tcx-strategy

Strategy Creator

Use this skill to author and manage strategy skills. A strategy skill is a standalone Codex-compatible skill whose name starts with strategy-, stored as a normal project skill under .agents/skills/strategy-*. It captures a user-approved investment strategy as an agent-readable operating procedure and should remain usable without importing TradingCodex service code.

Strategy skills guide judgment only. They never approve orders, execute orders, override policy, change MCP allowlists, bypass role boundaries, read secrets, or grant broker authority.

Each workflow must bind the selected strategy name and content hash before analysis. Later edits do not rewrite the strategy snapshot used by an earlier decision, replay, forecast, or postmortem. Historical and forward evidence may justify a draft change proposal, but it never silently changes or activates the strategy.

In native Codex, selection requires one exact explicit $strategy-* id, either as a plain invocation or a Markdown link whose label and target match the projected workspace skill. Repeated references to the same id are deduplicated; distinct multiple ids fail closed. An unprefixed name or natural-language similarity never selects a strategy. The hook accepts only an active validated workspace strategy and seals its content under the run before planning. The read-only viewer never selects a strategy.

The generated strategy body must be standalone. Do not mention platform names, platform role identifiers, subagent mechanics, MCP, approval gates, execution gates, or handoff mechanics inside the strategy skill. If a section has no user-provided rule, write not specified without adding a delegation sentence.

Managed Turn Admission

Creating, updating, activating, archiving, or deleting a strategy is a durable managed operation. Proceed only when the current root prompt has $tcx-strategy on its first meaningful line. Accept the plain token or a Markdown skill link only when its label and target match this workspace's projected tcx-strategy/SKILL.md; the concrete non-empty request may share the line or follow it. Do not combine it with $tcx-build, $tcx-brain, or an order marker. Use the normal trading-research profile; the marker admits only allowlisted Strategy lifecycle calls through the proof-protected manage_strategy MCP tool for this root turn. It does not grant generic Build, Brain, order, credential, publication, or global-config authority. Plan mode and subagents cannot issue or use the grant.

If the current turn lacks that admission, do not call shell or file-edit tools. Clarify or draft the proposed strategy in the conversation, then return a new root-turn prompt in this exact shape:

$tcx-strategy
<Create, update, activate, archive, or delete> <strategy-name> with <reviewed request>.

Managed admission remains bounded by the active Codex sandbox, the deterministic hook, workspace-only staging files, and the shared strategy service.

Workflow

1. Verify managed admission before any tool use. Identify whether the explicit request creates, updates, activates, archives, or deletes a strategy-* skill.

2. Choose a short lowercase hyphen-case name with the required strategy- prefix.

3. Draft a concise body directly from the required sections and strategy boundaries below. The pristine TradingCodex path must not depend on a host-global or plugin authoring skill. If the user explicitly selected an available external authoring skill, it may help draft the body within the same boundaries, but the shared strategy service remains the exclusive writer of the strategy bundle and projection.

4. For create or update, prepare a workspace-local body file outside protected TradingCodex paths, such as <strategy-name>.draft.md at workspace root, using native apply_patch. Put only the strategy body in that file; do not add frontmatter or agents/openai.yaml.

5. For each durable operation, invoke only its corresponding manage_strategy MCP action:

   action=create|update name=<strategy-name> description=<description> body_path=<root-level-workspace-body-file> language=<language> status=<draft|active|archived>
   action=inspect name=<strategy-name>
   action=list [active_only]
   action=activate|archive name=<strategy-name>
   action=delete name=<strategy-name> [force]
   

   For create or update, pass the reviewed body as body_path, plus an accurate description and language. Always pass --status; use active only when the current root request explicitly approves that exact reviewed content. Otherwise use draft, and never update an existing active strategy merely to stage an unapproved proposal.

6. Never manually edit .agents/skills, .codex/config.toml, generated indexes, or fixed subagent TOML. Never repair around a service validation error; report the error and revise only the staged body when appropriate.

7. Inspect the service result and, within the admitted Strategy turn, validate with manage_strategy actions inspect and list. Treat service-created bundle and projection metadata as the source of truth.

8. For an evidence-driven change, show the old and proposed rules, reason, affected scope, supporting and contrary cases, and validation status. Keep the current active version unchanged until approval. If an active strategy needs changed content, inspect and propose in one turn, then require a new exact $tcx-strategy root turn that explicitly approves the reviewed update; never carry the prior turn's managed grant across the reply.

9. Delete only when the root request explicitly names deletion. Do not add --force unless the user explicitly approved permanent removal of the active strategy after reviewing the consequence; otherwise let the service archive an active strategy.

10. Remove the staged body with native apply_patch after a successful service write unless the user explicitly asked to retain the draft file.

An exact create request with sufficient rules may create a draft in one turn; do not add a redundant confirmation. Activation, deletion, and replacement of active content require the original request to name that exact effect.

Service-Generated Skill Shape

The shared service must generate this folder shape:

.agents/skills/strategy-<name>/
  SKILL.md
  agents/openai.yaml

SKILL.md frontmatter must include these scalar fields, and name must match the strategy directory name. Do not write this frontmatter directly:

---
name: strategy-<name>
description: "<what this strategy does and when to use it>"
type: strategy
status: draft
language: <BCP-47 language tag or unknown>
owner: user
last_reviewed: unknown
---

Set status: active only after the user approves the strategy. Use draft, active, or archived.

Required Body Sections

Keep the body concise and include these headings:

• # <Strategy Name>
• ## Thesis
• ## Eligible Universe
• ## Preferred Setups
• ## Entry Criteria
• ## Exit Criteria
• ## Evidence Requirements
• ## Decision-Ready Standard
• ## Sizing Guidance
• ## Risk Controls
• ## Block Conditions
• ## Change Log

Use unknown or not specified for missing user input. Do not invent strategy rules. For ## Sizing Guidance, include only strategy-level sizing rules supplied by the user, such as max position size, leverage limits, loss limits, scaling rules, or cash/reserve constraints. If the user did not specify them, write exactly not specified.

Service-Generated Metadata

The service must generate agents/openai.yaml with:

interface:
  display_name: "<human strategy name>"
  short_description: "<25-64 character UI description>"
  default_prompt: "Use $strategy-<name> to apply this user-approved strategy."
policy:
  allow_implicit_invocation: false

The default prompt must mention the exact $strategy-<name> name.

Root Config Projection

The shared service projects active strategy skills into the strategy marker block in .codex/config.toml. Never edit the block directly. Its resulting shape is:

# BEGIN TradingCodex strategy skills
[[skills.config]]
path = "/absolute/path/to/.agents/skills/strategy-<name>/SKILL.md"
enabled = true
# END TradingCodex strategy skills

Do not add strategy skills to fixed subagent TOML files. The coordinator may read the selected strategy and pass compact context in assignment briefs, but that orchestration detail must not appear inside the strategy body.

Validation

Before finishing, confirm:

• The name starts with strategy-.
• The frontmatter name matches the .agents/skills/<name>/ directory.
• Required frontmatter is present.
• Required body sections are present.
• agents/openai.yaml has a valid default prompt with the exact strategy name.
• The shared service projected active strategies into the root strategy marker block.
• Fixed subagent TOML files do not reference the strategy.
• The strategy body has no platform coupling terms such as TradingCodex, role names, MCP, approval gates, execution gates, or handoff instructions.
• $tcx-build was not used as a wrapper, and no Brain or generic Build action was attempted with the Strategy grant.