docs(learning-tier): owner decisions 1–5 recorded (closes #34)

docs/architecture/learning-tier/plan-2026-09-19.md

@@ -9,9 +9,10 @@ outstanding:** §7. **Base:** `main` @ `4f8e3e0f`.
 **Item 1 — the mentor writes.** Today the mentor agent can *read* learning state on every harness but records
 nothing: no MCP teach-back writer exists (CLI only), kiro pre-approves one writer, claude grants none, the
 persona's only concrete "record progress" line points at a different tool. Result: zero writer calls in
-143,973 historical messages. This item adds the missing writer, grants the additive writers on the harnesses
-that have an in-repo grant grammar, gives every harness one machine-readable trigger table, and proves the
-pipe is open with a scripted learner on all six installed harnesses — claiming only what was observed.
+143,973 historical messages. This item adds the missing writer, names the additive writers in every harness
+definition (no new pre-approvals — owner decision 2, §7), gives every harness one machine-readable trigger
+table, and proves the pipe is open with a scripted learner on all six installed harnesses — claiming only what
+was observed.
 
 **Item 2 — one honest number for Jev.** A single pre-registered, directional measurement of TypeSafe's Jev
 as a *filter* over harness-proposed struggle candidates, on the repo's 13-session human-labelled gold read
@@ -25,7 +26,9 @@ Jev call before the pre-registration commit.
 | 1 | `feat/learning-tier-fed` | `~/code/personal/tools/studyloop-wt/tier` | own PR, mergeable independently |
 | 2 | `feat/jev-struggle-eval` | `~/code/personal/tools/studyloop-wt/jev-eval` | own PR (eval code + receipts only) |
 
-Both cut from `main` @ `4f8e3e0f`. No code dependency either way (item 2 scores frozen gold). Each stage:
+Item 1's branch is cut from `main` when S1-0 starts; item 2's is **not cut until milestone 0.5.x is closed**
+(owner decision 1, 2026-09-23 — item 1 is MVP work, item 2 a post-MVP measurement). No code dependency either
+way (item 2 scores frozen gold). Each stage:
 commit in coherent steps with why-bodies; full suite diffed against a clean control worktree (only the RED
 tests may flip); push only when the owner says so. The `feat/jev-judge` planning branch (Stage 0 receipt,
 this plan) merges first or is cherry-picked — owner's call (§7).
@@ -36,21 +39,28 @@ this plan) merges first or is cherry-picked — owner's call (§7).
 Record, from source, the grant mechanism per harness and freeze the writer sets:
 
 - `W_auto` = `log_topic`, `log_struggle`, `record_teachback` (new), `record_plan_learning` — additive, learner-
-  agreed or learner-stated. Pre-approved.
+  agreed or learner-stated. **Named in every harness definition; prompt-per-call everywhere** (owner decision 2,
+  §7). No new pre-approval on any harness this branch; kiro's existing `log_topic` allowlist entry stays as the
+  one recorded asymmetry (`execute_bash` is pre-approved there, so removing it buys no parity).
 - `W_srs` = `record_study_progress`, `log_review_outcome`, `record_topic_progress` — SRS mutators that accept
   an unverified `card_hash`/topic id (`mcp/tools.py:114,648`). **Stay prompt-per-call** this branch.
-- Grant grammar: kiro `allowedTools` (`agents/kiro/study-mentor.json`); claude `settings.json` permissions +
-  tool names in `socratic-mentor.md`; opencode already `studyloop *` (recorded as not least-privilege, unchanged);
+- Grant grammar: kiro `allowedTools` (`agents/kiro/study-mentor.json`, unchanged); claude `settings.json`
+  permissions (unchanged — no block) + tool names in `socratic-mentor.md` (**must name each `W_auto` tool**:
+  its `tools:` line lists `Read, Write, Grep, Bash` and no `mcp__studyloop__*` tool today, so naming is what
+  makes the writers reachable at all — naming is instruction, not approval); opencode already `studyloop *`
+  (recorded as not least-privilege, unchanged);
   codex/pi/grok have **no in-repo grant grammar** — canonical persona via session-dir `AGENTS.md`, approval is
-  harness-side (`adapters/{codex,pi,grok}.py`). No `agents/grok/` is created.
+  harness-side (`adapters/{codex,pi,grok}.py`). No `agents/grok/` is created. Parity is of the *instruction*
+  (trigger table + `W_auto` set, pinned by test), never of approval — the receipt records each harness's
+  approval cell in the capability matrix as it is.
 
 Finish: `docs/architecture/learning-tier/receipts/s1-0-capability-lock.md` committed with file:line per harness.
 
 ### S1-RED (1 day)
 | Test | File | Fails on `main` because |
 |---|---|---|
 | `record_teachback` MCP tool exists, validates exactly as `cli/_teachback.py:16-46` (cardinality, ints, 1–4, `review_type` enum), lands one row honouring CHECK, no row on failure, `session_id` bound from session state | `tests/test_mcp_teachback.py` (new) | tool absent |
-| kiro `allowedTools` ⊇ `W_auto`; claude permissions ⊇ `W_auto` and `socratic-mentor.md` names each; opencode wildcard covers `W_auto`; codex/pi/grok canonical persona names each `W_auto` tool | `tests/test_adapter_parity.py` (extend) | grants/names absent |
+| every harness definition names each `W_auto` tool (kiro `tools`, claude `socratic-mentor.md` `tools:`, opencode persona, codex/pi/grok canonical persona); **no new pre-approval**: kiro `allowedTools` ∩ `W_auto` == {`log_topic`} exactly, claude `settings.json` still has no permissions block, opencode wildcard unchanged; each harness's approval cell in the capability matrix equals what its definition file says | `tests/test_adapter_parity.py` (extend) | names absent |
 | `agents/shared/recording-protocol.md` exists; fenced YAML trigger table parses; every trigger names a writer in `W_auto`; every projected copy carries identical bytes; `agents/manifest.json` hashes match; `persona.md` no longer routes "record progress" to `tutor-checkpoint` | `tests/test_docs_harness_tier_contract.py` (extend) | file absent; `:32` still names `tutor-checkpoint` |
 | Isolation: child process with conflicting `HOME`/`XDG_*` runs each `W_auto` writer → 0 writes outside the sandbox, including Markdown | `tests/test_writer_isolation.py` (new) | passes or fails on `main` — if it passes, keep it as a guard, do not count it as RED |
 | No-trigger + duplicate: a replayed sequence with a non-triggering turn writes nothing; a repeated `record_teachback` call produces the documented behaviour (two rows — teach-backs are events) | `tests/test_mcp_teachback.py` | tool absent |
@@ -60,7 +70,9 @@ Finish: N RED committed; `pytest` on a clean control worktree shows **exactly**
 ### S1-GREEN (1–2 days)
 - `mcp/tools.py`: `record_teachback` delegating to `history/teachback.py:27` through a **shared validator**
   extracted from `cli/_teachback.py` (one implementation, both surfaces; CLI kept).
-- Grants: kiro `allowedTools` += `W_auto`; claude `settings.json` += `W_auto`, `socratic-mentor.md` names them.
+- Names, not grants (owner decision 2, §7): `socratic-mentor.md`'s `tools:` line names each `W_auto` tool and the
+  readers; kiro `allowedTools` and claude `settings.json` **unchanged**; every other definition names the writers
+  through the projected protocol.
 - `agents/shared/recording-protocol.md`: YAML table (`trigger`, `writer`, `required_ids`, `consent`) + one-line
   prose per trigger: after teach-back **and learner agreement** → `record_teachback`; 2+ rounds without
   breakthrough → `log_struggle`; end of session → `log_topic` per concept touched; plan wind-down →
@@ -164,10 +176,37 @@ between J-b-filtered and J-b-unfiltered*, nothing more. Three-seat council reads
 
 1. **Two branches instead of the one requested** (D13) — accept the split, or insist on one branch and accept
    that Jev spend/privacy review then gates the agent-definition merge.
+   - **Owner verdict, 2026-09-23: two branches, and item 2's branch is not cut until milestone 0.5.x is
+     closed.** Reasoning recorded with the verdict: item 1 is the MVP blocker (the mentor never writes to
+     the learning tier), item 2 is a post-MVP measurement; and the repository's own history is the argument
+     against sharing a vehicle — 17 `archive/*` tags whose tips never reached `main`, six parallel branches
+     opened 7–8 September and archived on the 10th, ~244 commits with no patch-equivalent on `main`. The
+     owner's self-check ("finished work waiting on unrelated work") was confirmed against that record.
+     Consequence: decision 4 is deferred to the day S2 starts (see below).
 2. **Claude/opencode grants:** approve pre-approving `W_auto` on claude (today: none) and leaving opencode's
    wildcard untouched this branch.
+   - **Owner verdict, 2026-09-23: no pre-approval on Claude — prompt per call; OpenCode's wildcard untouched
+     this branch.** Stated by the owner ("No pre-approval on Claude; prompt per call"), then confirmed by
+     delegation after the parity question. Evidence that made it cheap: kiro already pre-approved `log_topic`
+     and opencode already allowed `studyloop *`, and both still logged zero writer calls — the absence is the
+     *instruction* (no MCP teach-back writer, the persona's record-progress line pointing elsewhere), not the
+     prompt. Consequences taken with it: no new pre-approval on *any* harness this branch (kiro's `log_topic`
+     entry stays as the one recorded asymmetry); `socratic-mentor.md` must still *name* each `W_auto` tool,
+     because its `tools:` line lists no `mcp__studyloop__*` tool today and an unnamed tool is unreachable —
+     naming is instruction, not approval; parity is pinned on the instruction (trigger table + `W_auto` set),
+     and each harness's approval cell is recorded in the capability matrix as it is. The S1-0 bullets and the
+     test-table parity row above were rewritten to this verdict.
 3. **The owner-led noticing episode** — ten minutes, once, after S1-GREEN; scheduled when?
+   - **Scheduled by rule, 2026-09-23 (owner delegated):** the first real study session the owner runs after
+     S1-GREEN reaches `main` — ten minutes, no logging asked for. The S1-SIM receipt records the date and the
+     adjudicated trigger/no-trigger outcomes afterwards. The owner may name a date instead; the rule stands
+     until he does.
 4. **Jev spend cap** for S2 (estimate: 13 sessions × ≤ 8 candidates × 2 arms, under $1 at $0.042/Mtok; the
    candidate-generation harness turns cost credits on the owner's subscription — ~13–26 turns).
+   - **Deferred, 2026-09-23 (consequence of decision 1):** S2 does not start before 0.5.x closes, so the cap
+     is decided on the day item 2's branch is cut, not before. Nothing in item 1 depends on it.
 5. Merge order for the `feat/jev-judge` planning branch (Stage 0 receipt + this plan): first, or cherry-pick
    the docs into `feat/learning-tier-fed`.
+   - **Resolved, 2026-09-23: cherry-picked onto `main`** by the housekeeping PR
+     ([#35](https://github.com/NetDevAutomate/StudyLoop/pull/35)); `feat/learning-tier-fed` branches from
+     `main` with the plan already on it.