Repository navigation
Releases
0.1.14: end ended background jobs, refuse a second build in one copy, and count deleted tests
Acts on the second field report in #5: 0.1.13 on Linux, with four workers at once on a Rust workspace.
Update: /plugin marketplace update flash-agents, update the plugin, then run node install.mjs in the plugin folder. This release needs the install step: the build refusal lives in the guard, which only install.mjs copies into the profile. A clone: git pull, then node install.mjs.
Fixed
- Ended background jobs leave
current. A job leavescurrentwhen the Harness tells its worker the job finished, or when the process an earlier sample found is gone, whether or not the worker ever reads the job. A job whose process is not found no longer borrows another process (asleep 55) as itschild. It is called "may be stuck" only while the latest sample still finds its own process. 19 one-second edit scripts no longer show up as "running 20m0s". - Edit scripts are not "one command rerun". A command that carries its own program (a heredoc,
python3 -c) is left out of the "ran N times … with no file changed" warning: each one is a new program, and the files it writes are not seen as changes. - A worker in another slot's copy is named. A command or path that reaches into a sibling copy (
cd …/slots/12 && cargo test) gets a warning in status and in the result. The sandbox only confines writes, so it can't refuse that read. Workers are now told to stay in their own copy.
New
- The guard refuses a second build in one copy. A
cargo,go,gradleorswiftbuild is refused (CONCURRENT_BUILD) while the same worker's own build of that tool still runs, background or not. The refusal tells the worker to wait withjob_outputor stop the first withjob_kill.cargo fmt,go versionand other workers' jobs are not affected. Setguard.refuseConcurrentBuilds: falseinflash.config.jsonto turn it off. - Deleted tests and grown files beside the diffstat.
change.tests({ removed, added, byFile }) counts test definitions over the whole patch: Rust#[test], Pythondef test_…, Gofunc Test…,@Test, XCTest, xUnit attributes, and JSit(/test(. Two cases now produce a warning that names the files: a patch that removes more tests than it adds, and existing files that grew by 1,000 lines or more. - An idle stop names the last tool result, with its length and how long ago it reached the model, so a stalled model can be told apart from a result too large to read.
- Delegate skill: ask the worker to list every test that fails compared with the base, by file, and say in the brief when to run the whole suite.
Verification
1036 tests pass across the three suites (1025 in 0.1.13). Every fix has a test that fails without it. The new tests feed in the Harness's completion notice as dsh-tool-jobs writes it; no live Harness run has delivered one to the server yet, so the process-tree check also covers a job whose notice never arrives. The live suite was not run for this release.
0.1.13: count a command's process tree as activity, find the plugin's base folder from a shell, and start a copy from a diff file
Acts on the field report in #5: 0.1.12 in daily use on Linux, a Rust workspace, up to ten workers at once.
Update: /plugin marketplace update flash-agents, update the plugin, then run node install.mjs in the plugin folder. A clone: git pull, then node install.mjs.
Fixed
cleananddoctorfind the plugin's base folder. The plugin setting reaches only the server, so the server now records the base it runs with in the repository's data directory. Run from a shell, both commands read that record when no flag,.flash-agents.local.jsonorFLASH_STATE_BASEnames a base, and say which base they read and why.cleansays when nothing of the repository was found there.- A worker inside its own long command is not stopped as idle. CPU used by a command's process tree (5 CPU seconds between samples), or a new process in it, counts as activity. An idle stop that still fires while the worker is inside its own command says so.
childcounts its whole tree.treeCpuSecondsandtreeProcessessit beside the process's own CPU, and the progress line shows the tree's figure, so acargo testwhose test binaries do the work no longer reads as idle.currentnames what the worker is waiting on first. A background job nobody waits on comes after it; one its worker has run three commands past is markedoutlivedCommandsand warned about.- "Failed N times" only for real loops. A failure counts as a repeat only when it fails the same way with no file changed in between. A heredoc or
python3 -cwith different code is a new attempt.
New
basePatchFile: start a call from a diff file inside the repository — say a branch in another git worktree. The diff becomes part of the copy's base, so the stored patch holds only the worker's own change, andoverlapsno longer names every sibling.- Overlaps show where patches meet. Each overlap lists the meeting
hunks(ordisjointHunks: true, orhunksUnknownfor a binary file), withsameBase.flash_applylists them too, so a dry run tells you the order before anything is written. limits.maxCacheBytes(FLASH_MAX_CACHE_BYTES,--max-cache-bytes; off by default): bounds a repository's build caches. Stored caches go first, then the cache of the slot idle longest; a busy slot is never trimmed. Sizes are read off the event loop.- New warnings:
- one command rerun with four or more argument lists for twenty minutes or more, with no file changed between;
- a build started beside a running background build of the same tool (
cargo,npm,go,gradleand the like); - a load average above twice the hardware threads for at least a minute (not on Windows);
- a brief that names an existing file outside the root, in the background call's first answer and in the result.
- Workers are told to run one build at a time and to search with
git greporrginstead ofgrep -r … .over a build folder. The delegate skill says to name the helpers a brief should reuse, and to keep input files inside the repository.
Verification
1025 unit tests pass (1001 in 0.1.12); every fix has a test that fails without it. The tests no longer read the real volume's free space or use the real temp state base. The live suite was not run for this release.
0.1.12: read PowerShell in the guard, watch a command's own process tree, and keep slot paths across restarts
Acts on the field report in #5: 0.1.11 tested live on Windows 11 and Linux.
Update: /plugin marketplace update flash-agents, update the plugin, then run node install.mjs in the plugin folder. A clone: git pull, then node install.mjs.
Fixed
-
The guard reads PowerShell. It now checks PowerShell writes:
- cmdlets and their aliases, redirects, comma lists and
-Param:value; [System.IO.File]calls anywhere in a line.
Shells started as programs (
bash.exe -c,sh -c,pwsh -Command,cmd /c) are read with that shell's rules. Windows paths are resolved the PowerShell way (drive root, UNC,USERPROFILE, both slashes), and names match in any case. A PowerShell deletion of the workspace, its parent or home getsCRITICAL_PATH, likerm. Verdicts are the same on every host OS. - cmdlets and their aliases, redirects, comma lists and
-
The process watcher measures the command's own tree. CPU and new processes are counted from the process that runs the command down, leaving out bwrap, the dsh runner and other sandbox processes. Background jobs are watched and listed. Sampling needs no progress token, and a quiet warning stays in the result. Commands longer than 300 characters are matched. A worker that ends with a background command still running is named. The held-open warning is gone, because the sandbox already ends a call's children.
New
-
Slot paths stay the same across restarts. A slot lives at
<state base>/<digest>/slots/<n>, so absolute paths in build caches stay valid.A slot is locked by a file written complete in one step: a hard link, or an exclusive open where there are no hard links. The lock names the owner's pid and command line, so no clock is involved. Only the lock holder moves, empties or deletes a slot. A leftover copy is moved aside and salvaged, never erased, and the sweep never deletes a stable slot.
-
timingin every result: queued, prepare, setup, worker, collect and total, on failures too. -
Changes to the tools:
flash_resulttakesjobIds, andflash_patchandflash_applytakeexcludeBase;- a job that changed nothing stores no patch;
- idle runtimes stop after 10 minutes, read-only ones included;
- reports are written atomically and pruned.
-
.flash-agents.local.jsonmay setstateBaseper machine.doctornames where the setting came from and flags an invalid file.cleanhonours it and a symlinked root. -
flash_statusand the plugin settings: the job listing is shorter, the queue line names idle holders, and the slots setting allows up to 16. -
AGENTS.mdgives every worker on this repository its rules, and the delegate skill has a recipe for a warm SwiftPM build per slot.
Verification
1001 unit tests pass (866 in 0.1.11). The live suite passes 119 of 119 on macOS; it was not run on Linux or Windows for this release. Flash workers built the release, every part had an independent read-only review, and every finding is fixed with a test that fails without the fix.
0.1.11: fail a run that ends on a tool call instead of a report, keep build caches across restarts, and watch for stuck commands
Acts on two field reports in #5: a week of 0.1.10 on Linux and Windows 11, and the root cause of the Windows named-pipe failures.
Update: /plugin marketplace update flash-agents, update the plugin, then run node install.mjs in the plugin folder. A clone: git pull, then node install.mjs.
Fixed
- A run that ends on a tool call is a failure, not a success. A call that ended "ok" with an empty final message, or with its next tool call written out as text (DSML or invoke markup at a line start), now ends as
status: "error"withfinalMessageProblem. The patch so far is kept. A job's state inflash_statusis its result's status, so the two never disagree. - Three-way merges without a saved base. A file without one now merges from the base blob named in the patch's index line, read from your repository. Each file gets its own result (applied, merged, conflict, failed). If any file fails, nothing is written and the result lists the files that would apply. A patch that itself adds conflict markers gets a loud warning, on a dry run too.
- Failure warnings are clearer. They name the command after
nice,timeoutandenvprefixes, and group spellings of one command. They quote the last line of output and read exit codes above 128 as signals. A fail-before run that the brief asked for is reported as info.
New
- Every finished job's whole report is stored on disk by job id (
reportPath), read-only and discard jobs included.flash_resultreturns it after a restart, andflash_cancelreturns the patch it salvaged from the copy. - Build caches survive restarts.
cache:folders move into a per-root store at shutdown, or when the next server finds the old one dead (never while a process of the slot lives). New slots claim them by atomic rename. Stores are trimmed to the slot count and aged out.doctorlists slot, cache and store sizes, a call warns below 10 % or 5 GiB free, andclean --cachesempties them. - A process-tree watcher checks once a minute while a tool call is open, using
psorWin32_Process.current[].childnames the longest-running child. A tree with no CPU and no new process for five minutes is flagged. - Smaller additions:
labelonflash_taskandflash_batch, andreadOnly: true;overlapswith other unapplied patches;- a warning when
source: "head"leaves out uncommitted files; - a bad
state_baselogs its reason before the server exits.
- Worker instructions:
- no devices or whole-disk searches;
- report what was not done or not run;
- run checks after the last edit;
- no unverified "pre-existing";
- keep encodings;
- on Windows:
$env:does not carry over, ninja and parallel MSBuild hang, and cargo has no network.
Verification
866 unit tests pass (770 in 0.1.10). An independent review of the merged change found nine more defects in the cache moves, the marker check, the markup check and the process watcher. All are fixed, each with a test. The live suite was not run for this release, because the flash models it drives were out of credits.
0.1.10: keep a failed merge from writing half a patch, carry nested repositories, and apply settings without a reconnect
Acts on three reports in #5: two days and nights on 0.1.9 (Linux and Windows 11), 25 to 27 October on Linux, and the Python monorepo on Windows.
Update: /plugin marketplace update flash-agents, update the plugin, then run node install.mjs in the plugin folder. A clone: git pull, then node install.mjs. From 0.1.10 on no reconnect is needed: an idle runtime started before the install replaces itself, and both settings files are re-read on every call. That matters in Claude Code's VS Code extension, which has no /mcp reconnect.
Fixed
- A failed three-way apply no longer writes half a patch. Every merged file is read before the first one is written; if a write still fails part-way, the error names what was written. The reported
git cat-file --filters … failedwas any merged file over 1 MiB: Node's default output limit for a child process. Git output may now be up to 1 GiB. - A three-way apply keeps your line endings. With
core.autocrlf=trueand an LF checkout, every merged file came back CRLF. A merged file now matches your checkout as it is, and a worker's own line-ending change is kept. - A warm build cache survives a timed-out call. The slot that ran it is retired, and its
.flash-cache(cargo target,cache:folders) used to be deleted with it, so the retry compiled from cold. It now moves to the replacement slot; setup steps run once more there.isolation.buildCacheEmpty: truesays when a call starts with an empty cache. - A
basePatchalready in your tree is not applied again. The result warns, and the new patch then holds only the new work. - Edits inside a nested git repository are no longer lost. In a whole-root copy they are now carried in the patch, for a nested repository the top one neither tracks nor ignores, or one whose own folder it ignores. Repositories under an ignored folder such as
vendor/ornode_modules/are left as before. One that cannot be read is named inwarnings, and its copy is kept. - Unknown arguments are refused with the nearest known name:
unknown argument "file" for flash_patch; did you mean "files"?. Before, they were ignored, and a misspeltfilesreturned the whole patch.provider,reasoningEffort,agentOptions,permissionsandsandbox, which were ignored, are refused too. - The repeated-failure warning is fairer.
grep,rg,git grep,diffandcmpexiting 1 ("no match", "differs") no longer count. When a later run of the same command succeeded, the warning says so and does not suggest cancelling.
New
- Settings without a reconnect.
.flash-agents.jsonand.flash-agents.local.jsonare re-read at the start of every writing call. A call already running keeps the settings it started with. An invalid file keeps its last good settings and is named in every result until it is fixed. A changedslotEnvrestarts the tree's idle runtime, so the next call's commands see it. - The profile-version warning clears without a reconnect. An idle runtime started before
node install.mjsis replaced at its next use; a busy one says it will be replaced once its calls end. - Queue visibility. A queued call shows its place in line (first come, first served) and which jobs hold the trees;
flash_statushasqueue: { position, of, holders }. - Two worker instructions: list any change beside a stated rule ("only", "never", "one place") under its own heading,
Rules crossed; and on Windows, commands run in PowerShell (7 when installed, else Windows PowerShell 5.1), not bash. - Docs: the Windows sandbox limits (owner-only folders as Python's
tempfilemakes them, and named pipes, whichmultiprocessingand pytest-xdist need),state_baseon Linux for Unix socket path limits,uvfolders and a per-slot cargo registry throughslotEnv, and that a read-only call reads your live tree.
Verification
770 unit tests (716 in 0.1.9) pass. An independent review of the merged change found six more defects; all are fixed, each with a test that fails without the fix. The live suite was not run for this release, because the flash models it drives were out of credits.
0.1.9: keep a patch's base parent, apply a patch in parts, and settle settings per machine
Acts on the third field report in #5 (0.1.8 on Linux and Windows 11), and on the report from a night of reviews on 0.1.8.
Update: /plugin marketplace update flash-agents, update the plugin, then run node install.mjs in the plugin folder and restart Claude Code (or /mcp reconnect plugin:flash-agents:flash). A clone: git pull, then node install.mjs. From 0.1.9 on, a runtime that boots after the install re-reads the profile's stamp, so the version warning clears without a reconnect.
Fixed
drift.commitsBehindafter a restart. The patch record did not keep the base's parent, socommitsBehindwasnullfor every patch read from disk: after a restart, for a salvaged patch, and for a patch from another session. The record now keeps it.- The profile-version warning clears by itself. The stamp is re-read whenever a runtime boots, and a result carries the verdict of the runtime that served it. A runtime that booted before the install keeps its warning until it is replaced.
- A background call with bad arguments fails at once, with the same error a foreground call gives, instead of becoming a job that fails seconds later.
New
-
flash_applywithfiles. Apply only some files of a stored patch, named aschange.filesChangednames them (a rename by either path). Applied files are recorded one by one (appliedFiles, also onflash_patch's first page), so the rest can follow later. A whole apply after a partial one is refused with the count;force: truelifts it.This makes a review's proof tests reusable: run the review as a writing call with
apply: "none", read its tests withflash_patch, thenflash_applywithfilesnaming only the test files. They become the fix round's failing-first tests. -
Settings per platform and per machine.
.flash-agents.jsonmay hold aplatformobject keyed by Node's platform names (win32,linux,darwin, …), and a git-ignored.flash-agents.local.jsonholds one machine's paths, in the same shape. Layers apply in this order: shared file, its platform section, local file, its platform section, server options.slotEnvmerges by name (case-insensitively on Windows);nullremoves an earlier entry.copyExcludeandslotSetupadd up; the lastcopyExcludeDefaultswins.- The start log names the layers it applied. It warns once when the local file is tracked or not ignored by git, and per call when either file changed after start.
-
A start line in the log. Every launch logs
start pid …; flash-agents <version>; node …; <platform>; root …first, andanswered initialize <n> ms after the process startedonce the handshake is answered. A client that never launches the server leaves nostartline; a slow start shows in the number. -
Windows:
state_basepermissions. A folder made directly underC:\gives you "modify" but not the right to change permissions, and every shell command in a slot then fails withSetNamedSecurityInfoW failed (Win32 5).doctorhas a newstate-base-permissionsrow that tries the sandbox's grant on a test folder and names the fix:icacls "C:\fm" /grant "%USERNAME%:(OI)(CI)F". The README says so next to theC:\fmexample. -
node install.mjsalso marks the interactive profile's plugins folder as ES modules, but only when every plugin there is one. It says why when it leaves the folder alone, and never touches apackage.jsonit did not write. -
The description of
sourcenow says"head"applies to writing calls only.
Verification
716 unit tests (660 in 0.1.8) pass, and so do all 116 live checks with 0.1.9 installed in a throwaway profile. Two independent review rounds went over the change. They found defects in file names a patch record or a diff header cannot spell plainly (a tab, a newline, a leading quote, a path containing b/), in the ESM check, and in the order of background refusals. Every defect they confirmed is fixed, each with a test that fails without the fix.
0.1.8: place every path where its command runs, and say when the install drifted
Acts on the second field report in #5 (0.1.7 on Linux, and a Windows night on 0.1.5).
Update, with one more step this time: /plugin marketplace update flash-agents, update the plugin, then run node install.mjs in the plugin folder and restart Claude Code (or /mcp reconnect plugin:flash-agents:flash). A clone: git pull, then node install.mjs. A plugin update replaces the server but not the worker profile, and the guard lives in that profile. From 0.1.8 on, the server warns when the two differ.
Fixed
-
The guard places every path where its own command runs. It used to check redirects, write targets and secret reads against the folder a command line started in. So
cd server && … > ../logwas refused, andcd .. && echo x > ../elsewherewas not caught by the guard at all. It now followscd,pushd,popdandcd -. It also understands:- newlines and here-docs
( … ),{ …; },$( … )and backticksif/for/caseblocks>|,<>and redirects written without a space (echo hi>../f)- prefixes:
NAME=value,env,exec,timeout,nice,bash -lc,eval - more programs that write:
mkdir,chmod,truncate,tar -C,unzip -d,patch
rm -rf <copy>/with a trailing slash is now refused likerm -rf <copy>. A guard error, or commands nested more than 32 levels deep, denies the call withGUARD_ERRORinstead of letting it through. -
flash_applyno longer loses an edit silently. A patch that deletes or renames away a file you changed since the worker's copy was made is refused.threeWay: truemerges it and keeps your edit (marking any conflicts);force: truediscards the edit. -
Node's
MODULE_TYPELESS_PACKAGE_JSONwarning at every runtime start is gone.
New
- Installation drift is named.
install.mjsstamps both worker profiles with the release that installed them. If a profile is from another release or missing, the server warns at start and in every task result, anddoctorhas aprofile-versionrow. - A missing
dshis reported at start. The message names a launcher it finds under another nvm, nvm-windows or volta Node version. - An edited
.flash-agents.jsonis reported. If it changed after the server started, every task result warns until a reconnect. - Windows path length. The plugin option
state_base(FLASH_STATE_BASE,--state-base) moves the folder the copies live under; the path must be absolute. On Windows, a copy whose path is longer than 64 characters gives a warning and aslot-path-lengthdoctor row, because MSVCcl.exefails past 260 characters. flash_statusshows backgrounded commands. While a worker waits on a background command,currentnames that command (background: <job id>). A failing background command counts toward the repeated-failure warning.driftonflash_applyandflash_patch.commitsBehindcounts the commits on your HEAD since the copy was made;filesChangedSincelists the patch's own files you changed since. The apply warns when that list is not empty.- Three more worker guidance lines:
- run commands bare, so their exit code shows
- list anything removed or left out
- name files changed outside the task's files
Verification
660 unit tests (560 in 0.1.7) pass, and so do all 116 live checks with the new guard installed. Three independent review rounds went over the change; everything they confirmed is fixed, each with a test that fails without the fix.
0.1.7: salvage a crashed server's work, and show what a worker is running
Acts on the field report in #5, from a night of unattended work on Windows 11.
Update: /plugin marketplace update flash-agents, update the plugin, then restart Claude Code (or /mcp reconnect plugin:flash-agents:flash). A clone only needs git pull.
Fixed
- A crashed server's work is salvaged again. The restart sweep looked for a dead server's copies one folder level too high (
<pid>/slotsinstead of<pid>/<instance>/slots). It found nothing and deleted the uncommitted work. It now checks both layouts, and the tests build the folder layout a real server creates. A dead server's folder whose salvage fails is kept and retried, at most every 10 minutes and three times in all, instead of deleted. - No connect timeout after a crash. The server answers
initializefirst and sweeps afterwards. A dead server's directory is renamed totrash-*and deleted in the background.cleandeletes synchronously and includes trash in its report;trimleaves trash alone.
New
flash_statusshows what a worker is running: each open command and how long it has run (current). It warns when the same command fails three or more times in a call (warnings).- Automatic retry of early, empty failures: under 180 s and 12 model calls, nothing delegated or only empty reports, nothing changed in the copy. Such a call is retried once, after 10 s, and its result says so.
- Plain values in
slotEnv:"OPENSSL_DIR": "value:C:\\openssl".PATH,PATHEXT,NODE_OPTIONS,NODE_PATHandLD_*/DYLD_*are refused for every kind. - Short patch ids:
flash_apply,flash_patchandbasePatchaccept a unique prefix of at least 8 hex digits. - Patches record their task: a stored patch records its task (the brief's first line, job and session ids) and the worker's final report. A salvaged patch reads its task from the copy.
- A foreground writing call that has to wait for a busy copy says so. A
taskFilejob is labelled with the brief's first line. Workers get three more lines of guidance.
Verification
560 unit tests and 116/116 live checks pass. We also ran the crash case live: a server killed with SIGKILL mid-task, then a new server on the same project. The new server answered in 77 ms and salvaged the worker's change as a patch named after its brief.