Repository navigation
codr 0.11.2
This patch follows 0.11.1. It keeps the same MDK source pin, dependencies,
control protocol, bridge state version and storage cohort. No account migration
is required or authorised.
Phone questions and permissions
- OpenCode V2 multi-part questions show one part at a time. Answers survive
restarts. The final answer submits the complete form after a fresh identity check.
Multi-select fields and legacy multi-question dialogs still need the terminal. - Permission cards retain the scrolling command block. Empty metadata, duplicate
commands and visible approval codes are omitted. Send an ordinary1message
for Allow once, or2for Deny. No quoted reply is needed. - Replies stay bound to the exact session and pending request. A changed unanswered
request needs an extra confirmation before approval. Persistent approval stays
terminal-only. Uncertain submissions are not retried. - Permission status shows
Waiting for your permission..., thenAllowed once.
orDenied.after confirmation. This confirms the reply, not command execution.
History and delivery reliability
/recentsends conversations oldest-first and keeps each conversation's
continuations together. The next conversation waits five seconds after the
previous batch is confirmed. Live replies stay independent./lastand/recentcollect history outside the serial input loop. A bounded
durable queue survives restart. Results require the same account, chat and
native-session binding. Neither command resumes muted forwarding.- Late native acceptance clears stale delivery warnings and sends a durable
confirmation. Lost acknowledgements do not cause automatic prompt resubmission.
Genuinely unconfirmed messages remain paused. - An interrupted OpenCode local-service GET retries once. Failed reads do not
advance history checkpoints. HTTP rejections and submissions are not retried.
Codr does not start or replace the shared OpenCode server.
These read retries do not establish or repair the original cause of the service
interruptions in issue #22.
Performance and durability
- Unchanged bridge state skips file replacement, syncs and JSON encoding. Polling
clocks stay in memory. Real delivery, permission and setup changes still save
immediately through atomic private-file replacement and file/directory syncs. - Parsed forwarding controls use file identity and change metadata for cache
validation. Mute checks do not use a stale time-based cache. Locked mutations
read fresh controls. - Completed native reads update their session instead of repeating full discovery.
Authoritative snapshots, ended sessions and global checkpoints retain reconciliation. - Polling, manual history and desktop history share native readers. Immutable
history snapshots avoid deep copies. Desktop projections run only with subscribers. - Background history work has a shared concurrency limit. Fresh phone validation
remains independent, so unrelated background reads cannot occupy its capacity. - Recovery runs outside serial input dispatch. Payload memory stays within page
and channel bounds. Failures back off from two to sixty seconds.
Recovery still scans full retained history to find older late deliveries. Canonical
identity order can cross epochs with decreasing timestamps. No timestamp watermark,
receipt pruning, database migration or new dependency is introduced. Cursor identity
metadata and durable receipts still grow with history.
Local measurements
Before and after samples used 60-second windows with 31 native OpenCode sessions.
Forwarding remained muted. The earlier persistence-only fix was already installed
for the baseline.
| Metric | Before | Later post-deployment samples |
|---|---|---|
| CPU, percentage of one core | 16.9% | 4.6–4.8% |
| Logical reads | 14.49 MiB/s | about 0.26 MiB/s |
| Forwarding-file opens per minute | 8,427 | 0–1 |
These observations show about 72% less CPU and 98% fewer logical read bytes.
They are not controlled benchmark results. Session activity and host load changed.
Logical reads include cached data and sockets, not only physical-device reads.
File notifications include all local readers.
The later samples recorded four and two state replacements while session state
changed. They do not prove zero writes. Memory continued to warm after restart.
No sustained memory saving or memory-leak conclusion is claimed.
Upgrade and setup
Existing combined installations retain /update, then /update confirm, after
publication and verified release mirroring. These commands update the combined
executable without an account migration. Account keys, chats, receipts, setup drafts
and native sessions remain in their existing storage.
Fresh accounts still use the supported 0.9.0 setup and separately confirmed
cold-backup handover. Existing 0.9 installations retain their separate executable
update and handover confirmations. See the combined runtime guide
and installation guide. This release adds no fresh-account wizard.
Verification
The corrected 0.11.1-labelled candidate passed 492 Rust tests and 174 Python tests
before deployment. Eight local-data tests remained ignored. Two embedded-runtime
tests were excluded because managed temporary socket paths exceeded the local limit.
These two tests were not counted as passing. Strict Clippy, formatting and two
independent read-only reviews passed.
Deployment checks reran 260 focused Rust tests and all 174 Python tests.
Read-only checks read all 31 native sessions before and after deployment. The local
restart preserved all 36 chat mappings, receipts, both setup drafts and native
sessions. OpenCode was not restarted. No unexpected Codr restart occurred during
the measured intervals. The owner confirmed that phone-side operation works.
These candidate checks are not exact-source 0.11.2 platform certification. The
versioned source must pass GitHub CI and all four package lanes before publication.
The tag workflow builds Linux x86_64/aarch64 and macOS Intel/Apple silicon packages.
Published archives and checksum files must match their mirrors byte-for-byte.
An earlier local candidate must not be relabelled as a 0.11.2 release artifact.
Contributors
- Tom Dwyer (@tidley): maintainer-led phone permission
replies, multi-part questions, history ordering, delivery recovery, performance
fixes, regression coverage, local verification and release preparation. - Juraj Bednar (@jooray): retained baseline
contributions, including PR #20.
These are not new changes in 0.11.2. See the 0.11.1 credits
and original contribution records. - Marmot MDK contributors: the unchanged pinned runtime and retained security
backports. Packages retain their notices and licences.
Original authorship and baseline attribution remain separate. Keep these same
credits in mirrored release notes.