Commit History
-
Add CLI sandbox runtime and fix HF Jobs script paths (#237)
Unverified signature* Resolve bare Python job scripts Co-authored-by: OpenAI Codex <codex@openai.com> * Update HF Jobs script description Co-authored-by: OpenAI Codex <codex@openai.com> * Clarify HF Jobs script path prompt Co-authored-by: OpenAI Codex <codex@openai.com> * Add CLI sandbox tool runtime Co-authored-by: OpenAI Codex <codex@openai.com> * Document CLI sandbox tool runtime Co-authored-by: OpenAI Codex <codex@openai.com> * Wait for initial CLI sandbox preload Co-authored-by: OpenAI Codex <codex@openai.com> * Strengthen GPU sandbox preflight guidance Co-authored-by: OpenAI Codex <codex@openai.com> * Guard against text-only stops with unfinished plans Co-authored-by: OpenAI Codex <codex@openai.com> * Route sandbox deletion logs through tool events Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>b05b6f5
-
Create Hub artifact collections lazily (#245)
Unverified signature* Create Hub artifact collections lazily Co-authored-by: OpenAI Codex <codex@openai.com> * Cache lazy Hub collection slugs in child hooks Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>60474c1 -
Implement `/resume` command for CLI (#233)
Unverified signature* Add /resume command to CLI * Refine /resume feature * Refine /resume implementation * Address review comments --------- Co-authored-by: lewtun <lewis.c.tunstall@gmail.com>0321690
-
Skip sandbox spaces for artifact metadata (#241)
Unverified signature* Skip sandbox spaces for artifact metadata Co-authored-by: OpenAI Codex <codex@openai.com> * Address sandbox metadata PR feedback Co-authored-by: OpenAI Codex <codex@openai.com> * Avoid sandbox title metadata updates Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>1e1fe1f
-
Add CLI local model support (#228)
Unverified signature* Add CLI local model support Co-authored-by: OpenAI Codex <codex@openai.com> * Add shared local model endpoint fallback Support LOCAL_LLM_BASE_URL and LOCAL_LLM_API_KEY as shared fallbacks while preserving provider-specific local overrides. Co-authored-by: OpenAI Codex <codex@openai.com> * Address local model review feedback Clarify local probe failure behavior, add regression coverage for rejected local switches, and simplify local model validation. Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>4668dbd
-
Add ML Intern Hub artifact metadata (#225)
Unverified signature* Add ML Intern Hub artifact metadata Co-authored-by: OpenAI Codex <codex@openai.com> * Extend Hub artifact hooks to sandbox bash Co-authored-by: OpenAI Codex <codex@openai.com> * Create Hub artifact collections at session start Co-authored-by: OpenAI Codex <codex@openai.com> * Address Hub artifact PR review comments Co-authored-by: OpenAI Codex <codex@openai.com> * Respect Hub collection title limits Co-authored-by: OpenAI Codex <codex@openai.com> * Shorten collection session UUID fragments Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>4fc6e96
-
Add CI for tests and Ruff (#217)
Unverified signature* Add CI for tests and Ruff Co-authored-by: OpenAI Codex <codex@openai.com> * Enforce Ruff formatting in CI Co-authored-by: OpenAI Codex <codex@openai.com> * Document Ruff pre-commit checks Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>754345f -
fix(compaction): break infinite loop + truncate oversized messages (#213)
Guillaume SalouUnverified signature* fix(compaction): break the infinite-compaction loop, truncate oversized messages Sessions stuck in a compaction retry loop have been silently burning Bedrock budget while staying invisible in the session dataset. Pod logs from 2026-05-03 on prod-114 showed the pattern firing minute-after-minute on multiple replicas: Context compacted: 200001 -> 215566 tokens Context compacted: 215566 -> 215572 tokens ContextWindowExceededError — forcing compaction Context compacted: 200001 -> 215544 tokens ... Root cause: a single message in the "untouched" tail (typically a tool output of 80k+ tokens — bash dump, file content, CSV) keeps the post-compact context above the 90% threshold. Compaction triggers again, calls Bedrock again, gets the same 215k result, loops. ~$3 per Opus retry, indefinitely. Sessions never reach _run_session.finally → save_and_upload_detached never fires → cost is real but never lands in total_cost_usd. Cross-check: dataset Bedrock-only coverage on 2026-05-01 …d84b454
-
Auto-start CPU sandboxes for sessions (#200)
Unverified signature* Auto-start CPU sandboxes for sessions Co-authored-by: Codex <codex@openai.com> * Retry sandbox runtime visibility checks Co-authored-by: Codex <codex@openai.com> * Stabilize auto CPU sandbox creation Co-authored-by: OpenAI Codex <codex@openai.com> * Address sandbox PR review comments Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: Codex <codex@openai.com>8615c28 -
Add session YOLO auto-approval budget (#201)
Unverified signature* Add session YOLO auto-approval budget Co-authored-by: Codex <codex@openai.com> * Address YOLO approval review feedback Co-authored-by: Codex <codex@openai.com> --------- Co-authored-by: Codex <codex@openai.com>77324b8 -
Use HF username for personal trace uploads (#199)
Unverified signature* Use HF username for personal trace uploads Co-authored-by: OpenAI Codex <codex@openai.com> * Remove redundant HF token branch Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>71e1892 -
Auto-share session traces to a private HF dataset (#193)
Darshan Thakare +2: lewtun, Codex Unverified signature* feat: add share_traces toggle and per-user trace repo template * feat: support Claude Code JSONL format and per-target auth * feat: dual-upload sessions to private user trace dataset * chore: retry personal trace uploads on booting * feat: add /share-traces command to flip dataset visibility * docs: document HF trace auto-share and /share-traces * Use HF token owner for local dev auth Co-authored-by: Codex <codex@openai.com> * Rename personal session trace dataset Co-authored-by: Codex <codex@openai.com> * Add session dataset card metadata Co-authored-by: Codex <codex@openai.com> * Fix session trace upload review issues Co-authored-by: OpenAI Codex <codex@openai.com> * Preserve secret scrubbing before trace uploads Co-authored-by: OpenAI Codex <codex@openai.com> * Link ML Intern demo in dataset card Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: lewtun <lewis.c.tunstall@gmail.com> Co-authored-by: Codex <codex@openai.com>0bd7547
-
Add DeepSeek V4 Pro model option (#189)
Unverified signature* Add DeepSeek V4 Pro model option Co-authored-by: Codex <codex@openai.com> * Remove DeepSeek feature tests Co-authored-by: Codex <codex@openai.com> --------- Co-authored-by: Codex <codex@openai.com>1b922dd
-
feat(telemetry): track 5 untracked Bedrock call sites for full cost attribution (#179)
Guillaume SalouUnverified signature* feat(telemetry): track 5 untracked Bedrock call sites for full cost attribution Cost Explorer ($78,738 over 6 days) vs the session dataset's total_cost_usd (~$354/day attributed) showed the dataset captures only ~33% of real Bedrock spend. Root cause: out of 9 acompletion() call sites, only 2 (in agent_loop.py) emit the llm_call event that total_cost_usd sums. This wires telemetry into the 5 Bedrock-billing call sites that were flying blind, with a `kind` tag on each call so analytics can split spend by category: - research_tool.py × 3 → kind="research" (sub-agent loop) - context_manager.py → kind="compaction" (history summary) - effort_probe.py → kind="effort_probe" (cascade walk) Plus a fourth tag for the session-restore summary path (session_manager.py → kind="restore"). Plumbing changes: - telemetry.record_llm_call now accepts kind="..." (default "main" preserves existing behavior). - summarize_messages() and ContextManager.compact() take optional ses…7867a7a
-
Track Pro conversions + credits top-up; emit user_id + new KPIs (#174)
Aksel Joonas ReediUnverified signatureAdds two telemetry events the dashboard needs to answer "how many users converted to Pro?" and "how many came back from a billing block?": - pro_conversion: per-user Pro state lives in a new Mongo `pro_users` collection. mark_pro_seen upserts on every authenticated request and uses an atomic find_one_and_update to fire at-most-once when a user first appears as Pro after having been seen as non-Pro. Wired through session_manager.create_session via the existing `is_pro` signal from /auth. - credits_topped_up: fires from jobs_tool when an hf_job submit succeeds in a session that previously hit a billing-required tool_state_change. Guarded against re-firing within a session. build_kpis aggregates both as plain count columns. CLI/local users with no Mongo silently skip via the NoopSessionStore stub.2715896 -
Bill HF Jobs by namespace credits, drop Pro and org-join gating (#172)
Aksel Joonas ReediUnverified signature* Bill HF Jobs by namespace credits, drop Pro and org-join gating - Drop the Pro/Team subscription pre-check on hf_jobs. Any namespace the caller belongs to is eligible; HF returns a billing error at job creation time if the wallet is empty, which the agent surfaces as a "top up to launch" dialog pointing at /settings/billing. - Remove the "Join ML Agent Explorers" onboarding step. The org-join hook, /auth/org-membership endpoint, and the OAuth orgIds suggestion are gone. - Rework JobsUpgradeDialog: namespace picker shows all eligible accounts/orgs with avatars + select rings; billing mode opens the HF billing page; "Skip this tool call" stays available for both. - Backend pre-flight only forces the picker when the caller has multiple eligible namespaces; single-namespace users default to their personal account silently. - jobs_tool now catches HfHubHTTPError, detects billing failures by pattern + 402 status, emits a tool_state_change(state=billing_required) so t…bed4e38 -
Persist Space sessions in MongoDB (#170)
Aksel Joonas ReediUnverified signature* Persist Space sessions across backend restarts The hosted Space lost active conversations because sessions, approval state, event streams, and Claude quotas only lived in process memory. This adds a shared optional Mongo store for frontend sessions while keeping local and CLI runs on the no-op/local path when Mongo is absent. Constraint: Space deployment needs durable state without making PyPI/local CLI users configure Mongo. Constraint: Frontend deletes must hide sessions with a visibility tag, not erase traces needed for recovery or SFT export. Rejected: Restore every active session at startup | large user counts would create a thundering herd after deploys. Rejected: Store one giant session document | Mongo document limits make per-message snapshot docs safer. Confidence: medium Scope-risk: broad Directive: Keep session_messages as compactable runtime snapshots and session_trace_messages/session_events as append-only raw traces; do not collapse them into one collection wit…d9d9785
-
Recognize Pro OAuth users via isPro flag (#164)
Aksel Joonas ReediUnverified signatureOAuth whoami responses set type="user" and only signal Pro status with the isPro boolean. The previous code matched the type key first and only fell through to isPro when no plan/type/accountType string was present, so type="user" + isPro=true was classified as free and Pro OAuth users got blocked from running Jobs. Reorder the check so isPro takes priority. Add a regression test for the OAuth shape. Reported in smolagents/ml-intern Space discussion #21.895c557 -
Add web search so the agent can cite current sources (#159)
Aksel Joonas ReediUnverified signatureThe agent had research tools for papers, docs, and repositories, but no direct current-web lookup. This ports Claw Code's WebSearch behavior into a Python tool that searches DuckDuckGo HTML, extracts citeable title/URL hits, applies domain filters, and returns the same JSON-shaped result payload for model consumption. Constraint: The implementation was prepared in a writable checkout because the primary working tree was sandbox read-only during this session. Rejected: Add a new search API dependency | the Claw implementation uses the DuckDuckGo HTML endpoint and the repo already has requests. Confidence: medium Scope-risk: narrow Directive: Keep the output schema stable unless the agent prompt/tool consumers are updated together. Tested: pytest tests/unit/test_web_search_tool.py tests/unit/test_malformed_args_recovery.py -q Tested: python -m compileall agent/tools/web_search_tool.py agent/core/tools.py agent/tools/research_tool.py agent/tools/__init__.py tests/unit/test_web_sear…72f615f -
Add Slack gateway (#116)
Unverified signature* Add messaging gateway and notify tool Co-authored-by: OpenAI Codex <codex@openai.com> * Handle Bedrock streaming permission denials Co-authored-by: OpenAI Codex <codex@openai.com> * Revert "Handle Bedrock streaming permission denials" Co-authored-by: OpenAI Codex <codex@openai.com> * Add automatic completion notifications Co-authored-by: OpenAI Codex <codex@openai.com> * Auto-attach CLI notification destinations Co-authored-by: OpenAI Codex <codex@openai.com> * Defer CLI completion notifications until render Co-authored-by: OpenAI Codex <codex@openai.com> * Increase completion notification summary cap Co-authored-by: OpenAI Codex <codex@openai.com> * Add Slack user notification defaults Co-authored-by: Codex <codex@openai.com> * Increase Slack turn completion summary limit Co-authored-by: Codex <codex@openai.com> * Remove legacy auto notification event upgrade Co-authored-by: Codex <codex@openai.com> * Require session config instead of hard-coded model fallback C…6155b26 -
fix(session-uploader): forward user_id and total_cost_usd to dataset row (#153)
Guillaume SalouUnverified signaturePR #136 added user_id + total_cost_usd to Session.get_trajectory(), but the upload path builds session_row with hardcoded fields and drops anything extra. Result: rows in smolagents/ml-intern-sessions still show user_id=null even though the trajectory dict carries the values. Confirmed by kubectl exec on the running pod: session.py is patched correctly, but uploaded sessions still have null fields because they go through this hardcoded shape. Forward both fields explicitly via .get() so older sessions that lack them still upload cleanly with null.2d4ec20 -
Use `huggingface_hub.get_token()` for auth fallback (#126)
Unverified signature* Use cached HF login token for router auth fallback Let HF router requests resolve credentials from huggingface_hub's local login cache when INFERENCE_TOKEN, session token, and HF_TOKEN are absent. This keeps CLI and plugin flows working after `hf auth login` without requiring explicit env exports. Made-with: Cursor * Consolidate Hugging Face token resolution Centralize HF token cleanup and lookup so router calls, CLI startup, and backend request paths do not drift after adding huggingface_hub.get_token() fallback. The backend request helper intentionally keeps browser/user token precedence separate from local cached CLI auth. Constraint: huggingface_hub.get_token() already handles HF_TOKEN, HUGGING_FACE_HUB_TOKEN, and the local hf auth login cache. Rejected: Use cached login tokens for backend request auth | would let server-local credentials stand in for the browser user. Confidence: high Scope-risk: narrow Directive: Keep router-token fallback and request-user token extrac…7a76ad1 -
Preserve streamed thinking metadata with live model tests (#150)
Aksel Joonas ReediUnverified signature* Add opt-in live thinking model tests Add paid integration coverage for the concrete models requested for #143: Anthropic Opus 4.7 and OpenAI's current GPT-5.2 model. The tests load an explicit env file, run only behind ML_INTERN_LIVE_LLM_TESTS=1, and keep normal CI credential-free. Constraint: Live provider calls require local credentials and should not run by default in CI. Rejected: Make live provider tests unconditional | would fail or spend tokens anywhere credentials are absent. Confidence: high Scope-risk: narrow Tested: ML_INTERN_LIVE_LLM_TESTS=1 ML_INTERN_LIVE_ENV_FILE=/Users/akseljoonas/Documents/ml-intern/.env UV_CACHE_DIR=/tmp/uv-cache uv run --extra dev pytest tests/integration/test_live_thinking_models.py -q -rs Tested: UV_CACHE_DIR=/tmp/uv-cache uv run --extra dev pytest tests/unit/test_thinking_history.py tests/integration/test_live_thinking_models.py -q * Preserve streamed Opus thinking metadata Live Opus 4.7 exposed that LiteLLM surfaces streamed thinking bl…d408a51 -
Preserve thinking state across tool turns (#143)
Aksel Joonas ReediUnverified signature* Preserve thinking state across tool turns Anthropic thinking responses need their thinking_blocks and reasoning_content replayed with assistant tool-call messages. The loop was rebuilding assistant history from only content and tool calls, causing LiteLLM to strip thinking on continuation turns. Constraint: Non-thinking providers and responses without reasoning fields must keep the existing message shape. Rejected: Disable extended thinking for tool-using runs | avoids the warning by removing the feature that improves reasoning quality. Confidence: high Scope-risk: moderate Directive: Any future assistant-message reconstruction must preserve provider reasoning fields when present. Tested: UV_CACHE_DIR=/tmp/uv-cache uv run --extra dev pytest tests/unit/test_thinking_history.py tests/unit/test_dangling_tool_calls.py tests/unit/test_malformed_args_recovery.py * Replay thinking metadata only for Anthropic Review caught that reasoning_content is not safe to echo through OpenAI-co…59b2038 -
Keep repetition guard internal (#144)
Aksel Joonas ReediUnverified signature* Clarify repetition guard messaging The loop breaker is intentional recovery behavior, but the user-facing 'doom loop' phrasing reads like a crash. Rename the visible log and system hint to a repetition guard while preserving the existing detector behavior and historical SFT tag compatibility. Constraint: Existing trajectories may still contain the old wording, so the tagger must recognize both labels. Rejected: Remove the event entirely | users would lose visibility into why the agent changed strategy. Confidence: high Scope-risk: narrow Directive: Keep user-facing recovery logs operational and non-alarming; reserve internal jargon for code, not UI events. Tested: UV_CACHE_DIR=/tmp/uv-cache uv run --extra dev pytest tests/unit/test_doom_loop_polling.py tests/unit/test_sft_tagger.py * Hide repetition guard from user surfaces The repetition guard is an internal control-flow intervention, not a status event users need to act on. Remove the CLI/frontend tool_log emissions while …e8ed637 -
fix(doom_loop): normalize tool-call args before hashing (#119)
d 🔹 +2: d 🔹, Aksel Joonas ReediUnverified signatureThe doom-loop detector hashed raw `function.arguments` strings, so semantically-identical tool calls hashed differently when the LLM emitted them with different key orderings (`{"a":1,"b":2}` vs `{"b":2,"a":1}`) or whitespace (`{"a":1}` vs `{"a": 1}`). This silently broke `detect_identical_consecutive` and `detect_repeating_sequence`: the agent could be calling the same tool with the same args repeatedly and the detector would see three distinct signatures and stay quiet. Issue #61 P1 explicitly calls this out: > Add semantic-similarity or normalized-task matching for `research`. Fix: parse-and-redump JSON via `json.dumps(..., sort_keys=True, separators=(",", ":"))` before hashing. Falls back to the raw string when the input isn't valid JSON so non-JSON `arguments` strings (rare edge for some providers) keep the legacy behaviour and never raise. Tests: 23 new cases in `tests/unit/test_doom_loop.py` covering `_normalize_args`, `_hash_args`, `extract_recent_tool_signatures`, `detect_i…07c5699 -
fix(cli): persist HF user id in saved sessions (#146)
Unverified signature* fix(cli): persist HF user id in saved sessions Co-authored-by: OpenAI Codex <codex@openai.com> * test: remove submission loop user id unit test Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>2083476 -
feat(session): include user_id and total_cost_usd in trajectory dump (#136)
Guillaume SalouUnverified signatureAdd user_id (HF username from OAuth) to the Session object and propagate it from the SessionManager. Surface it in the trajectory JSON written by save_trajectory_local() and uploaded to the session dataset, alongside a total_cost_usd aggregate summed from existing llm_call events. This unblocks per-user cost attribution from the session dataset alone — today the dataset has session_id but no user binding, so 569Xlspend cannot be tied back to an HF account without correlating timestamps from Hub access logs (noisy, breaks for users with long-lived browser tabs). Cost data is already collected per call by agent.core.telemetry; we just propagate user identity and expose a rolled-up total.645964c
-
Run jobs under user or org accounts with upgrade / org-pick UX (#132)
Aksel Joonas ReediUnverified signature* agent/core/agent_loop.py: jobs access rewiring * agent/core/telemetry.py: jobs access rewiring * agent/tools/jobs_tool.py: jobs access rewiring * backend/dependencies.py: jobs access rewiring * backend/models.py: jobs access rewiring * backend/routes/agent.py: jobs access rewiring * frontend/src/components/Chat/ChatInput.tsx: jobs access rewiring * frontend/src/components/ClaudeCapDialog.tsx: jobs access rewiring * frontend/src/components/SessionChat.tsx: jobs access rewiring * frontend/src/hooks/useAgentChat.ts: jobs access rewiring * frontend/src/lib/sse-chat-transport.ts: jobs access rewiring * frontend/src/store/agentStore.ts: jobs access rewiring * frontend/src/types/agent.ts: jobs access rewiring * scripts/build_kpis.py: jobs access rewiring * tests/unit/test_build_kpis.py: jobs access rewiring * agent/core/hf_access.py: jobs access rewiring * frontend/src/components/JobsUpgradeDialog.tsx: jobs access rewiring * tests/unit/test_hf_access.py: jobs access rewirin…ff8c636
-
Fix CLI rendering corruption and split CLI/frontend model defaults (#121)
Aksel Joonas ReediUnverified signature* Stabilize CLI rendering and make surface defaults explicit The interactive CLI was interleaving live sub-agent redraws with streamed markdown output, which corrupted ANSI rendering and leaked raw control sequences into the terminal. The CLI and web app also shared one default model config even though they need different Anthropic routing defaults. Constraint: CLI default must use direct Anthropic credentials while web sessions must default to Bedrock Anthropic Constraint: Interactive terminal output must remain readable while sub-agent progress is live Rejected: Single shared config file with runtime overrides | keeps ownership of defaults implicit across surfaces Rejected: Keep background redraw ticker | concurrent terminal writers still corrupt streamed output Confidence: high Scope-risk: moderate Reversibility: clean Directive: Keep CLI and frontend default models in separate config files unless both surfaces intentionally converge again Tested: python -m compileall agent backen…3eec386 -
Switch from Bedrock to Anthropic endpoint as default. Include support for gpt-5.5 (#118)
Unverified signature* Default ml-intern to Anthropic Co-authored-by: OpenAI Codex <codex@openai.com> * Add direct OpenAI GPT-5 model support Co-authored-by: OpenAI Codex <codex@openai.com> * Raise probe budget for GPT-5 models Co-authored-by: OpenAI Codex <codex@openai.com> * Fix deps * Fix stale xhigh provider messaging Co-authored-by: OpenAI Codex <codex@openai.com> --------- Co-authored-by: OpenAI Codex <codex@openai.com>0545e40
-
feat(observability)
Aksel Joonas ReediUnverified signature2a2e170
-
fix: route bedrock/ model ids via LiteLLM Bedrock adapter (#89)
Guillaume SalouUnverified signature_resolve_llm_params treated any non-anthropic/openai prefix as an HF router model and wrapped it as 'openai/<id>' pointed at router.huggingface.co. 'bedrock/...' therefore fell through to HF router which rejected it with 'model does not exist'. Add a dedicated bedrock/ branch that passes the model id as-is so LiteLLM's Converse adapter picks it up with the standard AWS env creds.ac55773 -
feat: route Claude via Bedrock inference profile (#82)
Guillaume SalouUnverified signatureSwitch the Claude endpoint IDs from the direct Anthropic API to the US cross-region inference profile. Extend the model-provider substring check so both prefixes keep hitting the HF-org gate, the daily Claude quota, and prompt caching.2fac9ff -
feat: enable Anthropic prompt caching on system prompt and tools (#69)
Unverified signature* feat: enable Anthropic prompt caching on system prompt and tools Mark the rendered system prompt and the tool block with cache_control breakpoints when calling Anthropic models. The static prefix (~4-5K tokens of system prompt + 15+ tool definitions) was being re-billed at full input rate on every turn, every retry, and every research sub-agent iteration (up to 60 per task). With ephemeral cache breakpoints, subsequent turns within the 5-minute TTL are billed at cache-read pricing (~10% of input cost). Expected savings: 40-50% input tokens on multi-turn conversations, 60-80% on research sub-agent loops. Caching is GA in the Anthropic API and natively supported by litellm 1.83+ via cache_control blocks (no beta header required). Non-Anthropic models (HF router, OpenAI) are passed through unchanged. The helper does not mutate the caller's message list or tool list, so the persisted ContextManager.items history stays in its original string-content form. * refactor: hoist prompt_cac…5d357ba
-
fix /model switcher with effort params
Aksel Joonas ReediUnverified signature* /model: probe effort combo on switch, stop passing reasoning_effort to litellm Switching to claude-opus-4-7 with /effort set would 400 with "thinking.type.enabled is not supported" because litellm 1.83.0's Anthropic adapter substring-matches "4.6" to decide which thinking API shape to send, and doesn't know about 4.7's adaptive + output_config.effort contract. Rather than maintain a per-model capability table that rots every time a new Claude family ships, this change trusts the API itself: - llm_params.py: for anthropic/*, bypass litellm's reasoning_effort -> thinking mapping and pass thinking={type: adaptive} plus output_config={effort: ...} as top-level kwargs directly. litellm forwards unknown top-level params into Anthropic request bodies (extra_body does NOT work here — Anthropic rejects it as "Extra inputs are not permitted"). One localized monkey-patch widens litellm 1.83's hardcoded _is_opus_4_6_model check so effort=max isn't rejected synchronously on 4.7 — re…e2552e8 -
Make X-HF-Bill-To configurable via HF_BILL_TO env (default smolagents) (#57)
Aksel Joonas ReediUnverified signature0b64e00 -
Read max context from litellm + compact at 90% (#54)
Aksel Joonas ReediUnverified signature* Read max context from litellm.get_model_info + compact at 90% The local _MAX_TOKENS_MAP had Claude Opus 4.6 pinned at 200k, triggering compaction at ~20% of its real 1M window. Swap the hand-maintained table for litellm.get_model_info()['max_input_tokens'], which LiteLLM keeps in sync upstream (Opus 4.6=1M, GPT-5=272k, Sonnet 4.5=200k, etc.). HF router-only ids (MiniMax, Kimi, GLM) aren't in litellm's catalog and fall through to the 200k default — close enough to their advertised ranges and safe if the model lies. Also shifts compaction to fire at 90% of max_context instead of >100%. The old condition waited until context had already overshot, risking the next LLM call hitting ContextWindowExceededError before the compact finished. 90% gives headroom for the summary call + one more turn. * Drop the legacy -10k buffer — the 90% ratio is the headroom now ContextManager was subtracting a fixed 10k tokens from max_context on top of the new 90% compaction threshold, so a 1M-window Opu…28b8f2b -
Fix parallel-research display: per-call stats in CLI and web UI (#49)
Aksel Joonas Reedi +1: lee101 Unverified signatureWhen multiple research sub-agents ran concurrently, the CLI collapsed them into a single flickering slot and every web-UI tool card showed the aggregate across all agents — so 3 parallel calls looked like 1. CLI (agent/utils/terminal_display.py): - Replace the global SubAgentDisplay singleton with SubAgentDisplayManager that tracks each agent by agent_id and renders one block per agent with independent (tool count · tokens · elapsed). - 4 rolling tool-call lines per agent (was 2). - When 2+ agents are live, switch to a compact one-line-per-agent layout (label + stats + most-recent tool). Detailed 4-line rolling view only when a single agent is active. Keeps the live region small enough to fit on one terminal page so cursor-up/erase doesn't drift when content would otherwise scroll into scrollback. - Clip every live-region line to terminal width (ANSI-aware) so a wrapped line can't corrupt the cursor-up math on narrow terminals. - On completion, erase the live block and f…5ab7c4e -
Route HF inference through /v1 auto-router + add reasoning_effort knob
Users paste bare HF model ids (MiniMaxAI/MiniMax-M2.7, moonshotai/Kimi-K2.6) with an optional :fastest|cheapest|preferred|<provider> suffix; the router picks a provider and handles failover. /model does a live preflight against /v1/models and prints providers, pricing, context, tool support — warn-and- allow for unknowns with fuzzy suggestions. Friendly messages replace LiteLLM's raw traceback for model/provider mismatches, and the noisy 'Give Feedback' banner is suppressed. Adds a reasoning_effort config + /effort command (default high). OpenAI and Anthropic get the top-level param natively; HF router gets it via extra_body with minimal->low normalization for models like MiniMax M2 that require reasoning. Frontend + backend model selectors updated to the bare-id format.0a9e96d
-
Fall back to user HF token for router when INFERENCE_TOKEN is unset (#39)
Aksel Joonas Reedi +1: Claude Opus 4.7 (1M context)Unverified signatureThe HF router code path only read INFERENCE_TOKEN, which is the shared server-side key set on the hosted Space so inference is free for users. On the CLI / self-hosted path that env var is absent, so requests went out with no bearer token and the router returned 401 — surfaced to users as "Authentication failed" even with a valid HF_TOKEN (issue #36). Resolve api_key in this order: 1. INFERENCE_TOKEN env (unchanged Space behavior — shared billing) 2. session.hf_token (user's OAuth / CLI token) 3. HF_TOKEN env (belt-and-suspenders for CLI) Applied to _resolve_hf_router_params, research_tool._resolve_llm_params, and ContextManager.compact. Fixes #36 Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>182ddee
-
Refactor approval handling and enhance markdown table styles
4501765
-
Improve agent loop, tools, and research subagent for long-running autonomous tasks
- Add --no-stream and --max-iterations CLI flags for headless operation - Simplify context manager: remove staged pruning, track actual token usage - Add doom-loop detection and context budget to research subagent - Improve file tools: better enforcement, error messages, descriptions - Improve bash tool: longer timeouts, background execution guidance - Add training logging and HP sweep guidance to system prompt - Relax Python requirement to >=3.11 - Remove unused file_content_cache73882d9
-
feat: add headless CLI mode with local filesystem tools and rich terminal rendering
- Add argparse-based headless mode: `python -m agent.main "prompt here"` - Local tool specs override sandbox /app references for CLI mode - System prompt injects CLI-specific context (working directory, no sandbox) - Extract streaming/non-streaming LLM call helpers with LLMResult dataclass - Add shimmer thinking animation (truecolor gradient sweep) - Progressive markdown rendering via rich Live display - Clean MCP shutdown sequence, suppress asyncio teardown noise - Add rich dependency to agent extras577ec48 -
Deduplicate file re-reads via content hashing for local and sandbox tools
6d4f619 -
feat: LLM-based pruning for complex tool outputs (hf_jobs)
Tools in _LLM_PRUNE_TOOLS get a cheap LLM call (≤600 tokens) that extracts the answer from the raw output, preserving job IDs, metrics, and error messages exactly. Falls back to deterministic truncation if the LLM call fails. Three tool lists now control pruning behavior: - _PRUNE_SKIP_TOOLS: never pruned (research, plan_tool) - _LLM_PRUNE_TOOLS: LLM-summarised (hf_jobs) - Everything else: deterministic 80-char truncation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>d4b1495 -
Add deterministic pruning of old tool outputs before LLM compaction
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>64437d0 -
Add per-tool error limits and per-turn request cap
544a574
-
feat: add research sub-agent tool, slim down main agent system prompt
Adds a `research` tool that spawns a cheaper LLM in its own context window with read-only tools (github_find_examples, explore_hf_docs, etc.) and returns a concise summary. This keeps expensive research output out of the main agent's context. The system prompt's Phase 1 research section is replaced with a single `research({task, context})` call pattern — all the detailed research methodology (tool chains, correct patterns, examples) moves into the sub-agent's system prompt where it belongs.da0136f
-
feat: add doom-loop detection for repeated tool call patterns
Detects when the agent is stuck calling the same tools repeatedly (identical consecutive calls or repeating sequences like A→B→A→B) and injects a corrective prompt to break the cycle.d3bb2d2