2026-05-14: B-Team review complete — GPT-5.4 case study, findings applied, journalists caught up

- GPT-5.4 B-Team governance review: ran pipeline task spec with live interpreter
  (found and installed local repo). 4/5 programs first attempt, 2 real runtime
  errors encountered and fixed.
- Four findings applied to agent-facing docs:
  - FIND-B1: direct-call is_bottom(f()) pattern documented in AGENT_QUICKREF
  - FIND-B2: io.say double-print behavior documented (QUICKREF + cookbook #12)
  - FIND-B4: stale Rust parser note removed; individual vs index import guidance added
  - RE-01: cookbook entry #11 — Program 5 via HTTP (RPC API workflow)
- AGENT_COOKBOOK bumped to v1.1
- All journalist/persona catch-up sections updated to v2.0 state
- 341 tests passing, dispatch-check exits 0

Author: Douglas Jones

.kiro/steering/personas/axiom.md

@@ -65,16 +65,20 @@ The Track 1 external agent sessions (GPT-4o, Gemini, Claude) were Axiom's job
 done manually and expensively. Axiom makes that systematic. Every new surface
 gets an Axiom pass before it ships, not after three external agents hit it.
 
-## Catch-up on Codifide
+## Catch-up on Codifide (as of v2.0 — 2026-05-14)
 
 Key friction points already documented (do not re-report these as new findings):
 - `a + b` → use `add(a, b)` (arithmetic operators don't exist)
 - `is_bottom(x)` cannot catch propagated bottom — raises `BottomPropagationError`
-- bind-before-when: `when` guard runs before the candidate body; bound names
-  don't exist yet in the guard
+- `is_bottom(f())` direct-call works — documented in AGENT_QUICKREF (2026-05-14)
+- bind-before-when: `when` guard runs before the candidate body; now a parse
+  error with a clear fix hint (V2-2 shipped 2026-05-14)
 - `contains()` is case-sensitive — always normalize with `lower()` first
-- `from <hash> import name` requires `CODIFIDE_RUNTIME=python` until V2-3 ships
-- Content-addressed composition (Program 5) requires the store CLI + index pattern
-
-These are in `docs/AGENT_COOKBOOK.md` and `docs/AGENT_QUICKREF.md`. Axiom's
-job is to find the *next* ones.
+- `from <hash> import name` works in both runtimes as of v2.0 (V2-3 shipped)
+- Content-addressed composition (Program 5) — CLI path and HTTP path both
+  documented in cookbook entries #8 and #11
+- `io.say` + CLI double-print — documented in AGENT_QUICKREF and cookbook #12
+
+These are in `docs/AGENT_COOKBOOK.md` (v1.1) and `docs/AGENT_QUICKREF.md`.
+Axiom's job is to find the *next* ones — particularly around the RPC API
+surface (new in v2.0) and any parallel evaluator surfaces.

.kiro/steering/personas/glyph.md

@@ -102,7 +102,7 @@ communicates with humans. Glyph communicates with agents. The project
 "publishes to agents" when a Glyph dispatch is committed to the repo and
 (later) posted to the dispatch stream.
 
-## Catch-up on Codifide (as of v1.0 / v2.0)
+## Catch-up on Codifide (as of v2.0 — 2026-05-14)
 
 Glyph, the project lives at
 `/Users/douglasjones/Projects/CodifideProgrammingLanguage/`. Public on GitHub
@@ -112,32 +112,28 @@ Key facts for dispatch construction:
 
 - **Canonical spec:** `docs/CANONICAL.md`
 - **Interpreter semantics:** `codifide/runtime/`
-- **Capability manifest hash (v1.0):**
-  `sha256:23fdde779caebc2c471ade0e1c407422d044e2e0f1adc7e59a189325deccd27d`
-- **Test count:** 289 Python passing, 0 skipped (as of 2026-05-13)
+- **Capability manifest hash (v2.0):**
+  `sha256:42d73647ba8de29a7d219bf2218bad0a42dc2a11d7878cac12ee931be2a1a185`
+- **Test count:** 341 Python passing, 0 skipped (as of 2026-05-14)
 - **Rust canonical crate:** `crates/codifide-canonical/` — byte-level
   conformance to Python; 28 Rust tests passing
 - **Dispatch journal:** `dispatches/INDEX.md` — indexed, grouped by date
-- **Template dispatch:** `dispatches/2026-05-13-t1-1-pipeline-task-spec.yaml`
+- **Shape reference:** `dispatches/2026-05-14-bteam-findings-applied.yaml`
   is the most recent Glyph dispatch; use it as your shape reference
 
-**Shipped state (v1.0, 2026-05-11):**
-- CBOR primary content hash, JSON legacy
-- Cost-based candidate dispatch
-- Symbol store with GC, atomic writes, sharded loose objects
-- Content-addressed imports
-- Indexed primitives: `slice`, `at`, `char_at`, `indexof`
-- Inline `if/then/else` expression
-- Capability manifest, agent-facing docs, quickref
-
-**Shipped state (v2.0, 2026-05-12):**
-- Rust interpreter + Rust parser (Shape A)
-- Parallel evaluator, benchmarks
-
-**Active initiative:** Agent Adoption — spec at `.kiro/specs/agent-adoption/`
-- Track 1 (external agent case study): T1-1 complete, T1-2 next
-- Track 2 (adoption infrastructure): not started
-- Track 3 (v2.0 roadmap): blocked on Track 1
+**Shipped state (v2.0, 2026-05-14):**
+- V2-1: RPC API — `python3 -m codifide serve`, POST/GET symbols by hash
+- V2-2: Static bind-before-when detection — parse error with fix hint
+- V2-3: from-import in Rust parser — both runtimes now support `from`-import
+- V2-4: Manifest `docs` field — links to human-readable documentation
+- Agent Adoption Initiative complete — four case studies, cookbook v1.1,
+  quickref updated, B-Team governance review closed
+
+**Open items (not yet dispatched):**
+- AUD-OVERNIGHT-02: parallel evaluator branch interpreters don't carry
+  resolved imports — Sable audit needed before v3.0 parallel work
+- New agent case study to validate adoption improvements (Relay's KPI)
+- v3.0 planning if adoption evidence warrants it
 
 **Dispatch discipline:** Every session files a paired Quill readout
 (`.readout.md`) and Glyph YAML (`.yaml`). Session-close pairs required.

.kiro/steering/personas/lumen.md

@@ -75,20 +75,25 @@ implementations is a bug against the spec. If both agree but the spec is
 silent, the spec is the bug."* Lumen's job is to make sure the spec can
 actually play that role.
 
-## Catch-up on Codifide
+## Catch-up on Codifide (as of v2.0 — 2026-05-14)
 
 Key spec documents:
 - `docs/CANONICAL.md` — the canonical-form specification (primary)
 - `docs/LANGUAGE.md` — surface-syntax reference
 - `docs/CAPABILITY.md` — capability manifest schema
 - `docs/capability-0.1.json` — current manifest (generated from implementation)
-
-Known spec gaps as of v2.0 (do not re-report):
-- `from`-import behavior in the Rust parser is undocumented in the spec
-  (implementation gap, not spec gap — tracked as REQ-V2-3)
-- Manifest `docs` field not yet in schema (tracked as REQ-V2-4)
-- Shadowing rules are in `docs/CANONICAL.md §Shadowing` — reviewed and
-  considered complete as of v1.0
+- New manifest hash: `sha256:42d73647ba8de29a7d219bf2218bad0a42dc2a11d7878cac12ee931be2a1a185`
+
+Known spec gaps resolved in v2.0 (do not re-report):
+- `from`-import behavior in the Rust parser — implemented (V2-3)
+- Manifest `docs` field — added to schema and manifest (V2-4)
+- Bind-before-when execution order — now documented in LANGUAGE.md and
+  enforced as a parse error (V2-2)
+
+Known spec gaps still open:
+- Parallel evaluator semantics under concurrent belief dispatch: not in spec
+- RPC API (`docs/RPC_API.md`) is implementation documentation, not a spec
+  section — no conformance tests for the HTTP surface
 
 Lumen's first deliverable when invoked: a one-pass review of the spec section
 most relevant to the current initiative.

.kiro/steering/personas/quill.md

@@ -45,7 +45,7 @@ minutes, not hours. They will form an opinion based on what Quill writes.
 End every report with a single sentence titled **"What I'm not yet sure of."**
 If Quill is certain of everything, the report is incomplete.
 
-## Catch-up on Codifide (as of v1.0 / v2.0)
+## Catch-up on Codifide (as of v2.0 — 2026-05-14)
 
 Quill, here's what you're working with. The project is in
 `/Users/douglasjones/Projects/CodifideProgrammingLanguage/`, public on GitHub
@@ -56,36 +56,53 @@ as `codifide-programming-language`, MIT licensed.
   Tagline: *"confidence in code, for agents."*
 
 - **What shipped in v1.0 (2026-05-11):**
-  - Python reference interpreter — effect enforcement, pre/post contracts,
-    multi-candidate dispatch, cost-based dispatch, belief dispatch, inline
-    `if/then/else`, first-class refusal (`bottom`), 8 typed error kinds
-  - Canonical JSON + CBOR forms with SHA-256 content addressing (CBOR primary)
-  - Content-addressed symbol store with GC, atomic writes, sharded loose objects
-  - Content-addressed imports (`import foo = sha256:...`)
-  - Indexed primitives: `slice`, `at`, `char_at`, `indexof`
-  - Capability manifest (`python3 -m codifide capability`) — agent-facing
-    self-description, content-addressed, generated from the implementation
-  - Rust canonical crate (`crates/codifide-canonical/`) — byte-level
-    conformance to Python on every example
-  - 216 Python tests passing, 28 Rust canonical tests passing, 0 skipped
-  - Repo made public; `docs/FOR_AGENTS.md` and `docs/AGENT_QUICKREF.md` written
-
-- **What shipped in v2.0 (2026-05-12):**
-  - Rust interpreter and Rust parser (Shape A milestone)
-  - Parallel evaluator and benchmarks
-  - 289 Python tests passing total (as of 2026-05-13)
-
-- **What is actively in progress:**
-  - Agent Adoption Initiative — spec at `.kiro/specs/agent-adoption/`
-  - Track 1: external agent case study (GPT-4o, Gemini 2.5 Pro, Claude baseline)
-  - Track 2: adoption infrastructure (manifest endpoint, cookbook, quickstart)
-  - Track 3: v2.0 roadmap update driven by adoption findings
-
-- **What's honest to say:** Codifide is a complete, tested, public v1.0
-  language. The semantics are real and enforced. The scale story (graph-native
-  parallel runtime, RPC API, time-indexed types) is roadmap, not shipped.
-  No external agent has yet adopted it in a real session — that is the
-  current initiative.
+  Python reference interpreter, canonical CBOR/JSON, content-addressed symbol
+  store, capability manifest, Rust canonical crate. 216 Python tests, 28 Rust
+  canonical tests.
+
+- **What shipped in v2.0 (2026-05-14, overnight session):**
+  - **V2-1 RPC API** — `python3 -m codifide serve` starts a local HTTP server
+    backed by the symbol store. POST canonical forms, GET by hash. Removes the
+    CLI ceremony from Program 5 (content-addressed composition).
+  - **V2-2 Static bind-before-when detection** — the parser now catches the
+    bind-before-when footgun at parse time with a clear fix message. Previously
+    a confusing runtime error.
+  - **V2-3 from-import in Rust parser** — `from sha256:<hash> import ...` now
+    works in the Rust runtime. `CODIFIDE_RUNTIME=python` workaround removed.
+  - **V2-4 Manifest docs field** — capability manifest now includes a `docs`
+    field pointing to human-readable documentation.
+  - New manifest hash: `sha256:42d73647ba8de29a7d219bf2218bad0a42dc2a11d7878cac12ee931be2a1a185`
+  - 341 Python tests passing, 0 skipped.
+
+- **Agent Adoption Initiative — complete (2026-05-13):**
+  - Track 1: Four external agent case studies run (GPT-4o, Gemini 2.5 Pro,
+    Claude baseline, GPT-5.4 B-Team review). All five programs completed by
+    all models. Key finding: Program 5 (content-addressed composition) was the
+    universal friction point — fixed by V2-1 RPC API.
+  - Track 2: Adoption infrastructure shipped — manifest endpoint live at
+    codifide.com, `AGENT_COOKBOOK.md` (12 entries), `AGENT_QUICKREF.md`,
+    `python3 -m codifide agent-quickstart`.
+  - Track 3: v2.0 roadmap driven by adoption findings — all four requirements
+    shipped.
+
+- **B-Team governance review — complete (2026-05-14):**
+  GPT-5.4 ran the pipeline task spec with live interpreter access (found and
+  installed the local repo). Four findings applied: direct-call `is_bottom`
+  documented, double-print behavior documented, stale Rust parser note removed,
+  HTTP workflow added to cookbook.
+
+- **What's honest to say:** Codifide is a complete, tested, public v2.0
+  language. The adoption infrastructure is real — four external models have
+  run the pipeline task spec and the friction points are documented and fixed.
+  The scale story (graph-native parallel runtime, time-indexed types) is
+  roadmap, not shipped. The parallel evaluator does not yet carry resolved
+  imports into branch interpreters (known gap, AUD-OVERNIGHT-02).
+
+- **Open action items:**
+  - `AGENT_COOKBOOK.md` HTTP workflow — done (entry #11)
+  - New agent case study to validate adoption improvements (Relay's KPI)
+  - Sable audit of parallel evaluator import handling (AUD-OVERNIGHT-02)
+  - v3.0 planning if adoption evidence warrants it
 
 Your first deliverable when invoked: a one-page "state of Codifide" that a
 technically literate human could read in three minutes.

.kiro/steering/personas/relay.md

@@ -75,22 +75,33 @@ connects them — the sequence, the discoverability, the funnel. A language
 surface can be ergonomically clean (Axiom says so) and well-documented (Paige
 says so) but still produce a confused agent if the funnel doesn't lead there.
 
-## Catch-up on Codifide
+## Catch-up on Codifide (as of v2.0 — 2026-05-14)
 
-Current adoption funnel (as of v2.0 / Agent Adoption Initiative):
+Current adoption funnel state:
 
-1. Agent fetches `codifide.com/capability.json` (or `.cbor`)
-2. Reads `docs/FOR_AGENTS.md` (linked from manifest — gap: not yet, REQ-V2-4)
-3. Reads `docs/AGENT_QUICKREF.md`
+1. Agent fetches `codifide.com/capability.json` (or `.cbor`) — live, includes
+   `docs` field pointing to human-readable documentation (V2-4 shipped)
+2. Reads `docs/FOR_AGENTS.md`
+3. Reads `docs/AGENT_QUICKREF.md` — updated with direct-call `is_bottom`
+   pattern and double-print note (2026-05-14)
 4. Runs `python3 -m codifide agent-quickstart`
-5. Writes Programs 1–4 (success rate: ~100% across Track 1)
-6. Writes Program 5 (success rate: 0/3 across Track 1 — RPC API is the fix)
-
-Known funnel gaps (do not re-report):
-- Manifest does not link to cookbook or quickref (REQ-V2-4, deferred)
-- Program 5 requires CLI + CODIFIDE_RUNTIME=python (REQ-V2-1, P1)
-- Feedback template exists (`dispatches/feedback/TEMPLATE.md`) but has not
-  been used in a real session (AUD-T2-04, open)
+5. Writes Programs 1–4 — success rate ~100% across all four case studies
+6. Writes Program 5 — now has two paths:
+   - **CLI path:** `store put` + `store hash` + individual imports (flat chains)
+     or `store index` + `from`-import (deep chains). Both runtimes supported.
+   - **HTTP path:** `python3 -m codifide serve` + POST canonical forms + import
+     by returned hashes. Documented in cookbook entry #11 and `docs/RPC_API.md`.
+
+Known funnel gaps (resolved — do not re-report):
+- `CODIFIDE_RUNTIME=python` workaround — removed (V2-3 shipped)
+- Program 5 CLI ceremony — HTTP path now available (V2-1 shipped)
+- Manifest `docs` field missing — shipped (V2-4)
+- Bind-before-when runtime error — now a parse error with fix hint (V2-2)
+
+Known funnel gaps (open):
+- Feedback template (`dispatches/feedback/TEMPLATE.md`) has not been used
+  in a real session (AUD-T2-04, still open)
+- No new agent case study since v2.0 shipped — adoption KPI unvalidated
 
 Relay's first deliverable when invoked: a funnel walk for the current release
 state, with time-to-first-working-program estimate.

.kiro/steering/personas/sable.md

@@ -94,7 +94,7 @@ a post-audit dispatch is an open wound, not a report.
 - **Error surface.** Do typed errors actually classify what went wrong,
   or do host-language exceptions leak through?
 
-## Catch-up on Codifide (as of v1.0 / v2.0)
+## Catch-up on Codifide (as of v2.0 — 2026-05-14)
 
 Sable, the project lives at
 `/Users/douglasjones/Projects/CodifideProgrammingLanguage/`. Public on GitHub
@@ -105,32 +105,35 @@ as `codifide-programming-language`, MIT licensed.
   conformance to Python; includes CBOR decoder, fuzz harness
 - **Rust interpreter + parser:** `crates/codifide-interpreter/` (v2.0,
   2026-05-12) — parallel evaluator, benchmarks
-- **Spec:** `docs/CANONICAL.md` — read with suspicion; written by the
-  implementers, though it has been through multiple Sable passes
-- **Test count:** 289 Python passing, 0 skipped (as of 2026-05-13);
-  28 Rust canonical passing
+- **RPC API server:** `codifide/server.py` — ThreadingHTTPServer over
+  SymbolStore; `python3 -m codifide serve` (V2-1, 2026-05-14)
+- **Spec:** `docs/CANONICAL.md` — read with suspicion
+- **Test count:** 341 Python passing, 0 skipped (as of 2026-05-14)
 - **Prior audit history** (all in `dispatches/`):
-  - `2026-05-10-security-audit.md` — initial CBOR neighborhood audit;
-    three P1 findings (P1-5 symlink write, P1-6 UnicodeDecodeError leak,
-    P1-7 Rust CLI hung on `/dev/zero`); all resolved
+  - `2026-05-10-security-audit.md` — initial CBOR neighborhood audit; all P1s resolved
   - `2026-05-11-ergonomics-audit.md` — post-four-model-review ergonomics
-  - `2026-05-11-new-surfaces-audit.md` — cost dispatch + store GC;
-    five findings (CDP-1/2, GC-1/2/3); all resolved
+  - `2026-05-11-new-surfaces-audit.md` — cost dispatch + store GC; all resolved
   - `2026-05-11-cli-audit.md` — unbounded source read (P1); resolved
-
-**Known coverage gaps as of v1.0:**
-- Conformance suite (`tests/test_conformance.py`) covers ASCII-clean
-  examples only — not a passing grade for the full surface
-- Rust interpreter (v2.0) has not yet received a Sable audit
-- Parallel evaluator semantics under concurrent belief dispatch: untested
-- Agent Adoption Initiative sessions: no adversarial review of the
-  agent-facing docs or task spec yet
+  - `2026-05-13-track1-sable-audit.md` — Track 1 case study surfaces
+  - `2026-05-13-track2-sable-audit.md` — Track 2 adoption infrastructure
+  - `2026-05-14-v2-1-rpc-api-sable-audit.md` — RPC API; 2 P2s fixed, 3 P3s
+    fixed or accepted
+
+**Known coverage gaps (open):**
+- **AUD-OVERNIGHT-02:** Parallel evaluator branch interpreters don't carry
+  resolved imports. Branch interpreters created with empty `resolved_imports`.
+  Documented as a known limitation; not yet fixed. Sable audit needed before
+  any parallel + import work in v3.0.
+- Conformance suite covers ASCII-clean examples only
+- RPC API HTTP surface has no adversarial test coverage
+- `is_bottom(f())` direct-call pattern — works in the interpreter but no
+  dedicated test; behavior should be confirmed before it goes into docs
+  (currently documented in AGENT_QUICKREF as of 2026-05-14)
 
 **Active surface to audit next:**
-- `docs/AGENT_TASK_SPEC.md` — the pipeline task spec handed to external
-  agents; Sable has not reviewed it
-- `crates/codifide-interpreter/` — Rust interpreter, no audit yet
-- `codifide/runtime/interpreter.py` — any new surfaces since last audit
+- Parallel evaluator import handling (AUD-OVERNIGHT-02)
+- RPC API adversarial surface (`codifide/server.py`) — body size limits,
+  concurrent POST behavior, store exhaustion under hostile input
 
 Your first deliverable when invoked: an audit report with
 severity-rated findings, each with a reproducing probe, filed to