Platform Triage Ledger

Per-minor re-triage of native Claude Code capabilities against RDF’s own mechanisms — one dated block per minor release, newest first below the Conventions section. /r-ship Stage 1d greps the first dated block for the target release’s MAJOR.MINOR; a minor release without a current block is blocked (patch releases skip the check). Origin: docs/specs/2026-09-02-platform-alignment-spike-design.md §6 (D4).

---

Conventions

How to add a block at the next minor

Before the next minor ships, re-run the eight checklist lines of the newest block against current platform docs (skills.md, agent-teams.md, workflows.md, plugins-reference, etc.), then insert a new ## <MAJOR.MINOR> — <date> block immediately after this Conventions section — ahead of every dated block, which becomes history below it. Fill in the same five columns and update or carry forward the Product rulings. /r-ship Stage 1d greps only the first dated block, so the new block must be prepended, not appended.

Next minor — platform alignment follow-ons

Deferred from the spike’s Go/No-Go summary (spike §8):

• D1 — Workflow-backed /r-build behind a flag. Fix the tiers.md bugfix/min() contradiction, define a plan Status: form, add a deploy.sh workflows dir, a /r-build --workflow arg (with BATS coverage), an A/B run on one real plan, and an indicator-list contract.
• D2 — trigger-eval suite on claude plugin eval. The 3.8 re-triage found claude plugin eval GA (v2.1.269+), so D2 skips the local harness: an evals/*/case.yaml suite (37 positive + ~10 negative cases) run by claude plugin eval, a /r-ship 1e line, and an optional CI job.

---

3.8 — 2026-09-24

Re-run against Claude Code 2.1.280 docs (workflows.md, agent-teams.md, skills.md, plugins-reference, memory.md, plugin-evals.md, sub-agents.md, model-config.md, env-vars.md). Four rows are new this minor because 3.8 routes model/effort per agent and keys session state on the harness session id.

Line	Decision rule	Verdict	Evidence	Re-check trigger
Workflows	absorb when GA on the user’s plan and the RDF mechanism is a loop with no user input	absorb (gated spike, D1 — still deferred)	workflows.md: “available on all paid plans, with Anthropic API access, and on Amazon Bedrock, Google Cloud’s Agent Platform, and Microsoft Foundry” (unchanged)	D1’s A/B evidence lands
Agent Teams	absorb only when non-experimental, -p capable, worktree-isolated, nested	keep dispatcher (3.7 first case stands)	agent-teams.md: still experimental and off by default; “In non-interactive mode with the -p flag … Claude doesn’t spawn teammates”; “teammates cannot spawn their own teammates”	experimental flag drops, -p support lands, or per-teammate worktree isolation ships
Skills frontmatter	absorb keys the CC adapter can emit from skill-meta.json	keep (absorbed in 3.7; no new keys since 2026-09-02). Skill model/effort last one turn, so Fable for /r-spec//r-plan is a session-start advisory, not skill frontmatter	skills.md frontmatter reference	CC’s skill-frontmatter schema changes
Plugin components	absorb a component only when an RDF mechanism maps 1:1	keep	plugins-reference: “Plugin components include skills, agents, hooks, MCP servers, LSP servers, and monitors” (unchanged)	a new plugin component ships with an RDF-equivalent mechanism
Auto-memory	retire ~/.rdf/lessons-learned.md only when a native cross-project store exists	keep	memory.md: “Auto memory is machine-local. All worktrees and subdirectories within the same git repository share one auto memory directory”	Anthropic ships a native cross-project memory store
.claude/rules/ + @import	verify each minor that cc_generate_rules output still loads	keep as-is	memory.md (no behavior change); tests/rules-deploy.bats green	rules-deploy.bats regresses or CC’s @import behavior changes
Eval availability	switch the trigger-eval runner to claude plugin eval when the probe stops printing “early access”	trigger fired — claude plugin eval is GA (v2.1.269+); D2 is re-scoped to a claude plugin eval suite, no local harness (not built this release)	plugin-evals.md: requires “v2.1.269 or later”; the early-access message means “Your build predates general availability”; claude plugin eval --help on 2.1.280 has no early-access wording	D2 lands, or the command reports “currently unavailable” (switched off server-side)
Hazard CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS	a named subagent launches as a teammate when the flag is set	document “unset with RDF”	agent-teams.md (unchanged)	RDF intentionally wants Agent Teams semantics for a named subagent
Subagent model / effort (new)	absorb per-agent routing when the adapter can emit it from agent-meta.json	absorbed this release — model/effort per agent plus effort variants (rdf-engineer-focused, rdf-reviewer-challenge), since the Agent tool takes a per-call model but no effort	sub-agents.md: effort “Overrides the session effort level. Default: inherits from session”; model-config.md: frontmatter effort overrides “the session level but not the environment variable”	the Agent tool gains a per-call effort (variants become unnecessary)
Opus 5.5 effort default (new)	detect and advise; never write user settings	doctor harness scope WARNs	model-config.md: Opus 5.5 defaults to medium; “a top-level effortLevel in your user settings file doesn’t count for Opus 5.5”	the default or the precedence changes
omitClaudeMd (new)	absorb for agents that take everything from the dispatch payload, once measured	defer to context-economy spec B	sub-agents.md: skips user, project, and local CLAUDE.md; managed policy still loads; v2.1.271+	spec B lands
CLAUDE_CODE_SESSION_ID (new)	key session-scoped state on the harness session id; mint only outside Claude Code	absorbed this release (rdf_session_init)	env-vars.md: “Set automatically to the current session ID in Bash and PowerShell tool subprocesses, hook command subprocesses”; matches the hook session_id; updated on /clear; --resume <id> keeps it, --continue may get the startup ID	the variable’s scope or semantics change

Product rulings

1. No dual /r-build path this minor — carried from 3.7 (D1 still unbuilt).
2. No disable-model-invocation on lifecycle skills — carried from 3.7.
3. D4 gate strength: minor-only blocking — carried from 3.7.
4. Routing lives in agent-meta.json, not in dispatch calls. Canonical text dispatches by agent name; no per-call model: override (3.8). Effort that differs by mode is a variant agent, not a parameter.

---

3.7 — 2026-09-02

Line	Decision rule	Verdict	Evidence	Re-check trigger
Workflows	absorb when GA on the user’s plan and the RDF mechanism is a loop with no user input	absorb (gated spike, D1 — deferred to next minor)	workflows.md GA; spike §3	D1’s A/B evidence lands (see follow-ons below)
Agent Teams	absorb only when non-experimental, -p capable, worktree-isolated, nested	keep dispatcher (first case, below)	agent-teams.md; agent-teams-research.md:118-126	experimental flag drops, -p support lands, or per-teammate worktree isolation ships
Skills frontmatter	absorb skills/<name>/SKILL.md + disable-model-invocation on side-effect commands + allowed-tools when the CC adapter can emit them from skill-meta.json	absorbed this release (sibling spec 2026-09-02-skills-native-adapter-consolidation-design.md, Phases 1-6)	skills.md “Custom commands have been merged into skills”	CC’s skill-frontmatter schema changes
Plugin components	absorb a component only when an RDF mechanism maps 1:1	keep (only Workflows is a 1:1 candidate, tracked as D1)	plugins-reference component table	a new plugin component ships with an RDF-equivalent mechanism (MCP/LSP/monitors/output-styles currently have none)
Auto-memory	retire ~/.rdf/lessons-learned.md only when a native cross-project store exists	keep	no native cross-project store found (spike §6)	Anthropic ships a native cross-project memory store
.claude/rules/ + @import	already absorbed (3.4.0 --rules) — verify each minor that cc_generate_rules output still loads	keep as-is	adapters/claude-code/adapter.sh:96,140, tests/rules-deploy.bats	any minor where rules-deploy.bats regresses or CC’s @import behavior changes
Eval availability	switch the trigger-eval runner to claude plugin eval when the probe stops printing “early access”	local harness (D2 — deferred to next minor)	spike §2 probe	claude plugin eval output drops “early access”
Hazard CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS	a named subagent launches as a teammate when the flag is set, forming a team even when not asked	document “unset with RDF”	agent-teams.md	RDF ships a named-subagent workflow that intentionally wants Agent Teams semantics (none today)

First case — Agent Teams vs dispatcher

Verdict: keep the dispatcher; do not absorb.

1. “Agent teams are experimental and disabled by default” (agent-teams.md Warning); not enabled in this environment (spike §2).
2. “Spawning teammates also requires an interactive session. In non-interactive mode with the -p flag … Claude doesn’t spawn teammates” — no headless/CI path; RDF’s D2 harness and /r-build resume both need -p.
3. Shared working tree: “Two teammates editing the same file leads to overwrites” and worktrees are offered only as manual sessions “without automated team coordination”; RDF’s parallel path is worktree-isolated with a pre-commit scope hook (r-build.md:232-253,273, state/git-hooks/pre-commit, plan-schema.md Rule 8).
4. Gate machinery is hook-only (TeammateIdle/TaskCreated/TaskCompleted exit 2) — no engineer→QA→reviewer chain, no evidence grammar, no finding-resolution routing (dispatcher.md:457-536).
5. “No nested teams” and “Lead is fixed” contradict /r-build → dispatcher → engineer nesting (r-build.md:204-208).
6. “Task status can lag” is the exact failure the dispatcher’s status files exist to prevent.

Re-check triggers: the experimental flag drops, -p support lands, or per-teammate worktree isolation ships.

Product rulings

1. No dual /r-build path this minor. The Workflow prototype (docs/specs/support/rdf-build-workflow.prototype.js) stays an unwired artifact until A/B evidence exists — see D1 in the follow-ons below.
2. No disable-model-invocation on lifecycle skills. Measured usage shows the pipeline model-invokes /r-build 43×, /r-plan 75×, /r-spec 53×, /r-ship 5×, /r-save 25× (controller usage mining 2026-09-02: Skill-tool invocations across 10,453 Claude Code transcripts on the maintainer’s machine — not in the repo). Flagging them disable-model-invocation would remove them from Claude’s own suggestions — accepted as an AST03 risk (over-privileged skills, see spike §7) for now; revisit alongside allowed-tools scoping.
3. D4 gate strength: minor-only blocking. /r-ship Stage 1d blocks a x.y.0 release without a current top block; patch releases (x.y.z, z>0) skip the check unconditionally.