Repository navigation
Description
Outcome. The agent review check builds and runs on a pull request again, and it does not break at the next .NET patch release.
Why now. The trunk rule on main requires two status checks: pipeline (devcontainer) and agent review. The agent review workflow fails at its build step, so no pull request can land. #80's pull request and PRs #73 and #74 all wait on this.
Ground. AGENTS.md. tools/canon/global.json, Directory.Build.props, Directory.Packages.props and src/Canon/packages.lock.json. .github/workflows/agent-review.yml and validate.yml. .devcontainer/devcontainer.json. core/decisions/DDD-dec-101.yaml and the decisions that rule canon's build.
What was observed, 2026-11-18.
- The agent review run on PR The act map, a second act map and a process: DDD-dec-122, DDD-address-06 to -08 #78 failed at build with
NU1004: the package lock file is inconsistent with the project's dependencies. The lock file holdsMicrosoft.NET.ILLink.Tasksat 10.0.14 and the runner resolves 10.0.15. - The act map, scope and pull: what they are, and how step 1 starts from them #80's session saw the same from the other side: a local
dotnet runrewrotepackages.lock.jsonfrom 10.0.14 to 10.0.15. It discarded the change. validatestill passes (run 385, on7c1214e).
Read from the files at 84aa466, by Claude, UNVERIFIED — Emil review.
global.jsonasks for SDK10.0.100withrollForward: latestFeature, so any newer 10.0 SDK is accepted.Microsoft.NET.ILLink.Tasksis in the lock file as a direct dependency requested at[10.0.14, ). Nothing in the project files names it. The SDK adds it, at the version of the SDK in use.Directory.Build.propsturns locked mode on whenCIis set. So a newer SDK than the one that wrote the lock file fails the restore.- The agent review workflow installs .NET with
actions/setup-dotnet. The devcontainer installs it with the dotnet feature at version10.0. Neither names a patch. - So
validatepasses only while the devcontainer still holds the older SDK. When it picks up the newer one,validatefails the same way.
Deliverable.
-
The diagnosis confirmed or corrected: which SDK each workflow gets today, why the two differ, and where the dependency comes from. Exit: the session reproduces
NU1004and says in one sentence what causes it. -
Proposals, each with its weakest point and what it breaks. At least these two:
- regenerate the lock file for the newer SDK and change nothing else;
- pin the SDK to one exact version in
global.json, have both the workflow and the devcontainer install that version, and regenerate the lock file once.
Exit: each says what happens at the next .NET patch release.
-
On the ruling, the fix. Exit: both required checks pass on this issue's own pull request:
pipeline (devcontainer)andagent review. Thevalidateworkflow passes, and every commit carries aRefsline for this issue.
Gates. G1 after step 2: the principal rules which fix. G2 after step 3. Then the merge.
Out of scope. Any change to what canon checks. Any claim, decision text, method card or lifecycle data, beyond a decision record if the ruling needs one. The act map work (#80, #81, #82).
Opening position, [PROPOSED]: the second option. Regenerating alone fixes today and fails again at the next patch, and it fails validate as soon as the devcontainer is rebuilt.
A caution for the session: the agent review workflow runs from main (pull_request_target), so a change to the workflow file itself is not exercised by the pull request that makes it. Say how the fix is shown to work before it lands.
Corrections and rulings, 2026-11-18
Correction (Claude, found by this issue's session at its G1). The reading above is wrong on one point and silent on another.
validatedoes not pass because the devcontainer holds an older SDK. It holds the same one, 10.0.404. It passes becauseCIis never passed into the container, so the restore there is never locked.validatehas never enforced canon's lock file.- The deadlock. The agent review job checks out
mainand buildsmain's canon, and thetrunkrule requiresagent reviewwith an empty bypass list. So the check fails on every pull request, this issue's own included, untilmainbuilds. Deliverable 3's exit cannot be met as written.
Ruled by Emil at G1 ("agreed and accepted", 13:57), on the session's proposals and Claude's recommendation:
- The fix is option B: pin the SDK exactly in
global.jsonwithrollForwardset todisable, pin the same version indevcontainer.json, and regenerate the lock file once.agent-review.ymldoes not change. The version is 10.0.404. CIis passed into the devcontainer, sovalidaterestores in locked mode. This closes a gap in an existing rule (Directory.Build.props) and is not a new check.- Emil lifts the
agent reviewrequirement on thetrunkrule for this one landing and restores it straight after. No standing bypass is added. The waiver is recorded as a decision, whose text Emil rules before it is filed. - A throwaway workflow on the session's branch is allowed, to show the build on a real runner: push to that branch only, read-only permissions, no secrets, removed before the head to ratify.
- Deliverable 3's exit becomes:
validatepasses on the head with locked restore; the session's pre-landing proofs pass; every commit carries aRefsline. The first agent review run after landing is the final proof.
Successor items, not acted on here: a scheduled check that says when a newer .NET patch exists; and the deadlock itself, since a required check that builds main's canon with no bypass means a broken main can only be repaired by suspending the rule.
Generated by Claude Code
Metadata
Metadata
Assignees
Labels
Type
Fields
Priority
Effort
Progress
Activity
rulings.md: Emil's G1 rulings, verbatim, received 13:17 UTC
predictions.md: P-3 scored; P-4 to P-7 before G2's operations
canon's SDK pinned at 10.0.404; the lock file regenerated once; CI in…
Throwaway workflow: agent review's build steps on this branch, on push
Session canon-lock-file: G2 findings, the waiver draft, predictions s…
Remove the throwaway agent review build proof workflow
DDD-dec-123: the first agent review run after the landing, run 89
Hold at G2; successor items X-1 and X-2 for #82; the #82 correction
Record P-4.6's result: validate passed on 3d76c88; agent review faile…
DDD-dec-123: the three times; the live rule departed from 15:20 to 15:39
Merge pull request #86 from mindovermachine-dev/claude/canon-lock-fil…
Hold at G2; successor items X-1 and X-2 for #82; the #82 correction
Record P-4.6's result: validate passed on 3d76c88; agent review faile…
Session canon-lock-file: P-9 scored; the lift time's flag cleared
Sessions index: the canon-lock-file row brought up to what happened