RES — TFW-45 iter1: Cross-Platform Capabilities Audit¶

Date: 2026-06-15 Author: Researcher (Antigravity) Status: 🔬 RES — Complete Parent HL: HL-TFW-45 Mode: Deep (2 OODA loops in Gather)

Research Context¶

TFW-45 iteration 1 investigated cross-platform sub-agent spawn mechanics, system prompt hierarchies, and context inheritance across Antigravity, Claude Code, and Codex CLI. The user's #1 question: "Is invoking a slash command, skill, or workflow equivalent to a system prompt or not?" This research answered definitively: no, on any platform. Only rules-type files receive system-prompt-level enforcement. The research also mapped the complete instruction hierarchy per platform, identified 4 surviving configurations for swarm mode, and eliminated 5 configurations with structural reasons.

Briefing¶

Reference: 1_briefing.md. Focus: H2–H7 (cross-platform capabilities). User direction: equal depth across all 3 platforms, system prompt question is #1 priority. Mode: deep.

Decisions¶

#	Decision	Rationale
D1	Antigravity: custom sub-agent type for swarm mode (not `self`)	`self` type inherits parent's system prompt — cannot inject custom Mindset as system_prompt. Custom type allows Mindset = TRUE system prompt. Sub-agent must explicitly load TFW context via `system_prompt` instruction to read files. User confirmed this preference
D2	Workflows and Skills ≠ system prompt on ANY platform	Empirical (Antigravity: first-hand from own system prompt) + external (Claude Code, Codex docs). Workflows/Skills are file references loaded on demand. Only `.agent/rules/` (Antigravity), `CLAUDE.md` + `.claude/rules/` (Claude Code), and `AGENTS.md` (Codex) receive enforcement-level injection
D3	`swarm.md` is platform-agnostic protocol, not executable code	swarm.md describes WHEN to spawn, WHAT Mindset each stage agent gets, WHAT context to load, HOW results flow. Platform-specific spawn syntax lives in adapter layer (`.agent/workflows/`, `.claude/agents/`, `.codex/agents/`). Stress-tested in Challenge C5
D4	`.agent/rules/` is the ONLY user-controllable true system prompt in Antigravity	Wrapped in `<user_rules><RULE[filename]>` with preamble: "MUST ALWAYS FOLLOW WITHOUT ANY EXCEPTION. Take precedence over any following instructions." This is the strongest enforcement language. All critical TFW identity content must go here
D5	AGENTS.md has different roles per platform	Antigravity: placed in `.agent/rules/agents.md` = part of system prompt. Claude Code: fallback if no CLAUDE.md exists. Codex: primary/native instruction file. One file, three completely different injection mechanisms
D6	Claude Code sub-agents don't reliably inherit CLAUDE.md	External research + community reports confirm unreliable inheritance. Sub-agent rules must be explicit in agent definition file (`.claude/agents/*.md`). This is a design constraint, not a bug
D7	4 surviving configurations for swarm mode	C2 (Antigravity custom), C4 (Claude agent def), C7 (Codex TOML), C9 (single-agent fallback). 5 others eliminated with structural incompatibility evidence

Open Questions¶

#	Question	Status	Answer
Q1	Does Mindset-as-system-prompt produce measurably better behavior than Mindset-as-conversation-context?	Deferred to iter2 (H1)	Structurally plausible (system prompt = identity level, can't be "forgotten") but no empirical test yet
Q2	What are the token budget limits for `define_subagent.system_prompt`?	Deferred to iter2	Estimated ~300-400 words needed (Mindset + TFW loading + stage instructions). Likely within budget but untested
Q3	Can Claude Code `<system-reminder>` reliably replicate Mindset identity shift?	Deferred to iter2 (H8)	Functionally equivalent for TFW purposes (Challenge C3 verdict) but subtle priority conflicts possible with Anthropic's hidden system prompt

Hypotheses (from HL §10)¶

#	Hypothesis	HL Status	RES Status	Evidence
H2	`define_subagent` with custom `system_prompt` sufficient for Mindset + TFW context	needs-research	✅ Confirmed	First-hand: `system_prompt` param is TRUE system prompt. Default read tools available. Agent can read TFW files as first action. Challenge C1 + C4 stress-tested
H3	`self` type inherits `.agent/rules/` (TFW context auto-available)	needs-research	✅ Confirmed but NOT recommended	`self` inherits full parent config including rules. BUT: incompatible with custom Mindset injection — `self` clones parent's system prompt, can't replace it. C1 eliminated
H4	Workflows become system-level instructions when invoked	needs-research	❌ Refuted	First-hand: workflows listed by description in `<workflows>` block. Agent reads them via `view_file` on demand. NOT system prompt content. Same for Skills
H5	Antigravity changed since April 2026 — TFW-30 findings need correction	needs-research	✅ Confirmed	Claude Code added Agent Teams (Feb 2026). Codex formalized agent definitions via TOML. Antigravity API stable but Skills/workflow ecosystem matured
H6	Claude Code teamwork similar to Antigravity `define_subagent` — one swarm.md covers both	needs-research	🔄 Refined	Similar concept (isolated sub-agents with custom prompts) but different injection mechanism (`<system-reminder>` vs true system prompt). swarm.md IS platform-agnostic as protocol, but adapters differ per platform
H7	Codex has analogous spawn mechanism	needs-research	✅ Confirmed	TOML agent definitions in `.codex/agents/` with `model_instructions_file`. AGENTS.md auto-inherited. Cleanest model of the three

HL Update Recommendations¶

#	What to update	Source
1	§10 H3: Mark as "Confirmed but NOT recommended for swarm" — `self` type conflicts with custom Mindset injection	D1, Challenge C1 elimination
2	§10 H4: Change to "Refuted — workflows are NOT system prompt, they're on-demand file reads"	D2, G1 first-hand evidence
3	§10 H6: Change to "Refined — similar concept, different injection. swarm.md is platform-agnostic protocol, adapters differ"	D3, Extract E2
4	§3.2: Add constraint — TFW critical identity content MUST go in `.agent/rules/` (Antigravity), not workflows or skills	D4
5	§3.2: Add platform comparison table showing instruction hierarchy per platform	Gather G6

Fact Candidates¶

Cognitive mode: Pure reporting — record factual observations without interpretation or synthesis.

#	Category	Candidate	Source	Confidence
FC1	architecture	Antigravity system prompt is assembled from 10 XML-tagged layers: `<identity>`, `<user_information>`, `<mcp_servers>`, `<user_rules>`, `<workflows>`, `<skills>`, `<plugins>`, `<subagents>`, `<planning_mode>`, `<guidelines>`. Only `<user_rules>` (from `.agent/rules/`) has enforcement language ("MUST ALWAYS FOLLOW WITHOUT ANY EXCEPTION")	Gather G6, first-hand	★★★
FC2	architecture	Claude Code injects CLAUDE.md and .claude/rules/ as `<system-reminder>` XML tags within conversation history, NOT as API-level `system` role message. This is "high-priority context" but technically not system prompt. Anthropic's actual system prompt is fixed and hidden (for prompt caching optimization)	Gather G6, external research	★★☆
FC3	architecture	AGENTS.md serves fundamentally different roles: Antigravity = placed in `.agent/rules/` = part of system prompt; Claude Code = fallback if no CLAUDE.md; Codex = primary native instruction file. Same filename, three injection mechanisms	Gather G6, D5	★★★
FC4	architecture	Antigravity `define_subagent` custom type sub-agents have default read tools (view_file, grep_search, list_dir, search_web) without explicit configuration. Write tools, MCP, and sub-sub-agents require explicit opt-in	Challenge C4, first-hand (tool documentation)	★★☆
FC5	process	Claude Code sub-agents (`.claude/agents/*.md`) do NOT reliably inherit project CLAUDE.md or `.claude/rules/`. Multiple community reports + GitHub issues document this. Sub-agent rules must be explicit in agent definition file	Gather G2, G6, D6	★★☆

Strategic Insights (Research)¶

#	Category	Insight	Source	Confidence
SS1	architecture	The system prompt hierarchy across platforms reveals a consistent pattern: platform vendors keep their TRUE system prompt hidden/fixed, and provide a "user instruction layer" that is authoritative but technically sits below. Only Antigravity's `define_subagent.system_prompt` gives users TRUE system-prompt-level control for sub-agents. Claude and Codex both use high-priority-context workarounds. This makes Antigravity the strongest platform for Mindset-as-identity injection	G6 cross-platform comparison	★★☆
SS2	strategy	The adapter layer architecture (swarm.md = protocol, adapters = platform-specific) naturally maps to TFW's existing 3-tier structure: `.tfw/` (core) → `.agent/` or `.claude/` or `.codex/` (adapter) → project files. swarm.md belongs in `.tfw/workflows/research/`, not in the adapter layer	Extract E5, Challenge C5	★★★

Findings Map¶

    ┌─────────────────────────────────────────────────┐
    │         SYSTEM PROMPT HIERARCHY                  │
    │         (per platform)                           │
    └───────┬──────────┬──────────┬───────────────────┘
            │          │          │
    ┌───────▼──┐ ┌─────▼────┐ ┌──▼──────────┐
    │ANTIGRAV  │ │CLAUDE    │ │CODEX        │
    │          │ │CODE      │ │             │
    │rules/=SP │ │CLAUDE.md │ │AGENTS.md    │
    │wflow≠SP  │ │=reminder │ │=system ctx  │
    │skill≠SP  │ │rules/    │ │TOML agents  │
    │          │ │=reminder │ │             │
    │sub-agent:│ │          │ │sub-agent:   │
    │TRUE SP   │ │sub-agent:│ │instructions │
    │(custom)  │ │agent .md │ │file = SP    │
    └────┬─────┘ └────┬─────┘ └──────┬──────┘
         │            │              │
         └────────────┼──────────────┘
                      │
    ┌─────────────────▼──────────────────┐
    │         swarm.md (protocol)         │
    │  • WHEN to spawn                   │
    │  • WHAT Mindset per stage          │
    │  • WHAT context to load            │
    │  • HOW results flow                │
    │  (platform-agnostic)               │
    └─────────────────┬──────────────────┘
                      │
    ┌────────┬────────┼────────┬─────────┐
    │        │        │        │         │
    ▼        ▼        ▼        ▼         ▼
   C2       C4       C7       C9
  Antigrav  Claude   Codex    Fallback
  custom    agent/   TOML     single
  type+SP   .md      +instr   agent

Iteration Status¶

Iteration: 1 of 2 (min) / 4 (max)
Hypotheses tested: H2 (✅), H3 (✅ not recommended), H4 (❌ refuted), H5 (✅), H6 (🔄 refined), H7 (✅)
Hypotheses deferred: H1, H8, H9 (iter2 scope per iterations.yaml)
Gaps discovered: Token budget for system_prompt untested; H1 empirical validation needed
Superseded decisions: None

Open Threads (for iter2)¶

H1 empirical test: Does fresh agent with Mindset-as-system-prompt produce measurably better output than single-agent mindset switch? This is the foundational hypothesis — everything else is implementation detail.
H8 mode file vs execution_model: Should swarm be a mode file (swarm.md) or an abstraction layer (execution_model in project_config.yaml)?
H9 honesty vs context loss: Fresh agent = honest identity but loses conversation history. What's the trade-off boundary?
Token budget measurement: How many tokens can define_subagent.system_prompt accept before degradation?

Recommendation¶

[x] SUFFICIENT for iter1 — cross-platform capabilities mapped, surviving configurations identified
[ ] MORE NEEDED — iter2 required per iterations.yaml (H1, H8, H9 unresolved)
[ ] BLOCKED — N/A

Iter1 delivered what it promised: capabilities audit across 3 platforms. The 4 surviving configurations (C2/C4/C7/C9) provide a solid foundation for iter2's design validation. The H4 refutation (workflows ≠ system prompt) and the AGENTS.md role asymmetry (D5) are material discoveries that should update HL before iter2 begins.

Conclusion¶

This research determined the exact system prompt hierarchy across Antigravity, Claude Code, and Codex CLI through a combination of first-hand observation (Antigravity — reading own system prompt structure) and external research (Claude Code, Codex). The key finding: on no platform does workflow, skill, or slash-command invocation equal system prompt. Only rules-type files (.agent/rules/ in Antigravity, CLAUDE.md + .claude/rules/ in Claude Code, AGENTS.md in Codex) receive enforcement-level priority. For sub-agents, Antigravity's define_subagent.system_prompt provides TRUE system prompt — the strongest available mechanism. Claude Code uses <system-reminder> tags (functionally equivalent but technically not system prompt), while Codex uses TOML definitions with instruction files. The self sub-agent type was eliminated for swarm mode because it clones parent's system prompt and cannot accept custom Mindset injection. The research produced 4 surviving configurations and confirmed that swarm.md can be platform-agnostic (protocol-level), with platform-specific adapters handling spawn syntax. Without this research, the implementation would have likely assumed workflows = system prompt (wrong), used self type for sub-agents (incompatible with Mindset injection), and designed platform-specific swarm files instead of a single protocol + adapters.

RES — TFW-45 iter1: Cross-Platform Capabilities Audit | 2026-06-15