Implementation Plan: T5-P6-A5-WP1 Forward-Obligation Readiness Notes (#4125)

## Summary

Create the `docs/design/forward-obligations/` directory containing four
contract-impact markdown files (inspector, recording/replay, triage
portability, heterogeneous routing) and update `docs/design/README.md`
with table rows for each new document. This is a documentation-only
deliverable — no source code files are created or modified.

Closes #4054

## Implementation Plan

Plan file:
`/home/talon/projects/autoskillit-runs/impl-20260627-043714-780997/.autoskillit/temp/make-plan/T5-P6-A5-WP1-forward-obligation-readiness-notes_plan_2026-06-27_044600.md`

🤖 Generated with [Claude Code](https://claude.com/claude-code) via
AutoSkillit
<!-- autoskillit:pipeline-signature
steps=prepare_pr,run_arch_lenses,compose_pr,annotate_pr_diff,review_pr
-->

## Token Usage Summary

| Step | Model | count | uncached | output | cache_read | peak_ctx |
turns | cache_write | time |

|------|-------|-------|----------|--------|------------|----------|-------|-------------|------|
| plan* | opus[1m] | 1 | 4.3k | 9.6k | 411.2k | 55.1k | 20 | 62.3k | 4m
35s |
| verify* | sonnet | 1 | 1.8k | 9.5k | 545.8k | 55.5k | 27 | 36.9k | 5m
17s |
| implement* | MiniMax-M3 | 1 | 53.0k | 9.3k | 1.4M | 0 | 47 | 0 | 2m
40s |
| fix* | sonnet | 1 | 150 | 6.2k | 869.5k | 59.0k | 48 | 40.5k | 3m 30s
|
| audit_impl* | sonnet | 1 | 44 | 11.5k | 164.3k | 43.0k | 14 | 37.2k |
5m 34s |
| prepare_pr* | MiniMax-M3 | 1 | 44.1k | 2.1k | 195.0k | 45.5k | 13 | 0
| 44s |
| compose_pr* | MiniMax-M3 | 1 | 36.2k | 1.2k | 283.4k | 0 | 13 | 0 |
40s |
| **Total** | | | 139.6k | 49.3k | 3.9M | 59.0k | | 176.9k | 23m 2s |

\* *Step used a non-Anthropic provider; caching behavior may differ.*

## Token Efficiency

| Step | LoC Changed | cache_read/LoC | cache_write/LoC | output/LoC |
|------|-------------|----------------|-----------------|------------|
| plan | 0 | — | — | — |
| verify | 0 | — | — | — |
| implement | 259 | 5521.8 | 0.0 | 35.7 |
| fix | 6 | 144912.8 | 6747.2 | 1026.3 |
| audit_impl | 0 | — | — | — |
| prepare_pr | 0 | — | — | — |
| compose_pr | 0 | — | — | — |
| **Total** | **265** | 14714.5 | 667.7 | 186.0 |

## Model Usage Breakdown

| Model | steps | uncached | output | cache_read | cache_write | time |
|-------|-------|----------|--------|------------|-------------|------|
| opus[1m] | 1 | 4.3k | 9.6k | 411.2k | 62.3k | 4m 35s |
| sonnet | 3 | 2.0k | 27.1k | 1.6M | 114.7k | 14m 23s |
| MiniMax-M3 | 3 | 133.3k | 12.6k | 1.9M | 0 | 4m 3s |

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: Trecek

docs/design/README.md

@@ -7,3 +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 |
+| [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 |
+| [forward-obligations/heterogeneous-routing-impact.md](forward-obligations/heterogeneous-routing-impact.md) | Contract-impact note for heterogeneous per-item backend routing: `ToolContext.backend` single-backend constraint, `BackendSelector` proposal, `food_truck_capable` gate |

docs/design/forward-obligations/heterogeneous-routing-impact.md

@@ -0,0 +1,59 @@
+# Heterogeneous Per-Item Backend Routing Contract Impact
+
+| Field | Value |
+|-------|-------|
+| Status | Draft |
+| Date | 2026-06-27 |
+
+## Obligation
+
+Heterogeneous per-item backend routing means individual work-items in a fleet campaign can
+be dispatched to different backends (e.g., item A to `ClaudeCodeBackend`, item B to
+`CodexBackend`) rather than all items sharing a single campaign backend.
+
+## Current State
+
+- `BACKEND_REGISTRY` in `execution/backends/__init__.py:46–49` maps string names to backend
+  classes: `{"claude-code": ClaudeCodeBackend, "codex": CodexBackend}`.
+- `ToolContext.backend` (`pipeline/context.py:164`) is a single
+  `CodingAgentBackend | None` instance — one backend per session.
+- Fleet dispatch reads `ctx.backend` once per campaign.
+- `dispatch_food_truck()` (`execution/headless/__init__.py:349`) resolves
+  `self._ctx.backend` at line 416 — no per-item routing. It raises `RuntimeError` if
+  `food_truck_capable` is `False` (line 385).
+- `run_headless_core` supports a per-step `backend_override` string (lines 198–199) that
+  resolves to a local `step_backend` via `get_backend()`, but this override is call-local
+  and does not mutate `ctx.backend`.
+
+## Contract Gaps
+
+1. **`ToolContext.backend` is a single-backend constraint.** Heterogeneous routing requires
+   either making `backend` a per-item selection or introducing a routing callable.
+2. **`build_food_truck_cmd` and `dispatch_food_truck` assume a single backend.** The entire
+   fleet dispatch path reads `ctx.backend` without per-item indirection.
+3. **`BackendCapabilities.food_truck_capable`** (`_type_backend.py:80`) is the existing gate.
+   Both `ClaudeCodeBackend` (`CLAUDE_CODE_CAPABILITIES:226`) and `CodexBackend`
+   (`codex.py:556`) set it to `True`.
+
+## Required CodingAgentBackend Contract Additions
+
+1. Define a backend routing selector type:
+   `BackendSelector = Callable[[WorkItem], CodingAgentBackend]`
+2. `food_truck_capable` must remain `True` only on backends that satisfy the full dispatch
+   contract (command construction, output parsing, session lifecycle).
+3. `dispatch_food_truck` must accept a selector or per-item backend resolution rather than
+   reading `ctx.backend` unconditionally.
+
+## Key Constraint
+
+`BackendCapabilities` is a frozen dataclass. Adding a routing-selector field would need to
+be forward-declared with a tracking issue (added to `_FORWARD_DECLARED` in
+`test_capability_consumption.py`) until the fleet routing layer is refactored to consume it.
+
+## Key References
+
+- `execution/backends/__init__.py:46–49` — `BACKEND_REGISTRY`
+- `execution/headless/__init__.py:349` — `dispatch_food_truck`
+- `pipeline/context.py:164` — `ToolContext.backend`
+- `core/types/_type_backend.py:80` — `BackendCapabilities.food_truck_capable`
+- `fleet/` — campaign dispatch (reads `ctx.backend` per campaign)

docs/design/forward-obligations/inspector-contract-impact.md

@@ -0,0 +1,63 @@
+# Health Inspector Contract Impact
+
+| Field | Value |
+|-------|-------|
+| Status | Draft |
+| Date | 2026-06-27 |
+| Linked Issue | #3534 |
+
+## Obligation
+
+Both `ClaudeCodeBackend` and `CodexBackend` raise `CapabilityNotSupportedError` from
+`build_inspector_cmd()` when `inspector_capable` is `False` (the current default), and
+raise `AssertionError` if `inspector_capable` were ever set to `True` without a real
+implementation:
+
+- `claude.py:872–876` — `ClaudeCodeBackend.build_inspector_cmd()`
+- `codex.py:1117–1121` — `CodexBackend.build_inspector_cmd()`
+
+The Health Inspector feature (#3534) requires a real implementation on each backend.
+
+## Contract Gaps
+
+- `BackendCapabilities.inspector_capable` is `False` on both backends:
+  - `CLAUDE_CODE_CAPABILITIES` (`_type_backend.py:253`): `inspector_capable=False`
+  - `CodexBackend.capabilities` property (`codex.py:584`): `inspector_capable=False`
+- Neither backend provides a working `build_inspector_cmd()` — both have guard-then-assert
+  stubs.
+
+## Protocol Surface
+
+Already defined in `core/types/_type_inspector.py`:
+
+- **`InspectorEvidence`** (lines 15–24, frozen dataclass): `idle_seconds`, `stdout_path`,
+  `jsonl_lines`, `cpu_trend`, `rss_trend`, `connection_summary`,
+  `execution_marker_present`, `dispatch_context`
+- **`InspectorVerdict`** (lines 27–32, frozen dataclass): `action`, `reasoning`,
+  `confidence`, `elapsed_seconds`
+- **`InspectorCallback`** (line 35): `Callable[[InspectorEvidence], Awaitable[InspectorVerdict]]`
+
+The callback is threaded through `execution/process/_process_race.py`: the parameter is
+declared at line 189, guarded at 253, invoked at 279, and a timeout warning emitted at 281.
+Related `inspector_verdict` fields (lines 75, 102, 116) and inspector-state transitions
+(lines 262, 265, 287, 295, 540) complete the protocol surface in the same file. The callback
+also appears in `recording.py` at lines 141 and 432.
+The wiring point is `execution/headless/_headless_execute.py:270` where
+`inspector_callback=None` is hardcoded (no backend currently provides an inspector).
+
+## Required Steps
+
+1. Implement `build_inspector_cmd()` on `ClaudeCodeBackend` — return a `CmdSpec` that
+   launches an LLM-backed inspector subprocess.
+2. Implement `build_inspector_cmd()` on `CodexBackend` — or leave as `CapabilityNotSupportedError`
+   if Codex cannot support inspection.
+3. Set `inspector_capable=True` on backends with a working implementation.
+4. Wire the callback into `_process_race._run_race()` via the `inspector_callback` parameter
+   (currently hardcoded to `None` in `_headless_execute.py:270`).
+
+## Arch Enforcement
+
+`test_capability_consumption.py` (`tests/arch/`) verifies every `BackendCapabilities` field
+has a production consumer. `inspector_capable` already satisfies this via the
+`self.capabilities.inspector_capable` guard in both backends' `build_inspector_cmd()` methods.
+The field is NOT in the `_FORWARD_DECLARED` exemption set.

docs/design/forward-obligations/recording-replay-impact.md

@@ -0,0 +1,68 @@
+# Recording/Replay Contract Impact
+
+| Field | Value |
+|-------|-------|
+| Status | Draft |
+| Date | 2026-06-27 |
+
+## Obligation
+
+Recording/replay must remain backend-isolated. Cross-backend cassette replay is explicitly
+a non-goal (see `docs/design/recording-replay-accepted-degradations.md`).
+
+## BackendCapabilities Fields
+
+- `replay_capable` and `record_capable` are defined on `BackendCapabilities`
+  (`_type_backend.py:127–129`), both defaulting to `False`.
+- `ClaudeCodeBackend`: both set to `True` via `CLAUDE_CODE_CAPABILITIES`
+  (`_type_backend.py:249–250`).
+- `CodexBackend`: both set to `False` (`codex.py:580–581`). Codex replay is handled via
+  `CodexScenarioPlayer`.
+
+## Format Detection
+
+`_detect_backend_format()` in `recording.py:56–59` routes by the presence of
+`*/codex_stdout.ndjson` inside the scenario directory:
+
+```python
+def _detect_backend_format(scenario_dir: Path) -> str:
+    if any(scenario_dir.glob("*/codex_stdout.ndjson")):
+        return "codex"
+    return "claude"
+```
+
+This function is called by `build_replay_runner()` (`recording.py:494`) to select between
+`CodexScenarioPlayer` and `make_scenario_player` (Claude player).
+
+Recording dispatch in `RecordingSubprocessRunner.__call__()` (`recording.py:117–239`) uses
+`capabilities.pty_required` to distinguish backends: PTY path for Claude, non-PTY path for
+Codex (which writes `codex_stdout.ndjson` + `step_meta.json`).
+
+Consumption gates at the server layer (`_factory.py:223–252`):
+- `replay_capable` guards `build_replay_runner()` construction
+- `record_capable` guards `RecordingSubprocessRunner` construction
+
+## Contract Gaps
+
+Adding a third backend (e.g., opencode-via-ACP) requires extending
+`_detect_backend_format()` with a new sentinel file or NDJSON magic-byte check.
+
+## Forward Obligation
+
+Any new backend adding replay support must:
+
+1. Declare `replay_capable=True` in its `BackendCapabilities`.
+2. Provide a concrete `ScenarioPlayer` subclass.
+3. Register a format-detection branch in `_detect_backend_format()` with a unique sentinel
+   file.
+4. Add a `make_scenario_player` factory dispatch for the new format.
+5. If supporting recording: declare `record_capable=True` and implement the recording path
+   in `RecordingSubprocessRunner.__call__()`.
+
+## Key References
+
+- `execution/recording.py` — `_detect_backend_format`, `RecordingSubprocessRunner`
+- `execution/backends/codex_scenario_player.py` — `CodexScenarioPlayer`, `CodexStepRecord`
+- `docs/design/recording-replay-accepted-degradations.md` — accepted degradations
+- `core/types/_type_backend.py:127–129` — `replay_capable`, `record_capable` field definitions
+- `server/_factory.py:223–252` — capability-gated construction

docs/design/forward-obligations/triage-portability-impact.md

@@ -0,0 +1,65 @@
+# Triage Portability Contract Impact
+
+| Field | Value |
+|-------|-------|
+| Status | Draft |
+| Date | 2026-06-27 |
+
+## Obligation
+
+LLM triage (contract-staleness triage via `_llm_triage.py`) is gated by
+`BackendCapabilities.triage_capable`. Backends with `triage_capable=False` cause
+`_triage_batch` to return all hash-mismatch items as `meaningful=True` without launching
+an LLM subprocess.
+
+## Current State
+
+- `ClaudeCodeBackend`: `triage_capable=True` (set in `CLAUDE_CODE_CAPABILITIES`,
+  `_type_backend.py:229`). Uses a `claude -p` subprocess with Haiku model.
+- `CodexBackend`: `triage_capable=False` (`codex.py:559`). Triage is skipped entirely.
+
+## Contract Gaps
+
+`_triage_batch()` (`_llm_triage.py:80`) constructs the triage command at lines 123–131:
+
+```python
+triage_cmd: list[str] = [
+    backend.binary_name(),
+    ClaudeFlags.PRINT,
+    prompt,
+    ClaudeFlags.MODEL,
+    "claude-haiku-4-5-20251001",
+    ClaudeFlags.OUTPUT_FORMAT,
+    fmt.value,
+]
+```
+
+The command uses `backend.binary_name()` (not a hardcoded `"claude"` literal) but depends on
+`ClaudeFlags` constants (`PRINT`, `MODEL`, `OUTPUT_FORMAT`) which are Claude-CLI-specific.
+The model `"claude-haiku-4-5-20251001"` is hardcoded. Portability to a non-Claude backend
+requires either:
+
+- A backend-agnostic `triage_cmd()` protocol method on `CodingAgentBackend`, or
+- A dedicated lightweight triage backend decoupled from the coding-agent backend.
+
+## Required Steps for Portable Triage
+
+1. Add an optional `triage_cmd(prompt: str, model: str) -> list[str]` method to
+   `CodingAgentBackend` (or add a `BackendCapabilities.triage_model` field).
+2. Update `_triage_batch` to call `backend.triage_cmd()` instead of constructing the
+   command inline with `ClaudeFlags`.
+3. Gate on `triage_capable` (already done at `_llm_triage.py:107`).
+
+## Blast Radius
+
+`_llm_triage.py` is a standalone module at the package root. Changes are isolated and do not
+touch `BackendCapabilities` layout or the sub-protocol set. The `triage_capable` field on
+`BackendCapabilities` (`_type_backend.py:86`) already has production consumers (the guard at
+`_llm_triage.py:107`), so no `_FORWARD_DECLARED` exemption is needed.
+
+## Key References
+
+- `_llm_triage.py:33` — `triage_staleness()` entry point
+- `_llm_triage.py:80` — `_triage_batch()` with command construction
+- `_llm_triage.py:107` — `triage_capable` guard
+- `core/types/_type_backend.py:86` — `triage_capable` field definition