This is what a Context Receipt looks like.

Muninn Context Receipt (server-scored)

This is a context audit, not a code review. It is not a code-quality or security approval of the change.

Decision needed before merge

Your context tells the agent to use one package manager while the repo's lockfile implies another (2 such collisions this run).

Why it matters

The diff can pass CI and still install the wrong dependencies.

Evidence

• ·context surface path#194744c3d21031ca conflicts with lockfile path#5d2ae2d80560b2ae (reason: pkg_manager_mismatch)
• ·context surface path#944ee11b8cfa67c5 conflicts with lockfile path#5d2ae2d80560b2ae (reason: pkg_manager_mismatch)

What was scored

• ·files scored: 5
• ·risk breakdown: clear=1, flagged=0, high_risk=4
• ·admission lane hints: Active=0, Canonical=0, Probation=0, Quarantine=0, Reject=5, Retired=0
• ·2 policy collision(s) detected (context directive vs. lockfile)
• ·resource footprint: 1956 bytes, ~492 est. tokens

High-risk pointers (need review)

• ·path#194744c3d21031ca (lane hint: Reject)
• ·path#944ee11b8cfa67c5 (lane hint: Reject)
• ·path#194744c3d21031ca (lane hint: Reject)
• ·path#944ee11b8cfa67c5 (lane hint: Reject)

Flagged pointers (structural hygiene, not safety)

None this run.

Policy collisions (package manager mismatch)

• ·context surface path#194744c3d21031ca conflicts with lockfile path#5d2ae2d80560b2ae (reason: pkg_manager_mismatch)
• ·context surface path#944ee11b8cfa67c5 conflicts with lockfile path#5d2ae2d80560b2ae (reason: pkg_manager_mismatch)

What Muninn refused to conclude

• ·risk = high_risk: secret/injection-shaped signal matched by SHAPE only, on the redacted bundle the server received. The server cannot tell a live credential or a real injection from a quoted example, so it draws no conclusion either way and never proposes an automated fix. A human must verify the match in context before acting.
• ·policy_collision: a structural fact (a directive pattern matched in a context surface against which lockfile is actually present), not a semantic read of intent. Muninn does not conclude which package manager the project should standardize on, only that the two disagree today; a human must decide which side is correct.
• ·lane hints reflect Muninn's real admission engine run over only the signals a redacted bundle can honestly carry (no provenance data crosses the wire); a reject hint here is that engine's own fail-safe default given today's bundle schema, not an assertion that the underlying memory is bad.

Server-scored and signed: request 087615ce-ea95-4690-be5f-e9aa035f8c13, key cust_d328b1243f55f7f70a11ceda1222909a:v1, at 2026-07-22T02:19:35.899Z. This signature was verified by the client before this comment was posted. Findings are shown as stable pointers (path#hash), never raw filesystem paths, filenames, memory ids, or secret literals. Run muninn doctor --first-look --unredacted on your own machine for full local detail (never share that report).

Muninn does not approve code. It shows the context risk your reviewer should decide.