diff --git a/docs/design/README.md b/docs/design/README.md
index b4fdf157b..d554c549a 100644
--- a/docs/design/README.md
+++ b/docs/design/README.md
@@ -7,6 +7,7 @@ Design specifications for planned features and skills.
 | [env-setup-design.md](env-setup-design.md) | Design spec for the dedicated `setup-environment` skill — Docker vs micromamba-host decision tree, structured output tokens, and recipe integration |
 | [recording-replay-accepted-degradations.md](recording-replay-accepted-degradations.md) | Accepted degradations in P8 recording/replay: Claude PTY cassette format incompatibility with Codex replay, and unchanged Claude session recording path |
 | [acp-session-contract.md](acp-session-contract.md) | Normative reference for P6-A3-WP1: ACP session method mapping for the `CodingAgentBackend` lifecycle, recovery ladder from `RetryReason` to `session/resume`/`session/load`/`session/new`, capabilities translation of all 41 `BackendCapabilities` fields, and Codex shim deviations |
+| [paper-backend-n3-exercise.md](paper-backend-n3-exercise.md) | N=3 paper backend exercise: opencode-via-ACP classification of all 30 Protocol methods, 41 `BackendCapabilities` fields, 2 `BackendConventions` fields, and B3a/B3b probe categories with gap catalogue of Protocol change requests |
 | [forward-obligations/inspector-contract-impact.md](forward-obligations/inspector-contract-impact.md) | Contract-impact note for Health Inspector (#3534): `build_inspector_cmd` stubs, `inspector_capable` gaps, `InspectorCallback` wiring |
 | [forward-obligations/recording-replay-impact.md](forward-obligations/recording-replay-impact.md) | Contract-impact note for recording/replay: backend isolation invariant, format-detection extension points for future backends |
 | [forward-obligations/triage-portability-impact.md](forward-obligations/triage-portability-impact.md) | Contract-impact note for triage portability: `triage_capable` gate, hardcoded Claude CLI in `_llm_triage.py`, portable `triage_cmd` path |
diff --git a/docs/design/paper-backend-n3-exercise.md b/docs/design/paper-backend-n3-exercise.md
new file mode 100644
index 000000000..a1326ac15
--- /dev/null
+++ b/docs/design/paper-backend-n3-exercise.md
@@ -0,0 +1,423 @@
+# Paper Backend N=3 Exercise: opencode-via-ACP
+
+**Status:** Draft
+**Date:** 2026-06-28
+**Issue:** [#4052](https://github.com/TalonT-Org/AutoSkillit/issues/4052)
+**Candidate:** opencode-via-ACP (paper exercise — no implementation)
+**Normative ACP Reference:** [acp-session-contract.md](acp-session-contract.md) (produced by P6-A4-WP1, issue [#4053](https://github.com/TalonT-Org/AutoSkillit/issues/4053))
+
+## Overview
+
+This document classifies every surface of the autoskillit backend abstraction
+against the opencode (`github.com/sst/opencode`) coding-agent CLI as the
+N=3 paper backend candidate. The exercise walks:
+
+1. **Preamble** — candidate choice, scope, exclusions, and source citations.
+2. **Protocol Method Classification** — every method on `CodingAgentBackend`
+   and the four sub-protocols (`StreamParser`, `ResultParser`, `EnvPolicy`,
+   `SessionLocator`), classified as TRIVIAL / SHIM-REQUIRED / GAP.
+3. **BackendCapabilities Field Classification** — every field on
+   `BackendCapabilities`, classified using the three-category taxonomy
+   (ACP-Mappable / autoskillit-Local Extension / Forward-Declared) inherited
+   from `acp-session-contract.md` Section 3.
+4. **BackendConventions Classification** — both fields on `BackendConventions`,
+   classified and rationalised against opencode's filesystem layout.
+5. **Conformance Probe Classification** — the B3a (NDJSON/output schema
+   conformance) and B3b (Hook/guard efficacy) probe categories in the
+   context of opencode's output format and `--sandbox deny` blocker.
+6. **Gap Catalogue — Protocol Change Requests** — one named Protocol change
+   request (PCR-NNN) per GAP row across Sections 2–5, with grouping by root
+   cause.
+
+The classification is paper-only — no implementation work is planned until the
+upstream blockers documented in Section 6 are resolved. The exercise's purpose
+is to discover which Protocol surfaces are sufficiently flexible to absorb an
+opencode-class backend without extension, and which surfaces require new
+Protocol affordances or capability flags.
+
+### Candidate choice
+
+opencode is a Go-based coding-agent CLI distinct from Claude Code (a Node.js
+CLI shipping Anthropic's SDK) and Codex (a Rust CLI shipping OpenAI's
+Responses API). The choice isolates three independent axes of variation
+relevant to the autoskillit backend abstraction:
+
+- **Output format** — opencode emits bulk JSON or terminal-formatted output
+  rather than NDJSON, exercising the `StreamParser` / `ResultParser` shim
+  pathway rather than the existing fixture-based conformance probe pattern
+  in `tests/execution/backends/fixtures/codex_ndjson/`.
+- **Session storage** — opencode persists sessions in a SQLite database
+  (`~/.local/share/opencode/opencode.db`) keyed by ULID session IDs
+  (`ses_<ulid>`), exercising the `SessionLocator` shim pathway rather than
+  the per-backend filesystem-locator pattern used by both `ClaudeCodeBackend`
+  and `CodexBackend`.
+- **Sandbox policy** — `opencode run` hard-codes writes to `deny`
+  ([sst/opencode#13851](https://github.com/sst/opencode/issues/13851)),
+  exercising the GAP pathway for every `build_*_cmd` method that requires
+  headless write capability (`build_cmd`, `build_skill_session_cmd`,
+  `build_food_truck_cmd`).
+
+### Source citations
+
+The capability profile in Sections 2–5 derives from three sources:
+
+- **Issue #829** ("Open Code CLI Backend — Deferred") — captured the initial
+  capability profile, including the `--sandbox deny` blocker, SQLite
+  session storage, ULID session IDs, undocumented exit codes, and the
+  exclusion of the community `codex-acp` shim.
+- **`acp-session-contract.md`** (P6-A4-WP1, issue #4053) — provides the
+  normative ACP reference for lifecycle mapping, recovery ladder, and the
+  17 / 6 / 18 capability taxonomy.
+- **External research via `sst/opencode` repository** — confirmed the
+  output-format gap (bulk JSON or terminal, no NDJSON streaming) and the
+  AI SDK multi-provider model (`75+ providers natively via AI SDK`).
+
+> **Note on `external_landscape.md`:** The issue's Technical Step 3
+> references `docs/design/external_landscape.md`. That file does not exist
+> in the repository; equivalent opencode data was sourced from issue #829
+> and web research against the `sst/opencode` repository.
+
+### Explicit exclusions
+
+- **Community `codex-acp` shim** — explicitly excluded from scope. The
+  exercise evaluates native opencode only. Adapters that wrap Codex behind
+  the ACP surface are not a substitute for opencode-class N=3 coverage and
+  would mask the very gaps the exercise is designed to surface.
+- **Implementation work** — the document records classifications and
+  Protocol change requests only. No `OpencodeBackend` class, no fixture
+  files, no conformance probes.
+- **Token field derivation** — token fields stored in SQLite are noted as
+  "extraction pending" rather than modelled, since the opencode schema is
+  undocumented and changes daily with no version guarantee.
+
+---
+
+## Section 1: Preamble
+
+### 1.1 Candidate profile summary
+
+| Dimension | opencode value | Source |
+|---|---|---|
+| Binary | `opencode` | sst/opencode README |
+| Distribution | Single-binary Go release; also `brew install opencode` | sst/opencode README |
+| Provider model | AI SDK with 75+ providers (Anthropic, OpenAI, Google, Bedrock, etc.) | sst/opencode README |
+| Output format | Bulk JSON or terminal-formatted; no NDJSON streaming | issue #829 |
+| Session storage | SQLite (`~/.local/share/opencode/opencode.db`) keyed by ULID (`ses_<ulid>`) | issue #829 |
+| Resume / checkpoint | Undocumented; no `--resume` flag surfaced | issue #829 |
+| Skill injection | Undocumented | issue #829 |
+| Plugin system | None documented | issue #829 |
+| Exit codes | Undocumented; daily release cadence with no schema versioning | issue #829 |
+| Sandbox policy | `opencode run` hard-codes `--sandbox deny` for writes | sst/opencode#13851 |
+
+### 1.2 Normative reference relationship
+
+The `acp-session-contract.md` document (produced by P6-A4-WP1, issue #4053)
+provides:
+
+- **Section 1 (Lifecycle Mapping)** — the canonical 23-row `CodingAgentBackend`
+  method enumeration, which this exercise inherits as Section 2's row order.
+- **Section 2 (Recovery Ladder)** — the `RetryReason`-to-ACP-rung mapping
+  that bounds which rows in Section 6 require new Protocol affordances
+  (e.g., a GAP on `build_resume_cmd` implies a new `RetryReason`-to-rung
+  pathway for opencode's session model).
+- **Section 3 (Capabilities Translation)** — the 17 ACP-Mappable / 6
+  Forward-Declared / 18 autoskillit-Local = 41 total taxonomy that this
+  exercise's Section 3 adopts and re-applies per-field.
+
+### 1.3 Classification legend
+
+Each row in Sections 2–5 carries one of three classifications:
+
+- **TRIVIAL** — the surface accepts the opencode value as-is (or with a
+  constant literal). No adapter code beyond a constructor call.
+- **SHIM-REQUIRED** — the surface accepts the opencode value but requires
+  adapter code to bridge a format mismatch (NDJSON ↔ bulk JSON, filesystem
+  ↔ SQLite, etc.). The Protocol surface is sufficient; only the
+  implementation needs a shim.
+- **GAP** — the Protocol surface itself is insufficient to express the
+  opencode value. A Protocol change request (PCR) is filed in Section 6.
+
+---
+
+## Section 2: Protocol Method Classification
+
+The `CodingAgentBackend` protocol (`src/autoskillit/core/types/_type_protocols_backend.py`,
+23 methods) and the four sub-protocols (`StreamParser`, `ResultParser`,
+`EnvPolicy`, `SessionLocator`; 7 methods total) yield 30 rows below. The row
+order and per-method groupings follow `acp-session-contract.md` Section 1
+(Claude Code and Codex columns are omitted; the rightmost column is the
+opencode classification).
+
+| # | Protocol | Method | Classification | Rationale |
+|---|---|---|---|---|
+| 1 | CodingAgentBackend | `name` | TRIVIAL | Returns `"opencode"`. |
+| 2 | CodingAgentBackend | `capabilities` | TRIVIAL | Returns a `BackendCapabilities(...)` instance per the rows in Section 3. |
+| 3 | CodingAgentBackend | `conventions` | TRIVIAL | Returns a `BackendConventions(...)` instance per the rows in Section 4. |
+| 4 | CodingAgentBackend | `build_cmd` | GAP | `opencode run` hard-codes `--sandbox deny` (sst/opencode#13851); headless write execution is structurally blocked. Maps to PCR-001. |
+| 5 | CodingAgentBackend | `stream_parser` | SHIM-REQUIRED | No NDJSON streaming — adapter must consume bulk JSON or terminal-formatted output and emit `SessionEvent` values. |
+| 6 | CodingAgentBackend | `result_parser` | SHIM-REQUIRED | Aggregates adapted events into `AgentSessionResult`; token fields must be extracted from SQLite since they are not in stdout. |
+| 7 | CodingAgentBackend | `env_policy` | TRIVIAL | Standard env construction; opencode consumes provider config via env vars (e.g., `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`). |
+| 8 | CodingAgentBackend | `session_locator` | SHIM-REQUIRED | Sessions live in SQLite (`ses_<ulid>`), not filesystem JSONL/NDJSON. The three locator methods require a SQLite-backed adapter. |
+| 9 | CodingAgentBackend | `write_tool_names` | SHIM-REQUIRED | opencode's write tool vocabulary is undocumented; needs investigation before the `frozenset` can be populated. Codex returns `frozenset({"apply_patch", "Bash", "run_cmd"})`; opencode likely surfaces a different set. |
+| 10 | CodingAgentBackend | `binary_name` | TRIVIAL | Returns `"opencode"`. |
+| 11 | CodingAgentBackend | `build_resume_cmd` | GAP | No `--resume <session_id>` flag surfaced; SQLite session IDs are not first-class CLI arguments. Maps to PCR-002. |
+| 12 | CodingAgentBackend | `build_skill_session_cmd` | GAP | Blocked by `--sandbox deny`; even if resume were possible, the sandbox constraint forbids skill-session write execution. Maps to PCR-001. |
+| 13 | CodingAgentBackend | `build_food_truck_cmd` | GAP | Same sandbox constraint blocks orchestrator-level L2 sessions. Maps to PCR-001 and PCR-003. |
+| 14 | CodingAgentBackend | `build_interactive_cmd` | SHIM-REQUIRED | `opencode run` exists for interactive use; sandbox restrictions apply but interactive flows can tolerate a tighter sandbox. |
+| 15 | CodingAgentBackend | `validate_session_layout` | TRIVIAL | Returns `[]` (Codex pattern); no session directory to validate when sessions are SQLite-backed. |
+| 16 | CodingAgentBackend | `validate_skill_content` | TRIVIAL | Returns `[]` (Codex pattern); no frontmatter requirement when no skill injection is documented. |
+| 17 | CodingAgentBackend | `version` | TRIVIAL | `opencode --version` returns the release string; daily cadence but parseable. |
+| 18 | CodingAgentBackend | `list_plugins` | TRIVIAL | Returns `[]`; no plugin system documented. |
+| 19 | CodingAgentBackend | `ensure_pre_launch` | SHIM-REQUIRED | Pre-launch check that the SQLite database is initialised and the configured provider env vars are present. |
+| 20 | CodingAgentBackend | `translate_model` | SHIM-REQUIRED | AI SDK model identifiers differ from autoskillit's canonical aliases (`sonnet` / `opus` / `haiku`). Mapping table needed; precedent is `CODEX_MODEL_ALIASES` in `_type_backend.py:174–176`. |
+| 21 | CodingAgentBackend | `model_config_overrides` | TRIVIAL | Returns `()` or a minimal per-model override tuple (analogous to Codex's effort-mapping for `sonnet`/`opus`/`haiku`). |
+| 22 | CodingAgentBackend | `build_inspector_cmd` | TRIVIAL | Raises `CapabilityNotSupportedError` (same as both existing backends; `inspector_capable=False`). |
+| 23 | CodingAgentBackend | `setup_session_dir` | SHIM-REQUIRED | Bridges SQLite session storage to the filesystem session dir contract; copies the SQLite session row to a session-dir stub so downstream consumers see a directory. |
+| 24 | StreamParser | `parse_line` | SHIM-REQUIRED | Adapts non-NDJSON output (bulk JSON or terminal blocks) to `SessionEvent`. Per-line semantics do not apply; the parser operates on whole-output blocks. |
+| 25 | ResultParser | `parse_result` | SHIM-REQUIRED | Aggregates adapted `SessionEvent` values into `AgentSessionResult`; identical to existing backends once events are shaped. |
+| 26 | ResultParser | `parse_stdout` | SHIM-REQUIRED | Bulk JSON or terminal output parsing; token extraction deferred to SQLite lookup. |
+| 27 | EnvPolicy | `build_env` | TRIVIAL | Standard env construction; no denylist required unless opencode leaks through sandbox-bypass env vars. |
+| 28 | SessionLocator | `locate_session` | SHIM-REQUIRED | SQLite query by ULID session ID; returns a synthetic `Path` (e.g., the SQLite DB path) for callers that expect filesystem semantics. |
+| 29 | SessionLocator | `project_log_dir` | SHIM-REQUIRED | Maps to `~/.local/share/opencode/` (the SQLite directory); per-project logic is opaque since opencode is global-session. |
+| 30 | SessionLocator | `session_log_path` | SHIM-REQUIRED | SQLite-backed; no real path exists. Returning the SQLite DB path is the closest analogue. |
+
+### Method-cluster rollup
+
+| Cluster | TRIVIAL | SHIM-REQUIRED | GAP |
+|---|---|---|---|
+| Identity / introspection (`name`, `capabilities`, `conventions`, `binary_name`, `version`, `list_plugins`, `model_config_overrides`, `build_inspector_cmd`, `translate_model`) | 8 | 1 | 0 |
+| Command builders (`build_cmd`, `build_skill_session_cmd`, `build_food_truck_cmd`, `build_interactive_cmd`, `build_resume_cmd`) | 0 | 1 | 4 |
+| Validation (`validate_session_layout`, `validate_skill_content`, `ensure_pre_launch`) | 2 | 1 | 0 |
+| Parsing (`stream_parser`, `result_parser`) | 0 | 2 | 0 |
+| Session storage (`session_locator`, `setup_session_dir`, `write_tool_names`) | 0 | 3 | 0 |
+| Environment (`env_policy`) | 1 | 0 | 0 |
+| Sub-protocols (`StreamParser`, `ResultParser`, `EnvPolicy`, `SessionLocator`) | 1 | 6 | 0 |
+| **Total (30 rows)** | **12** | **14** | **4** |
+
+The 4 GAP rows map to 3 distinct root causes (sandbox blocker, resume
+absence, orchestrator session absence), consolidated into PCR-001, PCR-002,
+and PCR-003 in Section 6.
+
+---
+
+## Section 3: BackendCapabilities Field Classification
+
+The taxonomy below mirrors `acp-session-contract.md` Section 3:
+
+- **ACP-Mappable** — 17 fields with a direct or close ACP analogue.
+- **autoskillit-Local Extension** — 18 fields with no ACP analogue but
+  required for autoskillit's extended contract.
+- **Forward-Declared** — 6 fields with no current production consumer
+  outside the exemption set in `_FORWARD_DECLARED`
+  (`tests/arch/test_capability_consumption.py:26–63`).
+
+All 41 fields appear in exactly one row below. The "opencode value" column
+records what the field would hold for an `OpencodeBackend.capabilities`
+instance; the classification column is TRIVIAL / SHIM-REQUIRED / GAP with the
+same legend as Section 2.
+
+### 3.1 ACP-Mappable fields (17)
+
+| # | Field | opencode value | Classification | Rationale |
+|---|---|---|---|---|
+| 1 | `channel_b_capable` | `False` | TRIVIAL | No side-channel JSONL log; SQLite is the only persistence surface. Same shape as Codex. |
+| 2 | `pty_required` | `False` | TRIVIAL | opencode's pipe-based stdio; no pseudo-TTY allocation required. |
+| 3 | `session_resume_capable` | `False` | GAP | No documented resume mechanism; `build_resume_cmd` is a GAP. Maps to PCR-002. |
+| 4 | `skill_injection_capable` | `False` | GAP | No documented `--add-dir` or `--plugin-dir` equivalent. Maps to PCR-004. |
+| 5 | `supports_claude_format_stdout` | `False` | TRIVIAL | opencode emits bulk JSON or terminal-formatted output; not Claude format. |
+| 6 | `exit_code_is_terminal` | `False` | GAP | Exit codes are undocumented; cannot assert terminal semantics. Maps to PCR-005. |
+| 7 | `mcp_config_capable` | `False` (pending) | SHIM-REQUIRED | opencode may have an MCP-style config; research pending. Maps to PCR-006. |
+| 8 | `food_truck_capable` | `False` | GAP | Blocked by `--sandbox deny` and absence of orchestrator-level session support. Maps to PCR-001 / PCR-003. |
+| 9 | `completion_record_types` | `frozenset()` (pending adapter) | SHIM-REQUIRED | Depends on the bulk-JSON or terminal-output adapter from Section 2 row 5; cannot enumerate record types until the adapter is shaped. |
+| 10 | `session_record_types` | `frozenset()` (pending adapter) | SHIM-REQUIRED | Same dependency as row 9. |
+| 11 | `triage_capable` | `False` | GAP | No lightweight probe mechanism documented. Maps to PCR-007. |
+| 12 | `supports_context_exhaustion_detection` | `False` | GAP | No documented exhaustion signal in stdout or SQLite. Maps to PCR-008. |
+| 13 | `process_name` | `"opencode"` | TRIVIAL | Binary stem. |
+| 14 | `process_name_aliases` | `frozenset({"opencode"})` | TRIVIAL | Same as `process_name` for a non-shebang binary. |
+| 15 | `replay_capable` | `False` | GAP | No recording/replay support documented. Maps to PCR-009. |
+| 16 | `record_capable` | `False` | GAP | Same as row 15. Maps to PCR-009. |
+| 17 | `inspector_capable` | `False` | TRIVIAL | Same as both existing backends; raises `CapabilityNotSupportedError`. |
+
+### 3.2 autoskillit-Local Extension fields (18)
+
+| # | Field | opencode value | Classification | Rationale |
+|---|---|---|---|---|
+| 18 | `project_local_skills_capable` | `False` | GAP | No project-local skill discovery mechanism. Maps to PCR-004. |
+| 19 | `supports_tool_list_changed` | `True` | TRIVIAL | Default; kitchen reveal timing is unaffected. |
+| 20 | `required_skill_fields` | `frozenset()` | TRIVIAL | No frontmatter requirement when no skill injection exists. |
+| 21 | `applicable_guards` | `frozenset()` | TRIVIAL | No hook system documented; guards list is empty. |
+| 22 | `write_guard_tool_names` | `frozenset()` (pending vocabulary) | SHIM-REQUIRED | Same dependency as Section 2 row 9; vocabulary must be enumerated before the frozenset can be populated. |
+| 23 | `env_denylist_prefixes` | `()` | TRIVIAL | No env scrubbing required unless opencode leaks sensitive env vars; defaults to empty. |
+| 24 | `version_check_command` | `"opencode --version"` | TRIVIAL | Same pattern as Claude Code / Codex. |
+| 25 | `skills_subdir` | `"skills"` | TRIVIAL | See Section 4 caveat — the value is meaningless while `skill_injection_capable=False`. |
+| 26 | `hook_config_format` | `""` | TRIVIAL | No hook config; empty string. |
+| 27 | `write_detection_strategy` | `""` (pending) | SHIM-REQUIRED | opencode's write evidence format is unknown; strategy cannot be set until the adapter surfaces write events. Maps to PCR-010. |
+| 28 | `default_skill_sandbox_mode` | `"workspace-write"` | GAP | Mismatch: opencode hard-codes `--sandbox deny`; the field value is aspirational but cannot be honoured at runtime. Maps to PCR-001. |
+| 29 | `anthropic_provider_capable` | `False` | TRIVIAL | Multi-provider via AI SDK; not Anthropic-exclusive. |
+| 30 | `plugin_install_capable` | `False` | TRIVIAL | No plugin system. |
+| 31 | `supports_context_window_suffix` | `False` | TRIVIAL | AI SDK model identifiers do not carry `[1m]`-style suffixes. |
+| 32 | `has_unguarded_filesystem_access` | `True` (pending adapter) | SHIM-REQUIRED | Default `True` matches Codex behaviour; gated prompt supplements apply until the write-detection adapter surfaces the per-tool evidence needed to gate more precisely. |
+| 33 | `git_metadata_writable` | `True` | TRIVIAL | opencode's `--sandbox deny` is write-blocked at the user level, not at the kernel level; `.git/worktrees/` remains writable. |
+| 34 | `skill_sigil` | `"$"` (defaulted) | SHIM-REQUIRED | No documented invocation prefix; defaulting to Codex's `"$"` until research surfaces an opencode-native convention. Maps to PCR-011. |
+| 35 | `session_dir_persistent` | `True` | SHIM-REQUIRED | SQLite-backed sessions are persistent by construction; filesystem bridge must produce a persistent dir (analogous to Codex's `session_dir_persistent=True`). |
+
+### 3.3 Forward-Declared fields (6)
+
+| # | Field | Forward-declared issue | opencode value | Classification | Rationale |
+|---|---|---|---|---|---|
+| 36 | `supports_thinking_blocks` | #3497 (Claude + Codex pre-existing) | `False` | TRIVIAL | No thinking-block format documented. |
+| 37 | `required_session_files` | #3134 (Codex pre-existing) | `frozenset()` | TRIVIAL | No session-dir contract; SQLite-only. |
+| 38 | `session_dir_symlinks` | #3134 (Codex pre-existing) | `frozenset()` | TRIVIAL | No symlink targets. |
+| 39 | `patch_format` | #3776 (Codex pre-existing) | `""` | SHIM-REQUIRED | Write-detection adapter must surface patch evidence before a format string can be set. Maps to PCR-010. |
+| 40 | `min_version` | #3122 (Codex pre-existing) | `""` | TRIVIAL | Version policy deferred; daily cadence makes pinning fragile. |
+| 41 | `mcp_env_forward_vars` | #3458 (Codex pre-existing) | `frozenset()` | TRIVIAL | No MCP wiring; empty set. |
+
+### 3.4 Capabilities rollup
+
+| Category | TRIVIAL | SHIM-REQUIRED | GAP |
+|---|---|---|---|
+| ACP-Mappable (17) | 6 | 3 | 8 |
+| autoskillit-Local Extension (18) | 11 | 5 | 2 |
+| Forward-Declared (6) | 5 | 1 | 0 |
+| **Total (41)** | **22** | **9** | **10** |
+
+The 10 GAP rows consolidate into 9 PCRs (PCR-001 through PCR-009, plus the
+sandbox-policy mismatch in row 28 maps to PCR-001); the 9 SHIM-REQUIRED rows
+consolidate into PCR-006 / PCR-010 / PCR-011 plus the opencode-vocabulary
+adapter family.
+
+---
+
+## Section 4: BackendConventions Classification
+
+Both fields on `BackendConventions` (`src/autoskillit/core/types/_type_backend.py:38–49`)
+appear below. The classification legend matches Section 2.
+
+| # | Field | Classification | opencode value | Rationale |
+|---|---|---|---|---|
+| 1 | `skills_subdir` | TRIVIAL | `Path("skills")` | Conventional default; opencode could use `.opencode/skills/` or `.agents/skills/` if a skill-injection mechanism emerges. Caveat: the value is only meaningful if `skill_injection_capable=True` — which is a GAP for opencode. Setting it now is harmless but inert. |
+| 2 | `project_local_skill_search_dirs` | GAP | `()` | opencode has no documented project-local skill discovery. A Protocol extension that surfaces a registration hook (rather than a directory scan) would be required. Maps to PCR-004. |
+
+The dependency between row 1 (`TRIVIAL`) and the `skill_injection_capable`
+GAP (Section 3 row 4) is significant: the `skills_subdir` value can be set
+without Protocol change, but no consumer can read it until PCR-004 lands.
+
+---
+
+## Section 5: Conformance Probe Classification
+
+The autoskillit test suite carries two probe categories applicable to
+backend output formats and write-detection hooks. Each is classified against
+opencode's documented behaviour below.
+
+### 5.1 B3a — NDJSON / output schema conformance probes
+
+**Existing pattern.** `tests/execution/backends/fixtures/codex_ndjson/`
+contains sealed NDJSON fixtures that the Codex `StreamParser` is validated
+against. Each fixture is a JSONL file with named scenarios (e.g.,
+`context_exhaustion.jsonl`, `thread_started.jsonl`); the test deserialises
+each line, runs it through `parse_line`, and asserts the resulting
+`SessionEvent` matches the expected shape. New backends with NDJSON output
+add a fixture directory under `tests/execution/backends/fixtures/`.
+
+**opencode classification: GAP.** opencode has no NDJSON streaming — output
+is either bulk JSON or terminal-formatted. The fixture-based conformance
+probe category requires a new output-format adapter (Section 2 rows 5, 24)
+before any probe can be authored. The probe category itself remains valid
+as a contract surface; only the authoring pattern changes.
+
+**Required work before B3a probes can land:**
+
+1. Define a sealed output-format schema for opencode (bulk JSON or terminal
+   block) — analogous to the Codex NDJSON schema.
+2. Author fixture files under `tests/execution/backends/fixtures/opencode_<format>/`
+   with the same scenario coverage as the Codex suite (context exhaustion,
+   turn completion, turn failure, etc.).
+3. Extend `tests/execution/backends/test_stream_parser_conformance.py`
+   (or its successor) with an `OpencodeStreamParser` fixture-parameterised
+   branch.
+
+Maps to PCR-010 (output-format adapter) and PCR-012 (probe-authoring pattern).
+
+### 5.2 B3b — Hook / guard efficacy probes
+
+**Existing pattern.** `tests/execution/backends/test_hook_deny_efficacy_probe.py`
+validates that the backend fires `PreToolUse` / `PostToolUse` hooks for write
+operations, with the Codex-specific path labelled B4v. The probe asserts that
+the guard fires deterministically for known write tool calls.
+
+**opencode classification: GAP.** opencode's `--sandbox deny` hard-coding
+(sst/opencode#13851) means hook firing for write operations cannot be
+validated until the upstream blocker is resolved. Two failure modes apply:
+
+- Even with hooks configured, `--sandbox deny` may short-circuit writes
+  before the hook fires.
+- The hook config file path and format are undocumented; opencode may use a
+  different mechanism (e.g., shell-level sandbox) that bypasses the hook
+  surface entirely.
+
+Maps to PCR-001 (sandbox-write enablement) and PCR-013 (hook config
+discovery).
+
+### 5.3 Probe-category rollup
+
+| Probe category | TRIVIAL | SHIM-REQUIRED | GAP |
+|---|---|---|---|
+| B3a (NDJSON / output schema) | 0 | 0 | 1 |
+| B3b (Hook / guard efficacy) | 0 | 0 | 1 |
+| **Total** | **0** | **0** | **2** |
+
+Both probe categories are pure GAPs for opencode, blocking the corresponding
+test classes until upstream blockers resolve.
+
+---
+
+## Section 6: Gap Catalogue — Protocol Change Requests
+
+Each GAP row from Sections 2–5 produces a Protocol Change Request (PCR)
+entry below. Related rows that share a root cause are consolidated into a
+single PCR with multiple affected surfaces. The PCR ID is sequential.
+
+| PCR | Name | Source section | Gap description | Protocol impact | Upstream dependency | Precedent |
+|---|---|---|---|---|---|---|
+| PCR-001 | sandbox-write-enablement | §2 rows 4, 12, 13; §3 row 8, 28 | `opencode run` hard-codes `--sandbox deny`; every `build_*_cmd` method that requires headless write capability (`build_cmd`, `build_skill_session_cmd`, `build_food_truck_cmd`) is structurally blocked. `food_truck_capable=False` and `default_skill_sandbox_mode="workspace-write"` are aspirational only. | **Protocol change**: a new capability flag (`sandbox_policy_override_capable: bool`) that distinguishes backends whose sandbox can be relaxed via a flag from those whose sandbox is hard-coded. `build_*_cmd` returns must distinguish "capability blocked" from "config rejected". | sst/opencode#13851 | Codex applies `--dangerously-bypass-approvals-and-sandbox` (`codex.py` lines 98–107); Claude Code applies `--dangerously-skip-permissions`. Both have explicit bypass flags; opencode does not. |
+| PCR-002 | session-resume-mechanism | §2 row 11; §3 row 3 | No documented `--resume <session_id>` flag; SQLite session IDs are not first-class CLI arguments. `build_resume_cmd` cannot produce a `CmdSpec` that targets an existing session. `session_resume_capable=False`. | **Protocol change**: a new `ResumeSpec` variant (`SqliteBackedResume(session_db_path: Path, session_id: str)`) under `_type_resume.py`, or a generalised `backend-specific-resume` callable on the Protocol surface. | opencode resume feature | Claude Code uses `--resume <session_id>` flag; Codex uses `resume <session_id>` positional subcommand (`codex.py` lines 978–979). |
+| PCR-003 | orchestrator-session-support | §2 row 13; §3 row 8 | `build_food_truck_cmd` requires orchestrator-level L2 session support (driven by fleet dispatch); opencode has no equivalent and the sandbox blocker compounds. `food_truck_capable=False`. | **Protocol change**: a new `food_truck_capable` distinction between "no orchestrator support at all" (`False`) and "blocked by sandbox" (new variant or sub-flag). Same proposal as PCR-001 — likely merged into a single new capability. | sst/opencode#13851 | Codex supports food-truck via `build_food_truck_cmd` (lines 821–823) but discards `plugin_source` / `output_format` / `exit_after_stop_delay_ms`. |
+| PCR-004 | skill-registration-discovery | §2 rows 9, 12; §3 row 4, 18; §4 row 2 | No documented `--add-dir` / `--plugin-dir` / skill-discovery mechanism. `skill_injection_capable=False`, `project_local_skills_capable=False`, `project_local_skill_search_dirs=()`. | **Protocol change**: a new `skill_registration: Callable[[SkillManifest], None]` field on the Protocol, decoupling registration from CLI flags. | opencode skill-injection feature | Claude Code uses `--add-dir` / `--plugin-dir`; Codex has no equivalent and uses an empty frozenset. |
+| PCR-005 | exit-code-semantics | §3 row 6 | Exit codes are undocumented; daily release cadence with no schema versioning. `exit_code_is_terminal=False` is conservative but may yield false positives when opencode exits with a non-zero status on benign conditions. | **Protocol change**: a new `exit_code_is_terminal: Literal["true","false","unknown"]` (string union) to distinguish "verified terminal" from "undocumented". | opencode exit-code documentation | Codex has verified semantics (`exit_code_is_terminal=True`); Claude Code has verified semantics (`False`). |
+| PCR-006 | mcp-config-wiring | §3 row 7 | opencode may have an MCP-style config; the absence of documentation makes `mcp_config_capable` ambiguous. SHIM-REQUIRED rather than GAP because the Protocol surface is sufficient; the implementation simply awaits research. | **No Protocol change**. Implementation work only. | opencode MCP documentation | Codex: `mcp_config_capable=True`; Claude Code: `False`. |
+| PCR-007 | triage-probe-mechanism | §3 row 11 | No lightweight probe mechanism documented. `triage_capable=False` blocks `_llm_triage.py` from dispatching to opencode. | **Protocol change**: a new `triage_cmd_builder: Callable[[str], CmdSpec]` field that backends can implement to expose lightweight probes without requiring a full skill session. | opencode probe feature | Claude Code uses `claude -p` for triage (`_llm_triage.py`); Codex has no triage path. |
+| PCR-008 | context-exhaustion-signal | §3 row 12 | No documented exhaustion signal in stdout or SQLite. `supports_context_exhaustion_detection=False` blocks Codex-style `_headless_result.py` jsonl-context-exhausted handling for opencode. | **Protocol change**: a new `context_exhaustion_signal: Literal["stdout","sqlite","none"]` to distinguish the detection surface. | opencode exhaustion signal | Codex surfaces via `error_code == CODEX_CONTEXT_EXHAUSTION_MARKER` (`_headless_evidence.py` lines 105–108). |
+| PCR-009 | recording-replay-support | §3 rows 15, 16 | No recording/replay support documented. `replay_capable=False`, `record_capable=False` blocks api_simulator-based REPLAY_SCENARIO / RECORD_SCENARIO runners. | **No Protocol change**. Implementation work only — api_simulator can wrap opencode once the output adapter (PCR-010) lands. | opencode output-format documentation | Claude Code: `replay_capable=True`, `record_capable=True`. Codex: not yet wired. |
+| PCR-010 | output-format-adapter | §2 rows 5, 24, 26; §3 rows 9, 10, 27, 39; §5 B3a | opencode emits bulk JSON or terminal-formatted output, not NDJSON. The `StreamParser` / `ResultParser` pair require an adapter; `completion_record_types`, `session_record_types`, `write_detection_strategy`, `patch_format` all depend on the adapter's surface. | **Protocol change**: a new `output_format: Literal["ndjson","bulk_json","terminal"]` field on the Protocol so consumers can dispatch on the actual format rather than assuming NDJSON. | opencode output-format documentation | Both existing backends emit NDJSON; this is the first non-NDJSON candidate. |
+| PCR-011 | skill-sigil-convention | §3 row 34 | No documented invocation prefix. Defaulting to Codex's `"$"` until research surfaces an opencode-native convention. | **No Protocol change**. Implementation choice only. | opencode invocation convention | Claude Code uses `"/"`; Codex uses `"$"`. |
+| PCR-012 | conformance-probe-authoring | §5 B3a | The fixture-based conformance probe pattern (`tests/execution/backends/fixtures/codex_ndjson/`) assumes NDJSON; opencode requires a new authoring pattern. | **No Protocol change**. Test infrastructure extension only. | PCR-010 | Codex NDJSON fixtures. |
+| PCR-013 | hook-config-discovery | §5 B3b | Hook config file path and format are undocumented; opencode may use shell-level sandbox or another mechanism that bypasses the autoskillit hook surface. | **Protocol change**: a new `hook_config_paths: tuple[Path, ...]` field on `BackendCapabilities` that backends populate to declare the files the probe should write and observe. | opencode hook documentation | Codex: `hook_config_format="toml_nested"` (forward-declared); Claude Code: `hook_config_format=""`. |
+
+### 6.1 PCR rollup
+
+| Classification | Count |
+|---|---|
+| Protocol change required | 9 (PCR-001, 002, 003, 004, 005, 007, 008, 010, 013) |
+| Implementation-only | 3 (PCR-006, 009, 011) |
+| Test-infrastructure extension only | 1 (PCR-012) |
+| **Total** | **13 PCRs** |
+
+Note: PCR-001 and PCR-003 share the upstream sandbox blocker; a single
+sandbox-write-enablement change request in `sst/opencode` would unblock
+both. PCR-005 and PCR-008 are documentation-driven and may resolve without
+Protocol change once `sst/opencode` publishes exit-code and exhaustion
+semantics.
+
+### 6.2 Cross-reference summary
+
+| Section | Source of truth | File |
+|---|---|---|
+| §2 method enumeration | `CodingAgentBackend` Protocol + 4 sub-protocols | `src/autoskillit/core/types/_type_protocols_backend.py` |
+| §3 capability taxonomy | `BackendCapabilities` (41 fields) + `_FORWARD_DECLARED` | `src/autoskillit/core/types/_type_backend.py`, `tests/arch/test_capability_consumption.py` |
+| §3 capability categories | ACP-Mappable / autoskillit-Local / Forward-Declared counts | `docs/design/acp-session-contract.md` Section 3 |
+| §4 conventions | `BackendConventions` (2 fields) | `src/autoskillit/core/types/_type_backend.py:38–49` |
+| §5 B3a pattern | Codex NDJSON fixtures | `tests/execution/backends/fixtures/codex_ndjson/` |
+| §5 B3b pattern | Codex hook-efficacy probe | `tests/execution/backends/test_hook_deny_efficacy_probe.py` |
+| §6 PCR-001 / 003 blocker | `--sandbox deny` hard-coding | sst/opencode#13851 |
+| §6 PCR-002 / 005 / 008 documentation | Resume, exit codes, exhaustion signals | sst/opencode repository |
\ No newline at end of file
diff --git a/tests/docs/test_filename_naming.py b/tests/docs/test_filename_naming.py
index 4970ee2c5..0ffece4e9 100644
--- a/tests/docs/test_filename_naming.py
+++ b/tests/docs/test_filename_naming.py
@@ -22,6 +22,7 @@
         "recording-replay-accepted-degradations.md",
         "0001-prohibit-background-subagent-execution.md",
         "0002-ban-inline-shell-scripts-from-cmd.md",
+        "paper-backend-n3-exercise.md",  # 4 segments; prescribed by issue #4052
     },
 }