Owlibou/gnoma
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
52 Commits 2 Branches 10 Tags
95dfd0cf0ceac1c02fa0e17e066b3481e6b3b98b
Commit Graph

7 Commits

Author SHA1 Message Date
vikingowl 95dfd0cf0c feat: M1-M7 gap audit phase 3 — context prefix, deferred tools, compact hooks 
Gap 11 (M6): Fixed context prefix
- Window.PrefixMessages stores immutable docs (CLAUDE.md, .gnoma/GNOMA.md)
- Prefix stripped before compaction, prepended after — survives all compaction
- AllMessages() returns prefix + history for provider requests
- main.go loads CLAUDE.md and .gnoma/GNOMA.md at startup as prefix

Gap 12 (M6): Deferred tool loading
- DeferrableTool optional interface: ShouldDefer() bool
- buildRequest() skips deferred tools until activated
- Tools auto-activate on first model request (activatedTools map)
- agent + spawn_elfs marked as deferrable (large schemas, rarely needed early)
- Saves ~800 tokens per deferred tool per request

Gap 13 (M6): Pre/post compact hooks
- OnPreCompact/OnPostCompact callbacks in WindowConfig
- Called in doCompact() (shared by CompactIfNeeded + ForceCompact)
- M8 hooks system will extend these to full protocol
 2026-04-04 20:46:50 +02:00
vikingowl 704f3a7302 feat: M6 context intelligence — token tracker + truncation compaction 
internal/context/:
- Tracker: monitors token usage with OK/Warning/Critical states
  (thresholds from CC: 20K warning buffer, 13K autocompact buffer)
- TruncateStrategy: drops oldest messages, preserves system prompt +
  recent N turns, adds compaction boundary marker
- Window: manages message history with auto-compaction trigger,
  circuit breaker after 3 consecutive failures

Engine integration:
- Context window tracks usage per turn
- Auto-compacts when critical threshold reached
- History syncs with context window after compaction

TUI status bar:
- Token count with percentage (tokens: 1234 (5%))
- Color-coded: green=ok, yellow=warning, red=critical

Session Status extended: TokensMax, TokenPercent, TokenState.
7 context tests.
 2026-04-03 18:46:03 +02:00
vikingowl 9a78af7b05 feat: inject mode changes into engine conversation history 
Engine.InjectMessage() appends messages to history without triggering
a turn. When permission mode or incognito changes, the notification
is injected as a user+assistant pair so the model sees it as context.

Fixes: model now knows permissions changed and will retry tool calls
instead of remembering old denials from previous mode.
 2026-04-03 16:42:52 +02:00
vikingowl 24b5126d66 feat: wire permission checker into engine tool execution 
Tools now go through permission.Checker before executing:
- plan mode: denies all writes (fs.write, bash), allows reads
- bypass mode: allows all (deny rules still enforced)
- default mode: prompts user (pipe: stdin prompt, TUI: auto-approve for now)
- accept_edits: auto-allows file ops, prompts for bash
- deny mode: denies all without allow rules

CLI flags: --permission <mode>, --incognito
Pipe mode: console Y/N prompt on stderr
TUI mode: auto-approve (proper overlay TODO)

Verified: plan mode correctly blocks fs.write, model sees error.
 2026-04-03 16:15:41 +02:00
vikingowl 847735a9f7 feat: add router foundation with task classification and arm selection 
internal/router/ — core routing layer:
- Task classification: 10 types (boilerplate, generation, refactor,
  review, unit_test, planning, orchestration, security_review, debug,
  explain) with keyword heuristics and complexity scoring
- Arm registry: provider+model pairs with capabilities and cost
- Limit pools: shared resource budgets with scarcity multipliers,
  optimistic reservation, use-it-or-lose-it discounting
- Heuristic selector: score = (quality × value) / effective_cost
  Prefers tools, thinking for planning, penalizes small models on
  complex tasks
- Router: Select() picks best feasible arm, ForceArm() for CLI override

Engine now routes through router.Select() when configured.
Wired into CLI — arm registered per --provider/--model flags.

20 router tests. 173 tests total across 13 packages.
 2026-04-03 14:23:15 +02:00
vikingowl 09f102bdec feat: add security firewall with secret scanning and incognito mode 
internal/security/ — core security layer baked into gnoma:
- Secret scanner: gitleaks-derived regex patterns (Anthropic, OpenAI,
  AWS, GitHub, GitLab, Slack, Stripe, private keys, DB URLs, generic
  secrets) + Shannon entropy detection for unknown formats
- Redactor: replaces matched secrets with [REDACTED], merges
  overlapping ranges, preserves surrounding context
- Unicode sanitizer: NFKC normalization, strips Cf/Co categories,
  tag characters (ASCII smuggling), zero-width chars, RTL overrides
- Incognito mode: suppresses persistence, learning, content logging
- Firewall: wraps engine, scans outgoing messages + system prompt +
  tool results before they reach the provider

Wired into engine and CLI. 21 security tests.
 2026-04-03 14:07:50 +02:00
vikingowl 69f5dba091 feat: complete M1 — core engine with Mistral provider 
Mistral provider adapter with streaming, tool calls (single-chunk
pattern), stop reason inference, model listing, capabilities, and
JSON output support.

Tool system: bash (7 security checks, shell alias harvesting for
bash/zsh/fish), file ops (read, write, edit, glob, grep, ls).
Alias harvesting collects 300+ aliases from user's shell config.

Engine agentic loop: stream → tool execution → re-query → until
done. Tool gating on model capabilities. Max turns safety limit.

CLI pipe mode: echo "prompt" | gnoma streams response to stdout.
Flags: --provider, --model, --system, --api-key, --max-turns,
--verbose, --version.

Provider interface expanded: Models(), DefaultModel(), Capabilities
(ToolUse, JSONOutput, Vision, Thinking, ContextWindow, MaxOutput),
ResponseFormat with JSON schema support.

Live verified: text streaming + tool calling with devstral-small.
117 tests across 8 packages, 10MB binary.
 2026-04-03 12:01:55 +02:00