v0.8.1 — « Honesty » (patch release, no new features)

Patch release tracking the v0.8.0 audit.

🔒 Security

• Console binds 127.0.0.1 by default, --bind flag now honored
• Clean refusal when port is already taken, --bind validation

🐛 Fixed — Tool execution (the DeepSeek bug)

• Multi tool calls no longer overwrite the first call
• Tool calls drained on finish_reason: stop instead of discarded
• Ollama native queries /api/show for live capabilities and context size

🐛 Fixed — Markup recovery

• 4 new formats recovered: JSON fences, [TOOL_CALL], DeepSeek native tags, function.arguments JSON
• Fallback parser respects string="true", no longer trims values
• Markup detection requires both open AND close structure

✅ Fixed — Approvals

• Non-interactive mode refuses cleanly instead of blocking on stdin
• Approval card uses human language via humanize_tool_action
• Web approval broker expires after hard timeout, returns deny
• WaitingForApproval status visible in cockpit

🇫🇷 Fixed — Tool calling

• French anti-narration guard: detects « je vais créer… », « laisse-moi vérifier… »
• System prompts modulated by tier (trivial/small use lean prompt)
• Curator no longer creates skills from UI status text
• History filtered at serialization: no UI status lines re-injected

📊 Fixed — Numbers

• Real provider usage replaces estimates instead of adding to them
• duration_ms in OutcomeSummary, replay uses stored duration
• DeepSeek models receive non-zero pricing, re-enabling cost comparisons
• CLI only prints 'Done.' for completed runs

🎭 Fixed — Noise

• Router message is now a single clear line instead of franglais duplicate
• Verifier lane no longer falsely shows 'run closed' without verification
• Favicon inline, no more 404 on every page load
• reasoning_content captured from single source, preventing doubled context