Skip to content

fix(harness): classify OpenCode v2 errors and retain snapshots - #3466

Merged
yohamta0 merged 4 commits into
dagucloud:mainfrom
tbrandenburg:fix/opencode-v2-errors-snapshots
Nov 1, 2026
Merged

yohamta0 merged 4 commits into
dagucloud:mainfrom
tbrandenburg:fix/opencode-v2-errors-snapshots

Conversation

@tbrandenburg

@tbrandenburg tbrandenburg commented Oct 31, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Follow up on the merged OpenCode V2 adapter (#3367) with typed 404 handling and changed-file timeline preservation, adapted to current main after #3456.

Changes

  • Decode V2 _tag errors on session routes, including message history. Only SessionNotFoundError means the session is unavailable; missing commands, models, locations, permissions, and forms retain execution-error diagnostics. Malformed or oversized error bodies cannot establish session absence.
  • Preserve typed V2 errors during resumed-session setup and reconciliation, while retaining unavailable/waiting handling for host outages and missing sessions. Regression tests check that setup errors do not interrupt or replace user-supplied sessions.
  • Normalize assistant snapshot.files into a stable patch event. Append it after usage normalization so a snapshot appearing during refresh does not duplicate the usage timeline event. Repeated refresh tests verify updated files, stable sequence/timestamp, and no redundant progress update.
  • Extend real OpenCode 2.1.4 validation with an unknown command, a real file write and snapshot, and missing-session responses. Existing question/resume and cancellation-settlement checks remain in that test.
  • Link managed OpenCode sessions from the README action table.

Scope boundary

Command timeout policy is deliberately not changed. The old handoff removed the HTTP timeout using the caller context; current upstream instead uses context.WithoutCancel to learn the outcome of potentially accepted submissions. Removing its remaining bound would require a separate decision. Admission, deduplication, interruption, and settlement logic from #3456 is not replaced.

A separate temporary diagnostic against the new patch and OpenCode 2.1.4 confirms the remaining gap: a command containing a 31-second shell substitution returned after 30.568 seconds with unavailable, PromptSent=true, and zero provider calls. This diagnostic is not part of the submitted patch. The preexisting command-admission failure path's unconditional interrupt also remains outside this PR's scope.

Validation

New regressions were run before their fixes and failed for typed 404 classification, snapshot omission, resumed-host routing, swallowed history errors, malformed trailing JSON, and snapshot/usage timeline duplication.

Verified on Linux with Go 1.27.3 and OpenCode 2.1.4, using isolated temporary state and a local deterministic HTTP provider:

go test -race -count=1 ./internal/runtime/builtin/harness/... ./internal/opencodehost/...
go test -race -run '^TestV2' -count=20 ./internal/runtime/builtin/harness/
DAGU_TEST_OPENCODE_V2_EXECUTABLE=/path/to/opencode-2.1.4 \
  go test -race -run '^TestRealV2Run$' -count=2 ./internal/runtime/builtin/harness/
go vet ./internal/runtime/builtin/harness/... ./internal/opencodehost/...
go fix -embedlit=false -diff ./internal/runtime/builtin/harness/... ./internal/opencodehost/...

The affected package race suite passed; the real-V2 test passed twice; scoped vet/fix and formatting checks produced no findings. Repository-configured golangci-lint 2.13.2 reported zero issues. Independent review findings were addressed and re-reviewed.

Validation limitations: directory-boundary-corrected lint exposes the same 22 findings on untouched upstream and this patch; the changed-lines comparison reports zero new issues (see #3465). Windows/Darwin lint attempts exceeded three-minute package-loading budgets. Full make fmt/make check, repository-wide Go tests, frontend/API suites, and conformance were not run; this is a package-scoped harness change with no schema, frontend, workflow-language, or CLI changes.

Related Issues

Closes #3467 (tracking issue for the fixes in this PR).

Replaces the remaining applicable work from philo-x/dagu#1. Related: #3367, #3456, and the lint-exclusion follow-up #3465. This does not reopen or claim to resolve the closed cancellation issue #3448.

Checklist

  • Code follows the project style guidelines
  • Self-review of the code has been performed
  • Tests have been added or updated as needed
  • Documentation has been updated as needed
  • Changes have been tested locally

Summary by cubic

Classifies OpenCode v2 HTTP 404 responses so a session is marked unavailable only when the error is SessionNotFoundError. Missing commands, models, locations, permissions, and forms keep execution-error diagnostics instead, and typed errors survive resume and history reconciliation.

  • A run that cannot reach the host before it has a session now fails immediately, so the step can fall back to the CLI or retry; only resumed sessions wait for the host to return.
  • History polling fails a step only on typed 404s naming a missing resource; transient errors such as 5xx or unreadable pages retry on the next poll as with v1 hosts.

Assistant snapshot.files are normalized into a stable patch timeline event appended after usage normalization, so a snapshot appearing during refresh does not duplicate the usage event. Regression tests cover 404 classification, malformed and trailing error bodies, host-down and resume routing, transient vs. typed history errors, and snapshot/usage timeline stability. Real OpenCode 2.1.4 validation now includes an unknown command, a real file write with snapshot, and missing-session responses, and the README action table links managed OpenCode sessions via a relative doc link. Command timeout and admission/interruption policy are unchanged.

Written for commit 626c4b8. Summary will update on new commits.

View guided diff

Summary by CodeRabbit

  • New Features
    • Assistant messages now include file changes from content snapshots as patches.
  • Bug Fixes
    • Managed OpenCode sessions now report unavailable or missing sessions more clearly, while preserving other session and history errors.
    • Failed message refreshes on newer OpenCode sessions now surface their errors.
  • Documentation
    • The harness.run action description now links to documentation for managed OpenCode sessions.

Distinguish missing sessions from other V2 404 errors without losing typed failures during resume or history reconciliation. Preserve assistant snapshot files as stable timeline patches without duplicating usage events.

Add failure-first regressions and real OpenCode 2.1.4 validation. Keep the command submission timeout and cancellation/admission policy from upstream dagucloud#3456 unchanged.
@coderabbitai

coderabbitai Bot commented Oct 31, 2026 •

Copy link
Copy Markdown

Review in Change Stack →

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration
  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: Organization UI
  • Review profile: ASSERTIVE
  • Plan: Advanced
  • Run ID: 8112c585-089b-43d5-ba9d-c90f3d90d786

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • Trigger review

No actionable comments were generated in the recent review.

Recent review info
Run configuration
  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: Organization UI
  • Review profile: ASSERTIVE
  • Plan: Advanced
  • Run ID: ce677877-d170-49e8-a88e-27792b040cc5
Commits

Reviewing files that changed from the base of the PR and between ddda00c and 5a41ec7.

Files selected for processing: 5
  • README.md
  • internal/runtime/builtin/harness/opencode_client.go
  • internal/runtime/builtin/harness/opencode_managed.go
  • internal/runtime/builtin/harness/opencode_v2.go
  • internal/runtime/builtin/harness/opencode_v2_test.go

Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 7 remain after this review.

Walkthrough

OpenCode v2 now classifies typed API errors, propagates selected managed-session errors, and includes snapshot file paths in normalized assistant messages. The changes also add v2 run and error tests, plus a README link to managed-session documentation.

Changes: Managed OpenCode

Layer / File(s) Summary
V2 error classification and managed-session handling
internal/runtime/builtin/harness/opencode_client.go, internal/runtime/builtin/harness/opencode_managed.go, internal/runtime/builtin/harness/opencode_v2.go, internal/runtime/builtin/harness/opencode_v2_test.go
Non-2xx responses use typed error handling. V2 managed-session paths preserve or propagate selected errors. Tests cover tagged 404 responses, resume errors, history errors, and unavailable sessions.
Snapshot files in normalized messages
internal/runtime/builtin/harness/opencode_v2.go, internal/runtime/builtin/harness/opencode_v2_test.go
Assistant messages with snapshot files include a patch part. Tests check patch contents, identity, and refresh behavior.
Real-run coverage and documentation
internal/runtime/builtin/harness/opencode_v2_test.go, README.md
The real-run test initializes a Git worktree and checks command and tool-call behavior. The README links to managed OpenCode documentation.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

  • dagucloud/dagu#3367: This PR changes the OpenCode v2 routes and message normalization introduced there.
  • dagucloud/dagu#3331: This PR extends the normalized message path with snapshot patch parts and tests usage-event ordering.
  • dagucloud/dagu#3456: This PR refines error handling in the managed OpenCode session flows added there.
Merge Risk: Minimal · previous review 5a41e

No actionable merge-blocking issue was identified.

Pre-merge checks: Passed 4 · Failed 1

Failed Failed checks: 1 warning

Check name Status Explanation Resolution
Docstring Coverage Warning Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff with cyclomatic complexity of at least 2. Documentat… Write docstrings for the functions missing them to satisfy the coverage threshold.
Passed Passed checks: 4 passed
Check name Status Explanation
Title check Passed The title clearly and concisely describes the main changes: OpenCode v2 error classification and snapshot retention.
Linked Issues check Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check Passed Check skipped because no linked issues were found for this pull request.
Description check Passed The pull request description follows the required template. It includes a clear summary, detailed changes, scope boundaries, related issues, validation results and limitations, and a completed checkli…
Full details: Description check

Explanation

The pull request description follows the required template. It includes a clear summary, detailed changes, scope boundaries, related issues, validation results and limitations, and a completed checklist.

Full details: Docstring Coverage

Explanation

Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff with cyclomatic complexity of at least 2. Documentation-only edits and deletions within surviving functions are included; unchanged diff context and fully deleted functions are excluded. Analyzed 5 functions across 4 files. (1 skipped: 1 unsupported.)

Resolution

Write docstrings for the functions missing them to satisfy the coverage threshold.

Finishing Touches
Generate unit tests (beta)
  • Create a new PR
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

Session setup treated an unreachable host as an unavailable session, so
a run with no session yet ended waiting instead of failing. That
skipped the legacy CLI fallback and step retries. Host outages now wait
only for resumed sessions, which may still exist on the host.
A single 5xx or unreadable history page failed the step while its
OpenCode turn kept running. The poll now fails only on typed 404s that
name a missing resource, and retries other errors on the next tick as
it does for v1 hosts. Untagged 404 messages no longer end in "()".
@yohamta0

yohamta0 commented Nov 1, 2026

Copy link
Copy Markdown
Member

Thanks Tom, nice work on this, especially the real 2.1.4 coverage.

I pushed two small fixes on top:

  • A fresh run that can't reach the host fails again instead of waiting, so the CLI fallback and retries still work.
  • One flaky history read no longer fails the step. Typed 404s still do.

@cubic-dev-ai cubic-dev-ai Bot left a comment •

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

1 issue found across 6 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="internal/runtime/builtin/harness/opencode_client.go">

<violation number="1" location="internal/runtime/builtin/harness/opencode_client.go:180">
P2: The V2 create-session route is excluded from typed 404 decoding because its client path is exactly `/session`. A typed 404 from `POST /api/session` therefore loses its `_tag` and `errManagedNotFound` classification; include the exact create route in decoding while preserving the existing-session unavailable mapping only where appropriate.</violation>
</file>

Reply to a comment to ask cubic a question or push back. It learns from your replies.

View guided diff | Re-trigger cubic

Comment thread README.md Resolved
}

func (c *openCodeClient) apiError(resp *http.Response, method, path string) error {
if resp.StatusCode != http.StatusNotFound || !strings.HasPrefix(path, "/session/") {

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2: The V2 create-session route is excluded from typed 404 decoding because its client path is exactly /session. A typed 404 from POST /api/session therefore loses its _tag and errManagedNotFound classification; include the exact create route in decoding while preserving the existing-session unavailable mapping only where appropriate.

Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At internal/runtime/builtin/harness/opencode_client.go, line 180:

<comment>The V2 create-session route is excluded from typed 404 decoding because its client path is exactly `/session`. A typed 404 from `POST /api/session` therefore loses its `_tag` and `errManagedNotFound` classification; include the exact create route in decoding while preserving the existing-session unavailable mapping only where appropriate.</comment>

<file context>
@@ -180,6 +176,36 @@ func (c *openCodeClient) json(ctx context.Context, method, path string, body, ta
 }
 
+func (c *openCodeClient) apiError(resp *http.Response, method, path string) error {
+	if resp.StatusCode != http.StatusNotFound || !strings.HasPrefix(path, "/session/") {
+		return fmt.Errorf("OpenCode API %s %s returned %s", method, path, resp.Status)
+	}
</file context>

A root-absolute link breaks outside GitHub's renderer. Match the
README's other ./ links.

@yohamta0 yohamta0 left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks again, Tom. Careful work, and the real OpenCode runs made this easy to trust. Merging.

@yohamta0
yohamta0 merged commit d6e3f44 into dagucloud:main Nov 1, 2026
19 checks passed
connorwang added a commit to txehq/dagu that referenced this pull request Nov 7, 2026
…eping every fork behaviour (#230)

* ui(canvas): add canvas Flow view, step run panel and workflow panel (#3389)

* UI: workflow-workspace/messages, dags/components, visual-editor/run
* Fix minimap, cache and step name edge cases
* Ignore cancelled connect drags and wrap badges
* Fix reused-step label, waiting outputs, deep JSON

* ui(editor): add YAML document model, canvas model and run source logic (#3390)

* UI: visual-editor/doc, visual-editor/canvas, visual-editor/run
* Fix schedule, name and reuse edge cases
* Read quoted workflow defaults as plain params
* Fix schedule, chain diff and step reference edits

* ui(editor): Add edit session, YAML format and quick fixes, run actions, tab strip (#3391)

* UI: visual-editor/canvas, visual-editor/doc, visual-editor/(root)
* Fix formatter, dirty check and duplicate bugs
* Guard lookups by user keys and label sheet panel
* Fix compare, step format, clipboard and run label

* ui(editor): Add canvas graph model, Add step panel, issues popover, run bar (#3392)

* UI: visual-editor/canvas, visual-editor/run, dags/components
* Fix editor edge cases in issues, forms and canvas
* Fix step links, dead issue rows and env keys
* Make an inserted step precede an inferred target

* ui(canvas): Add layered layout, dependency edges, status styles, reference picker (#3393)

* UI: visual-editor/canvas, visual-editor/inspector, visual-editor/references
* Fix reference picks, continue_on alias and focus
* Fix reference picker focus and chip scrolling
* Fix continue_on off state, custom refs and layout

* ui(canvas): Add export, flow builder, upstream data panel, output capture (#3394)

* UI: visual-editor/run, visual-editor/references, visual-editor/canvas
* Keep step hashes complete and fix run views
* Cover the placeholder route name in tests
* Check the with block of run steps in the editor
* Fix step cache hashes, captures and run views

* ui(canvas): Add node cards, launchers, inspector sections, client issues (#3395)

* UI: inspector, canvas and dags components
* Clear the webhook URL copy timer on unmount
* Wait for secret visibility in the webhook tab test
* Fix review findings in the editor nodes and panels
* Preview cron runs in the scheduler time zone
* Ignore IME keys and clear the webhook secret flag
* Fix step edits that the server refuses

* ui(editor): Add visual editor, step inspector and YAML file panel (#3396)

* UI: visual-editor/inspector, dags/components, visual-editor/references
* Update spec editor E2E tests for the new editor
* Show the already-formatted note as info
* Ask before restoring a draft of an older version
* Keep save checks and editor state across updates
* Show the server steps when the YAML does not parse
* Log a failed load of the client issue checks
* Save the pending draft when the key changes
* Send the focused field on Ctrl+S and read 1e3

* ui(lists): Open runs on the workflow page; add tab and schedule helpers (#3397)

* UI: pages, dags/components, cockpit
* Expect the run page from a run progress card
* Load client issues before the node frame tests
* Trust the server next run and tighten UI tests
* Cover minimap and key-value fixes in tests
* Test that a nested object block stays open
* Test cancelled connect drags and expanded links
* Test the notice of a YAML the editor cannot read
* Ignore inherited keys in schedule descriptors
* Test key-value rows that share a key
* Test the exponent form of a number value

* ui(workspace): Add workspace URL state, legacy redirects and run list links (#3398)

* UI: dags/components, dag-runs, components
* Expect run list rows to open the run page
* Fix review findings in the run list and workspace

* ui(workspace): Add workspace data layer, run facts and Workflow panel (#3399)

* UI: dags/components, workflow-workspace/data, workflow-workspace/dock
* Fix status i18n, save race and test-run polling
* Keep the test-run poll alive across renders

* ui(workspace): Add run start controls, visible runs and reuse source (#3400)

* UI: workflow-workspace/bar, workflow-workspace/step, workflow-workspace/data
* Fix run checks of the reuse source
* Fix range, leave guard and cache wait edge cases
* Scope run links to the workflow they belong to
* Keep chosen and cached sources the server accepts

* ui(workspace): Add data provider, bar identity, continuous panel and step list (#3401)

* UI: workflow-workspace/continuous, workflow-workspace/bar, workflow-workspace/data
* Bound the samples cache and fix chart input
* Test that run history polls under re-renders
* Sync the webhook chip and fix bar and chart input
* Read a continuous step again when its run ends

* ui(logs): Add run actions cluster, cached logs and log popout page (#3402)

* UI: dags/components, workflow-workspace/bar, workflow-workspace/popout
* Mark the source-match note as info
* Fix stop-all, retry link, and run log details
* Show the stop-all toast as a note
* Let the log chevron and jump work as documented
* Fix stale log stream, reschedule and run links

* ui(workspace): Add Activity dock, step run view and workflow page (#3403)

* UI: dags/components, workflow-workspace/dock, workflow-workspace/step
* Expect runs to open on the workflow page in E2E
* Open run links of DAGs named apart from files
* Expect a sub-run link to drop its step
* Give sub-runs an address and show file-less runs
* Test the summary link of a waiting agent
* Read step reuse from the run record in step facts
* Fix draft step runs, foreach items and run links

* test(workspace): Add integration tests for the workspace shell and step view (#3404)

* UI: integration tests of the editor
* Expect runs to open on the workflow page in E2E
* Wait for the URL in the step run action tests
* Cover run chip and reused-inputs fixes in tests
* Tighten the workspace tests that could not fail
* Cover sub-run links and run view fixes in tests
* Clear the refresh feedback timer on unmount
* Test the progress text of the run details
* Test reused and waiting steps in step facts
* Test draft steps and foreach logs in step runs

* ui(pages): Retire classic run pages and open runs on the workflow page (#3405)

* UI: dags/components, pages, visual-editor/inspector
* Cover null-prototype issue maps in tests
* Keep sub-runs and workspaces on old run links
* Test __proto__ fields and IME keys in inspector
* Fix Mermaid retry, run link and Kanban Ctrl-click
* Resolve latest run links and trust the next run
* Test that Show in all workspaces selects them

* test(editor): Add integration tests for the editor, inspector and run overlay (#3406)

* UI: integration tests of the workspace
* Cover fixes of earlier PRs in tests
* Keep integration tests independent of order
* Test run with checks and IME keys in outputs
* Test the server steps of an unparsed YAML
* Test route edits, step labels and output cycles
* Test a step inserted on an inferred edge
* Test continue_on, step cache and inspector fixes
* Assert the step drawer closes in StepRunPanel test

* test(ui): Add repository-wide guard tests and ESLint boundary rules (#3407)

* UI: visual-editor/doc, src
* Expect a handler duplicate to be refused
* Close gaps in the UI guard tests
* Test renames of scalar targets and trimmed names
* Test big integers and trimmed names on paste
* Tighten UI guard tests and cancel stale CI runs

* e2e: Add workspace helpers and fixtures, update existing specs (#3408)

* e2e: shared helpers, fixtures and updated specs
* Split the e2e disk-conflict test by branch
* Fail e2e teardown when a DAG is not deleted

* e2e: Add workspace specs, part 1 of 2 (#3409)

* e2e: workspace specs (1/2)
* Tighten workspace e2e checks and share helpers
* Pin the run and Back target in workspace e2e

* e2e: Add workspace specs, part 2 of 2 (#3410)

* e2e: workspace specs (2/2)
* Drop dead e2e absence checks, assert step clear
* Expect a disabled run caret for executing roles
* Check profile scope and tab names in e2e

* examples: Add the workspace demo workflows (#3411)

* Examples: workspace demo
* Make the workspace demo DAGs run as described
* Keep JSON outputs intact in the demo workflows

* feat(harness): support managed OpenCode v2.1.4 (#3367)

* Move the correction messages into agentstep.AppendCorrection (#3421)

* Nix support (#3284)

Signed-off-by: Daniel Gomez <da.gomez@kernel.org>
Co-authored-by: Yota Hamada <yohamta@gmail.com>

* Add local registry support for DAG tools (#3268)

Co-authored-by: Yota Hamada <yohamta@gmail.com>

* Save scheduler state once per burst of events (#3450)

* fix(browser): read a page again when a navigation replaces its script context (#3455)

* Keep the text of params a schema does not type (#3459)

* Reload symlinked DAGs when their targets change (#3460)

* Fix lost dispatches and remote child exit codes (#3461)

The dispatch store deleted a task left unclaimed for the reservation
TTL, but the scheduler had already dequeued the run and a dag.run
parent waits on its child, so a run dispatched while its workers were
busy or just stopped stayed queued, or its child not started, forever.
A pending task now waits until a worker claims it; the coordinator
drops it only once its run no longer awaits it.

A dag.run child that failed on a worker returned no error, so the
parent step recorded exit code 0. The remote runner now returns the
child's last failed step error, as the in-process runner does.

Fixes #3033

* Treat backslash as path text in path fields (#3464)

A backslash before $ in a filesystem path was read as a Dagu escape, so a Windows path such as C:\done\${NAME} lost its separator and its reference. Path fields now read a backslash as path text: workflow paths (dotenv, working_dir, stdin, stdout, stderr, output paths) and the action inputs each executor declares as paths. Other fields keep the escape. Spec 003 defines the rule and lists the path fields.

Closes #3347

* Confirm managed OpenCode interrupts settle (#3456)

OpenCode accepts an interrupt before its cleanup settles, and can take
a prompt before its turn starts. Dagu sent a fire-and-forget abort,
recorded the session as aborted, and on retry created a new session in
the same checkout while the old turn could still run.

A cancelled, timed-out, or stopped managed run now interrupts its turn
once the turn has started and waits, within a bound cancellation
cannot cut short, for the host to report the session idle. It records
aborted or abort-unverified. A retry, restart, or DAG retry settles the
discarded session before creating a replacement, and an unstarted
prompt is redelivered to the same session instead. In a session the
user supplied, only Dagu's own turn is interrupted.

A stop also no longer starts a fallback provider or the legacy CLI.
A flaky browser test that required no error from a click closing its
window now judges the click as the operator does.

Closes #3448

* Record a dialog under the lock of its answer (#3468)

* Treat an act that closes its own window as done (#3462)

* build(deps): bump the actions-patch group with 2 updates (#3472)

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* Classify OpenCode v2 errors and keep snapshots (#3466)

A v2 session route returned 404 for a missing command, model, or
location as well as a missing session, and Dagu read all of them as an
unavailable session. A 404 now marks the session unavailable only when
OpenCode tags it SessionNotFoundError; other typed 404s fail the step
with their tag. Assistant snapshot files become a stable patch event in
the session timeline.

A fresh run that cannot reach the host still fails, so the CLI
fallback and retries apply, and a transient history error no longer
fails a running step.

Closes #3467

Co-authored-by: Yota Hamada <yohamta@gmail.com>

* build(deps): bump the go-patch group with 10 updates (#3471)

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Yota Hamada <yohamta@gmail.com>

* Cut idle polling and disk reads behind live updates (#3457)

* Peek at workflows and runs from the lists (#3453)

* Show the run history of each step (#3469)

* fix(dags): sort the DAG list by status and last run on the server (#3119)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Yota Hamada <yohamta@gmail.com>

* build(deps): bump github.com/go-chi/chi/v5 from 5.3.2 to 5.4.0 in the go-security group across 1 directory (#3322)

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

* Read the step for the l key from the written URL (#3474)

* Suggest harness providers in the YAML and visual editors (#3473)

* fix(browser): treat a page's analytics as noise, not saves (#3463)

* config: read server secrets from *_file keys (#3214)

Co-authored-by: Yota Hamada <yohamta@gmail.com>

* Select a run from its column in the History grid (#3475)

* Show a picked run from the history list at once (#3476)

* Preview the next runs of a DAG's schedules (#3431)

The DAG page showed one next run, and a schedule being edited showed
"No upcoming run", so the only way to check a schedule was to wait for
it to fire.

DAG details, the spec endpoint, and validation now return the next ten
fire times of the start schedules, each with its schedule entry and the
zone it runs in. The canvas schedule launcher lists them as the YAML
changes, and the MCP read and change tools include them. The list comes
from the definition alone; nextRun stays the scheduler-aware projection.

* Read unindented quoted strings in the visual editor (#3477)

* Open a step side panel on a click on the canvas (#3478)

A click on a step of the workflow page selected it and showed nothing
until the Run panel or the node view was opened. A pointer click on a
step card now opens a non-modal panel over the right of the canvas,
with the content of the former step drawer: run facts, Retry step and
Change status through the node menu path, the run details and the
step definition, then the Log, Output, History and Actions tabs. It
follows the selected step and run; Esc, the close button or an empty
canvas click closes it, and a double click still opens the node view.

* Let schedules follow business calendars (#3435)

A schedule could not skip holidays or pick business days of the month,
so payroll, settlement, and reporting jobs needed shell checks or
hand-kept cron exceptions.

A start schedule can name a calendar file, a list of holidays with
optional weekdays, extra business days, and includes, and keep only
chosen business days with business_days. The scheduler, catch-up, the
upcoming runs preview, dagu ls -n, and the new dagu calendar fires and
check commands all use the same rule, and DAGs reload when a calendar
they use changes. Spec 084 defines the behavior.

* Let clicks through the step panel only when needed

The step panel ignored the pointer for half a second after it opened,
so that the second click of a double click on a step under it still
reached the step. In that time a click on a tab or a button of the
panel went to the canvas instead, and could close the panel.

The canvas now gives the point of a step click, and the panel ignores
the pointer only when it opens over that point. A panel that opens
beside the click, or from a link, takes the pointer at once.

* Manage business calendars through the API (#3449)

* Add a Calendars page to browse and edit calendars (#3451)

* Sync calendars with Git (#3470)

* Keep the focus ring off the canvas after a menu click (#3479)

* Open the peek from a workflow name in the list

A click on a row of the workflow list showed its peek, but the name
link filled the whole Name column and opened the full page instead.
The name is where people click, so the peek was hard to reach.

The name now shows the peek on a plain click, like the rest of the
row. It stays a real link, so a middle, Ctrl or Cmd click still opens
the workflow page in a new tab, and the peek keeps its Open workflow
action.

* Remember whether the YAML panel is open

The YAML panel of the workflow page lived only in the URL, so each
workflow opened from the list or a link started with it closed, and
users who keep it open had to open it again every time.

The page now stores the panel state when it opens or closes, and a
workflow whose URL opens no other view shows the panel when it was
open last. A link to a step, a panel or a run snapshot keeps its own
view.

* Keep compressing proxies off live updates

The live update stream set only Cache-Control: no-cache. The UI dev
server compresses text/event-stream responses, and its gzip buffer held
every small event back, so the stream never delivered its first event.
The client gave up after 15 seconds, reconnected, and the workflow page
showed a run frozen at its first snapshot.

The stream now also sends no-transform, which compressing proxies honor
by passing the events through untouched.

* Preload the workflow page while the browser is idle

Opening a workflow from a list, a run or an artifact loaded the code of
the workflow page and of its canvas on the click. The old page stayed on
screen with no feedback until it ran: about 4 seconds on the dev server
before the page and 9 before the graph, about 1 and 1.5 in a production
build.

The app now loads both while the browser is idle after sign-in. The
canvas preload goes through the page chunk, so it loads the same chunk
that the page imports. On the dev server the page now shows in about
0.6 seconds and the graph in about 1.1.

* Show the workflows of a file as tabs

A file with child workflows offered them in a select at the right end of
the canvas toolbar, where they were easy to miss. Before the redesign
they were tabs above the workflow.

The parent and each child workflow are now tabs at the start of the
canvas toolbar row, before the canvas controls. A child tab has the
icon of a step that calls a workflow, and the arrow keys move between
the tabs.

* Open the workflow that a step calls

Before the redesign, a double click on a sub-workflow node of a run
opened its child run. The new canvas opened the YAML of the step
instead, and the only way into a child run was its context menu.

Opening a step that calls a workflow, by a double click, Enter or
Open step, now opens that workflow: its child run in the shown run
(a picker when there are several), else the tab of a child workflow of
the same file, else the page of the called workflow. Any other step
opens its step view as before.

* Show child runs in the form of the workflow page

A child run opened on the sub-run stack in the classic run details
view: a Mermaid graph, a table of steps and tabs that look nothing like
the workflow page that opened it.

On the workflow page the stack now shows the child run with the facts
of the run, the canvas graph with the state of each step, and the log
of the picked step in a pane over the graph. Opening a step that has
child runs opens them on the same stack, with a picker when there are
several. The classic pages keep their view.

* Leave Esc to the page when no step is selected

The canvas took every Esc pressed while it had the focus, to clear the
selection, also when nothing was selected. The page around it never saw
the key, so an Esc could not reach the views that close on it.

The canvas now takes Esc only while a step is selected.

* Open workflows and runs over the lists as the real page

A click on a row of the workflow and run lists opened a preview: a
read-only graph and a run summary that looked nothing like the workflow
page. Every action in it, a click on a step included, left for the
workflow page, which then had to load.

The row now opens the workflow page itself over the list, at its own
URL, with the step panel, the Activity panel, logs and artifacts. The
list stays below with its location and filters. Back, Esc, the backdrop
and the close button return to the list; the expand button keeps the
workflow page as the page, and a reload shows it alone. The workflow
names, run IDs and artifact buttons of the run lists open the same way,
and a plain link still opens a new tab on a modified click. A ?peek link
of the preview opens the workflow page.

* Make the workflow overlay read as a layer

The overlay took 88% of the window from the right edge with a faint
backdrop, so opening a workflow looked like a move to another page.

It is now a panel inset from the window edges, with rounded corners and
a border, at most 76rem wide, over a darker backdrop that covers the
whole page below. On a narrow screen it still fills the window.

* Keep the list mounted when a workflow opens over it

The routes of the page got a location only while a workflow showed
over it. With a location, the router renders the routes under a
provider of their own, so the switch mounted the page again: an
expanded group of the Executions page collapsed and a typed filter was
lost when a row opened.

The routes now always get a location, the current one or the one kept
below the overlay.

* Edit child workflows of a file on their tab

A child workflow of a multi-document file showed on its tab as a
read-only graph; only the YAML of the whole file could change it, and
the step panel of the parent stayed open over it.

The tab now opens the visual editor on the document of the child: an
edit goes back into the file at the place of that document, and Save
saves the file. Opening a step of a child shows its form. A child has
no run bar and no Run step, as it runs only from the step of its parent
that calls it, and a switch to its tab clears the step of the parent.
A node also hides Run step when the host says that the step cannot run.

* Let the width of the workflow overlay change

The workflow overlay had a fixed width, wide or narrow for every
window and every workflow.

Its left edge is now a resize handle: a drag or the arrow keys change
the width, and a double click gives the default back. The width is kept
across visits, stays between 640 pixels and the window less 96, and
applies from md up; a narrow window shows the overlay at full width.

* Step through the rows of a list in the workflow overlay

Before the redesign, Up and Down switched the run or workflow in the
modal of the Workflows and Executions pages to the next row. The
workflow overlay had no such step: each row had to be closed and the
next one opened.

The lists now give the overlay their rows in order: the workflows of
the table, the runs of the table, and the runs of the open groups.
Up and Down, and the arrow buttons of the overlay, show the row before
or after in place of the shown one, so close and Back still return to
the list. The page keeps the arrows that it uses, such as those of the
canvas, a text field, tabs or a pane that scrolls.

* Open the workflow overlay on the click

The overlay showed only once the router had moved to the workflow page,
after its code and its first render, so a click seemed to do nothing for
a while. The frame of the page below also took the shape of the
workflow page, as it read the URL: the breadcrumb of the list went away
and its padding changed.

The overlay now shows at the click with the loading frame of the
workflow, which the page fills when it is ready, and the frame of the
page below follows the location kept under the overlay.

* Show the times of a run on its start in the workflow bar

The workflow bar showed when a run started only as a relative time, such
as 7d ago, with the date and time of the start in a native title and
the finish nowhere.

A hover or keyboard focus on the start now opens a tooltip with the
scheduled, queued, started and finished times of the run that are set,
and its duration, as the stacked run facts have it.

* Show how to connect AI apps under remote access (#3481)

* Support defaults.max_concurrent for fan-out steps (#3482)

* Close UX gaps of the redesigned UI (#3483)

* fix(browser): tell unlabeled table inputs apart by column header (#3480)

Signed-off-by: Ryo Kusnadi <ryokusnadi@gmail.com>

* Let the run inspector work through Dagu Cloud (#3484)

* Fix the run selector of the Activity timeline (#3489)

* Open sub-runs from Activity and the child workflow tab (#3490)

* Publish npm packages through trusted publishing (#3492)

* Install the AI skill for the user, not the folder (#3493)

* Run the latest Dagu release from the Helm chart (#3494)

* Keep block scalars with trailing spaces as blocks (#3497)

* Send queue state to Dagu Console (#3495)

* Log each failing precondition and mask its secrets (#3501)

Failing preconditions are now logged one by one, with the condition
as written and its error, so the run log shows which condition
stopped a DAG or step. Lifecycle handlers now log with their step
tag. Secrets in condition errors are masked in the run log and in
stored run status, which also fixes a leak already present for
DAG-level preconditions.

Closes #3500

Co-authored-by: pommetjehorlepiep <8982487+pommetjehorlepiep@users.noreply.github.com>

* Show the run definition in the Activity panel (#3502)

* Resolve router targets by step id (#3511)

* Grow the bar of a running step in the Timeline (#3510)

* Let the YAML panel grow wide and maximize (#3509)

The YAML panel and the other right panels stopped at 960px, so a long
command line stayed cut off even on a 4K screen (#3507). A panel may
now take its whole zone but a 240px strip that keeps the canvas and
the resize handle in reach. The YAML panel also gets a Maximize
toggle that fills the zone and a Restore that brings back the stored
width.

All right panels now read one live width, so the canvas follows the
edge while it moves and a panel opened later matches the last
resize. The handle keeps the cap of a narrow window, and a click
without a drag no longer replaces a wider stored width.

The node view tests now clear the step view mode in sessionStorage,
which made five column tests fail when the run order changed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Pin a mail server certificate by SHA-256 with tls_cert_sha256 (#3503)

Signed-off-by: Ryo Kusnadi <ryokusnadi@gmail.com>

* Keep runtime-only references out of editor issues (#3513)

A reference that gets its value only during a run, such as
${context.paths.work_dir}, an env value from a profile, or a param
with no default, came back from the validate endpoint as an editor
issue. The editor drew it as a squiggle that read like an error,
against Spec 003: a runtime-only notice is not a validation warning
(#3512). The issues now carry defect notices only, and
valueReferenceNotices keeps every notice with its class.

The issues popover counted and listed every notice too. It now counts
and lists defects only, and keeps its Details button so runtime-only
notices can still be opened.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Let env secrets read runtime profile entries (#3517)

* Check that a killed process tree stopped, not that it vanished (#3519)

* Label start params by title, description first (#3521)

* Fix graph fit, dock and editor gaps in the new web UI (#3518)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* Let xlsx.write set a number format per column (#3522)

with.formats pins the number format of named columns, so an integer
amount can show as 17,500 instead of 17500. It applies to new sheets,
to rows appended below existing ones and under either style, and a
blank format is rejected up front.

update_rows with an empty rows list now succeeds and leaves the
workbook untouched, so a scheduled run with nothing pending no longer
fails or rewrites the file.

Github-Issue: #3043
Github-Issue: #3074
Co-authored-by: Yota Hamada <yohamta@gmail.com>

* Bound concurrent browsers in write gate tests (#3538)

* Speed up opening a workflow page (#3537)

* Detect PR changes against the current base tip

pull_request.base.sha can predate the base branch tip that GitHub
merges the PR into. Diffing it against the merge commit counted
commits that landed on main since then as PR changes, so a UI-only
PR ran the full Go test matrix after an unrelated Go change merged.

Diff against the merge commit's first parent, which is the base the
PR was merged into.

* Align the DAG schema with the parser (#3533)

* Resolve action file paths from the step working directory (#3534)

* Clear copy feedback timers on unmount (#3535)

* Make light warning badges readable (#3548)

* Slide the workflow overlay in from the right (#3550)

* Write the selected run with the workspace URL

The run selection wrote ?run with setSearchParams, which builds on the
URL of the last render. A write of the same tick, such as ?ndv=docked
from the YAML panel that the page restores, was lost. The selection now
goes through the URL writer of the workspace, which builds on a pending
write. Its URL is canonical, as for every other workspace write.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show the remembered YAML panel from the first render

The workflow page opened the YAML panel it remembered only once the
editor had loaded, so the graph first fit the whole canvas and then
shrank when the panel came. The page now decides in the first render
of a workflow, and the editor lays out the graph with the panel.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep the YAML panel open on Escape

Escape closed the YAML panel as it closes the expanded step view, so a
key that the user pressed for something else hid the panel. The panel
is one that the page keeps: Escape now closes only the step view, in
the editor and in the definition of a run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Keep legacy links through the run redirect

The run selection wrote through the canonical URL writer, which builds
the path and the query from the workspace state. The redirect to the
latest run then dropped a legacy path such as /history and keys such as
dagRunId before the legacy redirect could read them, and the History
tab of those links stayed closed. The run is now written as one query
key, with the path and the other keys as they are, on top of a write of
the same tick as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Collapse Activity before the editor mounts

With Activity open, the YAML panel that the page restores collapsed it
only after the panel had claimed its place. The graph first fit the
zone above the open Activity, then fit again when Activity collapsed,
and changed its size. The page now yields Activity before the first
paint for a panel of the first render, and mounts the editor once it
has.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Show the frame of the workflow page while it loads

The workflow page loaded with a bar and a spinner in the middle of the
whole page, which flashed for a load of a few hundred milliseconds, and
then every part of the page came at once. The loading view now shows
the frame of the page: the bar, the dots of the canvas, and the YAML
panel and Activity as the page will show them, at their stored sizes.
The spinner shows only for a load longer than 400 ms.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Leave quoted bare references in run to the shell (#3526)

* Retry the rename that re-points a symlink while Windows refuses it (#3556)

TestEntryReaderPollsSymlinkTargets/Recursive fails on windows-latest now
and then: the rename of a.yaml.staged over a.yaml returns "Access is
denied". The test re-points the link with a single rename, as nix
home-manager does, while the entry reader polls the link every 50 ms.

On Windows a rename over a file is refused while any handle to that file
is open, whatever its share mode and even when it only reads attributes,
as os.Stat and os.Lstat do. Probing on Windows 11, a goroutine looping
os.Stat on a regular file made about 8% of concurrent renames over it
fail with access denied, as did an attribute-only CreateFile with every
share mode including FILE_SHARE_DELETE; only os.ReadDir on the directory
never did. The reader has to open the link to learn where it points, so
no reader-side change can close the window; a tool that swaps files on
Windows retries the rename instead, as cmd/go's robustio does.

repointSymlink now renames through renameOverWatchedFile, which on
Windows retries a rename refused with access denied for up to five
seconds. Elsewhere a rename replaces an open file at once and the helper
renames once.

Fixes #3551

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* Open the Run panel tab that a link names

A legacy /history link redirects to ?dock=history, which opens the
History tab of the Run panel. The YAML panel that the page restores at
its first render then yielded the Run panel, and the tab stayed
collapsed. A link that names a tab of the Run panel now restores no
panel, as a link that names a step or a panel already did.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Glide the graph when the room of the canvas changes

A panel or Activity that opened or closed made the graph jump to its
new fit, and a change of both made it jump twice. The fit that follows
the room of the canvas, or a change of the graph, now glides there in
200 ms. A pan, a zoom or another view change ends the glide, and
reduced motion keeps the jump.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Switch right panels without an empty moment

A click on YAML closed the Workflow or launcher panel at once, while the
YAML panel showed only once the URL had changed. For that moment the
right of the canvas stood empty: Activity came back, and the graph moved
and moved back. The other panels now close when the YAML shows, in the
same render, so the graph stays where it is.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Remember the Workflow panel as the user left it

The workflow page remembered the YAML panel only: a Workflow panel that
the user kept open was closed on the next open. The page now stores
whether the Workflow panel is open and shows it from the first render
when no link opens another view, with Activity collapsed before the
editor mounts and the frame of the panel in the loading view. An Escape
of the page no longer closes it, as it no longer closes the YAML panel.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Drop the exit status of ending a browser whose processes have exited (#3555)

On Windows the browser runtime ends a launched browser with taskkill
unless it has already seen the browser exit. taskkill reports the
processes it could not end through its exit status, among them processes
that were exiting as it walked the tree, and the runtime returns that
status from Browser.Close although the browser is gone. The step's Close
then fails, which TestStagehandCloseKeepsProfileCookies hit on
windows-latest with "exit status 255".

closeBrowser already waits for the browser's recorded process tree once
the runtime returns. When that wait finds the tree gone, the exit status
of a command the runtime ran to end it says nothing more, so it is dropped
from the error; whatever else the runtime reports, such as a profile that
could not be removed, is kept. Without a trackable tree the runtime's
error is returned as before.

Fixes #3552

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* Restore run status, run actions and waiting actions (#3561)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* Report the machine's resources to Dagu Console (#3560)

* Add Settings and set up remote MCP when connecting (#3562)

Admins could not find, or easily turn on, the Dagu Cloud connection
and remote MCP access: Plan & features was reachable only from the
license badge, remote access appeared only after connecting, and the
default free connection left Dagu Cloud refusing MCP calls.

A Settings entry in the sidebar now opens one page with the plan,
Dagu Console connection, monitoring, remote access, and remote nodes.
Remote Nodes, External Monitoring, and Incidents leave the sidebar,
and /license and /remote-nodes redirect to Settings.

Connecting can turn on remote MCP access at a chosen role, Developer
first, and picks the trial with it. POST /license/connect takes the
level, and `dagu license activate --remote-access <level>` saves it
and prints the MCP URL. The remote access copy now speaks of MCP and
states what each level allows.

* Measure containers and report CPU since the last report (#3564)

* Make the workflow page usable on phones (#3568)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* Address review comments of the phone layout PR (#3569)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* feat(computer): decide act turns from the window's elements (#3545)

* Request stream usage from the local provider (#3567)

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>

* Refuse elevated-window input on press and fail closed (#3536)

Windows drops input sent to a higher-integrity window while SendInput
reports every event accepted. Refuse that input before it is sent, and
also when the target's level cannot be read. Only presses are checked,
so a release always goes out and nothing is left held down.

Refs #3527

Co-authored-by: Yota Hamada <yohamta@gmail.com>

* Use the configured SSH user for git key auth (#3584)

* Remove the openai-codex LLM provider (#3585)

* Honor the DAGU_HOME build arg in every image (#3586)

* Keep JSON param values as written (#3582)

* Rework the workflow page for phones and laptops (#3577)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* Remove the unused Redis global pool manager (#3590)

* Follow references when browsing a schema path (#3581)

* Use one pane for the selection of the graph

The trigger panel and the step panel were two panes that closed and
opened, with different sizes, and a step clicked right after a
trigger lost its panel. They now share the shell and swap at once; a
click on the empty canvas ends the focus on a trigger too; and with
the YAML of the file open, a trigger shows its part of the YAML.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Open the YAML and Activity by default on laptops

A laptop now opens both until the user closes them; a phone and a
tablet keep them closed. The media queries share one hook and one
set of breakpoints. The e2e specs start from both closed, and a new
spec covers the default.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Rename stale names and move phone checks to phone.spec

The dock is no longer a sheet on phones and Start is no longer an
icon only: the names say so. The phone check of a run link moves to
phone.spec.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Read the phone breakpoint from the shared constants

The dock check for a phone link to the YAML now uses WIDE_QUERY, and
the step panel test passes a click point of the right shape.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Open the pane of a trigger with no part in the YAML

A webhook, or a manual trigger with no params, has no key in the file:
with the YAML open, its click now opens its pane, which gives the YAML
back when it closes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Honor mark_success for non-retryable exit codes (#3554)

A step with a retry_policy that failed with an exit code outside
retry_policy.exit_code was marked failed without checking
continue_on.mark_success, so the run failed even though the step
asked to be marked successful.

Setup failures no longer match continue_on.exit_code either: a step
that never ran kept a reset exit code of 0, which an exit_code list
containing 0 could match.

Github-Issue: #3539
Co-authored-by: Yota Hamada <yohamta@gmail.com>

* fix(desktop): list no elements of a minimized window (#3542)

* feat(computer): show where a step acts on the desktop (#3547)

* Return rows for SQL that starts with a comment (#3580)

* Let Vault secrets fall back to VAULT_ADDR and VAULT_TOKEN (#3587)

* Validate the chat.completion provider in the schema (#3591)

* Run the dev image browser check on fewer pushes

Pushes to main triggered the check for any change under internal/cmd,
internal/service/frontend, internal/service/scheduler, and
internal/cmn/config, so most merges rebuilt the dev image twice for a
check unrelated to the change. Main now uses the same paths as pull
requests: the image, the entrypoint, the browser code, and the
shutdown path the smoke tests target directly.

* feat(computer): capture the step's app windows instead of the whole screen (#3546)

* feat(computer): hold the desktop only while a step sends input (#3549)

* Check out the fetched branch on an existing clone (#3583)

* Warn when a dry run cannot resolve secrets (#3588)

* Test only the latest main commit in CI workflows

Every merge to main started its own CI, Conformance, and Frontend CI
run, and none were cancelled, so bursts of merges re-tested code that
already passed on its pull request and held the org's runners. Runs
for a branch now share one group: one runs at a time and a newer
commit replaces the run still waiting.

* Drop CI triggers that repeat tested work

Path filters do not apply to tag pushes, so every release tag ran the
full CI suite on a commit main had already tested. No CI job is
specific to tags and the release workflow does not wait for CI.

Conformance tests never read specs/, so spec-only changes no longer
start them.

* Run only the golden test for non-UI frontend runs

Spec and YAML corpus changes start Frontend CI because the golden
round-trip test reads them, but each such run spent about ten minutes
on the full test suite and build. Without a ui/ change, the workflow
now runs only the golden test.

* Stop caching Go in the vulnerability scan

The scan runs weekly and on go.mod changes, so the 729 MB cache it
saved was never restored and only crowded the repository's 10 GB
Actions cache.

* Enforce the Redis max_result_size limit (#3589)

* Rename the pointer reach check to fix the build

Two computer step changes merged seconds apart each added a reach
method on run with a different signature, so main no longer compiled.
The check that a pointer action lands on the window the model saw is
now reachPoint; reach keeps the meaning its other callers expect.

* Report distributed workers to Dagu Cloud (#3592)

Heartbeats carry each worker's host and machine sample, and the
coordinator keeps both. On paid licenses and trials, cloud reports
carry how many workers have a fresh heartbeat and, from the runs
level, each worker within 256 entries and 64 KiB. Workers are left
out while no coordinator is registered, so an empty list always means
no live worker.

* fix(computer): let an act retry a smaller batch after a partial overrun (#3579)

Signed-off-by: CJstate <1507965754@qq.com>
Co-authored-by: Yota Hamada <yohamta@gmail.com>

* Show runs started with --name in the web UI (#3594)

* Fix web UI regressions from v2.24 (#3597)

* Remember the last Activity tab across workflows

The Activity panel kept the picked tab only while its workflow stayed
mounted. Moving to another workflow in the run modal, or opening the
modal again, mounted a new panel that fell back to the default tab.

Store the tab that the user picks and start a new panel on it, unless
a ?dock link names one.

* Page the run history of a workflow

The history endpoint returned a fixed 30 runs, so a client could not
reach older runs of a workflow. Add limit and cursor parameters and a
nextCursor in the response, read through the paged status query.

The status query can now return whole statuses: summaries from the
run index have no nodes, which the history needs. Ephemeral runs stay
on the first page, limited to those not older than its oldest run, so
that a cursor never skips a run.

* Load older runs in the History tab

The History tab showed only the runs that the run timeline drew, so a
timeline set to its last 5 runs left 5 columns in the grid. The tab
now shows the latest page of the history and reads older pages through
the cursor as it scrolls back: to the left end of the steps grid, or
to the end of the run list of a step. Left on the oldest run reads
more too.

The choices of the timeline about test runs and trigger types still
apply. A view scrolled back keeps its runs in place while older ones
come in, and only a view at the right end follows a new run.

* Mark the selected run in the History grid

A faint tint on a 16px column did not show which run was selected,
least of all for the newest run at the right edge. Give the column a
stronger band and ring its header dot, as the run timeline rings the
selected bar.

* Show every Activity tab in the tab row

Artifacts and Definition always sat in a More menu, even on a wide
screen with room to spare. Put every tab in the row: the strip already
shows icons only when the row is too narrow, and scrolls after that.

* Let Up and Down page the run modal from tabs

After a click on an Activity tab, Up and Down did nothing: the modal
left every key of a tablist to it, while a row of tabs moves with Left
and Right only. Only a vertical tablist keeps Up and Down now.

* Show the minimap only for a graph that runs out of view

The overview flashed on load and faded out once the first fit brought
the whole graph into view: the view before the fit, and the glide of
a refit, left out part of the graph. A graph that runs out of the view
now shows the overview only once the view has rested.

* Enqueue runs from the Start dialog by default

The Start dialog started a run at once unless the user checked
Enqueue. A run now waits in the queue by default, so it follows the
queue and concurrency limits of the workflow; unchecking Enqueue still
starts it at once.

* Draw step icons on the canvas in the foreground colour

The icon of a step card was a thin muted-grey stroke on a grey tile
over a tinted card, and it was hard to see in both themes. Draw it in
the foreground colour.

* List runs of the same second newest first

A run directory names its creation time in whole seconds, and the list
order broke a tie by run ID ascending. Generated run IDs grow with
time, so runs created in the same second came back oldest first: the
workflow history, read through this list since it became paged,
showed them in reverse in the run picker, the History grid and the
timeline. Break the tie by run ID descending.

* Open a step over the Workflow panel and give it back

A click on a step selected it, but its panel stayed under the Workflow
panel, which never closes on a canvas click, so the step details did
not show. Closing the full step view also left the Workflow panel shut.

A click on a step now closes the Workflow panel to show the step, and
the Workflow panel opens again once the step panel and the full step
view have closed, unless another panel opened in between.

* Keep status writes off the shared status cache (#3599)

Co-authored-by: SCcagg5 <sccagg5@gmail.com>

* Expose the repeat index in step context (#3596)

Expose a zero-based repeat index so workflows can vary each iteration.
Refresh step values on repeats and preserve the index across retries.

* Show the canvas focus ring only for Tab

The canvas drew its focus ring after any key press once it had focus,
so Escape after a click framed the whole canvas. The ring now marks a
canvas that Tab reached; a click, a key after it, or a focus that the
page moves shows none.

* Move the warning count off the Save button

Save carried an amber count of the warnings on its corner. Save is
disabled while there is nothing to save, so the count could not be
opened, and the badge did not read as a warning. The Issues button
beside Save now lists the warnings too: amber for warnings alone, red
when errors block the save.

* Stop restoring editor view state per model

Switching runs gave a run definition editor the model of a run it had
shown before, and the editor restored that model's view state. Monaco
then cancelled a pending word highlight and left the rejection
unhandled, which the development server showed as an uncaught
"Canceled" error over the page. The YAML editor no longer saves or
restores view state.

* Read a workflow's webhook from the webhook list

The workflow bar and the Webhook trigger on the canvas asked for the
webhook of the workflow, which the API answers with a 404 when there is
none. Every workflow without a webhook logged a failed request and the
query error in the browser console. Both now find the webhook in the
list of webhooks, which answers for every workflow, and the webhook
settings refresh that list after a change.

* Ask before the workflow overlay closes

Closing the overlay went back in history first, and the leave guard
caught that Back afterwards. The router had already shown the page
below, so the overlay vanished, came back, and only then asked about
the unsaved changes. The overlay now asks the workflow page before it
closes or steps to another row, and the guard lets the Back of an
accepted close pass.

* Stop leaking the YAML editor's model

DAGEditor disposed the Monaco editor in its own cleanup, which runs
before the cleanup of @monaco-editor/react. That one disposes the
editor's model, but a disposed editor has none, so the model and its
edited text stayed. The next editor of the same workflow reused it:
after leaving with unsaved changes and coming back, the YAML showed the
draft while the page still asked whether to restore it.

* Jump to the first fits of a canvas that just opened

Every auto fit glided. A page that opens gets its parts one after
another, such as the trigger nodes and the overlays of the canvas, and
each one fitted the graph again: the graph slid in from the side after
it first showed, against the workflow overlay that slides in from the
right. A fit within a second of the canvas first having a size now
jumps; later changes, such as a panel that opens, still glide.

* Name the run definition button in the workflow bar

The button that puts the definition of a stale run into the editor was
an icon with the name "Copy to editor", which did not say what it does.
It now reads "Revert to this run's YAML" on a wide bar, as the run
actions beside it do, and its tooltip says that nothing is saved until
the user saves.

* Put the Workflow and YAML panels in a right column

On a wide window the Workflow panel and the YAML of the file were
panels over the graph zone that opened, closed and replaced each other
as steps and triggers were clicked: the graph moved, and the YAML panel
closed on its own. They are now the tabs of a right column beside the
zone, which only the user switches, collapses to a rail or widens. A
clicked step or trigger shows in a temporary tab of the column, and the
expanded view of a step covers the zone and the column with the same
YAML editor. A phone keeps the panels.

* Fade the graph in at its first fit

A canvas that opened drew its graph at full contrast at once, which a
dark page made glaring now that the first fits jump instead of glide.
The graph layer stays transparent until its first fit and then fades
in, and a fit in its first moments happens while it is still faint.

* Shorten the fade of the graph to 150 ms

A fade of 300 ms made the graph show well after the workflow overlay
had slid in. At 150 ms it is fully drawn when the overlay arrives.

* Shorten the fade of the graph to 80 ms

The fade only has to soften the first draw of the graph, not make it
wait.

* Show the YAML editor once its text has settled

A YAML editor that mounted drew its text plain, coloured it, and a few
hundred milliseconds later inserted the "Open step" lenses, which
pushed the steps down. The editor now stays transparent until its text
is coloured and, for a text with steps, the lenses are drawn, then
fades in like the graph. It shows after 1.5 s at the latest.

* Show a skeleton of the YAML while its editor loads

The YAML area stood empty, or showed a spinner, until the editor had
loaded and settled, which made it look slow. It now shows a skeleton
drawn from the text itself: a bar for each line at its indent and
length, and one in the row of each "Open step" lens, at the metrics of
the editor. The editor fades in over it, with its text where the bars
were.

* Read ahead the rows beside the workflow overlay

Up and Down in the workflow overlay showed another workflow whose
details were not read yet: the page gave way to its loading frame and
then mounted again from nothing. The overlay now reads the details and
the spec of the rows before and after the shown one, with the keys of
the page, so a step shows the next workflow at once in the same page.

* Fit the graph of another file at once

A canvas that showed another file kept the view of the last one: the
new graph was drawn at it for a frame, then glided to its own fit. The
canvas now learns the file it draws and fits its graph before the
paint, as at a first draw, and the fits of its first moments jump.

* Settle the YAML editor again for another file

A YAML editor that took the model of another file showed its text at
once and then moved it for the "Open step" lenses. It now shows the
skeleton of the new text until the text has settled, as at its mount,
and hides the old text at once instead of fading it out.

* Give the Activity panel skeletons the shape of its views

While the runs loaded, the run timeline showed a strip of wide bars that
hung from its top, and the History tab only said "Loading runs...". The
timeline skeleton now draws the frame of the chart: the gutter of the
duration labels, the grid, the baseline, the time axis with its labels
and a bar of run width for each run. The History tab shows the grid with
its header and the steps of the workflow by name, with a mark for each
run column.

* Load catalogs and YAML parser on demand (#3600)

* Keep the width of the run actions for a stale run

The run YAML button took the width of its name on a wide bar only while
the run was stale, so the run actions grew when the run changed. On a
wide bar its slot now keeps the room of the named button in every
state, as the other slots of the bar keep theirs.

* Update README screenshots to the current UI

The screenshots showed the Web UI of v2.6, before the workflow page.
They now show the workflow page with its graph, YAML and Activity
panel, the log of a failed step, a Wiki runbook, the Artifacts tab and
the Start dialog with a typed parameter form, all from v2.25.7. The
parameter form is a PNG like the other screenshots.

* Show a run without its workflow file in the workspace

A run whose workflow file was deleted, renamed or could not be read
opened in the run view of the old run page, although every run stores
the definition it ran with. Such a run now opens in the workflow page,
drawn from that stored definition: its graph, its steps and their logs,
its outputs and artifacts. The page lists only that run and offers
nothing that acts on the file, such as Start, Save, Rename or Delete.

* Update the README Console screenshot

Recapture the Dagu Console Overview from a demo workspace on the
current Console, which reworked the page into Needs attention,
In progress, and Instances sections.

* [CC3-sd783e0] fix(api): TXE-5923 upstream v2.25's new run and DAG endpoints decide the owner as the fork's other reads do

The connector found two of upstream v2.25's new endpoints deciding only by
workspace under owner enforcement. An audit of all 39 operations upstream
added since the merge base found two more.

- Foreach reads of a root run (items, item, body-step log and download)
  use the per-run owner check (requireDAGRunStatusVisible), as the sub-run
  variant and every other run read do.
- /dag-run-activity aggregates runs, queues and schedules across the hub
  with workspace filtering only. It is now hidden wherever the fork hides
  its other whole-hub reads (txeHubHidden, TXE-4752), on both the REST and
  the SSE paths.
- The public-webhook endpoints apply txeDAGFileRoute: View for reading,
  Author for changes, as the existing webhook endpoints do.
- Pinning a DAG uses the per-DAG owner check, not the workspace alone.

The other new operations already went through a fork check, or are
admin-only, license-management or calendar settings. Step outputs, samples,
events and availability use findVisibleAttempt, which the merge already
moved to the per-DAG check.

Tests: /dag-run-activity joins the hub-records router test. A new router
test shows a legacy credential in enforce is refused a root run's foreach
read before anything of the run is read. Both fail without this change.

* [CC3-sd783e0] ci: TXE-5923 upstream's Helm appVersion check runs only in the upstream repository

The check compares charts/dagu/Chart.yaml appVersion with the newest v* tag
in the repository. This fork carries upstream's chart unchanged and holds
none of upstream's release tags (its own are txe-v*), so here the newest v*
tag is an old upstream one (v2.18.2) and the check can only fail.

* [CC3-sd783e0] fix: TXE-5923 calendar usage follows the owner scope; compression modules raised for two advisories

The calendar reads that upstream v2.25 added list the DAGs that use a
calendar, filtered by workspace only. Under owner enforcement a legacy
caller could see the DAG names that its DAG lists hide (connector, P2).
Calendar details and the usage counts now apply the fork's whole-hub list
scope. A delete still scans every DAG, so a calendar in use is never
deleted from under DAGs the caller cannot see, but it names only DAGs the
scope admits and counts the rest. A new router test fails without this
change.

govulncheck reports two advisories published after the merge, in
indirect compression modules: GO-2026-6888 in andybalholm/brotli (fixed in
v1.2.7) and GO-2026-6865 in klauspost/compress (fixed in v1.20.2). Both are
raised to those versions; nothing else moved, and go mod tidy is clean.

* [CC3-sd783e0] test: TXE-5923 CI fixes: lint, the fork's hidden Incidents tab in E2E, and apt in the Ubuntu image

- handler_hold_pinning_test.go: the switch over the pinning modes names
  every mode (golangci-lint exhaustive).
- e2e/workspace: upstream's specs expect the Incidents tab, which the fork
  deliberately does not offer (INCIDENTS_ENABLED, lib/fork.ts). The
  settings spec lists the three tabs the fork shows. In the deep-link
  table, /incidents keeps its wtab and shows the Overview. Every other row
  and assertion is unchanged.
- Dockerfile: this changes the built image. The tools layer upgrades only
  apt, which takes its dependencies with it. The Ubuntu image scan (Trivy,
  HIGH/CRITICAL, ignore-unfixed) failed on CVE-2026-109518 in apt 2.8.3,
  fixed in 2.8.4ubuntu0.1, which the ubuntu:24.04 base has not taken yet.
  The Dockerfile is otherwise unchanged from fork main, and is also the TXE
  image's build file (txe-image.yaml). The base stays the unpinned
  ubuntu:24.04 tag, as before, so the image drifts with the base and the
  archive exactly as it already did. The scan of this PR's head is the
  check that the fix takes.

* [CC3-sd783e0] feat(coordinator): TXE-5923 an unclaimed dispatch waits for a worker until its claim deadline, coordinator.unclaimed_dispatch_timeout (default 24h)

Option C, as Connor chose it. Upstream keeps an unclaimed pending task
until its run moves on, however long that takes. The fork removed an unheld
task after the 90s reservation time. Now the fork keeps it the way upstream
does, so a dispatch waits while every matching worker is busy, but only
until a claim deadline.

The setting: coordinator.unclaimed_dispatch_timeout
(DAGU_COORDINATOR_UNCLAIMED_DISPATCH_TIMEOUT).
- Default 24h, also when the value is empty or absent, and in a process
  that does not load the coordinator section.
- An explicit 0 (or 0s) means no ceiling.
- A negative, malformed, unitless or overflowing value refuses to load,
  with an error that names the key; it is never a warning with a fallback
  to the default. This is stricter than upstream's worker_wait_timeout,
  which warns and keeps its default; that rule is upstream's for its own
  key and is not copied here.

The deadline (dispatch store):
- It is stored with the task as claimDeadline (or claimUnbounded) and set
  once, at first enqueue.
- A coordinator restart, a released claim, a republication of the same
  execution or a later change of the setting never moves it. A new
  attempt, or a new execution of a reused attempt, gets its own.
- A record from before the field takes its deadline from its stored enqueue
  time, else the record's creation time, never from the present. With
  neither, it fails closed. A claim's release, which re-stamps the enqueue
  time, writes the derived deadline first.
- Past the deadline a task is never handed out. The store no longer removes
  unheld tasks itself (recycleExpiredPending is gone).
- ListPastClaimDeadline lists such tasks. DeletePastClaimDeadline removes
  only the record listed, while it is still that task and still past its
  deadline, using CompareAndDelete. A claim, or a newer attempt's task, is
  never touched.

The coordinator (expireUnclaimedDispatches, a new zombie-detector pass that
runs before the preparation journal is reconciled):
- An attempt the journal names loses its task record. Reconciliation then
  abandons it with the usual record, from the same evidence as before. So
  the fork's negative controls keep their fixtures and assertions; only
  their store now has a short claim deadline, and each runs this pass
  before reconciling. A crash after the removal leaves the journal entry,
  and the abandonment still happens.
- Any other attempt is failed first, while it still waits, and then loses
  its task, in the order root's review set. A crash between the two leaves
  a task nobody can claim; the next pass removes it without failing the run
  again.
- Neither path ever ends an attempt that has a worker, a lease or an
  active-run record, one that is not its run's latest, or one a queue still
  holds. In each case the stale task alone goes. A lookup that fails leaves
  everything for the next pass.

Unchanged:
- Enforce still refuses a dispatch with no machine of its owner at once.
- Held dispatches, cancel and orphan drop.
- The claim fences, including the job lifecycle check at claim time.
- A waiting task still holds no worker slot.

Upstream's TestExecution_QueuedDispatch_WaitsForBusyWorker and
TestSubDAG_WaitsForBusyWorker pass unchanged. TestDispatchTaskStore_
PendingOutlivesReservationTTL is upstream's again.

New tests run on a fake clock:
- deadline boundary;
- restart, and a setting change in either direction;
- a released claim;
- republication against a new attempt or execution;
- zero, and the 24h default;
- a held task;
- legacy records, including fail-closed;
- exact-record deletion;
- concurrent claims at the deadline;
- the journaled and unjournaled coordinator paths;
- crash between steps;
- leased, running, newer-attempt and still-queued attempts;
- strict config parsing.

Rolled back to txe-v2.18.2-19, a coordinator ignores the new fields and
expires unheld tasks after the reservation time, as it always did.

* [CC3-sd783e0] test: TXE-5923 the claim deadline bounds only the wait for a claim; rollback needs a drain

Root's follow-up on option C:
- The attempt id alone now keeps a newer attempt. A task with no execution
  marker, as a legacy queued task has, is told apart from a newer attempt
  of its run only by its attempt id. Before, the marker check also caught
  the case, so no test isolated the attempt-id check; one does now (with
  that check removed, the test fails).
- A claim made before the deadline stays alive. It is acknowledged after
  the deadline, and its lease and run are never ended by later passes,
  however late.
- A claim released past the deadline is refused to the next claimant. Its
  execution ends with a record, and the first claimant's late
  acknowledgement is refused.
- An unjournaled claim races the pass: claimers run while the clock
  crosses the deadline and the pass runs. Each run is either claimed and
  never failed, or failed and never claimed; both outcomes occur in every
  run.
- Docs (schema, Coordinator.UnclaimedDispatchTimeout, txe/contract): the
  setting bounds only the wait for a claim, and schedules are not affected.
  A rollback to txe-v2.18.2-19 does not keep the wait: that coordinator
  expires an unheld pending task after 90s. So pending dispatches are
  drained before a rollback; the rollback is not lossless for them.

* [CC3-sd783e0] refactor(store): TXE-5923 the claim deadline has one maker, one encoder and one claimer, and a test fails a new writer outside them

Fix the class, not the instance. The deadline used to be assigned at two
call sites, Enqueue and the admission bind. A future writer could forget it
and store a task with no deadline, which would wait for a worker forever.

- Instance: two writers each had to call assignClaimDeadline.
- Class: any writer of a pending task record.
- What makes the next case automatic:
  - newPendingPayload is the one maker of a new pending task's payload,
    whateve…
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

bug: OpenCode v2 misclassifies 404 errors and drops snapshot files

2 participants