release: 2026-09-25 (v0.8.5) — vendored sources you can trust, and nine contributor fixes - #184
Merged
Merged
Conversation
… skip The idempotency gate took the precise (trailer-matching) path whenever any commit in range carried an AIOS-Session trailer, including when this session's own SID was `unknown`. aios-commit omits the trailer when CLAUDE_CODE_SESSION_ID is empty, so for such a session MINE was 0 by construction and every close after the first skipped the block, the observed-context routing and the commit, silently. `unknown` is not an identity: provenance is unknowable, so the gate now falls back to the coarse rule it already uses for untagged ranges. Tests 17-18 prove the new case and reproduce the defect against the old logic. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
…arry, reconcile, plugin sync
Five places in /aios:update turned a measurement that never completed
into a confident verdict, each one silent:
- h_file / h_git returned rc=0 with an empty hash when `shasum` was
missing (a pipeline's status is its last command, `cut`), so an empty
LOCAL equalled an empty BASE and the three-way compare answered
"identical → overwrite silently, no backup". They now probe for the
tool and refuse on an empty hash — in both copies of h_file (the
self-update compare in Step 2.5 still carried the old one).
- The legacy .gitignore migration re-ordered the operator's rules
(`sort -u` + `comm`), so a `!exception` sorted above the pattern it
excepts and stopped working. The carry now subtracts the baseline
with `grep -vxF -f`, which keeps the operator's order.
- `{ cat canonical; awk operator; } > new && mv` took awk's status, so
an unreadable canonical .gitignore installed a file holding only the
operator's lines. The merge now refuses when canonical is unreadable
and chains cat's status into the group.
- `diff -rq vault/x clone/x 2>/dev/null` with vault/x MISSING printed
to stderr only (exit 2), dropped by the redirect and the trailing
`|| true`, so a bundled folder absent from the vault read as "0
drift" and the tracker advanced. A missing layer dir is now reported
as `Only in <clone>: <dir>`, the same shape the root-file loop uses.
- `cp $HOME/aios/...` with $HOME unquoted: a home directory containing a
space (common on Git Bash) split the path and nothing was copied, with
no error path. Quoted.
tests/update-measure-fail-closed.test.sh runs each replaced shape
against the new one so a case that cannot fail is caught as such.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
…-snapshot CLAUDE.md § Session End replaced the hand-rolled observed-context snapshot (copy, then pick the "next free letter" by hand) with hooks/aios-snapshot, which takes the decision under a lock and compares against every archive already written that day. /close-day and /aios:housekeeping kept the old instruction, so a session following the command ran the non-atomic protocol the helper retires. Both now call the helper and treat `identical` as success. tests/commands-snapshot-helper.test.sh anchors it: a command that names the dated snapshot destination must name the helper, and none may prescribe the by-hand letter step. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
- Bucket 18 compared the seven-character vendored hash with the forty-character sha the API returns, by equality: every source read as behind on every run. Compare by prefix. - Bucket 22's duplicate-frontmatter scan ran through `python3` (nothing on Windows, silence reads as clean) and printed nothing on an empty glob. Run through `uv`, always end with a count of files checked. - Bucket 13 skipped already-graduated patterns by a phrase that is not a substring of the mark it writes, so nothing was ever skipped. Match on the prefix the written mark carries. tests/housekeeping-checks-that-cannot-fail.test.sh runs each old shape next to the new one and executes the scan script extracted from the spec. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
…or reconcile filters, sweep both residue shapes - Two concurrent /aios:update runs shared one fixed clone path and each rm -rf'd it; the second wiped the tree the first was still comparing. A lock directory guards the clone — its own mtime is the start time, so there is no timestamp-writing window a peer could misread (no pid: the command runs as separate shell calls); a lock older than 30 minutes is reported with the command that clears it, never cleared by the run that finds it (two finders would both proceed); a refused run is told not to clean up what it never owned; released with the clone in Step 7. - Skill-folder dedup judged a stray folder by SKILL.md alone and deleted it whole. Compare the whole folder; back it up first when anything differs. - The reconcile's filters were bare substrings: `oauth` hid skills/anthropic/doc-coauthoring/ and oauth.json.template, whose drift was never reported. Anchored to whole path elements. - The residue sweep matched space-joined names only; zsh leaves newline-joined ones, and `find -print | read` fragmented them. Each directory is handed whole to a child shell and both shapes match. tests/update-clone-lock-skills-filters-sweep.test.sh extracts the lock block, the filter chain and the sweep from the spec and runs them beside the shapes they replace. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
…ecklist; one rule for empty fields
Three instructions a session obeys in order contradicted each other:
- Step 5 appends AND commits the session block through the helper,
which only appends; steps 8 and 9 then asked the operator questions
whose answers belong in that block. A new step 4.7 gathers both
(the comprehension recap-and-offer, "what was most useful?") before
the block is written; steps 8 and 9 now only record them.
- Step 9's checklist demanded direct observed-context writes with no
--auto exception, while --auto defers every such write to /close-day
because the observed files have no lock. The checklist now says the
candidates captured in the block satisfy those items under --auto.
- The block format said "Skip" for an empty field while § Rules said
every field must be present. Every field is present; an empty one
reads "None" ("None — not asked" for Most useful), in both Mode A
and the Mode B report.
tests/close-session-order-and-fields.test.sh anchors the three; run
against the previous copy it fails on every one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
…ry carry - The close-day precondition excluded only today's date, so a future-dated note won; a missing or unreadable calendar folder read as a first run. It now compares dates and reports both states. The guard hands /close-day the date it found, and /close-day closes exactly that note when given one — left to itself it picked today's, once /7plan had created it. - The backup probe checked `origin` unless a push remote was set, and read an empty URL as "ok". It now resolves the remote the way `git push` does (explicit, else a lone remote, else origin), checks every push URL, reports an empty one as "none", and names the remote to remove when it is the framework. - The last-note search covered two months; the drop check compared counts; three rules disagreed on escalating an untagged, undated carry. The search spans every month, the check goes item by item (parked and verified compass-hidden are the only exits), and the count-based table applies to exactly those carries. tests/today-carries-and-probes.test.sh runs both probes, extracted from the spec, under bash and zsh against temp vaults, and proves each carry rule can fail with a mutated copy. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
…A routed, reports by closed date, snapshots by section - The ecosystem.md redraw waited for `updated:` to pass 21 days, but every atomic append resets `updated:`, so a map receiving a note every few weeks was never redrawn. It now keys on its own `rederived:` date, which only a redraw moves (ecosystem.md only; profile.md has no full redraw defined). - Tier A runs first and excises the Reinforced entries it routes; the growth.md feed-in then read them from the buffer. It now also reads those routed today, where they landed, via the ROUTED markers. - Session reports were looked up by the clock's date; closing yesterday after midnight skipped yesterday's. Keyed on the date being closed. - The first project snapshot read `limit:40` per note and stamped the snapshot current. It now reads Current State and To-Dos in full. tests/close-day-derivation-and-sources.test.sh checks each property and proves each can fail with a mutated copy. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
- aios-commit scanned the committed paths but not the commit message; a token in -m went into history. The message is now scanned too. - aios-commit and pre-commit guarded the scan with `[ -x "$SCAN" ] &&`: a missing scanner, or one that lost +x, meant an unscanned commit. A missing scanner now refuses; one without +x runs through bash. - secret-scan.sh used `grep -I`, which skips a file with a NUL byte; a token next to one passed. Now `grep -a`. - aios-commit --vault discarded the sweep's git errors, and "no paths" read as "nothing to commit". Each failing sweep call now leaves a sentinel that refuses the commit and names why. tests/commit-scan-fail-closed.test.sh reproduces each defect against hooks rebuilt from a pinned pre-change commit (skipped, with a message, when that commit is absent) and uses a PATH git shim to fail one sweep call at a time. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
@modelcontextprotocol/server-github is deprecated on npm and was archived upstream. No AIOS command, agent or skill calls its tools; sessions use the gh CLI. Removes mcps/github-mcp, its setup block, the /mcps-setup walkthrough, its SECURITY.md row and the lint's allowlist entry. /aios:company now detects the GitHub substrate with `gh auth status` instead of an active MCP. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
…m-py 0.8.2 Every Python MCP in setup.sh was guarded by `[ ! -d .venv ]`, so it installed once and never again, whatever requirements.txt later pinned. notebooklm-py was not pinned at all and sat at 0.3.4 on a live machine while 0.8.2 shipped. Installs now stamp a CRLF-normalised hash of requirements.txt in the venv and reinstall when it changes; an existing venv with no stamp reinstalls once. notebooklm-py is pinned to 0.8.2. 0.8 moves the session under ~/.notebooklm/profiles/<name>/ on first run, so manual_login.py and save_storage.py now resolve that folder and fall back to the old root (which 0.8 then migrates). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
…w what runs Slack ran `npx -y @jtalk22/slack-mcp` unpinned (5.0.0 on a live machine) while the repo vendored a 3.2.5 subset no registration executed. The complete published 5.0.0 package is now vendored byte for byte, and every AIOS-authored invocation pins the same version. Its dependencies stay ^-ranged, stated in SECURITY.md and UPSTREAM.md. Atlassian is pinned to 0.23.1 (run.sh, setup.sh, README). Both versions were smoke-tested against live accounts before pinning; both leave the lint's known-unpinned list. vendored-pins learns two record fields: package=<name>@<version> for a runtime package (--upstream proves it against `npm pack`), and local= for the framework's own files beside a vendored copy. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
setup.sh stopped skipping skills/anthropic/ once it was measured that no marketplace provides them; setup.ps1 still skipped them, so on Windows every agent that declares one (doc-coauthoring, internal-comms, theme-factory, mcp-builder, skill-creator) silently got nothing. A parity test now fails if the two skip lists differ, and update.md's description of the scan matches what both scripts do. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
… too Reported privately by an operator after v0.8.4. --setting-sources user,project drops only the local layer, so a broad project rule such as Bash(python3 <hooks dir>:*) reaches every headless Bash call outside its --allowedTools; reproduced with a canary. MODEL-ROUTING.md now covers both layers and gives the two ways out with their cost: narrow the rule (interactive sessions then prompt more), or run the routine from its own folder with --setting-sources project (measured: user rules no longer apply). Bucket 29 reports broad interpreter/script-dir rules in both layers and names both options. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
§ Spawning said "Writing a file is not gated." A kill of a live session was held for the operator's approval in the session writing it; a resume naming a session that never existed was not (probed; the surface dead-lettered it as designed). The permission check judges what a request does. CLAUDE.md now says so in the same bytes; the orchestration-ladder skill carries the detail, what a brief planning a kill must say, and the optional inbox-write permission rule, which stays the operator's call. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
# Conflicts: # .github/workflows/validate.yml # CHANGELOG.md
…os-snapshot # Conflicts: # .github/workflows/validate.yml # CHANGELOG.md
… false # Conflicts: # .github/workflows/validate.yml # CHANGELOG.md
…keep every carry # Conflicts: # .github/workflows/validate.yml # CHANGELOG.md
…at Tier A routed # Conflicts: # .github/workflows/validate.yml # CHANGELOG.md
…p removes On top of #179. - Step 7's own cleanup line now removes the lock with the clone (clone first, lock last); before, only the closing Rules section did, so a model following Step 7 left the lock and blocked the next update. - A lock older than 30 minutes is reclaimed by an atomic mv to a name unique to the run, so exactly one of two racing runs wins. /today and /close-day fire this command unattended; a crash no longer wedges every later update until someone clears it by hand. - The residue sweep removes what it finds itself. Its report prints a newline inside a name as ^J, which cannot be typed back into rmdir. No {} inside the inner script: BSD find substitutes it mid-argument. - The test's sweep extractor ends on a fixed string; BSD awk read the old \{\} as an interval, so extraction always ran to the end of the spec. New checks: the reclaim, the two-run race, removal itself. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
…above it
Re-vendors 14 of the 15 v6.4.1 skills unmodified (5bf4e78, 56/56 files
proven against upstream) and adds the MIT LICENSE the v5 copy lacked.
diagnosing-superpowers is left out: it files bug reports upstream and
searches GitHub with session terms without an approval step.
v6 runs plans continuously and rules on ambiguities ("Rulings, not
stalls"). skills/_index.md and technical-cofounder now state that
INTENT.md ask/escalate items still stop the run, which upstream's own
precedence rule (CLAUDE.md outranks skills) supports, and name what it
still does unasked: dependency installs in using-git-worktrees, and bash
scripts in executing-plans / subagent-driven-development. Stale
"checkpoints" descriptions are corrected. Extensionless scripts/ files
pin LF.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
…writing Comprehension On top of #180. Its step 4.7 skipped both questions under --auto, which also skipped the scan behind the Comprehension field, so an unattended close could write "All operator-authored" and erase real comprehension debt. The scan now runs; only the question and the offer are skipped. "Everything the block must contain" narrows to "every answer the block needs from the operator" (Tier B candidates still form after step 5). A check pins the scan. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
…-closed gaps On top of #176. - Every hash helper (Step 3's and Step 3.4's) now uses shasum or, failing that, sha256sum. Refusing an empty hash alone left a machine without shasum backing up every file on every sync, permanently. - A missing vault/.obsidian, like a missing layer, now reads as drift instead of an empty (clean) compare. - A file with no local copy is added and never reported as backed up; the "any hash empty" row caught it before. - The test simulates "no hash tool" with a restricted PATH, since a function override is found by `command -v`. It adds the sha256sum-only case, and disables commit signing in its fixture repo. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
… empty repo On top of #183. - `--vault` in a repo with no commits yet: `git diff HEAD` has no HEAD, so the sweep read as failed and the first commit refused. It now reads the staged paths there. - `-a` reads a binary in full: a 200 MB attachment took ~12 s (main 0.04 s). One grep pass carries every pattern, and files over 20 MB (AIOS_SECRET_SCAN_CAP_MB) are scanned as text only and named in the output, so the gap is visible. Now 0.38 s. - Pre-existing: `| head -3` exited early, and once grep's output overflowed the pipe buffer grep died of SIGPIPE (141) under pipefail, so a real hit read as "FAILED to scan". Output now goes through `cut`, which reads it all and caps each reported line. - Tests: the empty repo, the size cap, and ~1.3 MB of hits (fails on main with exit 141). Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
- housekeeping (#178): an empty hash= was a prefix of every sha and read `current`; it now reads `unrecorded`. A package= record is checked against the registry's version, not a commit. - close-day (#177): the "already edited" case archives the committed version through the helper, which keys the archive by file name. - close-day (#181): without a date, the note to close is the most recent one dated on or before today, so the future-dated skeletons /7plan writes are never closed. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
…les on a seed On top of #182. - The redraw threshold points at observed-staleness.py's AGGREGATE_DAYS instead of restating "21 days". - An attempt that finds nothing to redraw (a seed ecosystem.md) sets rederived: alone, so a new vault is not redrawn at every close. - The duplicate "reset updated:" is gone, and the closing sentence no longer credits the staleness alarm with a redraw it no longer triggers. - Tier B finds routed entries by the marker prefix that is actually written and follows its [[target]] link; the quoted marker never matched the real one. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
… v0.8.0 Steps 1 and 6 of the release cut were skipped for v0.8.0–v0.8.4: both manifests still read 0.7.1 and the release map still said "Unreleased after 2026-09-13". Both catch up in this release. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
- context-load-audit.py recognised primary sessions by a regex of one operator's session names. Primaries now come from USER.md's `## Identity` table (its first GFM table; italic example rows skipped) plus --primary. With none declared it aborts and says how to declare one. - Its fit check (outward action without a declared/ read) only looked at the first --cap calls and ignored order. It now reads the whole transcript and asks whether a declared/ read came before the first outward action, within a single call too. --cap and --min-tools still measure the loading window. - context-rungs.py priced rung 1 as the last five entries of every observed file, while the floor emits a rule library's index and the newest entries by date. The floor's per-file selection is now one function, recent_slice(), and rungs imports it. Floor output is byte-identical. Tests: rung 1 equals the floor's emitted words per file; primaries, parser counterexamples and fit ordering cases; old-hook reproductions pinned to the pre-change sha. The parity test accepts the shared recency constant. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> AIOS-Session: 4ce35e79-7182-45d3-a479-d25e0004fda0
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ontributor fixes Replaces #175's early entry with one for the day. Action items are check-then-act: NotebookLM reinstall + login, GitHub MCP removal, Slack re-pin; session restart last. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
…ntity CI runners have no git identity, so the fixture commit failed there and read as the regression it guards against. Identity and no-signing are passed per command, never set globally; reproduced with HOME emptied. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
… price # Conflicts: # CHANGELOG.md
.ps1 is checked out with CRLF by design, so on Linux CI the parsed name ended in \r and never equalled setup.sh's. Stripped before parsing; checked against a CRLF copy. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
Found in a pre-merge rehearsal of this release against a copy of a live vault. In a UTF-8 locale macOS's `tr -d '\r'` stops at the first byte that is not valid UTF-8 and prints nothing, so every such binary (the Slack icon, a PDF in skills/anthropic) hashed as the empty stream. Any two of them compared "identical": an update to one never landed, and Step 3.4's deletion test could read an edited copy as untouched. All ten `tr` sites in update.md now run with LC_ALL=C. Test: two different binaries hash differently, to the whole CR-stripped file; both checks fail on the old helpers. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
- `trap release EXIT INT TERM` released the lock on a signal and the loop kept archiving the remaining files with no exclusion. INT and TERM now release and exit (130 / 143). - Copies went straight to the final name, so a copy cut short stayed as a valid-looking snapshot that every later run compared against. `put` copies to a hidden temp file in the same directory and renames it into place; the temp file is removed on failure and on release. tests/aios-snapshot.test.sh: a cp shim that fails halfway, one that SIGKILLs the archiver mid-copy, and TERM/INT during a slow copy (job control on so INT reaches the background job). Each case also runs against the hook pinned at the pre-change commit, where it reproduces the defect. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…or not at all # Conflicts: # CHANGELOG.md
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7
Member
Author
|
Released as v0.8.5. Main moved once. A rehearsal of Thank you for eleven contributions in one day (#175–#183, #185, #186). Each one fixed a check that could fail without saying so. |
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
2 participants
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.
Release branch for 2026-09-25 (v0.8.5). One merge to main; the CHANGELOG entry is written on the branch.
Contributions (#175–#183), each merged with authorship kept, fixed on top where review found gaps
Framework
🤖 Generated with Claude Code
https://claude.ai/code/session_01LCjeYxWVWosFKnRiCf8Ae7