codex-orchestrator 2.0.15 → 2.0.17

package/internal-workflow/skills/to-tickets/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: to-tickets
-description: Compile Plan's contextual PRD draft or approved product authority into the smallest executable ticket packet, run one fresh semantic packet review, obtain one approval, publish serially, verify authoritative tracker read-back, and stop before implementation.
+description: Compile Plan's contextual PRD draft or approved product authority into the smallest executable ticket packet, run fresh semantic review per settled revision, obtain one approval, publish serially, verify tracker read-back, and stop before implementation.
 ---
 
 # To Tickets
@@ -15,9 +15,8 @@ ticket contains only its local observable outcome, scope, blockers, proof, and
 Parent link. The approved ticket body plus Parent PRD is the final execution
 authority; no later planning pass is inserted.
 
-`$to-tickets` owns one fresh semantic reviewer, one explicit user approval,
-serialized publication, deterministic reconciliation, and authoritative
-tracker read-back for the complete packet.
+`$to-tickets` owns one fresh reviewer per settled revision, one explicit user
+approval, serialized publication, reconciliation, and tracker read-back.
 
 The primary route remains `planning-only` while compiling or publishing these
 artifacts. `ticket-graph` begins only after separate delivery authorization;
@@ -100,6 +99,14 @@ them exact. Acceptance criteria describe observable outcomes. Freeze an
 internal mechanism only when a public contract, ownership boundary, safety
 requirement, or inter-ticket dependency depends on it.
 
+Add `Binding implementation decisions` only when two locally valid
+implementations could conflict at a ticket edge, public contract,
+identity/cardinality, persisted or in-flight state, concurrency, shared
+behavior, or Parent proof. Record only confirmed boundary semantics; leave
+private structure to Implement. Omit the section when Parent and existing seams
+already determine the behavior. Return an unconfirmed product choice as a
+Decision Delta and missing technical evidence as blocking discovery.
+
 Keep unresolved product decisions with the user and unresolved external or
 technical discovery in a blocking AFK/HITL ticket. If the generated ticket
 would need a later planning artifact, it is not ready to publish.
@@ -130,14 +137,19 @@ ticket body, plus:
 This packet is contextual planning data, not a separately published artifact or
 workflow owner. Do not publish any intermediate PRD.
 
+Freeze it only after product decisions, ticket boundaries, and the deletion
+challenge are complete, with every unresolved technical or external gap made
+an explicit blocking discovery outcome. The expected review verdict must be
+`APPROVE`; never use review to finish inventory, decisions, or composition.
+
 ### 4. Run one fresh semantic packet review
 
-For the settled packet, launch exactly one fresh `standards_reviewer` without a
-history fork (`fork_context=false` on V1; `fork_turns="none"` on V2). Begin its
-brief with `Assigned role: standards_reviewer`, provide the full source authority
-and complete publish-ready packet, and require all three internal lenses in the
-same review activation: source fidelity, ticket executability, and minimum
-solution and scope conservation.
+For each settled revision composed by `$to-tickets` before approval or
+publication, launch exactly one fresh `standards_reviewer` without history
+(`fork_context=false` on V1; `fork_turns="none"` on V2). Begin with
+`Assigned role: standards_reviewer`, provide the full authority and packet, and
+apply all three lenses in the same activation: source fidelity, ticket
+executability, and minimum solution and scope conservation.
 
 - **Source fidelity:** every product claim, decision, scope boundary,
   consequence, and blocker is grounded in the source; no behavior or ownership
@@ -145,7 +157,9 @@ solution and scope conservation.
 - **Ticket executability:** every ticket is a cohesive vertical outcome that a
   fresh implementation context can execute and prove from its body plus the
   Parent PRD; dependencies are real, proof observes the claim, and no later
-  planning artifact is required.
+  planning artifact is required. Ask whether the implementer can execute it
+  without independently choosing product behavior or incompatible system
+  semantics; require confirmed binding decisions only when the answer is no.
 - **Minimum solution and scope conservation:** reject Speculative Generality,
   Middle Man indirection, unsupported new or duplicate owners, proof-only
   runtime machinery, acceptance criteria without Parent/user/repository
@@ -153,12 +167,37 @@ solution and scope conservation.
   Prefer deletion or narrowing before adding machinery. The reviewer may not
   introduce a new runtime mechanism merely to make proof more exhaustive.
 
+Require public contract detail only when a public boundary, inter-ticket
+contract, or existing invariant requires it. State proof gaps as observable
+obligations; never prescribe a private DTO, schema, service, rule engine, or
+exhaustive internal matrix to close them.
+
+Before verdict, traverse each applicable seam:
+
+- Parent obligation to owning ticket and proof;
+- producer and consumer across each edge: data, identity/cardinality, state,
+  timing, and failure semantics;
+- changed persistence or namespace to existing and in-flight state;
+- reused behavior to equal semantics or an explicit delta;
+- UI promise to its API or DTO source;
+- child-added choice to Parent, user, or repository authority.
+
+Return concise evidence for applicable seams and omit `N/A` seams. This is
+review output, not another packet artifact.
+
 The reviewer is read-only and returns `APPROVE` or `NEEDS_WORK` with exact
 source/artifact evidence. Capture a non-empty fresh child identity and wait for
 that same child. Do not split the lenses across children and do not launch the
-delivery Standards review. A missing, failed, or non-approving review
-blocks approval and publication. If repairs change a product decision, return
-that decision to the user rather than resolving it in review.
+delivery Standards review. After `NEEDS_WORK`, repair the packet and launch a
+new reviewer without prior review history over the complete revised packet. A
+same-context or findings-only recheck cannot approve. Repeat until the latest
+revision receives one full approval. Missing or failed review blocks
+publication; product-changing repairs return to the user.
+
+Before re-review, repair every verified blocker and sweep the complete packet
+for the same defect class. If a blocker exposes unfinished inventory,
+authority, or composition, return to drafting and pass the freeze conditions
+again.
 
 ### 5. Get one explicit approval
 

package/internal-workflow/skills/to-tickets/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "To Tickets"
   short_description: "Review and publish executable ticket packets"
-  default_prompt: "Use $to-tickets to compile the complete executable packet, run one fresh semantic review, obtain one approval, publish serially, verify tracker read-back, and stop before implementation."
+  default_prompt: "Use $to-tickets to compile and fully review each settled packet revision, obtain one approval, publish serially, verify tracker read-back, and stop before implementation."
 policy:
   allow_implicit_invocation: false

package/internal-workflow/skills/to-tickets/evals/evals.json

@@ -15,9 +15,9 @@
       "forbidden": ["create duplicate durable planning state", "request another execution artifact"]
     },
     {
-      "id": "complete-packet-gets-one-semantic-review",
+      "id": "complete-packet-gets-one-review-per-revision",
       "prompt": "Prepare a publish-ready PRD and three-ticket graph after Plan passed the combined PRD draft to you.",
-      "expected": ["launch one fresh semantic reviewer over the complete packet", "apply source fidelity and ticket executability in that one review", "obtain one explicit user approval before publication"],
+      "expected": ["launch one fresh semantic reviewer over each settled complete revision", "apply source fidelity and ticket executability in that review", "obtain one explicit user approval before publication"],
       "forbidden": ["review the intermediate PRD", "launch separate workflow owners for each lens", "request multiple approvals"]
     },
     {
@@ -86,6 +86,48 @@
       "prompt": "Semantic packet review finds that one acceptance claim is stronger than the available proof seam.",
       "expected": ["narrow the unsupported claim or use an existing proof seam", "prefer deletion or narrowing before adding machinery"],
       "forbidden": ["automatically add middleware", "introduce a framework or runtime mechanism merely for exhaustive proof", "invent custom tooling"]
+    },
+    {
+      "id": "simple-ticket-omits-binding-decisions",
+      "prompt": "Compile a small approved change whose behavior is already determined by the Parent and an existing public seam.",
+      "expected": ["publish the ordinary short executable ticket", "omit Binding implementation decisions"],
+      "forbidden": ["add a technical contract", "request another approval", "freeze implementation details"]
+    },
+    {
+      "id": "incompatible-semantics-get-minimal-binding-decisions",
+      "prompt": "Two implementations could both satisfy the observable criteria but choose incompatible persisted identity and failure semantics across existing runtime paths.",
+      "expected": ["record only the necessary confirmed choices under Binding implementation decisions", "state the material implementation freedom"],
+      "forbidden": ["write a file-by-file implementation plan", "guess an unconfirmed choice", "add another workflow stage"]
+    },
+    {
+      "id": "repair-gets-fresh-full-packet-review",
+      "prompt": "Semantic review returns NEEDS_WORK and the packet is repaired before user approval.",
+      "expected": ["repair every verified blocker and sweep the complete packet for the same defect class", "pass the settled packet conditions again", "launch one new reviewer without prior review history", "review the complete revised packet", "keep one user approval for the final packet"],
+      "forbidden": ["continue with the same reviewer", "approve from a findings-only recheck", "launch multiple reviewers for one revision"]
+    },
+    {
+      "id": "private-design-stays-with-implement",
+      "prompt": "A settled packet defines observable behavior through existing owners, but semantic review proposes a new DTO, service, exhaustive internal rule matrix, and another ticket for implementation convenience.",
+      "expected": ["require public contract detail only when demanded by a public boundary, inter-ticket contract, or existing invariant", "state an observable proof gap without prescribing private runtime design", "leave safe private implementation choices to Implement"],
+      "forbidden": ["use credible proof alone to require machinery", "add a private service, rule engine, or internal matrix"]
+    },
+    {
+      "id": "cross-ticket-cardinality-must-match",
+      "prompt": "A predecessor emits multiple identities while its successor contract consumes one identity.",
+      "expected": ["reject the packet until producer and consumer cardinality agree through a minimal confirmed binding decision"],
+      "forbidden": ["approve each ticket in isolation", "choose a product outcome without authority"]
+    },
+    {
+      "id": "changed-state-covers-existing-work",
+      "prompt": "A ticket changes a persisted state contract under an existing namespace while records and jobs may already exist.",
+      "expected": ["require an explicit outcome for existing state and in-flight work"],
+      "forbidden": ["invent a compatibility layer", "approve only the new-state happy path"]
+    },
+    {
+      "id": "ui-promise-requires-data-contract",
+      "prompt": "UI acceptance requires dynamic content absent from the proposed API or DTO.",
+      "expected": ["reject the packet until an existing or authorized data contract supplies the content"],
+      "forbidden": ["let the UI invent the data", "add a new owner solely for proof"]
     }
   ]
 }

package/internal-workflow/skills/to-tickets/references/publishing-details.md

@@ -9,6 +9,8 @@ planning artifact. Keep the body local and include only:
 - the local observable outcome and acceptance criteria;
 - in-scope and out-of-scope boundaries;
 - confirmed owner/public seam when needed for zero-guess execution;
+- only the confirmed binding implementation decisions needed to prevent
+  incompatible interpretations, when any exist;
 - blockers and native dependency links;
 - behavior proof or exact non-TDD proof;
 - AFK/HITL final state, which never authorizes implementation.
@@ -60,6 +62,16 @@ AFK: `ready-for-agent`; HITL: `ready-for-human`. State does not authorize
 implementation.
 ```
 
+Only when the Parent and existing seam leave incompatible implementations
+possible, insert this short section after `Owner and proof seam`:
+
+```md
+## Binding implementation decisions
+
+- <only confirmed choice required to prevent incompatible implementations>
+- Implementation freedom: <material choices intentionally left to Implement>
+```
+
 ## Serialized publication
 
 The single current root publisher performs every tracker effect. Do not create
@@ -100,8 +112,8 @@ approved packet. Verify:
 - exact AFK/HITL labels and absence of conflicting planning or delivery state;
 - absence of duplicate or partial packets;
 - every ticket remains executable from its body plus the Parent PRD;
-- the publication event trace contains one semantic review, one approval, and
-  no implementation action.
+- the final packet revision has fresh full semantic approval, one user approval,
+  and no implementation action.
 
 Post-publication read-back is integrity proof. It does not launch another
 semantic reviewer and does not request another approval.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codex-orchestrator",
-  "version": "2.0.15",
+  "version": "2.0.17",
   "description": "Reusable GitHub Issues runner for Codex.",
   "type": "module",
   "license": "MIT",