fix: default unknown models to the full prompt, single-source product identity and version

[konard]

Summary

Fixes #285 — three defects that made a session's behaviour unattributable, all fixed with the generalizations requested in the issue rather than another if.

1 — Unknown models no longer fall back to the without-todo prompt

js/src/session/system.ts now resolves the prompt through a rule table with an explicit default:

• default is anthropic.txt (12 mentions of todo) instead of qwen.txt (0);
• the resolution is logged on every run — system prompt: anthropic (default for unknown model formalai/formal-ai) (service system-prompt), so a run's prompt is attributable from its log;
• models that genuinely break on todo tools opt out explicitly via AGENT_SYSTEM_PROMPT=anthropic-without-todo (also usable to pin any other prompt: anthropic, beast, codex, gemini, grok-code, polaris). An unknown value is ignored with a warning. This also removes the "no workaround from the caller side" gap.

2 — The agent no longer says it is opencode

New js/src/branding.ts holds the product identity (name, repo, issues URL, docs URL) in one place and substitutes it at render time, so the prompt files stay byte-identical to their upstream originals and a rename is one edit instead of eight. Applied to every provider prompt, the summarize/title prompts, and the bash tool description (which also mentioned opencode).

Bug reports the model is told to file now go to https://github.com/link-assistant/agent/issues.

3 — Session records report the real version

New js/src/version.ts reads the package manifest and is the single source for --version, the process log and Session.createNext (js/src/session/index.ts previously hard-coded agent-cli-1.0.0).

How to reproduce the original behaviour

agent --model formalai/formal-ai --print-logs --verbose "say hi"

Before: the system message begins You are opencode, an interactive CLI tool…, contains no todo instructions, points users at sst/opencode/issues, and the stored session record says "version": "agent-cli-1.0.0".
After: the system message is the full anthropic prompt with todo discipline, all product references point at this repo, and the session record carries the real version (0.25.3).

Tests

js/tests/system-prompt.ts (new):

• unknown model ids resolve to the default prompt and that prompt contains todo instructions — this is the check that makes the class of regression impossible;
• known model ids keep their dedicated prompt;
• AGENT_SYSTEM_PROMPT opt-out works and invalid values are ignored;
• no rendered prompt matches /opencode/i, sst/opencode or opencode.ai — branding cannot regress when a prompt is re-synced from upstream;
• SystemPrompt.header still returns the anthropic OAuth spoof (api-surface guard).

js/tests/version.ts (new):

• VERSION === package.json version, is semver-shaped, and session records use the shared constant.

npm test
 623 pass  4 todo  0 fail  1274 expect() calls
 Ran 627 tests across 52 files.

npm run check (eslint + prettier + file size) passes.

[konard]

Working session summary

Done. PR: #286 (ready for review, all CI green — run 30571970748).

What changed — all three defects from issue #285:

1. Unknown models no longer get the without-todo prompt. js/src/session/system.ts now resolves the prompt from a rule table with an explicit default of anthropic.txt (12 todo mentions) instead of qwen.txt (0), logs the resolution on every run (system prompt: anthropic (default for unknown model formalai/formal-ai), service system-prompt), and supports an explicit opt-out / pin via AGENT_SYSTEM_PROMPT — which also fills the "no caller-side workaround" gap.

2. The agent no longer identifies as opencode. New js/src/branding.ts holds name/repo/issues/docs in one place and substitutes them at render time, so the eight prompt files stay byte-identical to their upstream originals and a rename is one edit. Also applied to the bash tool description, which mentioned opencode too. Bug reports now route to link-assistant/agent/issues.

3. Session records carry the real version. New js/src/version.ts reads the package manifest and is the single source for --version, the process log, and Session.createNext (was hard-coded agent-cli-1.0.0).

Tests (js/tests/system-prompt.ts, js/tests/version.ts): unknown model ids must get a prompt containing todo instructions; no rendered prompt may match /opencode/i, sst/opencode or opencode.ai; VERSION === package.json version.

One thing worth flagging: my first push broke every session — the refactor edit accidentally dropped SystemPrompt.header (the anthropic OAuth spoof), which the CI integration job caught. It's restored in 072d9d6 with an api-surface test guarding it.

---

This summary was automatically extracted from the AI working session output.

[konard]

🤖 Solution Draft Log

This log file contains the complete execution trace of the AI solution draft process.

💰 Cost: $3.586678

📊 Context and tokens usage:

• 81.8K / 1M (8%) input tokens, 24.0K / 128K (19%) output tokens

Total: (148 new + 68.0K cache writes + 4.6M cache reads) input tokens, 24.0K output tokens, $3.586678 cost

🤖 Models used:

• Tool: Anthropic Claude Code
• Requested: opus (claude-opus-5)
• Thinking level: off (disabled)
• Model: Claude Opus 5 (claude-opus-5)

📎 Log file uploaded as Gist (1832KB)

• View complete solution draft log

---

Now working session is ended, feel free to review and add any feedback on the solution draft.

[konard]

🎉 Auto-merged

This pull request has been automatically merged by hive-mind.

• All CI checks have passed

---

Auto-merged by hive-mind with --auto-merge flag