Skip to content

fix: confirm admit with a person and report approval waits - #64

Merged
abrichr merged 19 commits into
mainfrom
fix/ledger-wave2-2026-10-14
Oct 15, 2026
Merged

fix: confirm admit with a person and report approval waits#64
abrichr merged 19 commits into
mainfrom
fix/ledger-wave2-2026-10-14

Conversation

@abrichr

@abrichr abrichr commented Oct 14, 2026

Copy link
Copy Markdown
Member

This branch fixes ten problems in openadapt-agent and adds a guard test for one that main had already fixed. The two largest changes:

  • admit needs a person's answer in the MCP client's confirmation form. Before, an agent could accept its own draft.
  • A run that waits for a person to approve a save reports needs_review. Before, it read as running, then hit the timeout or read as a failure.

Fixes

Problem Change Regression test Ref
An agent could call admit with no arguments. Flow then recorded the draft as accepted by the OS user, and nobody was asked. admit takes no arguments. The server asks the person through MCP form elicitation and refuses on decline, cancel, or a client that can't show forms. A direct call without that confirmation refuses with admit_requires_host_confirmation. test_mcp_server.py::test_admit_requires_form_elicitation, test_authoring.py::test_admit_without_host_confirmation_refuses B074
Clients that declare elicitation: {} (form mode under MCP 2025-11-25, and every 2025-06-18 client) had every attended decision refused. An empty elicitation object counts as form support. Clients that declare only url, or no elicitation, are still refused. test_mcp_server.py::test_elicitation_capability_check_is_backward_compatible B083
serve --authoring without --url failed on every Mac. The tutorial extra had no pyobjc, and the pin searched for windows with an empty owner, which never matches. The tutorial extra includes openadapt-flow[macos]. The pin reads the front window's owner and title from Quartz, then checks that window is unique. Each failure names its cause and never includes a window title. test_authoring.py::test_pin_native_darwin_*, test_tutorial_extra_includes_macos_deps B077
Run records defaulted to ./runs in the folder the client started the server in, often a git repository. A server started elsewhere saw a different request-id ledger and pause queue. The default is ~/.openadapt/runs (or OPENADAPT_AGENT_RUNS_DIR), resolved once and created owner-only, and the startup line names it. If ./runs from an earlier version exists and no folder is named, the server refuses to start, so a resent request_id can't run twice. test_cli.py::test_serve_default_runs_dir_is_not_cwd, test_cli.py::test_serve_refuses_to_skip_an_earlier_versions_runs_folder, test_mcpb_entry.py::test_mcpb_defaults_to_local_read_only_server B082
A workflow with its own url input could never run through the deprecated run_<id> tools, and with recorded defaults the caller's value went to the app address. When a workflow declares url, the override is named target_url_override. test_bridge.py::test_declared_url_param_does_not_collide_with_override and two related tests B084
The MCPB had no way to set a certifying policy, and it listed decision tools to clients that could only get a refusal from them. New optional "Certifying policy" setting passed as --policy. Decision tools are hidden from clients without form elicitation. A call still refuses. test_mcpb_entry.py::test_mcpb_passes_the_certifying_policy, test_every_mcpb_variable_is_wired_to_user_config, test_mcp_server.py::test_attended_tools_are_listed_only_for_clients_that_can_confirm B086
Under Flow's approval modes, a run waiting for a person read running. The server killed Flow at --timeout during the wait, and when Flow gave up waiting the result said something didn't match. While a save waits, run_workflow and get_run return needs_review, reason waiting_for_approval, record_changed: "unknown". That live result is never stored. Time spent waiting for a person doesn't count toward --timeout (4 hours at most). After Flow gives up, the reason is approval_needed, and only on a proven-no-effect outcome. --approval-mode is set only by the operator and is recorded in every result. tests/test_approval.py (12 tests, two with a real child process), tests/test_approval_flow.py (end to end, opt-in) B272
serverInfo.version was the SDK's version under MCP 1 and empty under MCP 2. The server reports the package version. test_mcp_server.py::test_initialize_reports_the_package_version B091
docs/MAILBOX_CLI.md told pip users to pass a runner link to openadapt connect. That command is Cloud pairing and exits 2. The doc points at openadapt-agent authoring connect. test_distribution.py::test_runner_link_docs_do_not_use_cloud_pairing_command B088
No doc listed all 14 recording tools. docs/DESIGN.md has a table of all 14, linked from the README. A guard keeps the Beta label out of listings. test_distribution.py::test_docs_list_every_authoring_tool, test_listings_carry_no_beta_label B087
Documented commands didn't parse on the published 2.0.1. Main's docs are already right. Guard test only: every openadapt-agent serve, authoring and emit-skill command in the README, llms.txt and the shipped skill must parse, and every named extra must exist. It fails against the v2.0.1 parser. tests/test_readme_commands.py B072

Not fixed here

  • Releases. PyPI and the MCP registry still show 2.0.1, so B072, B077, B087 and B091 reach users only with the next reviewed openadapt-agent release. openadapt-flow's approval modes are merged (openadapt-flow #486) but not released, so CI's tests/test_approval_flow.py skips until they are.
  • Flow's side of B074. Flow's admit should require and record how the acceptance was obtained, and stop naming the OS user as the operator. This server already passes confirmation_channel="mcp_form_elicitation" once Flow's admit accepts it.
  • Copy after an approval times out. approval_needed tells the caller a person may "approve the save and resume the run, or end it". Flow's docs say a resumed run that opened its own browser (as --url runs do) writes nothing. The copy fails safe, since nothing is retried, but whether to tell the caller to resend with a new request_id is a product decision.
  • Which window the macOS pin takes. Without --url, the pin uses whatever window is in front when the server starts. That can be the MCP client or a terminal. This matches the documented behavior, and the pin doesn't refuse those windows.
  • Docs outside this repository (openadapt-ops get-started, openadapt-web agents.txt) and Desktop's approval screens, which don't exist yet. Until then next_action names openadapt-flow item-approval.

How it was tested

  • Each new regression test ran against origin/main (4b0b9bb) in a separate worktree that had only the test files copied in. Each one failed there because of the bug and passes on this branch. Some tests depend on helpers this branch adds (_run_flow_process, default_runs_dir, openadapt_agent.approval). For those, an equivalent test without the helper ran against main: B084 with the earlier subprocess stub, B082 by calling build_argv, and B272 with a child process that writes Flow's approval request. On main that child process was killed at the timeout and reported done while it waited, and Flow giving up read as unexpected_screen.
  • Full suite in CI's three dependency sets: Python 3.12 with current dependencies (413 passed, 8 skipped), Python 3.10 with openadapt-flow 1.26.0 and mcp 1.28.0 (412 passed, 9 skipped), Python 3.11 with mcp 2.0.0 (413 passed, 8 skipped). Also against openadapt-flow main 791a627d (414 passed, 7 skipped).
  • With OPENADAPT_AGENT_E2E=1 against openadapt-flow main, real Chromium and MockMed: tests/test_approval_flow.py 2 passed, tests/test_sandbox_flow.py 5 passed.
  • ruff check src tests scripts, uv lock --locked, python -m build, check_release_artifacts.py, check_dist.py, check_source_boundary.py --require-dist, and mcpb validate manifest.json (mcpb 2.1.2) all pass.

Before merging

  • Review one contract wording change. The needs_review rows in the README, llms.txt and the skills no longer say "Nothing was written" for every case. They say it unless record_changed is unknown, which happens only while a run waits for an approval. An earlier save in that run may have gone through.
  • The default runs folder moves from ./runs to ~/.openadapt/runs. A production or attended server that finds an earlier version's ./runs refuses to start until it's given --runs-dir. The release notes should say so.
  • admit now refuses on clients that can't show a form.

🤖 Generated with Claude Code

abrichr and others added 14 commits October 14, 2026 10:01
@abrichr @claude
Require a person's confirmation form before admit accepts a draft
7dec9f8 
admit accepted an empty or omitted confirm, so an agent could call it with
no arguments and Flow wrote admit.json naming the OS user as the operator
who accepted the draft. No person was asked. The tool description even
told the model which tokens accept.

admit now takes no arguments. The MCP layer asks the person through form
elicitation, with the draft's workflow id in the prompt, before it calls
the bridge, and only then marks the call as confirmed. A client without
form elicitation, a decline, or a cancel refuses and nothing is accepted.
AuthoringBridge.dispatch refuses admit with
admit_requires_host_confirmation unless the transport set
human_confirmed, so non-MCP callers can't accept a draft either. When
Flow's admit takes a confirmation_channel, the bridge passes
mcp_form_elicitation so admit.json can record how the acceptance was
obtained (the Flow side of this is separate work).

Ledger: B074

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Check that every documented openadapt-agent command parses
0c9f4bd 
The README and the agents pages showed `serve --allow-run`,
`serve --authoring` and the `[tutorial]` extra while PyPI shipped 2.0.1,
whose `serve` still required --bundles and had no tutorial extra, so a
partner's first command exited with code 2. Main's README now installs
from GitHub and says that 2.0.1 predates the sandbox, so the docs here
match the code again.

This adds a test that pulls every `openadapt-agent serve|authoring|
emit-skill` command from README.md, llms.txt and the shipped skill, and
checks that this code's parser accepts it and that every
`openadapt-agent[extra]` they name is declared in pyproject. The test
fails against the v2.0.1 parser and passes on main. Publishing a release
from main is still needed before the docs on PyPI-facing pages
(openadapt-ops get-started, openadapt-web agents.txt) are true.

Ledger: B072

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Accept clients that declare elicitation as an empty object
32d7de9 
The attended-action confirmation refused any client whose elicitation
capability had no form key. MCP 2025-11-25 says an empty elicitation
object means form mode, and clients written for 2025-06-18, which had no
form or url keys, always send it that way. Those clients could show the
form but every continue, skip, teach, escalate and reject was refused.

The check now lives in client_supports_form_elicitation: a client can
show a form when it declares form, or declares elicitation with neither
form nor url. A client that declares only url, or no elicitation, is
still refused before anything is submitted. admit uses the same check.

Ledger: B083

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Pin the frontmost macOS window for serve --authoring without --url
08efa53 
`serve --authoring` without --url exited on every Mac with a generic
"needs a locally pinned window" message. Two things were wrong. The
tutorial extra pulled Flow's browser extra only, so pyobjc (Quartz) was
missing and Flow's frontmost-window lookup quietly returned nothing. And
with pyobjc installed, the pin still failed: it looked the frontmost
window up with find_windows("", None), but Flow matches an exact owner,
so an empty owner never matches any window.

The tutorial extra now includes Flow's macos extra, whose pyobjc
requirements carry darwin-only markers. The pin reads the frontmost
window's owner and title from Quartz by window id, then asks Flow's
find_windows for that exact owner and title so the uniqueness check
still applies. Each failure now says what to do: pyobjc missing (names
openadapt-flow[macos]), no window in front, the owner unknown, titles
hidden until Screen Recording is allowed, or several windows with one
title. Errors never include a window title. uv.lock is updated for the
new extra.

Ledger: B077

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Keep run evidence in ~/.openadapt/runs, not the client's folder
e216143 
Outside the sandbox, `--runs-dir` defaulted to ./runs, relative to
wherever the MCP client started the server, which is usually the user's
project. Run reports, frames and entered values could land in a git
repository where `git add .` commits them, and a server started from
another folder saw a different Needs Attention queue and request-id
ledger, so earlier pauses were orphaned. RunnerConfig, the authoring
recording folder and the MCPB entry point (which fell back to "runs")
had the same relative default, while the MCPB manifest already said
${HOME}/.openadapt/runs.

There is now one default, default_runs_dir(): OPENADAPT_AGENT_RUNS_DIR
when set, else ~/.openadapt/runs. The CLI resolves the folder once to an
absolute path and passes it to runs, pauses and the ledger; authoring
and the MCPB entry point use the same default; FlowRunner creates the
folder owner-only. The startup line names the folder so an operator can
find their pauses. The sandbox keeps ~/.openadapt/agent-sandbox.

Ledger: B082

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Keep a workflow's own url input apart from the URL override
5786952 
The deprecated run_<id> tools popped "url" from the arguments as the
target-app override before checking them against the workflow's
parameters. A workflow with its own "url" parameter therefore always
looked like it was missing that input and every call was refused. With
--allow-synthetic-recorded-defaults and --allow-url-override, the run
reported success after typing the recorded URL into the workflow and
sending the caller's value to the app address instead. The schema also
wrote the override over the declared parameter's description.

bundles.url_override_key picks the override's argument name: "url" when
the workflow has no url parameter (unchanged for existing callers),
otherwise "target_url_override" (with a suffix if that is taken too).
The schema and the bridge both use it, so a declared url always reaches
the workflow's inputs and the override reaches --url only under its own
name and only with --allow-url-override.

Ledger: B084

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Pass the MCPB's policy and hide decision tools a client can't confirm
2b4c4f5 
The MCPB had no way to set a certifying policy: manifest.json had no
policy field and build_argv never passed --policy, so get_workflow
answered "Certification was not evaluated" unless the user happened to
pick a deployment YAML that carried one. And "Allow Needs Attention
decisions" listed continue, skip, teach, escalate and reject for every
client, although a client without form elicitation gets a refusal from
each of them.

The manifest now has an optional "Certifying policy" setting wired to
OPENADAPT_AGENT_POLICY, and build_argv passes it as --policy. The
deployment config and attended-decision descriptions say what each one
is for and that decisions need a client with form elicitation. The tool
list leaves the attended decision tools out for a client that can't
show a confirmation form; calling one anyway is still refused before
anything is submitted. get_workflow now says how to set a policy when
none is configured. A test ties every OPENADAPT_AGENT_* variable the
entry point reads to an mcp_config.env entry and a user_config key.

Ledger: B086

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Report runs that wait for a person's approval as needs_review
e1f59da 
openadapt-flow now stops before each save under the Standard and
Regulated profiles until a person approves it (each_item), and waits up
to 15 minutes by default. This server didn't know about that:

- run_workflow and get_run said "running" for the whole wait, with
  nothing telling the caller that a person had to act.
- The server killed Flow after --timeout (600 s by default) even while
  Flow was waiting for the person, so the person's time was cut short,
  the pause they were answering was lost, and the result read "The run
  didn't finish in time", a timeout dressed as a failure.
- When Flow itself gave up waiting, the result said "Stopped before
  saving because something didn't match", which is not what happened.

Now:

- While a save waits for an answer (an unanswered, unexpired
  item-approval request in the run folder), run_workflow and get_run
  return needs_review with reason waiting_for_approval. next_action says
  a person at that computer approves or declines the save and names
  `openadapt-flow item-approval show`. record_changed is "unknown",
  because an earlier save in the same run may have gone through, and the
  copy says only that this save hasn't been made. This live result is
  never stored. run_workflow returns as soon as the wait begins; a
  get_run that finds the run already waiting keeps waiting for the
  answer instead of returning at once.
- Time Flow spends waiting for a person no longer counts toward
  --timeout (Flow bounds the wait; the server allows at most 4 hours of
  person time in all). Flow runs go through _run_flow_process, a
  subprocess.run stand-in that does this; tests stub that function now.
- A report whose approval block says approval_required, with a
  transaction outcome that proves no effect, reads needs_review with
  reason approval_needed. An uncertain save stays not_sure_if_saved and
  a declined save stays stopped_by_person.
- The server never passes --approval-mode unless the operator starts it
  with `serve --approval-mode`, so not_gated is never chosen for them
  and the calling agent can't choose it. The mode is recorded in every
  result under technical.approval_mode (from Flow's report, else the
  operator's flag) and in the startup line. The flag is refused when
  this interpreter's Flow has no approval modes.
- The needs_review rows in the README, llms.txt and the skills no longer
  say "Nothing was written" unconditionally: they point at
  record_changed, which is "unknown" only while a run waits for an
  approval.

tests/test_approval_flow.py drives this end to end with real
openadapt-flow main, a real browser and MockMed (opt in with
OPENADAPT_AGENT_E2E=1; it skips until the installed Flow has approval
modes). Run locally against Flow main: with a 20 s server timeout and a
25 s approval wait, the server didn't kill the run and Flow stopped by
itself; an approval given through item-approval saved exactly once.

Ledger: B272

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Report the package version as the MCP server version
5f9d2f1 
Both Server(...) calls in build_server left out version=, so the
initialize response's serverInfo.version was the MCP SDK's own version
under MCP 1 (for example "1.28.0") and an empty string under MCP 2.
Clients, audit logs and support triage couldn't tell which
openadapt-agent release served a session.

build_server now passes version=openadapt_agent.__version__ on both SDK
branches. A new test checks create_initialization_options() and the
serverInfo a real client session receives, for the bridge and the
authoring server; it fails on MCP 1.28 and 2.3 without the change.

Ledger: B091

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Stop telling pip users to pass a runner link to openadapt connect
aa1bf63 
docs/MAILBOX_CLI.md offered `openadapt connect '<runner-link-or-pack-url>'`
as a "meta-package alias" and told the job page to show
`pip: openadapt connect '<runner-link>'`. The launcher's
`openadapt connect` is Cloud pairing: it takes only --pairing or --uri,
so a runner link exits 2 with "Got unexpected extra argument". A test
also required the README to contain "openadapt connect".

The doc now says there is no launcher alias, that `openadapt connect`
is Cloud pairing and rejects runner links, and that the job page
should offer `openadapt-agent authoring connect '<runner-link>'`. The
README already pointed at the right command. The test that required
"openadapt connect" is gone; a new test fails on any README, llms.txt,
docs, skills or source line that passes an argument straight to
`openadapt connect`.

Ledger: B088

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
List all 14 recording tools and keep the Beta label out of listings
609f163 
`serve --authoring` registers 14 tools (AUTHORING_TOOLS), but no doc
listed them all. The README named 7 with "such as", llms.txt named 4
"and more", and DESIGN.md named the 4 probe tools and `type`. A partner
engineer had no way to learn about continue_input, get_command_result,
set_coach, get_coach or bind_status short of reading the code.

DESIGN.md now has a "Recording tools" section with a table of all 14
tools in AUTHORING_TOOLS order, each marked probe or local with one
line on what it does, and the README links to it. A test requires the
table to match AUTHORING_TOOLS exactly and the README to link it.

The other two B087 claims are settled on main: pyproject, server.json
and manifest.json carry no Beta label (a new test keeps it out), and a
production-eligible VERIFIED report now reads as success, which
docs/DISTRIBUTION.md already says. PyPI and the MCP registry keep
showing 2.0.1 until the next release.

Ledger: B087

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Refuse to start when an earlier version's ./runs would be skipped
0901783 
The new ~/.openadapt/runs default (B082) moved the request-id ledger and
pauses away from ./runs. A server upgraded in place, with no --runs-dir,
would start with an empty ledger next to the old one: it wouldn't know
the request_ids sent before and could enter one of them again.

When no folder is named (no --runs-dir, no OPENADAPT_AGENT_RUNS_DIR) and
./runs/openadapt-agent exists, production and attended servers now exit
2 and name both folders, so the operator picks one with --runs-dir.

Ledger: B082

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Don't call a missing frontmost window a shared title
f14e063 
When the frontmost window was missing from its app's window list (Flow's
find_windows skips windows with no size), the pin said more than one
window had the same title. It now says macOS didn't list the frontmost
window, and keeps the shared-title message for the case it describes.

Ledger: B077

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@abrichr @claude
Test approval waits with a real child process
e3d58b2 
The approval tests stubbed _run_flow_process, so nothing checked that the
runner and the service find Flow's approval request in the run folder.
Two tests now start a child process that writes a request the way Flow
does: the runner doesn't kill it at --timeout while it waits, and
run_workflow returns needs_review with reason waiting_for_approval, then
done after the answer. Both fail on main.

Ledger: B272

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 14, 2026
Merged
Merged
abrichr added a commit that referenced this pull request Oct 14, 2026
@abrichr
ci(gitleaks): allow the synthetic bundle key in the approval test (#67)
e289381 
The scan reads every branch, so the fixture on PR #64's branch fails
gitleaks on every PR. The allowlist matches only that value in that file.

Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
@abrichr
Merge main into fix/ledger-wave2-2026-10-14
c089364 
Keep the version-pinned Server and the _list_v1 wrapper from this branch.

Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
@abrichr

abrichr commented Oct 15, 2026

Copy link
Copy Markdown
Member Author

Self-review. Not merging in this pass.

  • The branch conflicted with main in mcp.py, the version-pinned Server construction. I merged main in (c089364) and kept this branch's _list_v1 wrapper. 421 tests pass locally.
  • The gitleaks failure was the synthetic bundle key that main already allowlists (ci(gitleaks): allow the synthetic bundle key in the approval test #67). With main merged, a local gitleaks git . finds no leaks.
  • admit now takes no arguments and needs MCP form elicitation. A direct call refuses with admit_requires_host_confirmation. The design is right. Its limit: it trusts the client to really show the form to a person.
  • It's a large PR (+2.3k), so it should get its own pass once CI is green.

@abrichr
fix: call Flow admit once; drop a duplicate test name
e07a44b 
The admit fallback retried on TypeError, so a TypeError raised inside Flow's
admit would run it again. The call shape now comes from the signature.
Merging main left two test_initialize_reports_the_package_version
definitions; the second silently replaced the first. Keep the parametrized
one.

Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
@abrichr
Merge main into fix/ledger-wave2-2026-10-14
bac85db 
Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
@abrichr
test: stub _run_flow_process in the Seal runner test
43ba62a 
Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
@abrichr
build: keep .gitleaks.toml out of the MCPB bundle
246ed1b 
Signed-off-by: Richard Abrich <richard.abrich@mldsai.com>
@abrichr

abrichr commented Oct 15, 2026

Copy link
Copy Markdown
Member Author

Merge pass.

Admit fails closed. admit takes no arguments. AuthoringBridge._admit refuses with admit_requires_host_confirmation unless the MCP transport sets human_confirmed, and only mcp.py sets it, after elicit_form returns accept with confirmed: true. The other cases all refuse with no admit call: a client that declares no elicitation, declares url only, or can't send requests; a decline or cancel; extra arguments. A direct dispatch refuses too. Limit: the server can't prove the client really showed the form to a person, so a client that auto-accepts elicitation would pass. That is inherent to MCP. The tool description and server instructions say so plainly.

Fixed in this pass:

  • _admit retried on TypeError (kwargs, then "ok", then no args). A TypeError raised inside Flow's admit could run admit a second time. The call shape now comes from the signature, so admit runs once. Regression test added.
  • Merging main left two test_initialize_reports_the_package_version definitions (ruff F811), so one test never ran. Kept the parametrized one.
  • tests/test_seal.py (from feat: report sealed only when a Cloud Seal verifies against pinned keys #68) stubbed subprocess.run, which this branch replaced with _run_flow_process. It now stubs the new function.
  • The MCPB build failed because main's .gitleaks.toml landed in the bundle. Added it to .mcpbignore.

Also reviewed: approval-wait handling (only an unanswered, unexpired request counts as waiting; the time a person spends answering is capped at 4 h), the ~/.openadapt/runs default with a refusal when a legacy ./runs ledger exists, --approval-mode (fixed by the operator, never by the caller), and the url override key. No other issues. 424 tests pass, gitleaks is clean, CI is green.

@abrichr
abrichr merged commit b168f96 into main Oct 15, 2026
11 checks passed
@abrichr
abrichr deleted the fix/ledger-wave2-2026-10-14 branch October 15, 2026 03:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@abrichr