Compose the named-token credential graph #5
No reviewers
Labels
No labels
bug
duplicate
enhancement
help wanted
invalid
question
wontfix
bug
duplicate
enhancement
help wanted
invalid
question
wontfix
No milestone
No project
No assignees
2 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
allod/deploy!5
Loading…
Reference in a new issue
No description provided.
Delete branch "agent/named-tokens-composition"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
This PR completes the named-token feature by binding the deploy graph to the three reviewed implementation branches and teaching the composition canary the new credential projections. On a running machine nothing changes until pins are updated and rebuilt: the public registries are empty, so no credential or token exists yet. What changes is which framework revisions deploy composes and what the canary proves: empty public data stays inert, the synthetic per-token projection shape is exactly what secrets emits, an invalid default is rejected, and each follows edge still fails closed when sabotaged. What proves it is the composed-layer check built green over the pinned graph. If it is wrong, revert this PR or re-pin to the previous master revisions and the prior graph returns. Before merge, the three pinned branches must merge first and the pins here must move to their master merge commits.
Risk
R2 per the plan's slice table: this changes only the pinned input graph and eval-time canary assertions; no runtime behavior exists with empty public data, and rollback is a straight revert. The pins currently name the three agent branches (archetypes 9ba5883, nexus 0a9b30e via archetypes/nexus, secrets 9d637b1) per the established lock-only precedent for unmerged upstream work; they must be re-pinned to master merge commits before this PR merges.
Validation
nix build path:.#checks.x86_64-linux.composed-layer — green from this worktree over the pinned graph, exercising the empty public assertions, the synthetic multi-token fixture with null-default case, the provider join, the default sabotage rejection, and the four follows sabotage witnesses.
Closes allod/strategy#36
Review pass (required post-open review): read-only review by gpt-5.6-sol at reasoning high (author was gpt-5.6-terra). Both findings are fixed in the follow-up commit: the lock had silently reverted the transitive allod-tools pin while re-locking archetypes to its agent branch (restored — the lock now moves exactly the three reviewed revisions against origin/master), and the activation-projection half of the deploy contract is now discharged by building the archetypes pi-provider-lifecycle witness through this flake's follows-redirected inputs inside the composed-layer check. nix build path:.#checks.x86_64-linux.composed-layer re-run green by the managing agent from this worktree, building the composed lifecycle witness.