Changelog¶
Notable changes to Opsmith. The format follows Keep a Changelog; versions follow Semantic Versioning from 1.0.0. A release that changes the database schema says so under Upgrade notes.
Unreleased¶
1.5.3 - 2026-11-17¶
Changed¶
- Security problems are reported in the open, with a new Security problem issue form,
like any other bug, so that everyone who runs Opsmith can see what is known and when it is
fixed. Email to the maintainer remains for anyone who would rather not post details
publicly.
SECURITY.md, the security guide,CONTRIBUTING.mdand the bug form say so.
1.5.2 - 2026-11-17¶
Fixed¶
- Under
OPSMITH_BASE_PATH, the Compose stack did not start./healthzmoved under the path, so the image's health check failed, and the worker and scheduler, which wait for a healthy API, never started./healthzand/metricsare now also served at the root. - A task found only by a check could never be verified. After its merge the verification
job waited 48 hours and then failed as a dead job. It is now judged by the next run of the
same check. A later check run also no longer counts as re-auditing a task found by the
analyzers, which could have marked such a task
verifiedwithout anything checking it. opsmith applyignored misspelt tables and top-level keys ([projects.setings],[[projects.service]]), so what they held silently did not exist. Unknown keys are now errors at every level, as the docs said.
Added¶
SECURITY.md, issue forms for bug reports and feature requests, and a pull request template with the checklist fromAGENTS.md.- A check in
opsmith.example.toml.
Changed¶
- A review of every page against the code: about ninety corrections across the guides, the
developer docs, the README and the landing page. Among them: how to stop a worker without
killing an execution, backup commands, keeping the sandbox overlay on upgrades, the
Hostheader behind a proxy, revoking a key, the check stage's context size, which settingsopsmith applyreplaces, and whatsucceededmeans for a check run. AGENTS.mdis reorganized for agents: the invariants come first, then How to work in six phases (start from a known state, change with care, prove it, keep the docs true, ship, keep the workspace safe), with new rules to stay in scope, plan a cross-module change, verify claims before acting, count only runs you saw, scan what leaves before a push, clean up, know when to ask, and end with a report. It also gains the Claude API rules.CLAUDE.mdno longer repeatsAGENTS.md: it lists the steps most often skipped and what is specific to Claude Code.
Upgrade notes¶
opsmith applynow refuses a file with an unknown top-level or project key. If an apply fails after the upgrade, the message names the key: it was being ignored before.
1.5.1 - 2026-11-17¶
Changed¶
- The landing page, the README, the concepts, getting started, UI tour and operations guides,
and
AGENTS.mdnow describe checks. The Checks guide explains how to run one every few hours (a cron step restarts each day), and the README's "what has been tested" table says how far checks are tested: through the real gateway with a scripted model, not yet against a live model in CI.
1.5.0 - 2026-11-17¶
Added¶
- Checks: your own reviews, on a schedule. A check is an instruction in plain words
("review this code for missing authorization checks"), with the files it covers and a cron.
A model carries it out, and each problem it reports becomes a finding and then a task for
review, like any other. Use them for a nightly bug hunt, a monthly hardening pass or a
weekly security review. A finding is kept only if it quotes the file it names; the
instruction is sent outside the data, redacted; excluded paths, binaries and symbolic links
are never read; and every run is bounded by files, characters, findings and the budgets.
Checks need a model: without one, a run is recorded as skipped, with the reason. Admins can
run a check now with
POST /projects/{id}/trigger("job": "check"). See Checks.
1.4.3 - 2026-11-17¶
Changed¶
- Session rules: release every change.
AGENTS.mdandCLAUDE.mdnow require that each commit carries its changelog entry, that each change reachingmainis released at once (tag and GitHub Release), and that the docs and the documentation site describe a change before it is pushed, with the Pages deploy checked after.
1.4.2 - 2026-11-17¶
Fixed¶
- The secrets audit reported references as hardcoded credentials.
${DB_PASSWORD},$PASS,$(cat /run/secrets/x), a{{ template }},"$PWD:/src"in a shell command, and a variable or call in program code (this.password = password) each became a medium-severity finding and a task. On a real Java and TypeScript code base, every one of the 50 tasks the first audit drafted was of this kind or in a test. A literal is still reported: inside a string or a comment, unquoted in a configuration file, and as the default of${X:-...}. Redaction and the diff gate are unchanged. - The offline triage did not recognize tests named by convention:
Login.test.tsx,api.spec.ts,auth_spec.rb,AccountTest.java,AuthTests.kt,LoginIT.java. A credential in them now counts for less, and a placeholder one is dropped, as in atests/directory. - A value made only of mask characters (
"******") is marked as a placeholder.
1.4.1 - 2026-11-17¶
Fixed¶
opsmith runcould not start on a new SQLite database. Its API, worker and scheduler each created the schema at the same moment ("table alembic_version already exists"), and the database was left half made. Schema upgrades on SQLite now take the database's write lock, as they take an advisory lock on PostgreSQL, so one start migrates and the others wait. Starting withopsmith initfirst, as the guides do, was not affected.- Two processes opening a new SQLite file at once could fail with "database is locked": SQLite does not apply the busy timeout when a file is switched to WAL. The switch now waits for the other writer, for up to the same 30 seconds.
1.4.0 - 2026-11-17¶
Added¶
- A VictoriaLogs connector (
victorialogs), read-only, through the LogsQL query API. It reads each window oldest first, page by page, and takes an entry's level from the message itself when it states one, because Docker's syslog driver calls all of stderr an error. See Configuration. OPSMITH_BASE_PATH: serve the API and UI under a path of a shared domain (for example/opsmith) behind a reverse proxy. Links, redirects and assets follow the path, and the sign-in cookie is scoped to it. See Deployment.
1.3.3 - 2026-11-17¶
Added¶
- Music on both videos: an original instrumental track, no voice, composed for each and
ending on a resolved chord as the video fades out. The reel's is upbeat (112 BPM), the
demo video's calmer and quieter (92 BPM), under the captions.
scripts/soundtrack.pysynthesizes it from scratch, so there is nothing to license, and every run writes the same audio. The site's player now starts with the sound on when pressed.
Removed¶
- The narration support added to
scripts/demo_video.pyin 1.3.2. The videos have music and captions, and no voice-over, so the code had no use.
1.3.2 - 2026-11-17¶
Added¶
- A 30-second vertical video for phones and social media
(
docs/images/reel.mp4, 1080 x 1920, 2.2 MB, no sound): the Opsmith loop in seven beats, from "your logs already know what's broken" to "it never merges" and "verified", with real screens from the offline demo. It plays on the landing page of the documentation site and is linked from the README and the UI tour.scripts/reel_video.pyrenders it; every run gives the same video. scripts/demo_video.pycan narrate the demo video: each caption has a spoken line, the recording waits for each line to finish, and the lines are mixed in where their beats start. Without the recordings it records silently, as before.
1.3.1 - 2026-11-17¶
Changed¶
AGENTS.mdandCLAUDE.mdhold every session to the same rules: run the full build and tests before and after a change; test every change, starting a bug fix with a failing test; stay secure by default and prefer correct, simple code; never merge with a regression; and end every session by updating the Markdown docs, the documentation site,CHANGELOG.mdand, for significant decisions,docs/DECISIONS.md.CLAUDE.mdalso records the Claude API patterns introduced in 1.2.1.CHANGELOG.mdhas an Unreleased section for changes waiting for a release.
Added¶
- A work report on the releases 1.2.1 to 1.3.1.
1.3.0 - 2026-11-17¶
Added¶
- A documentation site at https://deadboyccc.github.io/Opsmith/: every Markdown page in
the repository, grouped by reader, with search, light and dark mode, a table of contents on
each page and copy buttons on code. It has a landing page with a quick start and a reading
order, and works on phones. It is rebuilt from
mainon every push, so its changelog is always this file. Pull requests build it strictly: a broken link fails the build (.github/workflows/docs.yml). make sitebuilds the site andmake site-servepreviews it. The tools are in a separatedocsdependency group, not in the application's dependencies.
1.2.4 - 2026-11-17¶
Documentation only. Every page was compared with the code as it now is.
Changed¶
- The README states what is now tested (coverage and its floor; Claude through the real SDK over a mocked transport) and links the security audit and the decision log. It no longer lists a connector compatibility matrix as open work: the matrix exists.
- Compatibility: how Claude is tested, and the FastAPI versions supported.
- Deployment: the values
OPSMITH_LOG_LEVELaccepts. - Operations: troubleshooting entries for the refusals added in 1.2.1 and 1.2.2 (a repository saved with an address or branch git must not be given, build output that trips the credential gate, a deferred audit, an oversized request), and the retries of a busy model endpoint.
1.2.3 - 2026-11-17¶
Tests and continuous integration. No change to what Opsmith does, apart from the one fix below.
Changed¶
- CI fails when test coverage drops below 93% (
fail_underinpyproject.toml). Coverage is 94.5% (it was 92% at 1.2.0 and 93% at 1.2.2, measured the same way). - CI audits the locked dependencies on every push with
pip-audit(newmake audit), so a newly published advisory turns the next build red. - Workflow actions are pinned to a commit, and the live-model workflow to an Ollama
release instead of
latest, so a moved tag cannot change what CI runs. - Coverage is measured by path, so
src/opsmith/migrations/env.py(which alembic loads as a file) is counted; it was reported as never run.
Added¶
tests/test_runtime.py: the worker and scheduler loops, including their recovery after an error and their stop signal; JSON logging; the object store; the Docker sandbox's failure paths; provider plugins; notifications without a project; thealembiccommand.- Tests for the GitHub provider's git commands, token handling, API failures and the commit-status fallback for CI.
Fixed¶
- A write to the object store that failed half-way left its
.tmpfile behind for good. It is now removed, and a complete file is synced to disk before it is published. - A test of the lease heartbeat tolerated less than a second of scheduling delay; it now tolerates nearly three, and still fails when the heartbeat stops.
1.2.2 - 2026-11-17¶
A security release: the result of an audit of the whole code base (Security audit). Upgrade if tenant admins other than yourself configure projects, or if the API port is reachable by anyone you do not trust.
Upgrade notes¶
- A
githubrepository'surlmust be anhttps://URL without credentials, and every repository'sdefault_brancha plain branch name. A repository that does not meet this is refused when it is next used (its audits and executions fail with a message naming the field) until it is corrected. - Request bodies are limited to 1 MiB, and to 32 MiB for pushed signals. A larger request
gets
413. - A file connector's
pathmay not contain... - Protected-path patterns cover more than before: a pattern now also covers everything below a directory it names. Nothing that was protected stops being protected. A pattern that can never match is refused when settings are saved.
- A change that adds a file containing something credential-like is now refused even when
the file is binary. A check command that leaves build output behind (compiled bytecode
holding a string from the source, for example) can trip this; ignore such files in the
repository's
.gitignore. - A review's
commentandinvestigationare limited to 20,000 characters. A redaction pattern must match at least one character.
Security¶
- A repository's address and branch were handed to git unchecked. A
githubrepository could name a directory on the worker, which made Opsmith clone any repository the worker can read, another tenant's included, pastOPSMITH_LOCAL_ROOT. A branch such as--upload-pack=<command>ran that command on the worker, outside the sandbox, at the repository's second audit. Both values are now checked when saved (API andopsmith apply) and again on every use; git is told where its options end and, for GitHub, may use https only. - A credential placed after a form feed passed the push-time credential scan. The diff was split at form feeds, bare carriage returns and the Unicode separators as well as at newlines, and the rest of the line was then never scanned.
- Protected-path patterns such as
infra/,/infra/**or./infra/**matched nothing, so a directory its admin believed protected was not. - A file-connector pattern could leave the local root with a
..after a wildcard, and the refusal then named the file it had found: file names outside the root could be read one pattern at a time. Contents were not exposed. - Request bodies had no size limit and are read before the key is checked.
/metricsgrew without bound: the HTTP method was a label as the client sent it.- Redaction missed credentials in common spellings: JSON carried inside a JSON string,
*_PASSand--pass,Authorizationwith schemes other than Bearer and Basic,Cookieheaders, OpenAI project keys, PGP private key blocks, and fields named in the plural. - A connector request had no total deadline, only a timeout for each read.
- The validation report quoted on a corrective retry could close its own data block. Review text was unbounded. A redaction pattern that matched the empty string was accepted.
Fixed¶
- Epoch timestamps in milliseconds were masked as credit cards one time in ten, which inflated the redaction counts and changed log templates from one line to the next.
Added¶
docs/SECURITY-AUDIT.md: what the audit looked for, what it found, what was fixed and what is still open.
1.2.1 - 2026-11-17¶
A bug-fix release. The whole code base was read for defects; each fix below has a test that failed before it. No database change.
Upgrade notes¶
- Some settings are now limited to the size of the column they are stored in:
audit_cronand a schedule'scronto 100 characters,executor_agentand a stage'sproviderto 50, a stage'smodelto 100. PostgreSQL already refused longer values, later and with a server error. A project whose stored settings exceed a limit (possible on SQLite only) must be corrected, throughopsmith applyorPATCH /projects/{id}/settings. OPSMITH_LOG_LEVELmust be a level name (DEBUG,INFO,WARNING,ERROR,CRITICAL, in either case). Anything else now stops the process at start with a configuration message instead of a traceback.- FastAPI 0.121 or newer is required (the lock file has 0.163).
- A connector created without
interval_secondsnow takes the project'singest_interval_secondsinstead of 900. With the default settings nothing changes.
Added¶
service_environmenton a connector inopsmith.toml, to link it to a service that exists in more than one environment.environmentin theconfigof theprometheusandlokiconnectors.
Fixed¶
Execution and review:
- With the real Claude SDK, every turn of the Claude coding agent failed before a request
was sent. The SDK refuses a non-streaming call above about 21,000
max_tokensunless the caller sets the timeout itself, and the agent's default is 32,000. A stage configured above that limit failed the same way. The timeout is now passed explicitly. The tests used a hand-written client, which is why this was not seen; one now runs the real SDK. - A tool call cut off by
max_tokensbroke the Claude agent's conversation. The half-writtentool_useblock was answered with "Continue.", which the API rejects because the block has notool_result. It is now answered as an error, and never run. A tool input that is not an object crashed the loop and lost the record of what the run had spent. - An agent edit rewrote every line ending of a CRLF file, so a one-line fix became a diff of the whole file, and the credential scan then treated every old line as newly added. Edits now keep the file's line endings.
- A work directory that could not be created left the task in
executingfor good. It now fails the attempt, and the task returns to review with a report. - The
execution.serializedaudit entry was never stored: it was written in the transaction that the deferral then rolled back. - A list or object as the value of an edited field was a server error (
priority,effort_estimate). A review'sreason_codewas only checked for rejections; any other decision stored it unchecked, which PostgreSQL refuses above 50 characters. - A plain Approve in the UI was recorded as an approval with edits when the task's title held a line break, and the title was rewritten.
Jobs and notifications:
- A dead job only notified its project for three of the nine job kinds. The others
(ingest, analysis, audit, execute, verify) carry no project id, so their
job.deadnotification was logged and never reached the project's webhook. - A job dead-lettered for being deferred too often kept its lease.
- Two audits of one repository could run on the same checkout (a manual trigger during a scheduled run), the second resetting the tree under the first. The second now waits.
- The per-project and per-service locks blocked everything that references the locked row on PostgreSQL. While a model drafted tasks, nothing could be ingested, analysed or reviewed in that project, and a review that opens a lead could deadlock with the aggregation. The locks no longer block inserts.
Ingestion. One record that an earlier stage accepted and a later one could not process used to stop a connector, or a service's bundles, for good: every retry and every later job met it again.
- A message holding half a surrogate pair or a NUL; a
file:linereference with thousands of digits; a severity such as²(a digit to Python, not a number); a service or environment name with a NUL; two metric points of 1e308;lag_secondsorbaseline_secondstoo large for a duration. Each is now cleaned, or dropped and counted inskipped_records. - A pushed timestamp at the edge of the calendar was a server error (which log shippers
retry forever). It is a
422. - About one in two hundred quoted epoch-millisecond timestamps was read as a date in the
year 17xx and the record lost without being counted: Python reads
1795011232359as the ISO basic format. - A connector linked to a service passed on only the service's name, so records without
an environment went to a new
productionservice of that name. The Prometheus connector could not produce metrics for any other environment at all. ingest_interval_secondshad no effect: every connector was created with 900 seconds of its own.- A few error lines with very long file paths removed every log template from a bundle.
- Each connector run opened an HTTP client and never closed it. A connector registered
through an entry point had no
type_name, so its error messages raisedAttributeError.
Models:
- An SDK's own exception (rate limit, overload, connection) passed the stages that handle model errors. The job was rolled back together with the usage of the calls already billed, and task drafting's fallback never ran. Every provider failure now reaches the stages as a model error.
- The built-in rules' own output could fail validation: a metric series with a few labels made a title longer than 300 characters, the bundle stayed unconsumed, and every later analysis of that service failed on it.
- A NUL in model output failed the run on PostgreSQL, after the call had been billed.
- One 429 or 5xx from an OpenAI-compatible endpoint failed the call at once, which for analysis is the whole slot. The client now tries twice more.
- A reply without token usage was recorded as free, so the budgets never saw it. Content given as a list of parts was read as the text of a Python list.
- Input that Claude reads from or writes to its prompt cache was not counted (no request uses the cache yet, so nothing was mis-billed).
Audits and source control:
- A file name that is not UTF-8 made its repository impossible to audit. The audit failed after every analyzer had run, on every attempt.
- Line numbers were wrong in a file that contains a form feed (or a Unicode line separator): findings pointed at the wrong line, carried another line's snippet, and the built-in fixes could not find their line.
- GitHub CI status read only the first page of check runs, so a commit whose failing or unfinished run was listed after the first 30 was reported as passing.
API, UI and command line:
- A write was acknowledged before it was committed. A commit that failed (a dropped
connection, a constraint) still answered
201with the id of a row that was never stored. /docswas blank: the strict Content-Security-Policy blocked the assets it loads./docsand/redocnow get a policy that lets them; every other path keeps the strict one.opsmith applychecked far less than the API. A value of the wrong type ended in a traceback or was stored as it was; services were looked up by name alone, so with one name in two environments every connector was linked to the last one in the file.- An id containing a NUL was a server error on PostgreSQL instead of
404. Deeply nested settings JSON in the UI was a server error.
1.2.0 - 2026-11-16¶
Security¶
- Code planted in a nested repository could run on the worker. If the repository has a
submodule, sandboxed code could fill it with a repository whose configuration defines a
filter driver; the platform's
git statusafter the checks descended into it and ran the driver outside the sandbox. A nested.gitanywhere in the workspace now refuses the execution before the platform runs git on it, and the agent cannot write one. - A
-difforbinarygit attribute hid a file from the credential scan. The staged diff is now read as text. - A tenant could send the platform's model API key to its own URL when the operator stored
that key as a file in the secrets directory (Docker secrets): a connector's
credentials_refcould name the same file. That file, and links to it, are now reserved. - The API accepted the UI's session cookie for state-changing calls without the UI's
Origincheck, so a form on a sibling subdomain could, for example, resubmit a task. The cookie now works for reads only; changes need the key in a header. - Values nested under a credential-named field were stored unredacted, for example
{"password": {"new": "..."}}. Everything under such a field is now masked. - An audited repository could reconfigure gitleaks and trivy. A committed
.gitleaks.tomlwith an allowlist of.*hid every leaked secret; a committedtrivy.yamlwithoutput:made trivy write its report to a path of the repository's choosing on the worker and leave nothing to parse, which read as "no findings". gitleaks now gets the platform's own configuration and trivy runs from an empty directory. Verified against gitleaks 8.30.1 and trivy 0.76.2. - Files whose name starts with
.opsmith-homeskipped the protected-path gate, and a sandbox home replaced by a link or file could be committed. The home is now matched by exact path and always removed before the gates.
Added¶
- A compatibility matrix: what Opsmith talks to, which versions
are verified and how. The Loki and Prometheus connectors now run against real servers in CI
on every push (Loki 2.9.17 and 3.7.10; Prometheus v2.53.5, v3.5.5 and v3.16.0), through
tests/test_connectors_live.py. - A one-minute demonstration video (
docs/images/demo.mp4, with a GIF preview in the README), recorded from the offline demo byscripts/demo_video.py: a reviewer approves a task in the UI and it goes through toverified. - A mascot and a social preview image (
docs/images/brand/).
Fixed¶
- A worker that had lost its lease could hand its job to a third worker. When its handler then failed or asked to be retried later, it requeued the job although another worker was already running it. Every write after the handler now checks, under a row lock, that the worker still owns the job, and lease recovery skips a lease that is being renewed.
- Two workers aggregating the same project at once created duplicate tasks (and so duplicate pull requests) on PostgreSQL: both read the same new findings and both drafted a task. Aggregation now locks the project, and an analysis run locks its service, so they run one at a time per project and per service.
- A task merged a second time was never verified. The verify job was keyed by task, so after a regression and a new pull request the second merge reused the first, finished job.
- Post-merge verification gave up silently after four hours without data, which with the
default daily audit left most code-only tasks
mergedfor good. It now waits up to 48 hours and then fails the job (dead, audited, notified) instead of succeeding quietly. - With the real Claude SDK, refusals, truncated replies and out-of-bound values bypassed the
budgets.
messages.parsevalidates the reply inside the SDK and raised a pydantic error before the provider could read the stop reason, so the billed usage was never recorded, the gateway's corrective retry never ran, and the whole job was retried (and billed) again. The provider now sends the same schema throughmessages.createand the gateway validates the reply. The tests run the real SDK over a mock transport instead of a hand-written fake. - Settings too large to use were accepted and then failed inside every later job, for
example
recently_closed_days = 1000000000(atimedeltaoverflow in aggregation) or a schedule interval beyond PostgreSQL'sINTEGER. Durations, intervals, counts and budgets now have upper bounds, checked when saved through the API oropsmith apply(Configuration). - The gitleaks analyzer never produced a finding with current gitleaks. It asked for the
report on
/dev/stdout, which gitleaks 8.30 does not write to; every audit recorded a gitleaks failure. It now uses-, gitleaks' documented name for stdout. - One unstorable record stopped a connector for good. A record nested more than 255
levels deep, or with a service name or environment longer than its column (on PostgreSQL),
failed the whole window, which was then retried until the job died, and nothing after it
was ingested. Such records are now dropped and counted in the run's
skipped_records. - Loki JSON lines lost fields. A line with both
messageandmsg(orlevelandseverity) dropped the second one; it is now kept in the attributes. - The Prometheus connector could read past the end of its window. It rounded the window's
times to the millisecond, so an end such as
…36.1766became…36.177and Prometheus returned a point that also belonged to the next window. Times are now truncated. Found by running the connector against real Prometheus servers.
1.1.0 - 2026-11-16¶
Upgrade notes¶
- Opsmith now migrates its database. Every process brings the schema to the newest revision when it starts; there is no separate step. A database created by 1.0.0 is adopted as it is, with its data. Stop all processes and take a backup before upgrading (Operations).
- A process now refuses to start, with a
database schema:message, on a database from before 1.0.0 or one migrated by a newer release.
Added¶
- Schema migrations with Alembic (
src/opsmith/migrations/; ADR-002). Revision0001is the 1.0.0 schema. On PostgreSQL, processes that start together are serialized by an advisory lock.tests/test_migrations.pyfails when the models and the revisions disagree, on SQLite and PostgreSQL. New dependency:alembic.
1.0.0 - 2026-11-16¶
The first stable release. A hardening release that follows the independent project review. It adds no product features. The review response maps the review's items to this work.
Upgrade notes¶
- The database schema changed (
reviews.scope_digest,schedules.custom). There are no migrations yet, so start from a fresh database. POSTGRES_PASSWORDis now required bydocker-compose.yml. Set it in.env.- The sandbox default is now
docker. For development with local repositories setOPSMITH_SANDBOX_BACKEND=local. Thelocalsandbox is refused for GitHub repositories. - File connectors and local repositories are disabled until you set
OPSMITH_LOCAL_ROOTto the directory they may read. file:credential references must live underOPSMITH_SECRETS_DIR(default/run/secrets), andenv:references may not nameOPSMITH_*, database or LLM-key variables.serviceAPI keys can no longer read anything. Use aviewerkey for read access.- Approvals recorded before this release are not honoured: they carry no scope digest. Approve the task again.
- Sandbox images must contain the project's dependencies. The sandbox has no network, so a check command that installs packages fails.
OPSMITH_LLM_PROVIDERwas never read and has been removed from.env.example.
Security¶
- Approvals are bound to the work. An approval stores a digest of the task's scope and is re-checked before the run and before the push (the second check used to come after the push). A changed task voids the approval and returns to review.
- The aggregator no longer rewrites approved tasks. New evidence, including model-chosen paths and severities, could change the paths, severity and priority of a task a human had already approved. It now waits until the execution is over.
- Protected-path gate bypass fixed. A file named like
dir/x -> ywas reported asy. Changed files are now read from git's NUL-delimited output. - Credential gate bypass fixed. An added line starting with
++was mistaken for a diff header and never scanned. - Files written by the platform's check run are now gated. They used to be committed after the gates had already run.
- Sandbox escape through git closed. Sandboxed code could plant hooks or configuration in the
checkout's
.git, which the worker executed when committing. Git metadata is now snapshotted before the agent starts and only the snapshot is used afterwards, with hooks disabled. - Sandbox policy is enforced.
dockeris the default;localis refused for non-local repositories; a missing Docker daemon refuses the execution instead of downgrading. - Docker sandbox hardened: read-only root filesystem, size-limited
/tmp, no swap, an init process, kill-by-name on timeout. The local sandbox kills the whole process group on timeout. - Compose fails closed on the database password, and
opsmith initno longer prints it. - Credential references are restricted so a connector or repository cannot send the platform's own secrets, or arbitrary files, to a URL its admin chose.
- Host filesystem access is confined to
OPSMITH_LOCAL_ROOTfor file connectors and local repositories. A tenant could previously ingest another tenant's stored signals by pointing a file connector at the data directory. - Connector URLs are validated; link-local and cloud-metadata addresses are refused, response sizes and timeouts are bounded, and upstream response bodies no longer appear in error text.
servicekeys are limited to pushing signals, as documented. They could read tasks, configuration and the audit log.- Prompt framing cannot be closed by its content. A log line containing the closing
</untrusted_data>tag ended the untrusted block early. - External scanners: file names can no longer be read as options by
rufforsemgrep;ruffignores configuration shipped in the repository; tools run with a scrubbed environment. - The secrets analyzer masks every secret on a line. A second secret on the same line was stored in clear text in the finding.
- Redaction covers more shapes: quoted JSON keys, prefixed keys such as
DB_PASSWORD, URL credentials without a username,Authorization: Basic, values of sensitive-named keys in structured attributes, and card numbers next to other digits. - UI: logout is a POST with an origin check; stored URLs are only rendered as links when they are http(s); login errors are fixed messages; the origin check works behind a TLS-terminating proxy.
- Reviews of the same task are serialized, every editable field is validated, branch names are
restricted to plain
opsmith/...paths, and pushes outside that namespace are refused.
Fixed¶
- The coding agent's spending was outside the cost budgets. It was recorded on the execution run only, so the daily cap and the per-run cost cap ignored the most expensive stage. It is now written to the same ledger as every other stage; an execution does not start a model-driven agent when the day's budget is spent, and the agent stops at the per-run cap or what is left of the day. A model call that fails part-way keeps what was already spent.
- The demo failed when run twice within one second against the same database (this made CI fail intermittently).
- Costs were lost when a model call failed. Failed, refused and truncated calls were billed
but not recorded, so budget caps could not trip. They are now recorded and the run is marked
failed. - An unlisted model id was priced at $0, disabling cost caps. It is now priced as the configured model, or at the most expensive known rate.
- The per-run token cap now holds for calls outside a run and reserves the full output allowance.
- A slow job could be run twice. Leases were not renewed, so a job longer than the lease was handed to a second worker. Workers now renew the lease while they are alive.
- A job that kept crashing its worker was retried forever. It is now dead-lettered.
- One unusable schedule stopped every schedule for every tenant. It is now disabled and audited. Impossible cron expressions and zero intervals are rejected when settings are saved.
- Schedule changes took effect one run late, and timing set through the API was overwritten by the next settings change.
- A refused execution left its task parked in
approved. It now returns to review. - A retry after a failed PR creation could never push to GitHub, because the branch already existed. Opsmith's own branches are now force-pushed.
- A rejected task was re-proposed on every audit after a severity re-rating, with nothing changed on it.
- Repeated detections of the same static finding no longer inflate evidence strength.
- Runtime evidence is no longer attached to a task in a different repository with the same path.
- An open task now wins over a newer closed one when both match new evidence.
- Blast radius and confidence are computed on attach the same way as at creation.
- Model output is bounded (line numbers, signal and path lengths), an out-of-range
finding_indexno longer binds a lead to another lead's finding,from_lead_idis honoured, and repeated evidence references no longer add strength. - Redaction in the gateway no longer makes the model's valid citations look fabricated.
- Ingestion: one malformed record no longer wedges a connector; non-finite metric values and naive timestamps no longer break bundling; Loki results beyond the page limit are no longer dropped; overlapping jobs no longer bundle or ingest the same data twice; late-arriving records are picked up through an ingestion lag; catch-up windows are capped; metric series whose labels are redacted stay distinct; a flat baseline no longer hides drops.
- Audits: one malformed file no longer discards an analyzer's findings for the whole repository; a failed analyzer no longer advances the audit baseline; scanner failures are no longer recorded as clean runs; PEM private keys are detected; the default policy glob matches root-level files; colliding findings are all kept; large repositories no longer overflow the scanners' argument list.
- UI: the tasks and jobs pages no longer show a macro object in the title; validation errors
say what is wrong; a plain Approve from a browser is recorded as
approve. - The API returns
409for duplicates instead of500, boundslimit, and refuses links across projects. Validation errors with custom messages no longer cause a500. opsmith applyprints a clear error for an invalid file and writes nothing.- Workers and the scheduler survive a temporary database outage, and a worker stops after its
current job on
SIGTERM. - The demo can be run twice against the same database.
Added¶
docker-compose.sandbox.yml: an explicit overlay that lets a containerized worker use the Docker sandbox. The runtime image now includes the Docker CLI.- No model vendor lock-in: an
openaiprovider and coding agent for any OpenAI-compatible endpoint (OpenAI, Azure OpenAI, Ollama, vLLM, LiteLLM, ...), per-stage prices (input_price_per_mtok,output_price_per_mtok), and provider plugins through theopsmith.llm_providersentry point. Verified live against Ollama. push_logsandpush_metricsconnector types for sources that push instead of being polled.- Settings:
OPSMITH_WORK_DIR,OPSMITH_LOCAL_ROOT,OPSMITH_SECRETS_DIR,OPSMITH_BASE_URL,OPSMITH_OPENAI_BASE_URL,OPSMITH_OPENAI_API_KEY_REF,OPSMITH_OPENAI_JSON_MODE. - Adversarial tests (
tests/test_hardening.pyand others), tests that start real containers, and the whole suite on PostgreSQL in CI. - A provider parity test: the acceptance scenario runs with every model stage and the coding agent going through the OpenAI-compatible HTTP path and must end exactly as the offline run.
- A UI tour with screenshots generated from the offline demo
(
scripts/ui_screenshots.py). - Documentation, split by reader (index): a user guide under
docs/guide/(getting started, UI tour, concepts, configuration, models, deployment, operations, API, security) and contributor pages underdocs/development/(architecture, testing), plusCONTRIBUTING.mdand this file.scripts/check_docs.pyruns in CI and compares the docs with the code: links, anchors, cited tests, environment variables, project settings, API endpoints, notification events, connector types, states, providers and commands. - A weekly and on-demand workflow that runs the OpenAI-compatible provider and coding agent against a real Ollama server.
0.1.0¶
The initial MVP: ingestion and ETL, chained batch analysis with carried-over leads, repository audits, aggregation and ranking, the human review gate, sandboxed execution, pull requests with CI tracking, post-merge verification, the REST API and UI, and the offline demo.