@webpresso/agent-kit 3.3.1 → 3.3.3

package/catalog/AGENTS.md.tpl

@@ -10,14 +10,12 @@ Prefer repo-local instructions when more specific.
 wp install && wp run setup:agent && wp sync  # setup and sync are separate, idempotent steps
 ```
 
-agent-kit's catalog is the single source of truth for generated agent surfaces.
-Agent-kit owns generated agent surfaces here; Webpresso CLI owns the end-user command surface.
+agent-kit catalog is SSOT for generated agent surfaces; Webpresso CLI owns the end-user command surface.
 
-- Optional agent tools can be WP-owned with `wp install codex|claude-code|opencode` or `wp install oh-my opencode`; `openagent` aliases the latter, and WP-owned scopes use `wp update`.
+- Optional agent tools: `wp install codex|claude-code|opencode` or `wp install oh-my opencode` (`openagent` aliases the latter); WP-owned scopes use `wp update`.
 - `wp setup` repairs the managed `.gitignore` block for regenerated surfaces.
-- Consumer repos use global `wp`, keep only `@webpresso/agent-config` locally, and do not add local `@webpresso/agent-kit`.
+- Consumers use global `wp` + local `@webpresso/agent-config` only (no local `@webpresso/agent-kit`).
 - Track repo-owned instruction sources (`AGENTS.md`, `agent-rules/`, `agent-skills/`); ignore generated/runtime surfaces (`.agent/`, `.agents/`, `.codex/`, `.opencode/`, etc.).
-
 - Keep the generated default `AGENTS.md` under 8 KB.
 
 Codex routing instruction surface:

package/catalog/agent/rules/pre-implementation.md

@@ -227,16 +227,23 @@ Ultragoal events, or approval frontmatter.
   killed writer is never overwritten. A complete content rejection is terminal
   for that unchanged target, while active provider limits/unavailability remain
   suppressed until their recorded retry window permits another attempt.
-- **Completed rejections have a two-subject circuit breaker.** Plan review
+- **Rejections have a convergence-aware subject budget (4).** Plan review
   lineage uses the canonical `wp-blueprint-v1` subject, so lifecycle/trust
   bookkeeping commits reuse the prior rejection and do not invoke another
-  reviewer. One semantically changed remediation subject may receive the final
-  formal review; its prompt includes the prior rejection artifact and requires
-  all remaining blockers together. After two distinct semantic subjects are
-  rejected, `wp review gate` returns `scope-exhausted` before provider
-  invocation. Stop automated retries and obtain an explicit human scope
-  decision before creating a successor blueprint. Plan and delivery purposes
-  have separate budgets; delivery retains its existing whole-tree subject.
+  reviewer. An identical resubmission of an already-rejected subject is likewise
+  reused for free (no reviewer call). Each genuinely different converging
+  subject — one that addressed the prior findings — is reviewed on its own, and
+  its prompt includes the prior rejection artifacts. After **four** distinct
+  semantic subjects are rejected, `wp review gate` returns `scope-exhausted`
+  before provider invocation. This is an anti-runaway cap, not a hard fork
+  point: automated runs must stop and escalate, but a human who judges the
+  attempt to be good-faith convergence may extend the budget **in place** by
+  setting `WP_REVIEW_SCOPE_CONTINUE=1` (recorded as an audit gate-issue) and keep
+  reviewing THIS blueprint — creating a successor blueprint only launders the
+  counter and is not required. The approval gate itself is unchanged: the
+  reviewer's verdict is still required under scope-continue. Plan and delivery
+  purposes have separate budgets; delivery retains its existing whole-tree
+  subject.
 - **`provider-event-too-large` means one captured provider JSON event exceeded
   the bounded event limit; it is not a reviewer verdict.** Keep the bound in
   place, retry the same gate id, and switch provider/model when possible because

package/catalog/agent/skills/ultragoal/SKILL.md

@@ -109,24 +109,11 @@ Resolve models from the live catalog every time:
 opencode models opencode-go
 ```
 
-Pick by purpose (same intent as the committed OpenCode reviewer policy / `opencode-go` skill):
-
-| Purpose                                     | Preferred family (live catalog)                                                                  | Why                        |
-| ------------------------------------------- | ------------------------------------------------------------------------------------------------ | -------------------------- |
-| Plan / architecture critique                | DeepSeek **Pro** (`opencode-go/deepseek…-pro`)                                                   | Deeper plan review depth   |
-| Coding / implementation quality             | Kimi **Code** first (`…kimi…-code`)                                                              | Coding quality lane        |
-| Cheap recheck / high-availability follow-up | DeepSeek **Flash** (`…deepseek…-flash`)                                                          | Fast secondary pass        |
-| Aggregate / unspecified                     | Follow `opencode-go` skill order: Kimi Code → DeepSeek Pro → DeepSeek Flash → remaining families | Default outside-voice lane |
-
-Resolve the actual model ID with catalog greps (never paste a fixed model string into agent instructions as the only option):
-
-```bash
-CATALOG=$(opencode models opencode-go)
-# plan critique example:
-MODEL=$(echo "$CATALOG" | grep '^opencode-go/deepseek' | grep -- '-pro$' | sort -V | tail -1)
-# coding example:
-# MODEL=$(echo "$CATALOG" | grep '^opencode-go/kimi' | grep -- '-code$' | sort -V | tail -1)
-```
+**Selection order and the purpose→family mapping live in one place: the committed
+reviewer policy (`packages/workflow-skills/src/opencode-reviewer-policy.ts`),
+rendered into the `opencode-go` skill and drift-gated against it.** Follow that
+skill; do not restate the ordering here. A second copy of a preference list is a
+second thing to update, and the copy that goes stale is the one an agent reads.
 
 Treat a provider-wide usage limit as terminal for every OpenCode Go model in the current review window; do not hop families to evade a usage block on the same provider.
 

package/catalog/agent/skills/verify/SKILL.md

@@ -35,18 +35,18 @@ the preferred skill (`claude` / `codex` / `grok` / one OpenCode Go), then stops.
 
 ## Tooling map
 
-| Gate                      | Prefer                                                          |
-| ------------------------- | --------------------------------------------------------------- |
-| Lint / types / tests / QA | `wp_lint`, `wp_typecheck`, `wp_test`, `wp_qa`                   |
-| Format                    | `wp_format`                                                     |
-| Audits                    | `wp_audit`, `wp_audits`                                         |
-| Blueprint                 | blueprint MCP tools + `wp audit blueprint-lifecycle`            |
-| PR / CI wait              | `wp_pr_status`, `wp_pr_wait` (not multi-sleep loops)            |
-| Release                   | `wp_release_readiness` when shipping                            |
-| Large output              | `wp_session_execute` / `batch_execute` / `execute_file`         |
-| Continuity                | `wp_session_capture` for verify summary; avoid paste-back walls |
-| E2E / act                 | `wp_e2e`, `wp_ci_act` when the change requires them             |
-| Outside authority         | `wp review gate` (not raw multi-agent review swarms)            |
+| Gate                      | Prefer                                                                                  |
+| ------------------------- | --------------------------------------------------------------------------------------- |
+| Lint / types / tests / QA | `wp_lint`, `wp_typecheck`, `wp_test`, `wp_qa`                                           |
+| Format                    | `wp_format`                                                                             |
+| Audits                    | `wp_audit`, `wp_audits`                                                                 |
+| Blueprint                 | blueprint MCP tools + `wp audit blueprint-lifecycle`                                    |
+| PR / CI wait              | `wp_pr_status`, `wp_pr_wait` (not multi-sleep loops)                                    |
+| Release                   | `wp_release_progress` for live Release phases; `wp_release_readiness` for package gates |
+| Large output              | `wp_session_execute` / `batch_execute` / `execute_file`                                 |
+| Continuity                | `wp_session_capture` for verify summary; avoid paste-back walls                         |
+| E2E / act                 | `wp_e2e`, `wp_ci_act` when the change requires them                                     |
+| Outside authority         | `wp review gate` (not raw multi-agent review swarms)                                    |
 
 ## Iron law — evidence before claims