prompt

Model prompt maintenance

Wingman's GPT prompts adapt Codex instructions to Wingman's tools and session lifecycle. Compare the effective model template, not just a file named *_prompt.md: current Codex renders model_messages.instructions_template from codex-rs/models-manager/models.json and supplies permissions, collaboration, project instructions, and environment context separately.

Comparison baseline

Re-reviewed on 2026-10-08 against local Codex commit be2e129ecd67d900c016eda12d5d193e33d9c3da. Every current model's instructions_template is unchanged from the September 30 baseline below; the catalog only removes obsolete null instructions_variables fields. Model-specific function-description prefixes are now supported upstream, but no catalog model supplies them, so there is no model guidance to import. Codex's partial_answer phase is adopted in Wingman's completion loop and retained through persistence and request replay. Partial answers continue the turn and cannot satisfy an empty finish_turn marker. Existing cancellation, CRLF preservation, Retry-After, and compaction-context coverage already handles the corresponding upstream fixes. Codex-only Guardian, incremental namespace, and idle multi-agent-v2 changes do not map to Wingman's harness.

The follow-up review also checked these integration changes:

The complete go test ./... suite passes in wingman-agent. The completion-loop and Claude ACP race suites also pass. The sibling backend's complete suite passes with ANTHROPIC_API_KEY, OPENAI_API_KEY, and GEMINI_API_KEY explicitly empty so paid reference comparisons skip. Its offline Claude Code/Bedrock fixture now injects a trailing system instruction deterministically instead of requiring every CLI version to emit one. The backend's existing Haiku 5.5 native/Bedrock thinking, sampling, output-limit, tokenizer, and conversation-replay support needed no further production change. The shared comparison preflight now skips when the backend model listing is unavailable and limits discovery to five seconds, preventing reference-provider calls before a usable backend/model has been established. Local mock tests cover available, absent, malformed, unauthorized, and unreachable listings plus caching.

Reviewed all current GPT catalog variants on 2026-09-30 against Codex commit 92bc601ad60542c92bf0bb1e7a2eb70b84ac49d2 using its pinned model catalog. All base instruction templates match the previous 2026-09-29 baseline at b1e72963c3b71a9265a551e54beff078384efed9, including GPT-6.1 Sol. The embedded external catalog now preserves the full new upstream entries, including updated descriptions, priorities, and auto-review metadata. Historical GPT-5.4 remains available with its retained entry. GPT prompt text required no changes.

The prior September 29 review introduced GPT-6.1 Sol and compared GPT-6, GPT-5.6, GPT-5.5, and historical GPT-5.1/GPT-5.2 against 83b56bc5ad9489a47a0eb3edc3e5852f597125be.

The review adopts GPT-6's structured multiline command-body guidance, restores the GPT-5.6 progress-update cadence, and restores concise GPT-5.5/GPT-5.4 collaboration guidance omitted from the earlier adaptations. GPT-6.1 Sol keeps its added guidance on direct wording, avoiding unnecessary apologies, and acknowledging meaningful mistakes. GPT-5.1, GPT-5.2, and GPT-5.4 Mini retain their existing adaptations; there is no revised upstream source to adopt.

Model routing uses the longest normalized prefix in VariantFor. Preserve distinct variants and their model-specific guidance, including GPT-6 Luna's explicit-request-only testing policy. Plan guidance and the independent unattended policy append Wingman's shared templates to the selected model's instructions. Plan's read-only restrictions take precedence over implementation guidance.

GPT-6.1 Sol uses Codex's low effort default and low verbosity. Its native effort choices are low, medium, high, xhigh, and max, as supported by the OpenAI API model documentation. Codex's ultra orchestration setting is not a native API reasoning effort. The full upstream GPT-6.1 Sol entry is copied into the embedded external Codex catalog; existing entries, including GPT-5.4, are retained.

Intentional harness adaptations

• Use Wingman's grep, glob, read, edit, exec_command, exec_session, and agent contracts. Do not import Codex-only tools or argument formats.
• elicit blocks. Resolve discoverable questions first, assume when a choice is cheap to correct, and ask only for a blocking user decision. Keep question formatting in the tool description instead of duplicating it in each prompt.
• Tool-execution approvals belong to Wingman's harness. Do not import Codex approval-mode names or its automatic reviewer instructions.
• Background command exits can arrive as notifications. Use bounded waits and avoid polling solely to discover an exit when notifications are available.
• Status questions and compaction preserve unfinished work. Hand off when the requested outcome is complete or a real blocker prevents further progress.
• Send useful progress updates; avoid thought-length-based update rules. Leave blank lines around Markdown headings and lists for consistent rendering.
• Keep personality guidance focused on collaboration and judgment. Omit product identity, claims about an inner life, and unrelated stylistic bans.
• Codex's optional persistent-mode prompt depends on its async messaging and scheduling tools. Keep Wingman's session and unattended-mode contracts.

BuildBaseInstructions contains model/mode guidance; BuildSessionContext contains project instructions, skills, memory, and environment data. The harness appends changed context snapshots to history without rewriting its cached prefix. Keep these shared sections out of individual model prompts.

The clarification and validation changes also follow the official GPT-6 prompting guidance: calibrate initiative and testing to the workload and available tools.

Validation

Run go test ./pkg/code/prompt ./pkg/code/agent for template routing, rendering, mode selection, and session integration. These checks do not measure model behavior. Use matched model/effort runs of bench/acpflow for behavioral or latency comparisons, and include clarification and post-compaction continuation cases when evaluating these instructions.

For model catalog changes, also run go test ./pkg/model ./pkg/external/codex ./pkg/acp/codex ./pkg/agent to cover discovery, backend IDs, Codex metadata, and request construction. Prompt text changes still require matched behavioral evaluations before claiming task-success or latency improvements.

Claude Sonnet 5.5

models/claude-sonnet-5-5/mode_agent.txt adapts the effective Claude Code prompt captured on 2026-09-28 with --model claude-sonnet-5-5. The source is Claude Code 2.1.284, release commit 2b8ce618c24de26410e4bdfc4e1d592accd61f61, built on 2026-09-28. Its temporary Darwin ARM64 binary was downloaded from Anthropic's release distribution and checked against the release manifest. SHA-256: 50a14c2f50f56668380fdda490167f1d3630d5cc18fb8aed3073c2c7ea7314fe.

Capture the first Messages API request's system blocks using a localhost mock API and a dummy API key, an empty working directory and CLAUDE_CONFIG_DIR, --setting-sources "", --strict-mcp-config --mcp-config '{"mcpServers":{}}', --settings '{"disableAllHooks":true,"autoMemoryEnabled":false}', and --disable-slash-commands --no-session-persistence --max-turns 1 -p. Disable nonessential traffic and automatic updates. This extracts the CLI's rendered instructions without asking a model to reproduce them or making an inference request to Anthropic. The same capture with claude-sonnet-5 confirms that the two models select different prompts in this release.

Sonnet 5.5 selects a lean prompt containing harness guidance, code style, careful actions, truthful reporting, context continuation, and acting on established information. Preserve this structure rather than inheriting the general Claude prompt's task and output sections or Opus's communication, delivery, and correction sections.

Wingman's adaptation replaces the product identity and dedicated tool names, describes its appended session-context messages instead of Claude Code's mid-conversation system turns, and retains Wingman's existing software-security and explicit-request-only Git rules. Billing metadata, the Claude model/product catalog, token countdown, and unrelated identity guidance are omitted. Project, skills, memory, and environment sections still come from the shared templates. The existing longest-prefix router selects this variant for Sonnet 5.5, including normalized, provider-prefixed, and dated IDs; plan and unattended instructions still use the shared modes.

Run go test ./pkg/code/prompt ./pkg/code/agent to check routing, rendering, and session integration. These checks do not measure Sonnet 5.5's behavior.

Claude Code review on September 30

Captured the rendered requests for Claude Sonnet 5/5.5, Opus 4.8/5/5.5, Fable 5/5.1, and Mythos 5/5.1 with Claude Code 2.1.285. The installed Darwin ARM64 binary matches the official release manifest: commit afb212976052ab038df25d5e871f6c049e094d3b, built September 29, SHA-256 51f09bd1e021d9fa8a1864c179799bd37cb39962a937935c5cf6823398e86db4. The capture uses the isolation described above, default CLI tools, a local mock API, and a dummy credential. It makes no model inference requests. The source captures and a dated report are retained in ../wingman-ops/runs/2026-09-30.

Translate Claude's system-turn/reminder wording to Wingman's appended session context, without granting ordinary tool results instruction authority. Keep software-security guidance and explicit-request-only Git rules. Omit billing, product catalogs, token countdowns, and unrelated identity guidance. The new 5.1 variants do not assume every session is unattended; the shared unattended and plan modes remain the policy owners. Longest-prefix routing distinguishes 5.1 from 5, including provider-prefixed, normalized, dated, and tagged IDs.

Validate routing, rendering, and mode composition with the existing focused Go suites. These checks do not establish a behavioral or latency improvement; matched live model evaluations were not part of this refresh.

Claude Haiku 5.5 review on October 8

models/claude-haiku-5-5/mode_agent.txt adapts the effective prompt captured from the installed Claude Code 2.1.294 with --model claude-haiku-5-5. The Darwin ARM64 binary matches the official release manifest: commit 8f033c6ebe3d82a87f502e199307f38f5d55ccca, built October 8, SHA-256 def0d15e64dd7d89621f88d28214f885b1c38b0ddd69762fb8593e34915d6d53. The capture uses the isolated localhost mock API procedure above and dummy credentials, with no Anthropic inference. Haiku 4.5 and Sonnet 5.5 were captured for comparison. Haiku 5.5 has its own lean prompt, distinct from both models; Sonnet 5.5 retains its existing structure.

Retain Haiku's new context-management guidance: effort controls reasoning depth rather than task scope; complete clear requests without unnecessary approval; distinguish questions from implementation requests; continue work that does not depend on a blocker; and verify a suspected failure before reporting it. Translate tools, identity, and session-context authority to Wingman's contracts. Preserve the established security and Git policies; omit billing, product catalogs, token countdown, and unrelated identity rules. The shared plan and unattended templates continue to own those modes.

Claude Code 2.1.292 did not recognize Haiku 5.5 and selected the general prompt; use 2.1.294 or newer when repeating this capture. Raw captures from this review are in /private/tmp/wingman-claude-prompt-review and are temporary artifacts. Validate with the focused prompt, agent, and Claude ACP suites. No live model behavior or latency improvement is claimed.

After the local backend was started, six direct Haiku 5.5 checks passed through its configured Bedrock model: default thinking, explicitly summarized adaptive thinking, signed replay, disabled thinking with xhigh, a forced tool with max, and Responses streaming with developer instructions in the input. A separate Claude Code 2.1.294 check selected the haiku alias, reported the resolved model as claude-haiku-5-5, read a temporary fixture with Read, and returned the expected calculation. These establish functional compatibility; they are not a matched prompt-quality or latency benchmark.

Gemini review on October 13

models/gemini/mode_agent.txt adapts gemini-cli's system prompt from packages/core/src/prompts/snippets.ts at commit 29315e53d0c2e9b78f301f3c24db235b0f3fa597 (2026-10-13, Apache-2.0). It replaces an earlier rewrite of VS Code Copilot Chat's geminiPrompts.tsx. Every Gemini model, including Gemini 4 Argon, uses this family prompt.

The adaptation follows gemini-cli's modern-model composition for an interactive session in Default approval mode, with grep, glob, sub-agents, and a git repository available, and with topic narration, the task tracker, the plan-mode tool, and sandboxing disabled. It keeps Core Mandates, Context Efficiency, Engineering Standards, sub-agent delegation, hook-context handling, Primary Workflows, Operational Guidelines, and the Git rules. Tools and parameters are translated to Wingman's contracts: explore stands in for codebase_investigator, batched edit calls replace the one-edit-per-file rule, and dependent calls wait for a later turn instead of wait_for_previous.

Product identity, the mode banner, GEMINI.md memory routing, skill activation, plan, YOLO, and non-interactive policies are omitted; the shared Memory, Skills, plan, and unattended templates own those. Sandbox recovery and /help//bug are omitted. Wingman's software-security guidance is retained.

Google's Gemini 4 Argon model page and prompting guide (both updated 2026-10-13) have no Argon-specific system-instruction guidance. They state that Gemini 4 ignores temperature, top_p, and top_k; this repository does not set them. The AI Studio developer guide renders client-side and could not be read here.

gemini-cli requires SWE-bench runs before changing its Context Efficiency section. The focused prompt and agent suites check routing and rendering only; use matched bench/acpflow runs before claiming behavioral improvements.