Spaces:
Running
Running
Commit History
-
Swap Groq default model to openai/gpt-oss-120b (llama-3.3-70b-versatile decommissioned)
muralipala15046bdf502
-
Revert "prompt: command-first structure + 20 hooks for retention improvement"
muralipala1504This reverts commit 1ed7f67fc414c4f4630a0a11cce9afc725b20345.7cb2f81 -
prompt: command-first structure + 20 hooks for retention improvement
muralipala15041ed7f67
-
Revert "prompt: add sentence length + hook delivery rules for TTS retention"
muralipala1504This reverts commit 93a8e188f6e6dd7bc3255a31efe11e9bdc0fd4b9.c9e68c0 -
prompt: add sentence length + hook delivery rules for TTS retention
muralipala150493a8e18
-
security: add daily prompt quota (5 per IP per day)
muralipala1504- Track prompt usage per IP per day in memory - Daily limit: 5 prompts/day per IP (resets at 00:00 UTC) - Return 429 if quota exceeded with clear message - Works with Cloudflare CF-Connecting-IP + X-Forwarded-For fallback - Keeps existing per-minute rate limits (10/min on streaming endpoint) - Quota dict auto-cleans entries older than 7 days - Thread-safe with quota_lock mutex0d92c98 -
security: add key masking, startup checks, .env.example template
muralipala1504- Add mask_key() utility to hide API keys in logs (shows first 4 + last 4 chars) - Add masked key logging to Groq and Cerebras client initialization - Add startup security check to log key status and warn about rotation - Create .env.example template with all required env vars documented - Never expose full API keys in logs or error messagescaa546b
-
feat: Google TTS EN voice (en-IN-Standard-B) — Indian male accent staging
muralipala150431492f8
-
Trainer prompt: app-level hook rotation + fix TTS language reset on clear
muralipala1504Hook rotation fix: - Previous prompt-level 'vary your hook' instruction failed - 3 consecutive Trainer responses all used the same library entry ('explaining to your manager why production is down'), confirming the model can't reliably self-rotate across independent stateless API calls - Moved rotation to the application layer instead: added HOOK_LIBRARY + get_rotating_hook() in llm.py, which picks a hook index at request time and avoids repeating the immediately prior selection - get_trainer_prompt_with_hook() injects the one selected hook into TRAINER_SYSTEM_PROMPT per-request, replacing the old approach of listing all 10 options and hoping the model varies on its own - __main__.py now calls get_trainer_prompt_with_hook() instead of using the static TRAINER_SYSTEM_PROMPT constant directly (both streaming and fallback call sites, lines 332/399) - Verified: 5 consecutive test calls produced 5 different hooks with no immediate repeats TTS language reset fix: - L…d4462df
-
Trainer prompt: add expanded hook library to fix overused '2am debugging' pattern
muralipala1504- '2am debugging' hook was appearing in 2/3 videos despite earlier variety instruction - explicit ban + concrete alternatives needed - Added 10-item hook library covering different consequence angles: career stakes, specific DevOps tools (pods, Terraform, Docker, CI/CD), peer comparison, production incidents - Model instructed to adapt wording to topic, not reuse verbatim, and not repeat the same library entry twice in a row - Not yet validated on dev - pending test batch across multiple topicsa2e4e54
-
Trainer prompt: add opening hook with variety, single-cycle structure, hard stop after pro tip
muralipala1504- Hook now leads every response (no more In this video preamble) - Rotates hook style (blunt statement, question, incident, consequence) instead of defaulting to 'Ever struggled with' - STRUCTURE RULE: one full teaching cycle per response instead of repeating the What it does -> Pro tip block for sub-concepts - HARD STOP RULE: stops cleanly after Pro tip, no more rambling conclusion/recap paragraphs - Attempted emoji removal from section headers (partial - model still adds emojis despite explicit instruction, left as-is for now)75a11e7
-
reduced the token cound and redundant chat conversation
muralipala1504eab7b08
-
Trainer: bump max_tokens to 4000, set 1300-word target in system prompt
muralipala150473d6b87
-
Add startup diagnostic for VISITS_LOG_TOKEN env var
muralipala15044f08793 -
Fix self-referential Content-Length bug: exclude /admin and meta paths from visit logging
muralipala1504d1b0d28 -
Add first-party visit logging middleware + analyzer
muralipala1504be6dc21
-
feat: add DeepShell favicon (ico, svg, 192px png)
muralipala150428d2c01 -
feat: add sitemap.xml and robots.txt endpoints
muralipala15049f7589a
-
Clean up: remove debug lines, restore health endpoint
muralipala15049fbc637 -
Debug: add repo_root to health endpoint
muralipala15047a954f4 -
Add About and Contact pages with navbar links
muralipala15042ed385e
-
security: use CF-Connecting-IP for real user IP rate limiting
muralipala1504cc8bcc3 -
security: harden endpoints — rate limit, CORS, admin key, IP logging
muralipala15043d7429d
-
feat: Piper Hindi TTS + backend translate/tts endpoints, drop direct LibreTranslate from frontend
muralipala1504b10a301
-
security: remove litellm (supply chain risk) + unused CLI deps, add openai
muralipala1504523ea15 -
feat: add Cerebras provider as Groq failover
muralipala1504a26a2bd
-
chore: remove legacy CLI files - web UI only
muralipala150468594d6 -
feat: add Ollama provider support - run fully local with zero API dependency
muralipala150414b51e4 -
feat: add Dry Run mode + TTS Voice + Replay button
muralipala15044e6e168
-
feat: add Trainer mode - structured DevOps teaching with toggle
muralipala1504b7a0abe
-
feat: add Docker support + fix __init__.py config import
muralipala1504596438d
-
Fix: Remove OpenAI CLI import from backend __init__.py
muralipala1504- Removed 'from .cli import app' which was triggering OpenAI setup wizard - Backend now imports only LLM client and config (Groq-only) - Fixes issue where importing deepshell would prompt for OpenAI API key - Backend now starts cleanly without any setup prompts1753af7
-
did changes to rate limits
muralipala1504dca7ec9 -
Add slowapi for rate limiting
muralipala15041dd02ae
-
feat: Add strict system prompt to refuse off-topic questions
muralipala1504- DeepShell now only answers DevOps/SysAdmin questions - Politely refuses general knowledge, entertainment, etc. - Focuses on Docker, K8s, Linux, Ansible, Terraform, CI/CD - Maintains professional boundaries6af529f -
feat: Remove execute button feature, add copy buttons for free version
muralipala1504b99a374
-
Add Phase 1 command execution: SSE streaming + Execute button with safety checks
muralipala150458ab20c -
Fix: Remove block logic, update to llama-3.3-70b-versatile, delete unused main.py
muralipala1504d03090c
-
refactor: move app to deepshell.main; keep main shim for python -m support
muralipala1504ef3975d -
chore: hard-block chat and show admin-only scope message; silence favicon
muralipala150427bfdc0
-
feat: wire Groq and serve UI from backend
muralipala15048574870 -
feat: wire ChatHandler to global LLM client (Groq/OpenAI)
muralipala1504ade4144 -
feat: replace wrapper with GroqWrapper (sync, non-stream)
muralipala1504945dbc6
-
Add Cases tab: browse, view, and download Clinic diagnoses
muralipala1504863efcd
-
Clinic: auto-archive diagnoses to clinic_cases/
muralipala150414a014a -
app.js,index.html, deepshell-back-end/wrapper.py changed to facilitatae Clinic Feature
muralipala150499051d0
-
feat(wrapper): harden domain filtering (block trivia, finance, and playful context) - v1.2.0
muralipala1504dc53e70 -
feat(wrapper): enforce strict default-deny domain filtering (block non-infra prompts)
muralipala150450af3d6
-
feat(wrapper): add domain filter + clean code-fenced outputs for copy button
muralipala1504d3376f3
-
Add fastapi and uvicorn to requirements.txt
muralipala150413ec373