JosephOIbrahim · JosephOIbrahim · May 22, 2026 · May 25, 2026 · May 29, 2026 · May 29, 2026
diff --git a/.claude/agents/critic.md b/.claude/agents/critic.md
@@ -0,0 +1,36 @@
+---
+name: critic
+description: Runs in exactly ONE mode per call — verify, red_team, or evaluate. Modes are never mixed.
+tools: Read, Grep, Glob, Bash
+---
+
+You are the **critic** under the Hanna orchestrator (`ORCHESTRATOR.md`). In D002 terms you are
+the **Compliance Reviewer** — mandatory, runs last and alone, never in parallel with workers.
+Your delegation specifies exactly one MODE. Do not blend modes within a call.
+
+## MODE: verify
+
+Does the artifact meet ACCEPTANCE? Read each acceptance criterion literally, check it, and return
+**pass** or **fail** with the specific criterion and per-item evidence. This is the EXIT gate —
+the orchestrator runs `critic[verify]` against EXIT_CRITERIA before terminating.
+
+## MODE: red_team
+
+What could break this? Adversarial pass over the diff: edge cases, contract drift, dead code,
+cross-file inconsistency, missed lockout / Rule 35 boundary gates, protocol desync. Return findings
+classified **BLOCKER / MAJOR / MINOR**. Hanna precedent: a round-3 review found 4 latent bugs that
+passing tests missed (belief c003) — assume the same and dig.
+
+## MODE: evaluate
+
+What is true given the evidence, and how confident? Return **claim + confidence (0.0–1.0) +
+provenance**. This is how the orchestrator sources belief deltas it can trust.
+
+## Hard constraints (inherited, D002 / §9)
+
+- Audit against `RULES.md` (especially 18, 34, 35, 36, 37), `docs/DECISIONS.md`,
+  `HANNA_BLUEPRINT.md` contracts, and the compliance greps.
+- Audit trailer hygiene: cloned files carry the attribution trailer (first 20 lines); fresh seeds
+  do not (D003/D004). Flag any model-id string in a committed artifact.
+- Read-only. Return artifact path + one-paragraph summary. You MAY propose belief deltas in
+  `evaluate` mode; you never write `state/beliefs.md` yourself.
diff --git a/.claude/agents/integrator.md b/.claude/agents/integrator.md
@@ -0,0 +1,29 @@
+---
+name: integrator
+description: Reconciles parallel diffs from N>1 concurrent workers into one coherent change. Only when N>1.
+tools: Read, Edit, Write, Grep, Glob, Bash
+---
+
+You are the **integrator** under the Hanna orchestrator (`ORCHESTRATOR.md`). In D002 terms you
+are step 5 — integration. You run ONLY when more than one worker ran concurrently and their diffs
+must be reconciled. For a single-worker task there is no integrator.
+
+## What you do
+
+- Merge parallel worker diffs into one coherent change: resolve overlapping edits, reconcile
+  shared schemas / imports, ensure the cross-file consistency that no individual worker could see
+  (e.g. two workers both touching `src/schemas.py`).
+- Do NOT introduce new behavior — you reconcile what workers produced. New behavior is a worker
+  task; surface it for the orchestrator instead of writing it.
+- Hand the reconciled artifact to `critic[verify]` before the orchestrator commits.
+
+## Hard constraints (inherited, D002 / §9)
+
+- Rule 34 / 35 / 36 / 37; preserve D003/D004 trailer hygiene across every merged file.
+- Canonical commit trailer (the orchestrator commits); no model-id in artifacts; no new
+  dependencies without a decision.
+
+## Output
+
+Reconciled artifact path + one-paragraph summary, naming any merge conflicts resolved and any
+belief deltas about integration risk. You never write `state/beliefs.md` yourself.
diff --git a/.claude/agents/planner.md b/.claude/agents/planner.md
@@ -0,0 +1,32 @@
+---
+name: planner
+description: Surveys the territory and produces a task graph with EXIT criteria. Runs on cold-start, stale plan, or belief reframe.
+tools: Read, Grep, Glob, Bash
+---
+
+You are the **planner** under the Hanna orchestrator (`ORCHESTRATOR.md`). In Hanna's D002 MoE
+taxonomy you are the **Architect** role, formalized as a Claude Code primitive.
+
+## What you do
+
+- Survey the territory: read `docs/ROADMAP.md` §5, `docs/DECISIONS.md`, `NEXT.md`,
+  `state/beliefs.md`, `state/open_questions.md`.
+- Produce a task graph — nodes, dependencies, acceptance per node. For buildout work the graph
+  IS `docs/ROADMAP.md` §5; reference lane IDs, do not duplicate the table.
+- Define GOAL + EXIT_CRITERIA (per-criterion, observable, verifiable) + CONFIDENCE_THRESHOLD
+  (default 0.8).
+- Record design choices WITH rejected alternatives. These become `docs/DECISIONS.md` D###
+  entries authored by the orchestrator — you propose, you do not write DECISIONS.md.
+
+## Hard constraints (inherited, D002 / §9)
+
+- Rule 34 (family-first lockout), Rule 35 (Harlo read-only / Octavius request-only),
+  Rule 36 (surface, don't decide), Rule 37 (never raise patent topics).
+- No new dependencies without a `docs/DECISIONS.md` entry.
+- Read-only: you never edit `src/`, `tests/`, or `state/`.
+
+## Output
+
+Artifact path (a proposed plan / decision block) + a one-paragraph summary. You MAY propose
+belief deltas `{claim, suggested_confidence, evidence}`; the orchestrator validates and writes
+`state/beliefs.md`. You never write it yourself.
diff --git a/.claude/agents/worker.md b/.claude/agents/worker.md
@@ -0,0 +1,32 @@
+---
+name: worker
+description: Executes one bounded task — the smallest change that satisfies acceptance. The most common role.
+tools: Read, Edit, Write, Grep, Glob, Bash
+---
+
+You are a **worker** under the Hanna orchestrator (`ORCHESTRATOR.md`). In D002 terms you are a
+builder expert (Bridge / Computation / Stage / Brief-Composer / MCP-Surface) — your INPUTS tell
+you which lens applies. You read cold: only your declared INPUTS, never the parent conversation.
+
+## What you do
+
+- Execute exactly ONE bounded task. Make the SMALLEST change that satisfies ACCEPTANCE.
+- One concern per delegation. If you discover out-of-scope work, note it in your summary for
+  `parked.md` — do not do it.
+- Write your diff / tests under `state/tasks/<TASK_ID>/` if the contract's OUTPUT says so;
+  otherwise edit in place per the contract.
+
+## Hard constraints (inherited, D002 / §9)
+
+- Rule 34 lockout gate at any publish / MCP-tool site; Rule 35 boundary (never call Harlo
+  `store`/`stage_reload`/`resolve_verifications`/`trigger_cognitive_recalibration`); Rule 36
+  voice (surface, don't decide); Rule 37 silence on patent topics.
+- D003/D004 trailer hygiene: cloned-from-Harlo files carry the attribution trailer in their
+  first 20 lines; fresh-seed files with no Harlo ancestor carry NO trailer.
+- The orchestrator commits, not you. Never place a model-id in any artifact.
+- No new dependencies without a decision.
+
+## Output
+
+Artifact path + one-paragraph summary. Include belief deltas `{claim, suggested_confidence,
+evidence}` when your work changes what is known. NEVER write `state/beliefs.md` yourself.
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -21,22 +21,73 @@ jobs:
           pip install -e ".[dev]"
       - name: Run pytest
         run: pytest tests/ -v
-      - name: Rule 35 — no forbidden Harlo writes
+      - name: Rule 35 — no forbidden Harlo writes (name-anchored first line of defense)
         run: |
-          if grep -rE "harlo\.(write|store|author|mutate)" src/ scripts/ ; then
+          if grep -rE "harlo\.(write|store|author|mutate|commit|persist|save|update)" src/ scripts/ ; then
             echo "::error::Rule 35 violation: write surface to Harlo found."
             exit 1
           fi
+          if grep -rE "octavius\.(write|store|author|mutate|commit|persist|save|update)" src/ scripts/ ; then
+            echo "::error::Rule 35 violation: write surface to Octavius found."
+            exit 1
+          fi
       - name: Rule 35 — no forbidden cognitive-twin store/stage/verify calls
         run: |
           if grep -rE "\b(stage_reload|resolve_verifications|trigger_cognitive_recalibration)\b" src/ scripts/ ; then
             echo "::error::Rule 35 violation: forbidden cognitive-twin call."
             exit 1
           fi
-      - name: Rule 37 — no patent topics
+      - name: Rule 35 — enumerated bridge surface (authoritative allowlist per D001)
+        run: |
+          python3 - <<'PY'
+          import re, sys, pathlib
+          allowlist = {"status", "coach", "recall", "query_past_experience", "patterns"}
+          path = pathlib.Path("src/harlo_bridge.py")
+          text = path.read_text(encoding="utf-8")
+          # Match self._call_tool("X" — capture X.
+          pattern = re.compile(r'self\._call_tool\(\s*["\']([^"\']+)["\']')
+          found = set(pattern.findall(text))
+          violations = found - allowlist
+          if violations:
+              print(f"::error::Rule 35 violation: src/harlo_bridge.py invokes _call_tool with non-allowlisted tool(s): {sorted(violations)}")
+              print(f"Allowlist (per D001): {sorted(allowlist)}")
+              sys.exit(1)
+          if not found:
+              print(f"::error::Rule 35 check: no _call_tool sites found in src/harlo_bridge.py — pattern broken?")
+              sys.exit(1)
+          print(f"Rule 35 allowlist OK: found tools {sorted(found)}")
+          PY
+      - name: Rule 37 — no patent topics (code, scripts, tests, docs)
+        run: |
+          # Meta-references that discuss the rule itself (rather than raising patent
+          # topics as content) are exempted. The exemption pattern matches: explicit
+          # "Rule 37" mentions, the grep-recipe self-reference, the literal "patent
+          # topic(s)" / "patent-topic(s)" phrasing the rule uses to name itself, and
+          # the upstream PATENTS.md filename reference.
+          hits=$(grep -rEi "patent|provisional|uspto|claim[[:space:]]+language" src/ scripts/ tests/ docs/ \
+            | grep -vEi "Rule 37|grep.*patent|patent[- ]topic|patent[- ]adjacent|PATENTS\.md|no patent" \
+            || true)
+          if [ -n "$hits" ]; then
+            echo "::error::Rule 37 violation: patent topic found in code or docs."
+            echo "$hits"
+            exit 1
+          fi
+      - name: Rule 37 — no patent topics in commit messages (last 50)
         run: |
-          if grep -rEi "patent|provisional|uspto|claim[[:space:]]+language" src/ scripts/ tests/ ; then
-            echo "::error::Rule 37 violation: patent topic found in code."
+          # Need enough history; checkout above uses default depth=1, so fetch more.
+          git fetch --deepen=50 origin 2>/dev/null || true
+          # Same exemption posture as the docs grep: rule-discussion language is
+          # not a Rule 37 violation. Historical commits reference the brief
+          # composer's negative-space examples ("Patent-section paragraph",
+          # "Patent application formation") as rule-enforcement framing, not as
+          # patent-topic content.
+          hits=$(git log -50 --format=%B \
+            | grep -iE "patent|provisional|uspto|claim[[:space:]]+language" \
+            | grep -vEi "Rule 37|grep.*patent|patent[- ]topic|patent[- ]adjacent|patent[- ]section|patent application formation|no patent" \
+            || true)
+          if [ -n "$hits" ]; then
+            echo "::error::Rule 37 violation: patent topic found in recent commit messages."
+            echo "$hits"
             exit 1
           fi
       - name: Rule 1 — no sleep() calls

diff --git a/CLAUDE.md b/CLAUDE.md
@@ -21,6 +21,12 @@ Hanna conventions accrete as sessions resolve open questions. See [`docs/CONVENT
 
 ---
 
+## Orchestration
+
+Multi-agent work in this repo runs under the orchestrator operating manual in [`ORCHESTRATOR.md`](ORCHESTRATOR.md). Subagent roles live in `.claude/agents/` (planner / worker / critic / integrator); durable belief and open-question state lives in `state/`. `/hanna-dispatch-next` is one workflow under that orchestrator (see ORCHESTRATOR.md §8). This file does not duplicate the manual; it points at it.
+
+---
+
 ## Commit trailer (canonical)
 
 Every commit in this repo ends with the canonical Claude Code format:
@@ -73,5 +79,6 @@ Session docs live under `docs/SESSION_NN_*.md`.
 - Not a duplicate of the global `~/.claude/CLAUDE.md`. Behaviors, body-first protocols, coworker blend, RSD guidance — those live globally. This file is project scope only.
 - Not a duplicate of `RULES.md`. Rules live there; this file points at them.
 - Not a duplicate of `HANNA_BLUEPRINT.md`. Architecture lives there.
+- Not a duplicate of `ORCHESTRATOR.md`. The orchestration manual lives there; this file points at it.
 
 Each file has one job.
diff --git a/NEXT.md b/NEXT.md
@@ -1,22 +1,34 @@
 # NEXT — for tomorrow-you
 
+## 2026-05-22 (post-merge) — buildout chapter closed
+
+PR [#1](https://github.com/JosephOIbrahim/Hanna/pull/1) merged to `main` at `e08aebb619526c031b4d895484a1ae1bc1b51a23` (merge commit; method: merge — preserves the 18-commit lane-by-lane history). Branch `claude/hanna-mcp-review-ZsorY` is now consumed against `main`; 18 commits landed.
+
+Buildout-session checkpoint preserved below as the source-of-truth handoff for the next session.
+
+---
+
 ## 2026-05-22 session outcome — post-ratification buildout (L1–L4a + L3b)
 
-The `/hanna-dispatch-next` harness advanced five lanes on `claude/hanna-mcp-review-ZsorY` via D002 MoE dispatches. PR [#1](https://github.com/JosephOIbrahim/Hanna/pull/1) covers the full buildout against `main`.
+The `/hanna-dispatch-next` harness advanced five lanes on `claude/hanna-mcp-review-ZsorY` via D002 MoE dispatches; three rounds of CodeRabbit review landed in three separate fix batches.
 
-### Lane commits (top of branch, in order)
+### Lane commits (top of branch, in order before merge)
 
+- `298f50c` — fix: CodeRabbit round 3 (4 latent bugs + 2 test polish) — HarloTimeout fallback in PoC, frame coalescing in bridge `_read_frame_with_timeout` (with new `self._recv_buffer`), ISO-datetime parsing in `_parse_forcing_function`, README §"Rules" layer-count contradiction, brittle wall-clock assertion removed, `FrozenInstanceError` narrowed.
+- `8e83296` — docs: 4 doc/code contradictions reconciled (post-PR diagnostic — README Status table, CONVENTIONS delegate references, BLUEPRINT path, bin dangling reference).
+- `2e68877` — docs: D003 + D008 contradictions in BLUEPRINT + README + RULES + NEXT.md refresh (CodeRabbit round 2).
+- `8bafe27` — ci: pin actions/checkout + actions/setup-python to SHAs + persist-credentials false (CodeRabbit round 2).
 - `06effc8` — **L3b** D005 bridge hardening (composition scope + selectors-based read timeout + stderr drainer with `deque(maxlen=64)`; **+21 tests**; ROADMAP §5 atomically updated).
 - `0f35f33` — Composer voice fix (Rule 36 honesty: `_portfolio_line` surfaces counts by status; active-first display order; **+5 tests**).
 - `ecd465b` — **L4a** D007 product files + composer rewrite (`ProductFile` + `BriefPayload` schemas + `compute_brief_priority` pure function + four `data/products/*.md` stubs; **+25 tests**; ROADMAP §5).
 - `04af5da` — **L3a** Session 03 phase bodies (six branches complete; PoC `try/except NotImplementedError` catch deleted; **+18 tests**; ROADMAP §5).
 - `2f44e52` — **L2** Substrate hygiene (`pyproject.toml` + `tests/conftest.py` + `.github/workflows/ci.yml` + `.gitignore` patches; ROADMAP §5).
 - `ec6752a` — **L1** D008 propagation (BLUEPRINT §4 table renamed + §5 strikethroughs + §10 lane diagram; README mermaid; RULES.md applicability note).
-- `413e7ad` — docs+fix: CodeRabbit PR #1 review batch 1 (D001 anchors, NEXT.md status alignment, midday/evening phase fallback).
+- `413e7ad` — docs+fix: CodeRabbit PR #1 review round 1 (D001 anchors, NEXT.md status alignment, midday/evening phase fallback).
 
 ### Test count
 
-73 tests pass on `06effc8` (52 baseline + 21 from L3b). Up from 7/7 at Session 02 close.
+**74 tests pass on HEAD** (`298f50c`). Progression: 7 (Session 02) → ≥22 (post-L3a) → 47 (post-L4a) → 52 (voice fix) → 73 (L3b) → 74 (round-3 ISO datetime test).
 
 ### Substrate-decision tree at end of session
 
@@ -26,10 +38,10 @@ The `/hanna-dispatch-next` harness advanced five lanes on `claude/hanna-mcp-revi
 | D002 | resolved | every MoE dispatch in this session followed the protocol |
 | D003 | resolved | clones carry trailer only; fresh seeds carry no trailer |
 | D004 | resolved | reviewer audits trailer placement within first 20 lines |
-| D005 | resolved | **landed via L3b** (`src/harlo_bridge.py` + `tests/test_harlo_bridge.py`) |
+| D005 | resolved | **landed via L3b** (`06effc8`); frame coalescing patched in `298f50c` (`self._recv_buffer` preserves trailing bytes across `_read_frame_with_timeout` calls) |
 | D006 | resolved | **pending L4b implementation** (next lane) |
-| D007 | resolved | **landed via L4a** (`src/schemas.py` `ProductFile` + `BriefPayload`; `data/products/*.md`; composer rewrite) |
-| D008 | resolved | **landed via L1** (BLUEPRINT + README + RULES docs propagation) |
+| D007 | resolved | **landed via L4a** (`ecd465b`); `_parse_forcing_function` ISO-datetime split patched in `298f50c` (canonical `: ` delimiter; preserves `2026-06-01T10:30:00-04:00` timezone offsets) |
+| D008 | resolved | **landed via L1** (`ec6752a`) |
 
 ### ROADMAP §5 status
 
@@ -47,9 +59,9 @@ The `/hanna-dispatch-next` harness advanced five lanes on `claude/hanna-mcp-revi
 
 ## Where you are
 
-- **Branch:** `claude/hanna-mcp-review-ZsorY`.
-- **PR:** [#1](https://github.com/JosephOIbrahim/Hanna/pull/1) open against `main`. CI passing on `06effc8`. CodeRabbit reviewed; this session addressed the four open threads (CI workflow hardening + D003/D008 doc-contradiction reconciliation).
-- **Branch is 14+ commits ahead of `main`.**
+- **Branch:** `claude/hanna-mcp-review-ZsorY` — consumed against `main` at merge commit `e08aebb`. Branch still exists in the remote; not deleted (Joe's discretion).
+- **PR:** [#1](https://github.com/JosephOIbrahim/Hanna/pull/1) **merged**. CI was green on HEAD `298f50c` (success in 13s). All 13 CodeRabbit threads addressed across 3 rounds.
+- **`main` HEAD:** `e08aebb` (merge commit; 18 commits incorporated).
 
 ## Next session entry point — L4b
 
@@ -67,10 +79,16 @@ After L4b: L5 (schemas 2–5: `OverrideToken`, `JoeStateSnapshot`, `FormationReq
 ## Open questions still parked
 
 - **§C.2** — Octavius IPC PoC (deferred until `octavius_bridge` lane / L7).
-- **§C.3** — Harlo MCP-client precedent — **closed by L3b.** D001 + D005 ratifications + L3b implementation cover the surface; the bridge now sustains long-lived callers, hung-subprocess timeouts, and stderr backpressure.
+- **§C.3** — Harlo MCP-client precedent — **closed by L3b + round-3 patch.** D001 + D005 ratifications + L3b implementation + `298f50c` frame-coalescing patch cover the surface; the bridge now sustains long-lived callers, hung-subprocess timeouts, stderr backpressure, and frame coalescing.
 - **§C.4** — `LockoutResponse` shape (needed before L6 `mcp_tools` lane).
 - **§C.6** — RED override in delegate dispatch — **void per D008.1** (delegate Cut; Layer 2 collapsed into Layer 3 per-tool lockout check in L6 `mcp_tools` lane).
 
-## Staleness flag — carry forward
+## Lessons from this session (for future buildout sessions)
+
+- **Invocation surfaces what mocked tests miss.** The L4a Rule 36 voice bug ("4 threads in flight" when only 1 was) surfaced via `PYTHONPATH=. python3 scripts/first_hanna_brief.py`, not via the 47 passing tests. Always invoke the PoC at edges after each lane.
+- **Deeper review catches what your own confidence misses.** CodeRabbit's round-3 scan caught 4 real bugs in L3b/L4a (frame coalescing, HarloTimeout fallback, ISO-datetime split, README contradiction). All 4 were latent — passing tests + my self-audit didn't surface them.
+- **Doc-drift accumulates faster than code-drift.** The post-PR diagnostic survey caught 4 contradictions in README Status, CONVENTIONS, BLUEPRINT, and bin/ that CodeRabbit's grep missed. Plan for one diagnostic survey per multi-lane session, not just before PR creation.
+
+## Staleness flag — resolved 2026-05-25
 
-`docs/SESSION_01_RECON.md` §G claims the 33 rules "do not exist in Harlo, synthesize from distributed sources." Still wrong per Session 01.5's direct extraction from `Harlo/CLAUDE.md` lines 37–194. Joe's call: correct or leave as session-stamped historical artifact.
+`docs/SESSION_01_RECON.md` §G (the "rules don't exist in Harlo" recon-time misread) was preserved as a session-stamped historical artifact during the Phase 4 doc-drift sweep — a `> Historical note` block at the top of §G points to the live state (Session 01.5's direct extraction from `Harlo/CLAUDE.md` lines 37–194). The narrative below the note stays unedited as the recon-time record. q003 / q009 closed.