Skip to content

fix: apply synthetic safety guards to retry execution evidence - #3

Draft
dinakarjs wants to merge 4 commits into
review/retry-evidence-basefrom
fix/guarded-retry-evidence
Draft

dinakarjs wants to merge 4 commits into
review/retry-evidence-basefrom
fix/guarded-retry-evidence

Conversation

@dinakarjs

Copy link
Copy Markdown
Owner

Supplemental draft for coordination with the existing confidential-workflow harness owner under agentrust-io#199. Addresses the independent review reproducer: agentrust-io#199 (comment). Based on main 5f9b48e.

Previously retry_lineage emitted execution=established even when its own authorization was contradicted, or the shared inputs were revoked, replayed, below the version floor, or lacked execution evidence. The overall main-lineage result still refused or remained unknown, but the retry observation overstated its synthetic outcome.

Share admission and execution checks between main and retry paths. Add retry receipt/admission observations and bind retry execution causally to admission. Seven paired controls cover workload, key, policy, downgrade, revocation, replay and missing evidence; each includes a valid twin and its gate-weakening control. Document that a retry after main timeout is a separate synthetic observation and does not overwrite the main outcome.

Validation: all 24 existing/new harness tests pass on Python 3.12.14; targeted Ruff E,F,W excluding E501 passes; git diff --check passes. A process-only mutation forcing retry execution back to established kills all seven negative/unknown controls. No mutation committed. Existing timeout, lineage and causal-parent controls pass.

Synthetic adapter only; no live retry, protocol API, package target, hardware or conformance change. Fork draft for owner review/integration, not a competing upstream implementation.

Signed-off-by: Srinivasa Dinakar Jainti <dinakarjs@gmail.com>
@github-actions

github-actions Bot commented Oct 13, 2026 •

Copy link
Copy Markdown

🟡 Contributor Check: MEDIUM

Check Result
Profile MEDIUM
Credential LOW
Overall MEDIUM

Automated check by AgenTrust Contributor Check.

@github-actions github-actions Bot added the needs-review:MEDIUM Contributor check flagged MEDIUM risk label Oct 13, 2026
Signed-off-by: Srinivasa Dinakar Jainti <dinakarjs@gmail.com>

Copy link
Copy Markdown
Owner Author

Hosted CI follow-up: lint, documentation, contributor and integrity checks passed on 9eba4dd. Validate integrations stopped at main's inherited Marketplace count mismatch (46 recorded, 47 listed), before reaching the harness test step. Added the generator's count-only correction at 69d3ef6; all five generator tests and --check pass locally. This is the same baseline correction already submitted in upstream agentrust-io#315. CI is rerunning; the functional retry patch remains a draft pending owner coordination.

Copy link
Copy Markdown
Owner Author

Hosted CI at final head 69d3ef6 is now green: all seven returned workflows passed, including Validate integrations, Lint, documentation links and four integrity checks. The inherited catalog blocker is resolved. This draft remains pending coordination on retry semantics with the existing agentrust-io#199 owner.

dinakarjs and others added 2 commits October 14, 2026 18:50
Integrate the non-overlapping regression controls from Altru patch dff4a395;
shared refusal precedence already landed at 6ac5fff. Clarify synthetic
permitted-execution outcomes do not prove physical side-effect absence.

Co-authored-by: Val Rukhaylo <altrudev@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

needs-review:MEDIUM Contributor check flagged MEDIUM risk

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant