Repository navigation
feat(governance): settle the campaign-governance rules that were contradictory or undefined - #9006
Conversation
…radictory or undefined PR 2 of 3 for #8733: the normative decisions of the Governance working group, confirmed by the maintainer on the issue. - A1: a sync that reduces budget.total below committed spend is rejected with VALIDATION_ERROR; the training agent gets the sync-time check - A4: an authenticated caller that fails approved_sellers or delegations gets a denied verdict with a finding and a check_id; PERMISSION_DENIED stays for a missing or mismatched identity - B1: plans with human_review_required: true escalate in every mode - C1: a plan_id names at most one plan within a governance agent - C2: execute_only acts only on a pre-approved action binding - C3: no or country-level geography is outside a regions-restricted plan - C4.3: an unresolvable registry policy ID is REFERENCE_NOT_FOUND - C4.4: "the same action" means the same consultation_context No wire field, enum member or plan_hash rule changes. Refs #8733 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…ed callers in every mode Follow-up within PR 2 of #8733, where the accepted wording sat against rules already in the specification. - C1: plan_id is unique per owner principal, as the reference agent has implemented since #6214. A sync never reads, changes or reveals another principal's plan, which keeps the plan-ID enumeration rule. - A4: governance mode does not relax who may call. A buyer-side caller that is neither owner nor current delegate, and an execution caller that is not the preserved audience, are denied in every mode. Refs #8733 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…e target seller Follow-up within PR 2 of #8733. The audit-trail mode examples showed an unapproved seller's execution check approved in advisory and audit mode, which contradicts the rule that a caller without authority is denied in every mode. They are the buyer's intent checks naming an unapproved target seller: a finding about the target follows mode. The finding text now reads "Target seller", as the reference agent emits. Refs #8733 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated schema paths touched without required review approval.
This PR modifies published protocol surface under static/schemas/source/** (delegation-authority.json, governance-mode.json, sync-plans-request.json). These are gated paths under a hard, deterministic approval gate, and review_decision is REVIEW_REQUIRED (not APPROVED). Per decision rule row 2, the gate is not satisfied and the PR cannot be auto-approved — human/CODEOWNERS review is required on the gated files.
The reviewer surfaced no Critical/High/Medium findings: schema↔docs coherence verified (prose-only enumDescription/description edits, no field/type/required/enum-member changes), a correct minor changeset, no wire-shape change, and the authority-denial refactor was confirmed equivalent and leak-free. On the merits this is a clean spec-settling PR. However, the diff touches gated schema source and is a protocol-policy/normative-classification call (Breaking vs Normative), which is the arbiter's/human's judgment and is never auto-approved absent an APPROVED review.
Blocking findings: none.
Escalation reasons:
- Modifies gated path
static/schemas/source/enums/delegation-authority.jsonwithreview_decision: REVIEW_REQUIRED— human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/enums/governance-mode.jsonwithreview_decision: REVIEW_REQUIRED— human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/governance/sync-plans-request.jsonwithreview_decision: REVIEW_REQUIRED— human/CODEOWNERS approval required.
Why human review
- Modifies gated path static/schemas/source/enums/delegation-authority.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/enums/governance-mode.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/governance/sync-plans-request.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- This PR touches a path under a hard, non-overridable approval gate (static/schemas/source/enums/delegation-authority.json (modified) matches
static/schemas/source/**; static/schemas/source/enums/governance-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/sync-plans-request.json (modified) matchesstatic/schemas/source/**) and the current GitHub review decision is 'REVIEW_REQUIRED', not APPROVED. This is a hard gate enforced in code — Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded, regardless of how clean the diff is.
|
@lukasz-pubx thanks for PR 2. A1, B1 and C3 match the decisions. Four places differ from what was settled on #8733 (my comment there: #8733 (comment), plus the working-group record in the issue description). Could you align them?
Two smaller things:
Since the 3.3 RC is cut Monday, this will land as an RC fix. That's fine for errata-level alignment like this. |
|
A question on point 2 of the review above, before @lukasz-pubx aligns to it. Point 2 gives a delegation denial to every authenticated caller, including one with no access to the plan. Doesn't that conflict with Three smaller questions on the alignment:
|
|
@KonstantinMirin thanks. The conflict is real. Point 2 and @bokelley that is yours to decide. I have the alignment to your four points ready and am holding the push until you answer, because the answer changes what the specification sentence for point 2 says. C1 has the same shape on a smaller scale: On the three smaller questions:
|
|
@lukasz-pubx @KonstantinMirin thanks, you're both right; the uniform-response rule wins.
Please push the alignment with these. I'll update the record on #8733 so it matches. |
Applies the review on adcp#9006 and the decision record update on adcp#8733 of 2026-11-15. - A4: a failed approved_sellers or delegations evaluation is denied in every governance mode. The mode section, the mode summaries and the governance-mode enum descriptions name it as a second case that mode does not change, next to human_review_required. - A4, uniform response: a caller with no access to the plan receives PLAN_NOT_FOUND, the same response as for an unknown plan_id, and the attempt writes no audit entry. The denied verdict for a failed delegations evaluation is for a caller that can access the plan. - C1: plan_id stays scoped per owner principal. No PERMISSION_DENIED on reuse, which would reveal that an ID is in use. - C2: an execute_only delegate acts only on an action that already has an approved intent check, which includes accept_proposal. The rule is worded as an obligation on the governance agent. - sync_plans error table: an explicit human_review_required: false on a plan that requires human review is rejected with VALIDATION_ERROR. - audit-trail: the three mode examples use a budget finding, because an unapproved seller is now denied in every mode. Training agent: check_governance no longer approves a failed approved_sellers or delegations evaluation in advisory or audit mode. For a plan that does not exist or that the caller cannot access it returns the PLAN_NOT_FOUND error and stores nothing; it returned a denied verdict with a plan_lookup finding. Refs #8733. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated schema paths require human/CODEOWNERS approval (review_decision is REVIEW_REQUIRED).
This PR modifies three files under static/schemas/source/**, which are declared under a hard, non-overridable approval gate (gated_paths: true). The current review_decision is REVIEW_REQUIRED, not APPROVED, so the gate is not satisfied. Per decision rule row 2, the outcome must be escalate until a human/CODEOWNERS approval is recorded.
Escalation reasons
static/schemas/source/enums/delegation-authority.json(modified) is under a hard approval gate; review_decision is REVIEW_REQUIRED, not APPROVED.static/schemas/source/enums/governance-mode.json(modified) is under a hard approval gate; review_decision is REVIEW_REQUIRED, not APPROVED.static/schemas/source/governance/sync-plans-request.json(modified) is under a hard approval gate; review_decision is REVIEW_REQUIRED, not APPROVED.
No blocking findings exist. The reviewer confirmed a clean pass: schema changes are description-only and coherent with the updated docs (no wire field/enum/required/plan_hash change), a correctly-scoped minor changeset is present, and there are no oneOf/immutability/patch-eligibility issues. The governance-handlers.ts logic matches the budget-period and binding tests. There are no Critical/High/Medium findings.
This is consistent with the prior escalate: the gate persists purely because human approval has not yet been recorded, not because of any defect in the diff. Once review_decision becomes APPROVED, this clean diff can fall through to approve.
Why human review
- Modifies gated path static/schemas/source/enums/delegation-authority.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/enums/governance-mode.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/governance/sync-plans-request.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- This PR touches a path under a hard, non-overridable approval gate (static/schemas/source/enums/delegation-authority.json (modified) matches
static/schemas/source/**; static/schemas/source/enums/governance-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/sync-plans-request.json (modified) matchesstatic/schemas/source/**) and the current GitHub review decision is 'REVIEW_REQUIRED', not APPROVED. This is a hard gate enforced in code — Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded, regardless of how clean the diff is.
|
@bokelley thanks, pushed in
Three things to know before merging:
|
…escriptions PR 4 for #8733: defects found while preparing PRs 2 and 3. Each edit follows a rule that is already stated in the specification or schemas. No schema property, enum member or plan_hash rule changes. - name all five human-review triggers on the trust, known-limitations and FAQ pages; the FAQ says an omitted flag is accepted from 3.3 - the requires_human_review descriptions say the governance agent sets plan.human_review_required - reallocation_threshold is required only without reallocation_unlimited - overview: advisory mode returns approved; no budget is reserved - drop a stale reference to a closed issue; mode is agent configuration - the training agent's plans[].mode is labelled a sandbox test control Rebuilt on the aligned PR 2 (#9006). The human_override declaration is removed from this PR, and the PLAN_NOT_FOUND change now lives in #9006. Refs #8733 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
KonstantinMirin
left a comment
There was a problem hiding this comment.
Thanks @lukasz-pubx. Point 2, C1, A4 and C2 now match the ruling, but two sentences at specification.mdx:215 need a fix before merge. The reference agent gives an inaccessible plan and a missing one the same bytes and writes nothing, as the uniform-response rule requires.
Blocking
- Define plan access. Lines 215 and 885 choose between
PLAN_NOT_FOUNDand thedelegation_authoritydenial by whether the caller "can access the plan", and no page defines access. The reference agent reads it as the owner, a current delegate, or a caller that presents agovernance_contextorconsultation_contextissued for the plan (governance-handlers.ts:336,:1462). Today two agents can answer the same call differently. Please state that definition in the specification. - Scope the
PERMISSION_DENIEDsentence. "returnsPERMISSION_DENIEDonly when the caller is unauthenticated or its authenticated identity does not matchcaller" excludes an expired or invalidgovernance_context.error-code.json:217lists that case underPERMISSION_DENIED, the reference agent returns it (governance-handlers.ts:1737), and the new test attraining-agent-governance-intent-binding.test.ts:985asserts it. Suggested: "A faileddelegationsorapproved_sellersevaluation is neverPERMISSION_DENIED."
Non-blocking
- The list after "
deniedin every mode:" at line 885 reads as complete, but it leaves out a delegate acting outside itsbudget_limit,marketsorexecute_onlyscope. The reference agent denies those in every mode. - Some mode summaries still say audit mode approves everything:
campaign/index.mdx:31andgovernance/overview.mdx:240.safety-model.mdx:62andgovernance-mode.json:14keep "never block" next to thedeniedexception, and the audit risk column still says "Zero." - The test at
training-agent-governance-intent-binding.test.ts:949is named "outside delegations", but its fixture is a delegate over itsbudget_limit.
Local runs at 1d11b5d: test:schemas and lint:additive-minor pass, the 14 governance and training-agent test files pass (1,072 tests), and CI is green.
|
@lukasz-pubx thanks for I'd define it as: the plan's owner, a current delegate, or a holder of a governance context issued for that plan. The last group covers a delegate whose delegation lapsed mid-flow, who then gets |
Second revision of adcp#9006 after review. - specification.mdx: define which callers can access a plan. A plan is accessible to its owner principal, to a caller named in a delegations entry that has not expired, and to a caller that presents a governance_context or consultation_context issued for that plan. Any other caller gets PLAN_NOT_FOUND. This is what the reference agent does. - specification.mdx: the PERMISSION_DENIED sentence no longer says "only". It says a failed delegations or approved_sellers evaluation is never PERMISSION_DENIED and names the authentication cases as examples, so an expired governance_context is not excluded. - specification.mdx: the list of outcomes that are denied in every mode adds a delegate that acts outside the authority, budget_limit or markets of its delegation. - check_governance.mdx: the PLAN_NOT_FOUND row states the access rule. - index.mdx, overview.mdx, safety-model.mdx, governance-mode.json: the audit mode summaries no longer say that the mode approves everything or never blocks. - Rename one test to match its fixture, a delegate over its budget_limit. - Changeset: state the access rule. No behaviour change in the training agent. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated schema paths require human/CODEOWNERS approval.
This PR modifies three files under the hard approval gate static/schemas/source/**, and review_decision is REVIEW_REQUIRED (not APPROVED). Per decision-table row 2, the outcome must be escalate until a human/CODEOWNERS approval is recorded. The clean diff cannot lift this gate on its own.
The reviewer found no Critical/High/Medium findings: this is a coherent documentation-and-wording refinement pass (PLAN_NOT_FOUND vs PERMISSION_DENIED model stated consistently, the "two outcomes hold in every mode" exception aligned across index/safety-model/overview and the governance-mode enum, a test rename matching the added delegation-scope exception). Schema↔docs coherence holds; no enum member, wire field, or plan_hash change. The prior escalate also carried no blocking findings — so escalation here persists solely because the gated-path approval is still outstanding, not because of code concerns.
Escalation reasons
- Modifies gated path
static/schemas/source/enums/delegation-authority.jsonwhilereview_decisionisREVIEW_REQUIRED— human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/enums/governance-mode.jsonwhilereview_decisionisREVIEW_REQUIRED— human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/governance/sync-plans-request.jsonwhilereview_decisionisREVIEW_REQUIRED— human/CODEOWNERS approval required.
Once a human/CODEOWNERS approval is recorded, this PR is clean and can fall through to approval.
Why human review
- Modifies gated path static/schemas/source/enums/delegation-authority.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/enums/governance-mode.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/governance/sync-plans-request.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- This PR touches a path under a hard, non-overridable approval gate (static/schemas/source/enums/delegation-authority.json (modified) matches
static/schemas/source/**; static/schemas/source/enums/governance-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/sync-plans-request.json (modified) matchesstatic/schemas/source/**) and the current GitHub review decision is 'REVIEW_REQUIRED', not APPROVED. This is a hard gate enforced in code — Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded, regardless of how clean the diff is.
|
@bokelley @KonstantinMirin thanks, you both asked for the same sentence. It is in 1. Plan access. The specification now defines it, in the identity paragraph:
This is the definition bokelley gave, and it is what the reference agent already does. One detail: the sentence names both context kinds. An intent check cannot carry a 2.
The cases are examples, so an expired Non-blocking points from the review
|
KonstantinMirin
left a comment
There was a problem hiding this comment.
Thanks @lukasz-pubx. 84cbf74 fixes both points from my last review: plan access is defined at specification.mdx:215, and the PERMISSION_DENIED sentence no longer excludes credential failures. I have one question about C1 before merge.
Which plan does a shared plan_id name? With plan_id unique per owner (sync_plans.mdx:114), brands A and B can both sync q1-2027 and delegate the same agency. The agency's intent check carries only plan_id (check-governance-request.json:233), so I can't find what decides whose budget it runs against. get-plan-audit-logs-request.json:21 ("Plans uniquely scope account") also stops being true.
Could we make plan_id unique instead of scoping it? Two ways:
- A UUID, the pattern
idempotency_keyalready follows ("Use a fresh UUID v4"). - An ID the governance agent mints, the way a seller mints
media_buy_id.
Either way the buyer gets a separate field for its own label. IDs can't collide, so neither per-owner scoping nor a tie-break rule is needed. @bokelley, does that fit your C1 ruling?
Smaller points
- A
governance_contextissued for another plan reads asPLAN_NOT_FOUNDat:215, buterror-code.json:217makes itPERMISSION_DENIED. Could the spec say the context is verified before access is resolved, and add the case to the row atcheck_governance.mdx:744? sync-plans-request.json:383says an expired delegate's actions are denied, while:215gives that delegatePLAN_NOT_FOUND.- The list at
:885misses an execution check whose original buyer has since lost its delegation. index.mdx:146still says "approve everything", andoverview.mdx:240still has advisory mode returningdenied.- The rc.1 notes need the notice basis for A1 and A4, as
3-2-to-3-3.mdx:936records one. - No storyboard grades A4 or
PLAN_NOT_FOUNDbyte identity yet. A follow-up is fine.
Local runs at 84cbf74: test:schemas and test:additive-minor pass, and the 14 governance and training-agent files pass (1,072 tests). CI is green.
…ch the specification The row listed two cases as if exhaustive; the specification now gives them as examples of authentication and credential failures. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
@lukasz-pubx thanks, the access definition and the rescoped sentence are right. I pushed one line in |
…d mode wording - Verify a presented governance_context before resolving plan access; a context that fails verification or names another plan is PERMISSION_DENIED, matching error-code.json. - The every-mode denied list is non-exhaustive and includes an execution check whose original buyer lost its delegation. - A conditions re-check returns the same consultation_context. - Mode wording in the campaign index and overview no longer contradicts the specification. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated schema paths require human/CODEOWNERS approval.
This PR modifies protocol source-of-truth files under static/schemas/source/**, which fall under a hard, deterministic approval gate. The current review_decision is not APPROVED (unknown/none), so the gate is not satisfied and auto-approval is prohibited regardless of how clean the diff is.
Findings: None. The reviewer found no blocking or medium issues. This subsequent pass addressed only a 13-line docs clarification (the PERMISSION_DENIED row in check_governance.mdx broadened to cover authentication/credential failures), coherent with the adjacent PLAN_NOT_FOUND row and the A4 decision — no spec drift.
Why escalate (not approve): Row 2 of the decision table — gated_paths: true AND review_decision ≠ APPROVED. The gate lifts only when a real human/CODEOWNERS approval is recorded. The prior decision was also escalate for the same reason.
Escalation reasons:
- Modifies gated path
static/schemas/source/enums/delegation-authority.json— human/CODEOWNERS approval required - Modifies gated path
static/schemas/source/enums/governance-mode.json— human/CODEOWNERS approval required - Modifies gated path
static/schemas/source/governance/sync-plans-request.json— human/CODEOWNERS approval required
A human reviewer with CODEOWNERS authority must approve the schema-source changes before this can merge.
Why human review
- Modifies gated path static/schemas/source/enums/delegation-authority.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/enums/governance-mode.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/governance/sync-plans-request.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- This PR touches a path under a hard, non-overridable approval gate (static/schemas/source/enums/delegation-authority.json (modified) matches
static/schemas/source/**; static/schemas/source/enums/governance-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/sync-plans-request.json (modified) matchesstatic/schemas/source/**) and the current GitHub review decision is 'unknown', not APPROVED. This is a hard gate enforced in code — Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded, regardless of how clean the diff is.
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated path requires human/CODEOWNERS approval.
This PR modifies files under static/schemas/source/**, which are under a hard, non-overridable approval gate (gated_paths: true). The current review_decision is unknown/none, not APPROVED, so the gate is not satisfied (decision table row 2).
The reviewer pass was clean: no critical/high/medium findings. The delta is four governance docs (72 lines) plus schema enum/description reconciliation, and the reviewer confirmed schema↔docs coherence (governance-mode.json and delegation-authority.json enum descriptions match the delta text) with no spec drift. The prior run also escalated solely on the gated-path trigger with no actionable findings. Row 6 (sticky escalation) does not independently apply since this run surfaces no fresh critical/high/medium findings — but the gated-path gate (row 2) fires regardless and holds until a human/CODEOWNERS approval is recorded.
Escalation reasons
- Modifies gated path
static/schemas/source/enums/delegation-authority.jsonwhilereview_decisionis not APPROVED — human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/enums/governance-mode.jsonwhilereview_decisionis not APPROVED — human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/governance/sync-plans-request.jsonwhilereview_decisionis not APPROVED — human/CODEOWNERS approval required.
The diff is mechanically clean; once a human/CODEOWNERS approval is recorded the gate lifts.
Why human review
- Modifies gated path static/schemas/source/enums/delegation-authority.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/enums/governance-mode.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/governance/sync-plans-request.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- This PR touches a path under a hard, non-overridable approval gate (static/schemas/source/enums/delegation-authority.json (modified) matches
static/schemas/source/**; static/schemas/source/enums/governance-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/sync-plans-request.json (modified) matchesstatic/schemas/source/**) and the current GitHub review decision is 'unknown', not APPROVED. This is a hard gate enforced in code — Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded, regardless of how clean the diff is.
|
@bokelley thanks for
|
…ble plan Per-owner plan_id scoping lets a delegate of two owners see two plans with the same ID, and an intent check carries only plan_id. The governance agent now rejects that check with VALIDATION_ERROR and evaluates it against neither plan, and buyers SHOULD use globally unique plan IDs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated schema paths require human/CODEOWNERS approval.
This PR modifies files under static/schemas/source/**, which are under a hard, non-overridable approval gate (gated_paths: true), and the current review_decision is REVIEW_REQUIRED, not APPROVED. Per decision rule row 2, the outcome must be escalate until a human/CODEOWNERS approval is recorded.
Escalation reasons
- Modifies gated path
static/schemas/source/enums/delegation-authority.jsonwhilereview_decisionis not APPROVED — human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/enums/governance-mode.jsonwhilereview_decisionis not APPROVED — human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/governance/sync-plans-request.jsonwhilereview_decisionis not APPROVED — human/CODEOWNERS approval required.
No blocking findings. The reviewer's subsequent pass is clean: the delta is 3 docs-only files adding one normative decision (a check_governance caller that can access more than one plan under the same plan_id returns VALIDATION_ERROR), which composes cleanly with existing PLAN_NOT_FOUND and PERMISSION_DENIED cases with no contradiction or cross-principal leak. The prior pass was also clean. This escalation is driven purely by the unsatisfied gated-path approval requirement, not by any code defect — the gate lifts once review_decision becomes APPROVED.
Why human review
- Modifies gated path static/schemas/source/enums/delegation-authority.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/enums/governance-mode.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/governance/sync-plans-request.json while review_decision is not APPROVED — human/CODEOWNERS approval required
- This PR touches a path under a hard, non-overridable approval gate (static/schemas/source/enums/delegation-authority.json (modified) matches
static/schemas/source/**; static/schemas/source/enums/governance-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/sync-plans-request.json (modified) matchesstatic/schemas/source/**) and the current GitHub review decision is 'REVIEW_REQUIRED', not APPROVED. This is a hard gate enforced in code — Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded, regardless of how clean the diff is.
…sync_plans Generated by build:compliance; the PR bundle check requires it committed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
Ladon verdict: Escalate to human review
Escalate — gated schema paths require human/CODEOWNERS approval.
This PR modifies files under static/schemas/source/**, which is a hard, deterministic approval gate (gated_paths: true). The current review_decision is REVIEW_REQUIRED, not APPROVED, so the gate is not satisfied (decision table row 2). Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded.
Review quality: The diff is clean. The reviewer found no critical/high/medium findings. The only delta since the prior run is a single editorial change in docs/governance/campaign/tasks/sync_plans.mdx — converting a bare VALIDATION_ERROR reference into a cross-reference link to /docs/building/verification/compliance-catalog#error-code-validation-error, with the anchor verified to exist and the link format matching established convention. Schema↔docs coherence, changeset scope, and oneOf discriminator checks surfaced nothing actionable.
Escalation reasons:
- Modifies gated path
static/schemas/source/enums/delegation-authority.jsonwhilereview_decisionisREVIEW_REQUIRED— human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/enums/governance-mode.jsonwhilereview_decisionisREVIEW_REQUIRED— human/CODEOWNERS approval required. - Modifies gated path
static/schemas/source/governance/sync-plans-request.jsonwhilereview_decisionisREVIEW_REQUIRED— human/CODEOWNERS approval required.
Once a human/CODEOWNERS approval lands and review_decision becomes APPROVED, this clean diff can fall through to approve.
Why human review
- Modifies gated path static/schemas/source/enums/delegation-authority.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/enums/governance-mode.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- Modifies gated path static/schemas/source/governance/sync-plans-request.json while review_decision is REVIEW_REQUIRED — human/CODEOWNERS approval required
- This PR touches a path under a hard, non-overridable approval gate (static/schemas/source/enums/delegation-authority.json (modified) matches
static/schemas/source/**; static/schemas/source/enums/governance-mode.json (modified) matchesstatic/schemas/source/**; static/schemas/source/governance/sync-plans-request.json (modified) matchesstatic/schemas/source/**) and the current GitHub review decision is 'REVIEW_REQUIRED', not APPROVED. This is a hard gate enforced in code — Ladon cannot auto-approve until a human/CODEOWNERS approval is recorded, regardless of how clean the diff is.
…escriptions PR 4 for #8733: defects found while preparing PRs 2 and 3. Each edit follows a rule that is already stated in the specification or schemas. No schema property, enum member or plan_hash rule changes. - name all five human-review triggers on the trust, known-limitations and FAQ pages; the FAQ says an omitted flag is accepted from 3.3 - the requires_human_review descriptions say the governance agent sets plan.human_review_required - reallocation_threshold is required only without reallocation_unlimited - overview: advisory mode returns approved; no budget is reserved - drop a stale reference to a closed issue; mode is agent configuration - the training agent's plans[].mode is labelled a sandbox test control Rebuilt on the aligned PR 2 (#9006). The human_override declaration is removed from this PR, and the PLAN_NOT_FOUND change now lives in #9006. Refs #8733 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…escriptions (#9015) * fix(governance): correct stale campaign-governance pages and schema descriptions PR 4 for #8733: defects found while preparing PRs 2 and 3. Each edit follows a rule that is already stated in the specification or schemas. No schema property, enum member or plan_hash rule changes. - name all five human-review triggers on the trust, known-limitations and FAQ pages; the FAQ says an omitted flag is accepted from 3.3 - the requires_human_review descriptions say the governance agent sets plan.human_review_required - reallocation_threshold is required only without reallocation_unlimited - overview: advisory mode returns approved; no budget is reserved - drop a stale reference to a closed issue; mode is agent configuration - the training agent's plans[].mode is labelled a sandbox test control Rebuilt on the aligned PR 2 (#9006). The human_override declaration is removed from this PR, and the PLAN_NOT_FOUND change now lives in #9006. Refs #8733 Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> * fix(governance): apply review corrections to the follow-up errata Second revision of adcp#9015 after review. - policy-category-definition.json: the requires_human_review description no longer says that a policy can override the category's setting. A policy adds the requirement, and no policy removes a category's. - get-plan-audit-logs-response.json: the entries[].mode description no longer mentions a governed_actions[].mode field, which does not exist. - specification.mdx, audit-trail.mdx: state one strength for mode on audit entries, the one the schema has. A governance agent SHOULD record the mode on check entries, and outcome entries carry none. The specification said MAY and the audit-trail page said every entry. - overview.mdx: the image alt text no longer says budget is reserved. - specification.mdx, sync-plans-request.json: the reallocation table and the reallocation_threshold description point to reallocation_unlimited: true for unlimited reallocation, because a threshold equal to total tightens when total changes. - Changeset: cover the edits above. No schema property, enum member or plan_hash rule changes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> * fix(governance): say that mode is reported on the check_governance response The specification said that governance mode is "an internal implementation detail of the governance agent, not a protocol-level field". The check_governance response has carried an optional mode property since 3.1, the task page documents it, and the reference agent returns it. - specification.mdx: mode is configuration of the governance agent. No request carries it, and it is not a plan field. The agent MAY report the mode in force on the check_governance response. Callers still MUST NOT branch on it. - Changeset: cover the edit. No schema, wire field or agent behaviour changes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> * fix(governance): align the delegation expiry description with the plan-access rule The delegations[].expires_at description in sync-plans-request.json said that the governance agent denies an expired delegate's actions. Since adcp#9006 a caller with an expired delegation and no issued context gets PLAN_NOT_FOUND. - sync-plans-request.json: after expiration the agent is no longer a current delegate. Its intent re-check that presents a consultation_context issued for the plan, and an execution check on an action it proposed, are denied with a delegation_authority finding. Any other check from that agent returns PLAN_NOT_FOUND. - Changeset: cover the edit. Description only. No schema property or validation outcome changes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
Summary
PR 2 of 3 for #8733: the normative decisions, as confirmed in this comment. Where the specification contradicted itself or was silent, this PR states the rule the Governance working group chose and brings the reference agent into line for A1 and A4. It was revised after review, and follows the decision record update of 2026-11-15 for A4, C1 and C2. No wire field, enum member or
plan_hashrule changes.Refs #8733. PR 1 was #8931. PR 3 (B2, C4.1, C4.2) was #9007.
What changes
campaign/specification.mdx, training agentbudget.total(or a periodamount) below the amount already committed is rejected withVALIDATION_ERRORand leaves the plan unchanged. The training agent gets the sync-timebudget.totalcheck, next to the existing period check.campaign/tasks/check_governance.mdx,campaign/specification.mdx,campaign/audit-trail.mdx,enums/governance-mode.json, training agentapproved_sellersordelegationsevaluation for an authenticated caller is adeniedverdict with a finding and acheck_id, in every governance mode. A caller with no access to the plan getsPLAN_NOT_FOUND, the same response as for an unknownplan_id, and writes no audit entry.PERMISSION_DENIEDstays only for an unauthenticated caller or an identity that does not matchcaller. The audit-trail example of a rogue seller's call becomes an execution check. The three mode examples now use a budget finding.campaign/specification.mdx,enums/governance-mode.json,safety-model.mdx,index.mdx,tasks/get_plan_audit_logs.mdxhuman_review_required: trueescalate in every mode. The same sentences now name the A4 denial as the second case that mode does not change.campaign/tasks/sync_plans.mdx,governance/sync-plans-request.jsonplan_idis scoped per owner principal. Another principal'ssync_planswith the same ID creates or updates that principal's own plan, and never reads, changes or reveals the first one.campaign/specification.mdx,enums/delegation-authority.jsonexecute_onlydelegate acts only on such an action, which includesaccept_proposal. The governance agent MUST deny an intent check from such a delegate that would create a new binding or increase committed spend.campaign/specification.mdxconditionsat intent,deniedat execution.campaign/tasks/sync_plans.mdxREFERENCE_NOT_FOUND.campaign/tasks/check_governance.mdxconsultation_context.campaign/tasks/sync_plans.mdxhuman_review_required: falseon a plan that requires human review is rejected withVALIDATION_ERROR, as decided on #9007.All doc paths are under
docs/governance/.Notes for review
What the revision changed. The commit
1d11b5dapplies the review and the decision record update. It sits on a merge ofmain, whose only conflict was two adjacent rows of thesync_planserror table.human_review_required: trueescalate, and a failedapproved_sellersordelegationsevaluation isdenied. The mode summaries onindex.mdx,safety-model.mdx,tasks/get_plan_audit_logs.mdxand inenums/governance-mode.jsonsay the same.PLAN_NOT_FOUND, aserror-handling.mdx("Uniform response for inaccessible references") and thePLAN_NOT_FOUNDdefinition require. The specification sentence and thecheck_governanceerror row say so, and say that the attempt writes no audit entry. Thedeniedverdict with adelegation_authorityfinding is for a caller that can access the plan.PERMISSION_DENIEDon reuse.accept_proposaland words the prohibition as an obligation on the governance agent.sync_planserror table states theVALIDATION_ERRORfor an explicithuman_review_required: false.Second revision,
84cbf74, after this review and this comment:delegationsentry that has not expired, or that presents agovernance_contextorconsultation_contextthe governance agent issued for that plan. This is the definition given in review, and it is what the reference agent already does. Any other caller getsPLAN_NOT_FOUND.PERMISSION_DENIEDsentence no longer says "only". It now says that a faileddelegationsorapproved_sellersevaluation is neverPERMISSION_DENIED, and names the authentication cases as examples. The earlier wording excluded an expiredgovernance_context.authority,budget_limitormarketsof its delegation.index.mdx,safety-model.mdx,overview.mdxandenums/governance-mode.jsonno longer say that audit mode approves everything or never blocks.budget_limit.The mode examples in
audit-trail.mdxchanged. They showed an unapproved target seller that enforce mode denies and advisory and audit mode approve. That is no longer true, so the three examples now show a buy that asks for more than the plan's total budget. One sentence under them says which two outcomes do not move with mode.Reference agent after this change.
seller_complianceordelegation_authorityfinding is no longer approved in advisory or audit mode. That includes a delegate that exceeds itsbudget_limitormarkets. Four new test cases cover the two modes.check_governancereturns thePLAN_NOT_FOUNDerror and stores nothing. It returned adeniedverdict with aplan_lookupfinding and stored a check record. A test asserts that the two cases give byte-identical responses and that neither the owner's session nor the request's session holds a check. No compliance storyboard depended on the old shape.delegations[].authority, andtraining-agent.test.tshas anexecute_onlydelegate opening fresh intents that are approved. So the agent does not yet enforce the C2 sentence. That fixture is not changed here.Smaller points:
budget.totalunchanged on a plan that is already over-committed is accepted, because concurrent approvals can legitimately produce that state (specification.mdx:391). The check comparescommittedBudget, which only settled outcomes add to, so the two storyboards that cut a plan from 100,000 to 10,000 before any outcome is reported still pass.mode: "enforce". They are issued before the plan's mode is read.safety-model.mdx,index.mdxand the advisory wording repeat the losing sentence, so they change too.plan.regionsand never reads it, so the existing rule atspecification.mdx:487is already unimplemented there. Not changed here.error.field. It names thepolicy_idsarray parameter, aserror-handling.mdxrequires forREFERENCE_NOT_FOUNDon an array.Changeset
minor. The surface is experimentalgovernance.campaign, and a previously conformant agent can fail the new sentences. The changeset gives before and after for A1 and A4.Not settled here: the experimental-surface rule asks for six weeks of notice before a breaking change lands. A1 and A4 change behaviour. Earlier 3.3 changes recorded a waiver in the migration guide; this PR does not edit the release docs.
Verification
Run locally on Node 24.22.1:
npm run test:schemas,test:json-schema,test:examples,test:docs-nav,test:owned-links,test:additive-minor,test:mcp-schema-projection,test:compliance-snippets: passnode scripts/check-docs-mdx-syntax.mjs,node scripts/docs-json-field-audit.cjs: pass, no new findings/governanceis 59 clean with 183 steps,governance/budget_periods7 passed, and the fivemedia_buy_seller/governance_*scenarios pass with their required countsExisting tests changed, because they asserted the old error:
training-agent-governance-intent-binding.test.ts(non-audience execution caller) andtraining-agent-governance-agent-credentials.test.ts(two service-audience assertions). The identity-mismatch assertions beside them are unchanged.Changed in the revision: four new cases in
training-agent-governance-intent-binding.test.tsfor an unapproved target seller and an out-of-scope delegate in advisory and audit mode; the unknown-plan assertion intraining-agent.test.ts; and the no-access case intraining-agent-governance-caller-binding.test.ts, which now expectsPLAN_NOT_FOUND.The second revision,
84cbf74, changes no training-agent or compliance file. The pre-commit hook passed on it. The pre-push hook did not rerun the storyboard matrix for it, so the matrix result above is from1d11b5d.🤖 Generated with Claude Code