Chat, modes & context lanes

ContextMint chat is a governed copilot — not a generic web chat. You choose how much autonomy the model has (mode) and what evidence it may use (context lane).

Opening chat

  • Activity bar → ContextMint chat view
  • Shortcut: Ctrl+Shift+Alt+L (Windows/Linux; macOS may differ per keymap)

Modes: Ask, Plan, Agent

Mode Behavior
Ask Q&A and explanations; no autonomous file edits
Plan Structured plans and todos; still human-gated before apply. Optional Plan Studio / EPA planes stay off until enabled — see Planning below.
Agent Proposes patches and assisted actions; gate 2 preview before writes

Context lanes: Repo, Work, Hybrid

Open the compose ⋯ menu → Context to select a lane. The choice persists per workspace.

Lane Retrieval Best for
Repo (default) Knowledge objects when available, then hybrid search (vector + BM25), window expansion, workspace grep, and dependency-graph neighbours “What does our code do?” — answers only from assembled evidence
Work Session packs + attachments + editor selection — no vector repo search Planning, stakeholder comms, concepts — job-relevant but not file lookups
Hybrid Full evidence assembly (same as Repo) plus labeled recommendations “Should we ship X?” — repo evidence plus labeled recommendations

How evidence is assembled (Repo / Hybrid)

ContextMint works on any indexed workspace — not one repo layout. For structural questions (for example route wiring), the server uses index-time knowledge objects when the registry has a hit. For semantic questions, it runs hybrid search, then expands around hits: file windows, workspace grep, and dependency-graph neighbours. Optional LLM evidence selection is tier-gated and off by default — inference is last resort, not the primary path. See Settings → Evidence assembly.

Provenance chips (answers)

After a response, you may see a small label such as:

  • From your codebase — Repo lane with retrieved chunks
  • Work notes — Work lane (packs/attachments, not a repo search)
  • Repo + recommendations — Hybrid lane
  • No repo matches — Repo lane found little indexed context

Work-lane auto-suggest

On Repo lane, if retrieval is weak for a planning-style question, ContextMint may show a banner: “Little repo context matched — retry in Work lane?” Click Use Work lane to re-send the same question without vector search.

Attachments & context files

  • Use the attach control to pin files as chips before sending
  • Attached paths get full-fidelity chunks in the assembler prompt
  • Open editor files are boosted in search and indexed with priority

Image evidence (screenshots)

Paste, drag-drop, or pick PNG, JPEG, or WebP images in the compose bar. Image chips appear beside file chips. Context Lens shows thumbnails and the vision route (local by default) before you send.

  • Local Ollama vision is the default — no cloud API key required to attach images
  • Repo and Hybrid lanes combine images with retrieved code; Work lane combines images with packs
  • Use Run visual audit in Lens for ARGUS structured findings (separate from chat Q&A)

Full guide: Image evidence & ARGUS.

Chat sessions

Start a new session from the session list; history is stored per workspace up to configured limits. Sessions are local to your machine (or your team server in enterprise mode).

Model routing

Compose bar model selector: Auto, Local (Ollama), or Cloud when BYOK is enabled. The routing badge on answers shows which tier ran.

Plan mode → Studio → EPA → Save (defaults off)

Plan mode can emit structured PlanIR. Optional planes (all default off) sit between proposal and persist:

  • Save Plan / PAEP — only path that persists an authoritative plan under .contextmint/plans/
  • Plan Studio — verify EXISTING claims, sanitize unsupported assertions, freeze exportable PlanIR snapshots under .contextmint/studio/. Improves proposal cleanliness — not authority. Enable contextmint.planStudio.enabled (extension) and chat.plan_studio.enabled (server).
  • EPA (Engineering Plan Arbitration) — compare ordinary writer PlanIRs, re-observe FACT, record a RESOLUTION. Does not Save. Enable contextmint.planArbitration.enabled / chat.plan_arbitration.enabled.

Studio freeze is not an EPA decision. EPA must re-observe FACT independently. PAEP alone authorizes persistence.

When chat is blocked

Send is blocked only when readiness is offline, starting, or error. Indexing does not block send — answers improve as more files finish indexing.

Deep dive on retrieval: Indexing & readiness. Problems: Troubleshooting → Chat.