prompt

Model prompts

This package builds the instructions used by Wingman's built-in agent. Choose a model with /model; its matching prompt variant is selected automatically. Project guidance comes from AGENTS.md or CLAUDE.md, and reusable workflows come from skills. See customization.

Working on templates

Model-specific instructions live in models/<variant>/mode_agent.txt. Routing uses the longest normalized model-name prefix, with shared instructions as the fallback. Shared Plan and unattended templates control those modes.

BuildBaseInstructions composes the model template, shared runtime/tool fragments, and finally the active mode restrictions. section_runtime.txt owns editing, process lifecycle, authorization, evidence, and completion rules. section_delegation.txt owns the model-independent delegation policy; code intelligence and web research have their own shared fragments.

Tool-specific guidance is rendered from SectionData.ToolNames, populated from the actual catalog after configuration, mode, and role filtering. Optional fragments disappear entirely when their tools are absent. Search and fetch are independent (DisableWebSearch/WINGMAN_DISABLE_WEBSEARCH and DisableWebFetch/WINGMAN_DISABLE_WEBFETCH); shell guidance follows DisableShell/WINGMAN_DISABLE_SHELL. The standard file search/read tools are always present. Editing guidance is omitted in Plan mode and read-only child roles. Unattended sessions omit question-tool guidance. Do not copy these contracts into each model template or assume a tool is present from the model ID.

BuildSessionContext adds project instructions, skills, memory, and environment context as a replaceable session-context message. BuildAgentContext supplies shared runtime/tool guidance, project/environment context, and active mode restrictions to the child's role prompt. It uses the child's filtered tools, so children do not receive delegation instructions. This scoped context is saved with the durable child specification at launch. Keep shared sections in their existing templates when adapting a model prompt.

Use Wingman's tool names and argument formats. Approval handling, session continuation, and background commands follow Wingman's runtime behavior.

Source references

Variants adapt serialized CLI requests, including runtime instructions that do not appear in a model catalog. Preserve model-specific behavior, but translate tool contracts, identity, authorization, and mode handling to Wingman.

GPT / Codex and Claude / Claude Code

Refreshed from local request captures on 2026-10-16:

scripts/capture-code-prompts.py points each CLI's API provider at a localhost mock with dummy credentials and returns synthetic text without tool calls. Codex uses a custom Responses provider, its bundled model metadata, and flags that exclude user configuration and rules. Claude uses ANTHROPIC_BASE_URL and a temporary CLAUDE_CONFIG_DIR, safe mode, and an empty MCP configuration. It deliberately does not use bare mode or force the simplified prompt: either would change the prompt being measured. Both run in temporary workspaces with telemetry disabled. The mock never forwards requests, and request headers are not saved.

Reproduce from the repository root (Python standard library and the CLI required):

python3 scripts/capture-code-prompts.py codex \
  --model gpt-6.1-sol --model gpt-6-astra --model gpt-6-sol \
  --model gpt-6-luna --model gpt-5.6-sol --model gpt-5.5

python3 scripts/capture-code-prompts.py claude \
  --model claude-sonnet-5-5 --model claude-haiku-5-5 \
  --model claude-opus-5-5 --model claude-opus-5 --model claude-opus-4-8 \
  --model claude-fable-5-1 --model claude-fable-5 \
  --model claude-mythos-5-1 --model claude-mythos-5 --model claude-sonnet-5

The printed directory contains executable provenance, per-run manifests, request JSON, extracted base/system/developer text, tool schemas, and CLI logs. --output accepts a new directory and --binary selects a specific installed executable. Localhost listening is required. Isolation flags are version-specific; inspect CLI errors when a future release changes them. Raw captures stay outside the repository because they include transient paths and CLI-specific context.

The newer Codex request format puts the base prompt in the first developer message and tool schemas in additional_tools input items. The script extracts both that format and top-level instructions/tools. GPT-6.1 Sol, GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, and GPT-5.6 Sol base instructions match the embedded catalog baseline be2e129ecd67d900c016eda12d5d193e33d9c3da. Their captured runtime adds an explicit restriction on unsolicited delegation. Reviewing Codex source at b9775a0d57 (2026-10-16), especially codex-rs/core/src/session/multi_agents.rs, shows this is runtime policy selected by configuration/catalog overrides and reasoning effort, not an intrinsic model trait. Wingman uses one shared default: delegate bounded independent work when useful, subject to user/project/skill restrictions, independent of model or effort. GPT-5.5's small checklist wording difference from the embedded catalog was already reflected in Wingman's template. The newer source retains the same base prompts for these six captured model IDs.

This Codex version reports fallback metadata for GPT-5.4, GPT-5.4 Mini, GPT-5.2, and GPT-5.1. The script marks those runs unsuccessful even when the mock response completes. Their existing model-specific templates remain in use, with Wingman runtime improvements; generic fallback captures are not model-specific source evidence. All older GPT models remain in the catalog. GPT-5.6 Terra/Luna and other IDs sharing a family template were not separately captured.

Claude Sonnet/Opus 5.5 and Opus 4.8 use lean prompts; Haiku 5.5, Opus 5, Fable/Mythos, and general Sonnet 5 retain their respective longer structures. The refresh removes a stale pre-mutation stop rule from Opus 4.8 and Fable/Mythos 5 and replaces Opus 5's delegation restriction with Wingman's shared policy. The general Claude template no longer treats arbitrary reminder tags in tool output as trusted instructions. Shared family routing does not imply every older Claude ID was captured.

Fallback (Gemini, Qwen, Nemotron, GLM, Kimi, MiniMax, Grok, DeepSeek, Mistral)

Only Claude and GPT keep variants, each adapted from a captured vendor CLI request. Every other model uses the fallback. The former Qwen, GLM, Kimi, MiniMax, and Grok templates were paraphrases rather than vendor captures: Qwen from Qwen Code, the others from VS Code Copilot's per-vendor prompts. Their sources contain no model-version-specific instructions: Qwen Code (0.25.2) varies only few-shot tool-call examples by model; Copilot applies its GLM prompt only to GLM 4.6/4.7 and gives later GLM releases its default prompt. NVIDIA publishes no coding-agent prompt for Nemotron.

The fallback keeps the behavior-changing rules those templates shared: real tool calls instead of prose, no invented paths or APIs, no placeholders or omission markers in files, efficient reads, and a stop-and-reassess rule for repeated failures (from Copilot's Kimi and xAI prompts and Qwen Code's loop detector). It reuses earlier observations for follow-up questions, as Qwen Code's 2026-10-04 change does to avoid repeated investigation on local hosts. From GPT-6.1 Sol it takes shell quoting and destructive-target checks; personality, writing style, and update cadence stay GPT-specific. Formatting differences such as GLM 4.x's capitalized MUST rules and section names were not kept.

All model variants and the fallback receive the same shared runtime contract. Templates do not restate rules it already covers, such as commit limits, preserving user changes, parallel reads, and claims about checks. GPT-5.1/5.2 no longer suggest asking to run validation that the task already authorizes. The fallback also clarifies action requests, persistent authorization, task steering, and completion checks. Shared Plan mode follows the same delegation policy, restricted to read-only roles. Wingman's Plan mode forbids builds and tests; this is intentionally stricter than Codex's allowance for checks that do not mutate repository files. Vendor skills, permission mechanisms, token budgets, environment data, and marketing copy are not imported as Wingman runtime instructions.

Gemini / Antigravity

Gemini models use the fallback. The removed Gemini template adapted requests captured from Antigravity CLI 1.4.0 (Homebrew build 5281842794987520) on 2026-10-16; the executable SHA-256 was a4502aaf310b2fd6eb657fcfec2fcc31a690fbbd0c2d14c8eefbb556409024ce. Unlike the Claude and GPT variants, it was a Wingman rewrite rather than a close vendor adaptation. Once it stopped restating shared runtime rules and took the fallback's context-reuse and focused-change guidance, it was mostly a subset of the fallback.

The capture uses the CLI's Gemini API-key provider and its documented GOOGLE_GEMINI_BASE_URL override, pointed at a localhost mock with a dummy key. Separate configuration and workspace directories isolate the run. Synthetic text responses let the CLI finish without model inference or agent tool execution. Re-capture a newer release to check whether it adds model-specific instructions worth a variant.

Reproduce from the repository root (Python standard library and agy required):

python3 scripts/capture-agy-prompt.py \
  --model gemini-3.1-pro-low --model gemini-3.1-pro-high \
  --model gemini-3.7-flash-high --model gemini-3.8-flash-high

The printed directory contains the executable hash, per-run manifests, request JSON, extracted system text, tool schemas, and CLI logs. The capture requires a localhost listener. --output accepts a new directory; --mode plan captures planning mode. Directory-isolation flags are implementation details of this CLI version and may change. No real API key is needed, and headers are not recorded.

Observed API-provider requests:

After normalizing temporary paths and conversation IDs, Pro's two prompts match, and Flash's two prompts match. Pro adds a dedicated-tool preference and a literal thought-prefix instruction; otherwise the prompts are identical. The shared runtime's tool preference covers the former; the private output protocol is not used. Effort differences belong to request configuration, not prompt routing. The captures also contain a separate short title-generation request; it is not the coding agent's system prompt. No Gemini 4 or signed-in account-provider prompt was captured.

Request capture and prompt tests establish provenance and compatibility. They do not demonstrate improved model behavior or benchmark performance.

Check a change

From the repository root:

go test ./pkg/code/prompt ./pkg/code/agent

For model-catalog changes, also run:

go test ./pkg/model ./pkg/external/codex ./pkg/acp/codex ./pkg/harness

These tests check routing, conditional fragment rendering, shared tool/schema references, and serialized child requests with mode and tool restrictions. They require localhost mock-provider listeners but no model inference. Use matched model and effort runs of the ACP benchmark to compare task behavior or latency.