feat(intake): add human_review — the declaration-only work-type

agents/conductors/intake/AGENTS.md

@@ -34,6 +34,7 @@ must not duplicate:
 |-------|------|-------------------|
 | `/route` | infer work-type, **dispatch** (starts dev now) | intake **files a prompt**, defers |
 | `triage/` | the unclassified-prompt bucket | intake **writes into** it on low confidence — reuse, don't reinvent |
+| `human_review/` | shipped work a human asked to sign off | intake files it **only when declared**; it is never inferred, never demoted to `triage/`, and never counted as backlog |
 | `create_issue` | prompt → GitHub issue + registry | intake runs strictly **before** it |
 
 ## What it decides (the IntakeDecision)
@@ -48,7 +49,12 @@ Proposed path · Header (the block written verbatim) · Risks · Next action
 
 - **Work-type** — scored by keyword signal into `feature/bug/refactor/docs/test/
   release/maintenance/research/experiment`; a tie across dissimilar types or zero
-  signal → `triage`.
+  signal → `triage`. `human_review` is the one type the classifier may never
+  select: it means *a human has decided this shipped work needs their eyes*,
+  which no keyword carries, so only an explicit `Type: human review` declaration
+  reaches it (`MANUAL_ONLY_WORK_TYPES` in the sizing faculty). Nothing files a
+  review automatically and no workflow requires one — review is opt-in, not a
+  lifecycle stage.
 - **Target** — resolved from `@RepoName` mentions *and* bare repo names, incl. the
   **organism repos** (`pyautomind/pyautobrain/pyautoheart/pyautobuild/pyautomemory`)
   — the gap that made the Feature Agent mis-route a `pyautobrain` target.
@@ -100,7 +106,7 @@ the page's Hygiene section), and an un-themed prompt falls back to `Target:`.
 | **classify** | `intake "<text>"` / `intake classify --file P` | classify one raw input; `--apply` writes the prompt |
 | **ideas** | `intake ideas` | scan `ideas.md`, propose one prompt per bullet; `--apply` writes them |
 | **census** | `intake census` | inventory every filed prompt (work-type/target/difficulty/status + hygiene flags); always read-only |
-| **dashboard** | `intake dashboard` | render the census as the Mind **task** page — picks, in flight, parked, planned, backlog, bundles, recent, epics; `--apply` writes `PyAutoMind/dashboard.md`, `--check` exits 1 on drift |
+| **dashboard** | `intake dashboard` | render the census as the Mind **task** page — picks, in flight, human review, parked, planned, backlog, bundles, recent, epics; `--apply` writes `PyAutoMind/dashboard.md`, `--check` exits 1 on drift |
 | **formalise** | `intake formalise [prefix]` | retroactively header the prompts census flags — derive the missing fields, insert in place, prose untouched; `--apply` writes |
 | **reconcile** | `intake reconcile [prefix]` | rank backlog prompts that look already-shipped (vs the `complete/` records / `active/`), and pair live prompts that look like the same work filed twice; always read-only — retiring stays human |
 | **reconcile --repo** | `intake reconcile --repo <target> [prefix]` | **also** read the target repo's source: identifiers the prompts name that exist upstream, and lines they quote that are **gone** — the two signals that see a prompt with no Mind-side trace. Opt-in; the default path is offline |

agents/conductors/intake/INTAKE_TAXONOMY.md

@@ -25,6 +25,12 @@ prefix match) and picks the strongest:
 | `experiment` | spike, proof of concept, prototype, sandbox |
 | `feature` | add, implement, support, introduce, enable, extend (the default verb set) |
 
+`human_review` is deliberately absent from that table, and is filtered out of
+the signal sets rather than merely left out of them. It means *a human has
+decided this shipped work needs their eyes* — a judgement no keyword carries —
+so it is reachable only through an explicit declaration and can never be
+inferred. See §7 below.
+
 **Confidence & triage.** `high` when the winner has ≥2 signals, `medium` at 1.
 A tie across *dissimilar* non-feature types, or zero signal, is genuinely
 ambiguous → `triage` (low confidence), written to `triage/<slug>.md` for a human
@@ -134,7 +140,41 @@ suspects high/medium/low with the evidence shown. Always read-only: the final
 verification (target repo git log / merged PRs) and the retirement to
 `issued/` stay human.
 
-## 7. What intake does NOT own
+## 7. Human review — the one type only a human may set
+
+`human_review/` holds prompts about work that has already **shipped**: a
+completed task a human wants to read and sign off before calling it done. It is
+a work-type like any other in the layout (`draft/human_review/<target>/<slug>.md`,
+same light header) and unlike any other in three ways:
+
+- **Never inferred.** It is in `MANUAL_ONLY_WORK_TYPES` (sizing faculty), so the
+  classifier skips it entirely. The only route in is a declaration —
+  `Type: human review`, `human-review` or `human_review`, in a header line or
+  mid-sentence; `norm_work_type` folds all three to the folder key. Prose *about*
+  reviewing shipped work classifies as ordinary work, as it should.
+- **Never demoted.** An unresolved target sends an ordinary prompt to `triage/`
+  ("nobody classified this"). Here somebody did, and a review's subject is
+  shipped work whose repo may live in a completion record rather than in an
+  `@RepoName` the body repeats — so it files flat at
+  `draft/human_review/<slug>.md` instead.
+- **Never drift.** For every other draft prompt a merged PR in the body, or a
+  `Status:` saying shipped, means the lifecycle stalled. For a review both are
+  the premise, so the drift checks skip it.
+
+Nothing puts a task here automatically and no workflow requires it: review is
+opt-in, not a lifecycle stage. An empty section means nothing has been flagged,
+not that nothing shipped.
+
+**On the dashboard** it is its own top-level section, directly under *In flight*
+— both are live obligations, and a review sunk below a 140-prompt backlog would
+never be read. It is kept out of `census()["records"]` entirely, which keeps it
+out of the backlog count, the pick lists, the work-type sections, the bundler and
+the epics in one move. Its 📋 hands out a read-and-report prompt rather than a
+`/start_dev` (there is nothing to start) that ends by naming both exits — sign
+off and retire the prompt, or file the follow-up with `/intake` — because a
+review that stops at "looks fine" leaves the row on the board forever.
+
+## 8. What intake does NOT own
 
 - The difficulty heuristic, prompt parsing, repo sets → the **sizing faculty**.
 - The work-type → agent routing → `PyAutoMind/ROUTING.md` + the conductors.