapshots (behaviour change)
maw gc --recovery-snapshotskeeps pins of workspaces that still exist (bn-wxg28), including the default workspace's dirty-trunk pins;--include-livedrops them too.- Risky sweeps need
--force(bn-wxg28). With--older-than 0, a pin younger than a day, or a live workspace's pin, gc deletes nothing and lists each ref with its workspace, a LIVE marker and its age, plus the exact commands to inspect, preview (--dry-run) or rerun with--force. Scripts that ran--older-than 0must add--force.
Verification (the v1.0 gate)
- The production-tier deterministic-simulation tests now run on the consolidated layout (the default since pre.2) as well as legacy v2, with reachable prepare/validate failpoints and real SIGKILLs at commit sites (bn-1jfui).
- The SG1 soak now drives a real trunk worktree through the production
update_default_workspace, writes uncommitted trunk edits of every kind, and judges them with preservation, displacement and replay-faithfulness oracles (bn-1h9ue, bn-2zubk). A pre-release review of the harness fixed six ways a seed could pass unjudged or overstate coverage (bn-3adck); the remaining known blind spots are documented incrates/maw-assurance/src/trunk.rs. Oracle fixes for gc-drained snapshots and the destroy-record path (bn-m7kjy). The soak campaign restarts on this release's harness. just checknow runs the assurance-feature clippy lane that CI runs (bn-2zubk).
v1.0.0-pre.17
Seventeenth dogfood pre-release, focused on crash recovery, the replay of uncommitted trunk edits across a merge, fail-closed configuration, and the integrity of the SG1 soak (the v1.0 release gate). It includes a fix for a ws sync failure reported from the bones repo. A pre-release sweep (one slice run by codex / gpt-6-astra) found and fixed further defects in code added this cycle. Upgrading is recommended.
Merge crash recovery
- A merge that crashed after its commit step is recovered instead of wedging the repo (bn-1fcox). Previously the journal could only be deleted by hand, and FF-absorb,
doctor --repair,epoch syncandundoall refused while it existed.maw ws merge --recovernow finishes a merge whose ref update provably landed (op record, sibling rebase, target checkout,--destroyof sources still at their frozen HEAD, journal cleared) and aborts one that provably did not; it refuses and keeps the journal when the refs prove neither.--abortfinishes an already-landed merge (and points atmaw undo), and the nextws mergerecovers automatically. Every refusal message now points at--recover.fsck --repairno longer deletes post-commit journals. - A crash between the target checkout and the replay no longer strands your edits in a recovery ref (bn-15fzo, bn-1bkr0). A durable checkout intent lets the next run replay the original edits, including after a second crash during recovery. The intent is now written before any tree change, residual edits are pinned before cleanup, and a corrupt intent stops recovery instead of being treated as absent.
- Recovered merges report and record like live merges: no false local-vs-merge conflict, no duplicated Merge op, the target's Snapshot op is written, and skipped post-merge hooks are printed with their commands (bn-28s78, bn-1losw). If the manifold config is invalid, recovery warns and skips the sibling rebase instead of applying defaults.
Your uncommitted trunk edits across a merge
- An executable bit committed by a merged workspace is no longer reset to 644 (bn-3fcbu). Your own mode change still wins.
- Symlink/file and file/directory type changes no longer abort the replay (bn-2ygs0, bn-3jqfk). They are reported as
type_change/directory_changeconflicts that keep the merged side on disk and print one exact restore or inspect command for yours. The fidelity check no longer writes through a symlink, and the pre-merge capture and fallback recovery pin store a dirty symlink as a symlink rather than as its target's contents.
Sync
- Syncing a workspace whose commits move a directory works again (bn-1ijl). gix's rename tracker reports a whole-directory rename in addition to the per-file renames; maw read the directory as a blob and failed with "blob expected, found tree", leaving the workspace unable to merge. Directory-level renames are dropped (their files are still reported). A submodule replaced by a regular file no longer loses the file's bytes.
Configuration fails closed
- An invalid
.maw/manifold/config.tomlor.maw.tomlmakes commands that change state refuse (bn-hcbc8, bn-qi5br). Previously they silently used defaults — e.g.ws mergeabsorbed trunk commits with default policy andepoch syncre-pointed the epoch atmain. Refusals name the file, line and parse error with a fix line; read-only commands warn once. Recovery commands never refuse because of config on the consolidated layout (recover --restore-file), andrecover --to/migraterefuse before changing anything (bn-15ebo). Unknown.maw.tomlkeys warn with the nearest valid key (bn-1losw).
LFS
- Smudge decodes exactly what git-lfs smudge decodes (bn-hcbc8), so non-canonical pointers written by other tools are materialized, while size mismatches and extension pointers stay as pointers as git-lfs leaves them. maw never discards bytes past a pointer's first 1024 bytes (a deliberate difference from git-lfs).
Other
maw ws destroyexplains an "unmerged changes" refusal when the content is already in the epoch (e.g. cherry-picked) and suggests--force, which still snapshots (bn-2eszz).maw ws history defaultworks on the consolidated layout, andmaw ops logorders same-millisecond ops stably (bn-15fzo).
SG1 soak (v1.0 gate) integrity
- The pre.6 soak campaign (96% of 1e8 op-steps) is abandoned and not counted: its two "violations" were disk-quota errors, its driver counted harness errors as clean, and its in-process destroy never ran (an invalid recovery ref name was rejected and ignored), so destroy, recover and the gc recovery drain were never exercised (bn-30v6e, bn-25pac). The driver now fails closed with a
HarnessErrorverdict and a per-seed vacuity guard, infra I/O errors are retried rather than counted, pinned regression seeds must reproduce the named oracle finding, op-steps are counted exactly, and a violation can never be reported as DONE or overwritten by an infra halt (bn-2qamr). Two new crash points guarding the fixes above are crash-tested in DST (bn-1sbjf). The campaign restarts on this release's harness.
v1.0.0-pre.16
Sixteenth dogfood pre-release, focused on formal verification. The existing Kani proofs, merge property tests and Stateright protocol model had silently stopped compiling or running; they are revived, wired into just check and a new verify CI workflow, and extended to the code that actually produced past defects: path predicates, the FF-absorb sibling decision logic, the LFS pointer codec, merge-phase transitions, and the multi-process protocol. That work, a cargo mutants pass and a three-way pre-release sweep (one run by codex / gpt-6-astra) found and fixed several data-loss and corruption defects, most of them present since pre.10 or earlier. Upgrading is strongly recommended.
Data loss and corruption fixes
- K-way conflicts list every overlapping workspace (bn-1lwk; since pre.10). When three or more workspaces edited the same region, the conflict listed only the first two. Later workspaces were missing from
sides,maw ws conflicts,maw merge last-conflictand the markers, and--resolve cf-X=<ws>rejected them, so every available resolution silently discarded their committed edits. Participants are now every workspace whose edits pairwise-conflict with another's, on both the plain and AST paths. - Resolving a conflict keeps other workspaces' edits to the same file (bn-28os; since pre.10). If
waandwbconflicted on line 1 andwcedited line 7 of the same file,--resolve cf-X=wa(and--resolve-all, and atom resolution) usedwa's whole file, droppingwc's committed edit while reportingwcas merged. Conflict records now carry these "bystander" workspaces and every resolution folds their edits back in; if that fold conflicts, the bystander becomes a side instead of being dropped.content:PATHwarns when the supplied file omits a bystander's edit. - Atom-level
--resolve cf-X.N=<ws>no longer joins lines or reverts clean edits (bn-34zv). The splice dropped the region's trailing newline (l1\nwbl2\nl3\nbecamel1\nwbl2l3\n), used a wrong line offset after earlier clean edits, and reverted every clean edit outside the conflict. Atom resolution now replaces each conflict block with the chosen side's exact bytes (CRLF and missing final newline included). - AST-merge atom resolution no longer drops clean edits or overwrites following code (bn-22nk, found by the codex sweep). AST atoms omitted a participant's clean edit to another function, and added-node atoms used the variant's coordinates as base ranges. AST atoms are now used only when they provably reconstruct every side; otherwise maw falls back to diff3 atoms. Atom resolution also refuses files containing literal conflict-marker lines (which could parse to empty content); whole-file choice still works.
- File/directory clashes between shared paths no longer commit a corrupt tree (bn-2jml). If two workspaces added file
aand two others addeda/b,maw ws mergereported OK and advanced the branch to a tree containing both a blob and a tree nameda(git fsck: duplicateEntries), then failed to check it out. Clashes are now detected across all non-deleted paths, the conflict names every workspace involved, and tree building refuses any file/directory clash or duplicate entry as a last line of defence. - FF-absorb can no longer silently revert merged work or overwrite an agent's edits (bn-302v). The sibling's epoch ref was written before its files and HEAD; a crash or a warning-only failure in between left the ref ahead of HEAD, and the sibling's next merge cleanly reverted the absorbed epoch. The fast-forward also held no sibling lock and used a dirty-path list computed earlier, so a concurrent agent edit or commit could be overwritten. Each sibling is now try-locked, re-classified on fresh state, materialized, moved with a HEAD compare-and-swap (git's own
HEAD.lock), and only then has its epoch ref written. A sibling whose HEAD cannot be read is left stale instead of fast-forwarded. - Quarantine promote publishes the fix, not the failed candidate (bn-3rhz). If an agent fixed a quarantine with
git commit, promote validated the fixed worktree but published the original candidate, then deleted the worktree holding the only ref to the fix. Promote now builds on the quarantine's HEAD (which must descend from the candidate). - Removing a quarantine always leaves a recovery snapshot (bn-jfj2).
maw merge abandondeleted the quarantine worktree without a recovery pin, so fix commits and edits made inside it were unrecoverable. Abandon (and promote's cleanup, when the worktree holds anything the promoted commit does not) now snapshots it underrefs/manifold/recovery/merge-quarantine-<id>/with a destroy record, somaw ws recover merge-quarantine-<id>lists it and--show/--restore-file/--towork; if the pin fails nothing is removed. The library refuses to remove a quarantine worktree without proof of a pin. The redo path ofmaw ws undolikewise no longer re-destroys a workspace whose recovery pin failed. maw merge abandoncan no longer delete an ordinary workspace (bn-ggo5, bn-1cth).abandonjoined the id into paths unvalidated (../..escaped.maw/manifold/quarantine/and could delete an outside directory holding astate.json), and in the consolidated layout it deleted.maw/workspaces/merge-quarantine-<id>even when no quarantine was recorded — which maw's own sync/merge refusal message told users to run. Ids are validated, unrecorded ids are refused (pointing atmaw ws destroy, which snapshots), andmerge-quarantine-is reserved at workspace creation.- Concurrent
maw ws createno longer loses a workspace to a racing prune (bn-1u93). The worktree admin entry was written before the worktree directory existed; a prune from any concurrent create/destroy deleted the half-built workspace and the create failed withindex.lock: No such file or directory. Admin dirs are now marked with git's ownlocked"initializing" marker during creation, prune skips locked dirs, andgitdiris written last via rename. gc --recovery-snapshotsno longer collects fresh pins of old work (bn-3maj). Pins were aged by the pinned commit's timestamp, so adestroy --forcetoday of a workspace whose last commit was two months old lost its recovery point at the next gc. Pins are now aged by their creation time (from the ref name, else the destroy record).
Crash-safety and concurrency
doctor --repaircan no longer move the epoch backwards (bn-32g8) — it now holds the epoch lock and advances with a compare-and-swap.- Operations refuse to move the epoch under an unfinished merge journal (bn-302v, bn-3rhz). After a crashed merge, FF-absorb,
doctor --repair,maw epoch syncandmaw ws undocould advance the epoch, after which neither--abortnor the next merge could recover the journal. They now refuse and print the recovery command. - Quarantine promote is locked and atomic (bn-3w2b): it holds the epoch lock, moves epoch and branch in one atomic compare-and-swap, and refuses while an unfinished merge journal exists.
maw changes closedeletes the branch with a compare-and-swap against the checked commit. The COMMIT phase andepoch_afterare written in one journal write. - Merge-state writes fsync the parent directory; the unsafe
Cleanup -> Abortedtransition is removed;SequenceGeneratorno longer overflows atu64::MAX(bn-2l63, bn-s8ti).
Layout, config and LFS
- One manifold config location per layout (bn-2dyz). Merge read
.maw/manifold/config.tomlwhile FF-absorb, auto-rebase, backend selection and others read.maw/config.toml, so a setting in one was silently ignored by the rest. Everything now reads<manifold_dir>/config.toml; keys set only in the legacy file are still honoured with a warning.[merge.validation]in.maw.tomlis honoured (bn-ila3). - Merge quarantines are handled as quarantines (bn-1cth, bn-ila3, bn-2dyz). In the consolidated layout they were created inside the trunk checkout (
<root>/ws/); they now live under.maw/workspaces/, are skipped by auto-rebase and FF-absorb, are shown with promote/abandon commands, and are refused byws sync/ws merge.maw merge promoterefreshes the root checkout in the consolidated layout. Validation no longer creates a stray.manifold/at the root. - LFS pointer handling matches git-lfs (bn-z9t3, bn-ggo5). The parser now accepts exactly what git-lfs accepts and writes (it previously accepted malformed pointers such as keys out of order or unknown keys), and write/parse round-trip. The clean path passes content through unchanged exactly when git-lfs would, so non-canonical pointers written by other tools are no longer wrapped in a pointer-to-a-pointer, and a file that merely starts with the version line is stored in LFS. A batch-client panic on non-ASCII oids is fixed.
- The workspace name
epoch-delta(used internally for epoch-delta injection) is reserved at creation (bn-2l63).maw ws undono longer accepts an empty or.path, which would have removed the whole workspace (bn-2wav).
Verification
just checknow also runs the no-default-features build, the merge property tests (just proptests) and a fast multi-process Stateright model check (just formal-fast). A newverifyworkflow runs the Kani proofs (41 harnesses across the root crate, maw-core and maw-lfs) and the full Stateright model on PRs and pushes to main (bn-86hd).- The Stateright model now tracks file content across merge, sync, destroy, auto-sync, doctor, quarantine promote, agent edits, trunk commits and crashes, with the epoch lock and per-workspace locks. Every historical defect class (bn-mq3b, bn-rah2, bn-p3m9, split CAS, pre-bn-38vw, reversed lock order, and those fixed in this release) is kept as a mutation the checker must still catch (bn-3ppf).
- FF-absorb sibling classification is a pure function in maw-core, proven exhaustively with Kani (bn-27n7); ...
v1.0.0-pre.15
Fifteenth dogfood pre-release, focused on merge validation and quarantine recovery.
- Verbose validation no longer deadlocks (bn-15tj). Merge and quarantine promotion drain stdout and stderr while validation commands run. Each stream retains its last 1 MiB, with explicit truncation notices. Timeouts preserve partial diagnostics and terminate the command's process group on Unix. Output collection also has a deadline when descendants retain pipe handles.
- Consolidated-layout merges use the correct workspace path (bn-1uif). The active-change ancestry guard uses backend-resolved paths for merge,
--check, and--plan. Unbound workspaces with unmerged active-change ancestry still cannot promote those commits to trunk. - Quarantine promotion refreshes the default worktree (bn-emj1). A successful
maw merge promotechecks out the promoted epoch and preserves uncommitted edits. The default worktree now agrees with the advanced branch and epoch refs. - Maintenance. Updated repository security-review instructions, corrected expressions rejected by Rust 1.98 Clippy, and pointed the manual formal-check recipe at the relocated model tests.
v1.0.0-pre.14
v1.0.0-pre.14 (2026-08-26)
Fourteenth dogfood pre-release. Three data-loss-class defects closed in the FF-absorb path (the third and fourth since pre.11), a destroy-time stat-cache-blindness fix that closes the last known window in that class, a new opt-in DST corruption primitive, and two fresh-eyes sweeps that found and fixed eleven further substantiated defects across sync, checkout, recovery, and LFS materialization.
FF-absorb data-loss fixes (third and fourth in this class, after bn-p3m9 and bn-286g)
- Stale dirty siblings no longer silently revert an epoch's hunks (bn-mq3b, bn-2fto). FF-absorb could advance a dirty sibling's HEAD to the target epoch even when the sibling was more than one epoch behind with an uncommitted edit on a path that changed in an earlier epoch — the
dirty ∩ ff_pathscheck missed it because the path wasn't in the global fast-forward set. The worktree kept the stale blob behind a HEAD claiming the new epoch, so the sibling's next commit silently reverted the epoch's changes. A newff_dirty_stale_conflict()predicate (fails closed on unreadable HEAD/tree) now leaves such a sibling untouched instead of moving it, andmaw ws historyrecords the skip and any zero-replay FF-absorb advance so both are visible after the fact. - Directory/file prefix collisions are now caught (bn-26wt). The FF-absorb safety predicate only checked exact-path overlap; a directory replacing a file (or vice versa) at a shared prefix slipped past it. The central safety predicate, the stale-dirty classifier, and the sibling materializer filter now all reason about prefix collisions, not just exact paths.
- Dirty-path inspection fails closed (bn-vzj1). A repo/status error during FF-absorb's dirty-path check used to return an empty set — treated as "nothing dirty" — which could let a mutation proceed against unverified state. Errors now propagate, and all call sites reuse one verified dirty-path snapshot instead of re-querying (and re-risking silent failure) mid-operation.
- Index-write retries no longer blame the wrong failure (bn-vzj1). Index I/O retried on any error, including non-contention I/O failures it misattributed to "another process holds the lock." Retries are now scoped to the actual
PermanentlyLockedcase.
Destroy-time stat-cache blindness closed (bn-2k9e)
maw ws destroy's recovery snapshot was blind to stat-cache-masked stale worktree bytes in two places:list_dirty_paths(whether to snapshot at all) andcapture_dirty_worktree'sgit add -A(which bytes actually get staged — git's own add trusts the stat cache too). A masked file could be destroyed with its recovery snapshot silently pinning the wrong (HEAD) blob instead of the real on-disk content.- Fix reuses bn-154g's tree-diff-based hidden-divergence detector plus a chunked force-rehash pass before staging. Detector failure now widens the rehash to every tracked file (or refuses outright on the non-force path) rather than trusting the stat cache; pin failure aborts the destroy before any record or removal happens. The
MaskedStalePreservationDST oracle now judges destroy operations directly — previously carved out as "unjudged."
New DST corruption primitive (bn-22jy)
- A new opt-in fault class injects stat-cache-masked worktree corruption during DST runs (
corrupt_weight, 0 by default — byte-identical to prior runs, pinned by op-stream digest tests), verified by a newMaskedStalePreservationoracle and a regression plan that turns RED when the bn-154g guard is neutered.
Fresh-eyes sweeps (two rounds, eleven substantiated defects)
- Append-only merge-driver gate hardened (bn-3vvm). Invalid
[merge].append_onlyglob patterns were silently ignored instead of erroring;--check/--planskipped the append-only gate entirely; FF-absorb applied directory/file-swap paths in lexical order, leaving directory-to-file siblings dirty; materialization rejected non-UTF-8 Unix symlink targets it should have preserved verbatim. .gitattributesand migration detection (bn-1b0t).worktree_state_commitignored a modified or deleted.gitattributeswhen computing LFS/clean-filter behavior; the brownfield-migration notice failed to account for an existing.maw/manifolddirectory.- Recovery containment and LFS materialization (bn-1b3n).
maw ws recover --restore-filefollowed parent symlinks outside the intended destination; pointer-shaped files that weren't actually LFS pointers were smudged anyway; materialization verification depended on an externalgit-lfsbinary being present; repair and FF-absorb paths could write raw LFS pointer stubs into the worktree instead of smudged content. - Auto-rebase verification race (bn-jhd7). Materialization verification after auto-rebase ran outside the sibling lock, and corruption failpoint targets accepted relative paths that could resolve outside the intended scope; both are now scoped correctly, with cross-process and parser regressions added.
- Destructive checkout paths (bn-3s8u). The sync pre-overwrite guard failed open when authoritative content hashing errored, and native checkout could leave deleted non-UTF-8 tracked paths behind instead of removing them.
Other
- Ref-update errors from push/release are now classified by a typed
GitError::RefConflictinstead of matching a"lock"substring in the error text (bn-2gbv); this deliberately drops generic lock-contention messages down to a plain retry hint rather than misreporting them as ref conflicts. AttrsMatcheris now memoized per HEAD tree OID during blob writes instead of re-walking the HEAD tree per file — 10-38x faster on 300-2000-file snapshots (bn-2fps).checkout_detach_removes_stale_tracked_filesno longer assumes a git behavior that was never guaranteed; it now checks out in place instead of reusing a populated worktree directory (bn-24xy — not a regression in the underlying git version).- Severity wording, config-path messages, and timestamp formatting cleaned up across
doctor/fsckoutput (bn-34wr).
maw v1.0.0-pre.13
Thirteenth dogfood pre-release, focused on preserving user decisions during conflict resolution, hardening working-copy recovery, and making release automation reliable across maw's workspace layouts.
Conflict resolution and recovery
- Structured
--keepnow refuses to overwrite a path that diverged from its recorded placeholder. - Added explicit
--accept-currenthandling with header validation, marker rejection, exact manual-byte preservation, path-limited auto-commits, and honest verification guidance. - Reconciled conflict state across
ws resolve --list,ws conflicts, and merge gates, including stale metadata and placeholder-only states. - Hardened snapshot/replay handling for file-directory transitions, staged-work isolation, unsafe relative paths, and symlink escapes.
Release and workflow tooling
maw release prepareandpreflightnow support single-version Cargo workspaces and internal path dependencies correctly.- Publish dry-runs patch internal crates to local paths so CI validates the actual package graph.
- Improved consolidated-layout help, default-rebase guidance, and the
ghoverride path used by change workflows.
v1.0.0-pre.12
Twelfth dogfood pre-release. The largest batch since pre.1: the "30 ideas → top 5" reliability push, full closure of a second mess field report (bn-1m4d, ~30 workspaces / 147 commits, zero data loss), the release-automation tooling used to cut this very release, and a fresh-eyes sweep across merge, clean, history, and fsck.
Reliability top-5
- Repo-level epoch lock (bn-13rc). All epoch mutations (sync, merge, advance, epoch-sync) now serialize on
.maw/manifold/locks/epoch.lock(fs4 flock, holder JSON metadata, exit 75EX_TEMPFAILon contention,[lock]config +MAW_LOCK_NO_WAIT/MAW_LOCK_WAIT_SECS). Racing concurrent syncs now serialize and both succeed instead of racing the epoch ref. maw fsck(bn-1uot). New offline deep state verifier: a 14-invariant catalog covering ref/record/recovery coherence;maw doctor's 6 coherence checks now delegate to it.--repairre-pins unpinned recovery refs and clears only provably-orphaned merge-state, exit 0/1/2. A fresh-eyes follow-up (bn-1283) hardened--repairto never treat an unreadable merge-state as repairable — deleting it could kill a live merge — plus fixed a disabled-audit JSON omission and added an advance-orphan regression test.- Prime-Invariant runtime auditor (bn-2rnq). Capture/audit brackets now run inside the epoch lock on every merge, sync, advance, and epoch-sync, printing a proof line (
INVARIANT: verified N sibling workspace(s)) and, on violation, pinning a recovery ref and blocking loudly with a nonzero exit. A fresh-eyes follow-up (bn-lf1d) extended the audit boundary to cover FF-absorb itself, so a source workspace whose committed work is lost during epoch absorption is caught instead of being hidden by the final merge audit's source exclusions. maw undo+maw ops log(bn-117s). Undoes the last epoch-mutating operation (epoch + branch CAS rewind, sources restored from recovery snapshots, dirty trunk byte-preserved), pinning the result atrefs/manifold/recovery/undo/<ts>; undo-of-undo is redo. Refuses on epoch-advanced/diverged/pushed/workspace-recreated states.- DST escape-path coverage (bn-2bcx). New out-of-maw-commit / dirty-trunk-write / gc fault classes plus sibling-ref-faithfulness and trunk-dirty-preservation oracles in the production DST tier, proven against a deliberate revert of the bn-rah2 fix (fires at step 7).
Field report 2 — closed end to end (bn-1m4d)
post_synchooks (bn-1lhb). Configured[hooks] post_synccommands now run after any sync or auto-rebase replay, signal-only, surfaced as a NOTE in the merge summary andhook:FAILinws list/JSON.- Comment-syntax-aware conflict headers (bn-drk3). Structured-conflict headers now use the right comment syntax per file type (
//for.rsetc.,--for.sql/.lua), with the legacy#form still parseable everywhere. (add/add and modify/delete headers remain#-only.) maw ws clean(bn-auu5). New command snapshots before deleting (refs/manifold/recovery/<ws>/clean-<ts>), supports--paths/--dry-run/--ignored, and defaults to requiring--force. The oldcleanwas renamedclean-build. Along the way, a real gix bug was found and fixed: linked worktrees only read their own.git/worktrees/<n>/info/excludeand missed the common-dir.git/info/excludethat git itself honors —maw-gitstatus now honors it too. Two fresh-eyes follow-ups hardened this further:cleanandclean-buildnow honor a configured non-default default-workspace name instead of only matching the literal namedefault(bn-3uiq, bn-ado3), and clean's capture path now stages into a temporary index rather than the caller's real index — the old force-stage-then-reset sequence could destroy work the caller had already staged. That same fix also bounds hook process trees: timed-outpost_synchooks now run in their own process group so aSIGKILLon timeout reaches the whole descendant tree instead of leaking orphans that block the output-reader threads.- Merge JSON completeness (bn-20fp).
--format jsonmerge output now includesmerged_sha,epochs,sources,destroyed, per-siblingsiblings[](action + overlap + hook result),warnings[], the invariant result, andrecovery.pinned_refs/cwd_destroyed; text output ends with a sentinel line → JSON hint → destroy-cwd warning, in that fixed order.
Release automation (bn-1obp) — used to cut this release
maw release prepare vX.Y.Z: lockstep version bump across everyCargo.toml(workspace version + all internal path-depversion = "…"strings, ~19 strings), Cargo.lock regen, and a CHANGELOG section scaffold. Idempotent; leaves everything uncommitted for review.maw release preflight [vX.Y.Z]: read-only gate checking version consistency, CHANGELOG presence, and a clean tree. Also runs in PR CI via the new publish-dryrun workflow (paths-filtered on manifest changes): preflight +cargo publish --dry-run --no-verifyover the full 6-crate publish chain, so a broken publish chain fails on the PR instead of at tag time.- The bare
maw release vX.Y.Z(tag + push) is unchanged. AGENTS.md's release flow now documents prepare → notes → check → commit → preflight → tag.
CI (bn-1q7x, bn-1n2b)
just checknow also runscargo test -p maw-cli— previously thecrates/maw-cli/tests/suite (undo, epoch-lock CLI tests, etc.) and the maw-cli lib unit tests were entirely CI-ungated. Landed serialized (--test-threads=1) to dodge the then-unexplained flock flakiness below; the bn-1d22 fix later removed that serialization.sg1-assurance-clippyis now wired intodst-faithful.yml(it existed but ran in no workflow).- New
tests/ci_lane_wiring.rsmeta-test enforces that every*-clippy/*-checkJustfile recipe is either wired into a workflow or explicitly marked# ci: local-only, with the convention documented in the Justfile header. - bn-1d22, the flaky-flock-test mystery, is solved: the intermittent
maw-clilock-test failures under parallel execution were never a lock bug or timing race — they were fork fd-inheritance: a sibling test'sCommand::spawndups the flock'd lockfile fd acrossfork(), and since BSD flock lives on the open file description, a just-released lock reads as held in the child until it execs orCLOEXECclears the fd (a microseconds-wide window). The fix is one-directional (a free can spuriously look held, never the reverse): tests let post-release observations settle (bounded, 5s) while must-fail-while-held assertions stay immediate. The--test-threads=1crutch from pre.11-era CI work is removed —maw-clitests run parallel again (55+ green worker runs, 9/9 lead including a 32-thread stress run).
Other fixes
- Cleanup lockstep with recovery refs (bn-3uou).
maw gc --recovery-snapshotsnow prunes destroy records in lockstep with recovery refs it removes;maw doctorsplitsabandoned-with-snapshotfrom the newdestroy-record-unpinned. - Bench-feature rot fixed +
sg2-bench-clippyCI lane (bn-1nmh). ws list/ws ls/maw ls --names(bn-3kiv): pipeable name-only listing, order matching JSON output.maw ws historyno longer hides oplog corruption behind a silent git-log fallback (bn-100i, fresh-eyes). A corrupt oplog chain now surfaces as an error instead of quietly rendering a shorter, wrong history.
v1.0.0-pre.11 — FF-absorb is ref-faithful to siblings
Eleventh dogfood pre-release. Closes a second data-loss-class field report from mess (bn-rah2), hit on pre.10 but latent since ~v0.60.8.
FF-absorb no longer orphans committed-ahead sibling workspaces (bn-rah2)
- When a merge absorbed out-of-maw trunk commits (the epoch-behind fast-forward case), the absorb loop reset every path-disjoint sibling workspace's worktree HEAD to the absorbed tip with a raw HEAD-file write. For a sibling with committed work that silently orphaned its commit: HEAD back at base, the work demoted to uncommitted/untracked changes, no output, and the next
maw ws syncrefusing with "uncommitted changes" — a state that invited destroying reviewed work. The safety predicate only checked touched-path overlap, never committed-ahead-ness. - The absorb now classifies each sibling first: committed-ahead + clean siblings are replayed onto the absorbed epoch through the guarded rebase path (HEAD lands at the rebased commits — work stays committed, tree stays clean, the next
ws syncreports up to date); committed-ahead + dirty siblings block the absorb before any ref moves, with the usualmaw epoch syncguidance; dirty-only siblings fast-forward as before with their edits preserved. Replays run before the epoch ref advances, so a failure aborts the absorb cleanly — no sibling is ever left half-moved. - The absorb is now loud: one NOTE line per affected sibling ("replayed N commit(s) onto absorbed epoch" / "fast-forwarded to absorbed epoch") alongside the aggregate "Absorbed N upstream commit(s)".
- Both raw HEAD-file writes in the absorb path (sibling and merge-target) were retired in favor of a guarded, reflog-writing native
set_headprimitive — restoring the single-choke-point invariant for HEAD movement (bn-8flz) at the one site it had missed.
ws sync refusals name the offending paths (bn-3rst)
- "Workspace has uncommitted changes that would be lost by sync" (and the rebase-path and auto-sync-skip variants) now list the dirty files with status letters, capped at 10 with an "…and N more" tail. An operator can now tell orphaned work from junk scratch at a glance. (The reported
.git/info/excludeissue did not reproduce — maw's dirty check already honors gitignore/exclude exactly likegit status.)
v1.0.0-pre.10 — trunk dirt is never lost + multi-agent signal clarity
Tenth dogfood pre-release. Resolves the mess multi-agent field report (bn-38nz) end to end: one data-loss-class fix at the merge/trunk boundary, plus a sweep of conflict-signaling and output-clarity improvements for orchestrated agent workflows.
Merge can no longer silently clobber uncommitted tracked trunk files (bn-1xmk)
- The trunk preserve-and-replay cycle in
maw ws mergehad three silent failure modes that could restore an uncommitted tracked file at the repo root to its committed content — the reported incident wiped a session's append-only.bones/eventsjournal that no workspace had touched. - A durable recovery snapshot (
refs/manifold/recovery/default/<ts>) is now pinned before the trunk checkout whenever the trunk has uncommitted tracked changes. It survives a clean replay and is visible inmaw ws recoveras a pinned row. - Dirty trunk files whose committed content changed across the merge (including via absorbed out-of-maw commits) now replay through the driver-aware 3-way merge —
merge=unionjournals keep both the merged commits and the uncommitted appends — instead of a bare stash apply that could resolve to the committed side. git stash createreturning nothing on a dirty tree now errors loudly instead of being treated as "clean" (mirrors the bn-3mpx destroy-path fix).- A post-replay fidelity check verifies every pre-merge-dirty file whose committed content the merge did not change ends up byte-for-byte intact; on mismatch it self-repairs from an in-memory pre-merge capture and prints the exact
maw ws recovercommand. - Merges announce
preserving N uncommitted trunk file(s) across merge (recovery snapshot pinned), andmaw ws merge --checklists dirty trunk files informationally (never blocking).
Conflicted syncs are unmistakable (bn-mq6j)
maw ws syncconflict summaries (new and residual committed conflicts) are nowWARNING:-prefixed and duplicated to stderr. Exit code stays 0 — conflicts are data, not errors.- Sibling auto-rebase results with conflicts render as a distinct
CONFLICT:line with the exactmaw ws resolvecommand, and the merge summary lists any siblings left conflicted. maw ws sync --format json(and--json) added: workspace, action, commits replayed, conflict count, and conflicted paths — mechanical detection for agents.
Auto-rebase semantic-risk hint (bn-2cvx)
- When a sibling is auto-rebased over epoch commits touching files that sibling also touches, the merge output line and the workspace's auto-rebase notice now append: "replayed over commits touching N file(s) this workspace also touches — re-run its tests before merging" (with a capped sample of paths in the notice JSON).
ws resolve --keep union (bn-nmu7)
- New resolution mode for the common "two workers each added a line to the same list" conflict: like
--keep both, but conflicted hunks get a per-hunk line union with cross-side dedup and stable order (ours' order first, theirs-only lines appended in their original order). Binary / N>2 / no-base cases fall back exactly like--keep both.
Clearer conflict-gate message for leftover headers (bn-1etl)
- When a committed file was hand-resolved (markers gone) but still begins with the maw conflict header, the merge gate now says exactly that — delete the leading
#header lines, commit, re-run — instead of the tamper-flavored placeholder message. Both cases remain hard-blocking and not bypassable by--force.
Output & UX (bn-3eew, bn-1kop, bn-1aey)
maw ws merge --check's epoch-drift note is now informational when the drift is a safe fast-forward ("will be absorbed automatically when you merge") and only warns on blocked or truly diverged states; the drift classification rides in the check's JSON output.- Merge success ends with a stable, grep-able sentinel:
[OK] merged <ws> into <branch> @ <sha>. maw ws list --helpdocuments the full lifecycle-state vocabulary (clean,stale,conflicted,committed-unintegrated, …) and mentions--format json.- Destroying a workspace (directly or via
merge --destroy) while your shell's cwd is inside it now prints a note telling you tocdback to the project root; a failed cwd lookup hints that the directory may have been deleted.