diff --git a/docs/upstream/aihero-course.md b/docs/upstream/aihero-course.md index fa9fa2c4e..ad34080bc 100644 --- a/docs/upstream/aihero-course.md +++ b/docs/upstream/aihero-course.md @@ -17,8 +17,9 @@ Vetting ran as six lanes on the `pocock-course-lanes` contract steering lanes ([#2909](https://github.com/melodic-software/claude-code-plugins/issues/2909), [#2910](https://github.com/melodic-software/claude-code-plugins/issues/2910), [#2911](https://github.com/melodic-software/claude-code-plugins/issues/2911)). Lanes decide but -never implement: every ADOPT row points at a filed work item; changes execute through the normal -pipeline. +never implement: every plain ADOPT row points at a filed work item (the ADOPT terms sub-form, +defined with the row schema below, executes via the lane 6 term-verdict table instead); changes +execute through the normal pipeline. **Record provenance and supersession.** Lanes 3–5 were each closed by two parallel session chains — one writing rows into this document's copy on the contract branch @@ -45,7 +46,9 @@ the basis it was derived against, the as-of date, and a recheck trigger. Columns | Lesson claim | Ours | Verdict | Reasoning, basis, as-of | Verdicts: **ADOPT** (with the filed work item), **REJECT** (with reason), **TRACK** (on a named -event), **COVERED** (already present at parity or stronger, with evidence). +event), **COVERED** (already present at parity or stronger, with evidence). One sub-form: +**ADOPT terms** marks a vocabulary adoption whose execution path is the term-verdict table +(lane 6) rather than a filed work item. **Recheck trigger, all rows (fixed at creation):** divergence at re-fetch. A read-time re-fetch of the row's basis (the lesson text, or the upstream artifact the row cites) finds the source no @@ -189,9 +192,9 @@ regression-gating the lock/auto-guard defenses — which stands. | Lesson claim | Ours | Verdict | Reasoning, basis, as-of | |---|---|---|---| -| Agents rush from prompt to implementation with no alignment step (the asset rush) | `planning:interview` is the pipeline's pre-clarity stage: a contract locked before exploration, planning, or execution | **ADOPT** (convergent) | The critique is the design rationale the pipeline already embodies; behavior-change work is interview-first by default, with auto-detect keeping it cheap. As-of 2026-08-17 | +| Agents rush from prompt to implementation with no alignment step (the asset rush) | `planning:interview` is the pipeline's pre-clarity stage: a contract locked before exploration, planning, or execution | **COVERED** (convergent) | The critique is the design rationale the pipeline already embodies; behavior-change work is interview-first by default, with auto-detect keeping it cheap. No change needed, so no work item (reclassified from ADOPT per schema, PR #3008 review). As-of 2026-08-17 | | Plan mode is a buffer between exploration and implementation: a plan document you review, then proceed | Plan mode is treated as a permission/safety gate; alignment is owned upstream by the interview contract | **COVERED** (position) | The harness feature is not rejected, it is repositioned: a permission mode cannot produce shared understanding, and nothing in the family asks it to. As-of 2026-08-17 | -| "Plan mode is still rushing": the plan reads like the implementation; the decisions are already made and written down | Two structural answers: the Brief locks intent BEFORE `/planning:plan`, and the plan skill itself refuses to lock decisions inline (Open Decisions surfaced before the plan body; a confidence gate routes judgment calls back to interview rounds; user approval gate before any code; fresh-context devils-advocate stress-test) | **ADOPT** (convergent, already answered structurally) | His diagnosis names exactly the failure the pipeline's two gates exist to prevent; no change needed. As-of 2026-08-17 | +| "Plan mode is still rushing": the plan reads like the implementation; the decisions are already made and written down | Two structural answers: the Brief locks intent BEFORE `/planning:plan`, and the plan skill itself refuses to lock decisions inline (Open Decisions surfaced before the plan body; a confidence gate routes judgment calls back to interview rounds; user approval gate before any code; fresh-context devils-advocate stress-test) | **COVERED** (convergent, already answered structurally) | His diagnosis names exactly the failure the pipeline's two gates exist to prevent; no change needed, so no work item (reclassified from ADOPT per schema, PR #3008 review). As-of 2026-08-17 | | Root cause is the sycophantic trait: told to produce, the agent produces | Interview stance: recommendations-first but facts-are-mine/decisions-are-the-user's; the auto-guard forbids resolving a genuine user choice; `/planning:audit-answers` is the producer-not-critic control | **COVERED** | The trait is countered by structure, not exhortation; "sycophancy" and "asset rush" go to lane 6 as term candidates. As-of 2026-08-17 | | The design concept (Brooks): shared understanding is not an asset; conversation sharpens it | The general-domain interview terminal drives to a shared understanding and STOPS: no Brief, no artifact, no pipeline handoff | **COVERED** (embodied) | Q33: the endpoint already exists; the term maps to our "shared understanding" (lane-6 adoption candidate); no mechanism change earned. As-of 2026-08-17 | | Walkthrough: cycle `shift+tab` until plan mode is on | Verified harness behavior | **CONFIRMED** | Verdict C7 (two-pool); no fixed press count is taught. As-of 2026-08-17 | diff --git a/docs/upstream/mattpocock-skills-v12-map.md b/docs/upstream/mattpocock-skills-v12-map.md index d4739f0eb..4173c218d 100644 --- a/docs/upstream/mattpocock-skills-v12-map.md +++ b/docs/upstream/mattpocock-skills-v12-map.md @@ -1,5 +1,14 @@ # Full map — mattpocock/skills (v1.2.3, HEAD 84fdeff, 2026-08-06) ↔ melodic-software/claude-code-plugins (main a89a4a33) +> **Correction (2026-08-18, lane 6):** this map's "Claude-only private marketplace" +> characterization (Codex-sidecar row and the marketplace row of the cross-cutting table) was +> stale; the repository is PUBLIC (verified via the repo listing on 2026-08-17 during the +> pocock-course-lanes effort; the row-level notes below carry the same 2026-08-18 correction +> date). The Claude-only part stands; the private part does not. The map is otherwise a +> point-in-time record as of its header date. A separate 2026-08-17 recheck at HEAD `068b6e0` +> found the 35-skill inventory and all relation rows structurally intact (record: +> `aihero-course.md` effort + the pocock-course-lanes contract). + Sources: his-repo full inventory (35 skills, every SKILL.md read), our-repo provenance sweep (git log + grep + docs), v1.2.0 release notes, v1.2 changelog article, video transcript. As-of 2026-08-08. Legend — **Relation**: DERIVED (attributed port), PARTIAL (specific ideas taken, attributed), CONVERGENT (same territory, no provenance), NONE (no counterpart). **v1.2+ delta**: what changed upstream since our port / what's new.