7b4a9becc2e1fa81df027fe2234cf89b1a67b168
8 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 marfrit
|
955bd82efb |
safety + repl: wire secrets into safety.lua (closes #52)
Closes the last #13 gap — Norris broker call + is_destructive LLM
second-opinion probe were the two egress points NOT covered by the
scrub-at-egress design in commit
|
||
|
marfrit
|
17e62c0326 |
safety: permission policy DSL — allow/confirm/deny rule lists (closes #9)
The confirm_cmd boolean was too coarse: true interrupts every harmless
ls; false ungates everything. Most workflows want trust for read-only
ops while still gating writes/network/sudo.
New config:
permissions = {
allow = { "^ls%s", "^cat%s", "^git status" },
confirm = { "^rm%s", "^git push", "^docker%s", "^sudo%s" },
deny = { "^ssh%s+root@", "^curl%s+http[^s]" },
}
Verdict order: deny > confirm > allow. First match in the chosen
category wins. Unmatched defaults to "confirm". Patterns are Lua
patterns (not regex) per PHASE0.md §3 — no compiled extensions.
Verdict behavior in the interactive CMD: loop:
- allow → run without prompt
- deny → status line, skip
- confirm → [y/N] prompt (same UX as legacy confirm_cmd=true)
Backward compat:
- permissions unset + confirm_cmd=true → always confirm
- permissions unset + confirm_cmd=false → always allow
- permissions set → policy table is authoritative
Scope deliberately limited to the interactive AI-suggested CMD: gate.
Norris autonomous mode keeps its own safety.is_destructive machinery
(combining the two would double-gate or replace the LLM probe — both
non-obvious behavioral changes that belong in their own issues).
User-typed shell-routed lines (`router.classify → "shell"`) and
:exec also bypass the policy by design — those are direct user intent.
New introspection:
:perms list — show the configured rule lists
:perms check <cmd> — report verdict + matching rule (debug)
safety.classify_command is exported and unit-tested with 12 cases
covering each category, priority order (deny > allow on overlap),
and both fallback paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|
marfrit
|
11b1f566b3 |
safety: norris_step planner (Phase 3 commit #4)
Phase 3 commit #4 per docs/PHASE3.md §12. Single-iteration planner. The driver loop in repl.lua (commit #5) calls this in a while loop, advancing step_n on every "continue" return. M.norris_step(ctx, model_cfg, helpers, opts): 1. One broker.chat_stream round-trip — text + tool_calls collected, text streamed via helpers.render_assistant_delta. 2. Parse actions from response: tool_calls (already collected), CMD: lines (via helpers.extract_cmd_lines), GOAL: complete sentinel (line-level exact match per R-C5). 3. Record the assistant turn (with tool_calls if any) and log it. If no actions AND no goal_done → status="stalled". 4. Dispatch tool_calls (structured route first): - is_destructive check on serialized call. - If destructive → halt_fn(proceed/skip/abort). - Else → auto_approve lookup; absent → halt for consent (R-C6: Norris is conservative; auto_approve is the only consent bypass). - On skip: synthesize role:tool turn "[aish] tool call skipped by user" — alternation preserved per C5/C7. - On abort: return status="aborted". - On proceed: dispatch via helpers.dispatch_tool, append role:tool turn with result content. - Argument JSON parse failure also synthesizes a tool turn (same alternation rationale). 5. Dispatch CMD: lines (legacy route): - is_destructive check. - Destructive → halt_fn. - Non-destructive → run directly (Norris user accepted autonomy for non-destructive shell). - skip → ctx:append_exec_output "[aish] CMD skipped by user". - proceed → exec via helpers.exec_cmd, frame via render_exec_begin/end. 6. Skip-budget escalation (R-C1): after dispatch, if ctx.norris_consecutive_skips >= 3 → escalation halt; abort exits, proceed resets counter. 7. Goal-done check AFTER all dispatch (R-C2 / Q25 resolution). 8. Budget check: step_n >= max_steps → status="budget_exhausted". 9. Otherwise → status="continue", driver advances. Helpers are passed in as injected functions rather than directly requiring repl/renderer/executor — keeps safety.lua's coupling clean and norris_step testable with a mocked helpers table. State carried across iterations on the ctx: - ctx.norris_consecutive_skips (resets on any successful proceed) - ctx.norris_goal / ctx.norris_active (set/cleared by the driver) Existing test_safety.lua corpus (87 cases) still passes — norris_step addition doesn't touch is_destructive's behavior. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
marfrit
|
2abd5da3a6 |
safety: LLM second-opinion + session cache (Phase 3 commit #2)
Phase 3 commit #2 per docs/PHASE3.md §12. Adds the LLM-probe gate on top of commit #1's static patterns. Together they form is_destructive. broker.lua extension: - opts.max_tokens (A2) — passed through to the request body. Phase 3 probes cap at 4 tokens for YES/NO replies. - opts.timeout_ms — overrides model_cfg.timeout_ms per-call. Probe uses 15000ms cap regardless of the model's normal timeout (the user's deep model has 1800000ms for long generations; the probe must stay snappy). - M.chat now accepts an opts table (same shape as chat_stream's). Backwards compatible — existing callers passing (cfg, msgs) unaffected. safety.lua additions: - llm_probe(cfg, system, cmd): single broker.chat call returning "YES"/"NO"/"YES_FAILSAFE"/"YES_UNPARSEABLE" — fail-safe defaults. - llm_second_opinion(cmd, cfg): two-probe protocol per R-B2. Probe 1: "Is this destructive?" — YES → flag. Probe 2 (only if probe 1 said NO): "Is this safe?" inverted question — NO → flag (disagreement = HALT). Both NO → safe. - Session-scoped cache _llm_cache keyed by normalized command (lowercased + whitespace-collapsed). Mitigates Q23 latency for repeated commands within a Norris run. - Model-selection precedence: cfg.safety.llm_model (explicit) → cfg.models.deep (independent local class) → cfg.models[default]. Fail-safe YES if none configured. - is_destructive(cmd, cfg): runs static patterns first (always), then LLM if cfg present + not explicitly opted-out. cfg=nil yields static-only mode (handy for tests). End-to-end verified against hossenfelder using qwen-coder-7b-32k as the deep probe (qwen3-30b-a3b-instruct in repo's config.lua isn't currently loaded on the local backend): cat /etc/hostname → hit=false (LLM: NO, NO inverted = safe) rm /tmp/x.log → hit=true (LLM flagged; static missed because no -r/-f flags) cp /etc/passwd /tmp/passwd.bak → hit=false (safe copy) cache: second probe on same cmd → 0s wall time static-only (cfg=nil): rm -rf /tmp/x → static hit, no LLM call opt-out (llm_second_opinion=false): cp x y → hit=false, no probe Test corpus (test_safety.lua, 87 cases) still all pass — cfg=nil preserves the static-only behavior. Note: production config.lua currently has `deep = qwen3-30b-a3b-instruct` which isn't loaded on the proxy backend right now; Norris users will hit the fail-safe (everything flagged destructive) until either the deep model is brought up OR cfg.safety.llm_model = "cloud" is set to route the probe through anthropic/claude-haiku-4.5. Update the config or model deployment for production use — covered by Phase 3 verify test case. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
marfrit
|
bd59ce7243 |
safety: is_destructive static pattern matcher (Phase 3 commit #1)
Phase 3 commit #1 per docs/PHASE3.md §12. Static-pattern destructive-op heuristic; no LLM second-opinion yet (lands in commit #2). Implementation: - 34 patterns in DESTRUCTIVE_PATTERNS table, grouped: 9 shell-wrapper patterns (R-B1 — bash -c / sh -c / zsh -c / eval / python -c / perl -e / pipe-to-sh both forms / pipe-to-bash both forms / xargs ... rm). HALT on the wrapper itself; user reads the inner before proceeding. 10 filesystem destructive (rm -rf, find -delete, dd to device, mkfs, shred, wipefs, truncate -s 0, ...). 5 version-control destructive (git push --force/-f, git reset --hard, git clean -fd, git branch -D). 5 database/process (DROP TABLE/DATABASE, TRUNCATE TABLE, kill/pkill -9). 2 permission (chmod 777, chown on root path). - ci=true flag for case-insensitive SQL patterns; rule patterns must be lowercase when ci is set (matcher lowercases input). - pkill -9 ordered BEFORE kill -9; kill rule uses %f[%w] frontier so "pkill -9 nginx" reports "pkill -9" not "kill -9" substring match. - M._patterns exposes the rule table for :safety patterns meta (Phase 3 commit #5) and for the test corpus. - M.norris_step stub stays — lands in commit #4. Test corpus (test_safety.lua, 87 cases): - 49 destructive cases across all categories (incl. all 11 wrapper forms, the canonical curl|sh end-of-string bypass, sudo-prefixed rm -rf, etc.). - 38 safe cases (read-only commands, non-destructive variants of risky verbs like "git push" without --force, "find" without -delete, "chmod 644", "kill 1234" without -9, etc.). - Documented one accepted false positive: echo "rm -rf /" matches the rm pattern by substring — Norris user can proceed after reading; tradeoff between false positives and false negatives, biased toward false positives per §5. - Run from repo root: `luajit test_safety.lua`. Exit 0 on pass. - Verified all 87 pass at commit time. R-C4 / readline rebind, broker opts.max_tokens, LLM second-opinion, norris_step planner, repl driver, and the wider Norris UX land in subsequent commits per §12. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
marfrit
|
f26cbd9a3a |
phase2 amend: __ separator (Bedrock-safe) + post_sse error diagnostics
Phase 7 verify finding from TC #26 against :model cloud: HTTP 400 from openrouter→Amazon Bedrock: "tools.0.custom.name: String should match pattern '^[a-zA-Z0-9_-]{1,128}$'" Anthropic via Bedrock validates tool names against that regex and rejects dots. PHASE2 originally chose "." as the namespace separator ("boltzmann.list_dir"); OpenAI tolerated it, Bedrock does not. Separator switched to "__" (two underscores) everywhere — internal API matches on-wire shape, no transformation layer: - repl.lua: - tools_schema builds "alias__name" - dispatch_tool_call splits via "^(.-)__(.+)$" (non-greedy → leftmost __) - :mcp tool parser uses same split - :mcp tools formatter prints "alias__name" - HELP block shows <alias__name> - safety.lua confirm_tool_call: alias.* glob → alias__* glob - config.lua example block: keys rewritten - docs/PHASE2.md: amendment header added; §1, §2 row, §3 config.lua row, §5 wire-shape JSON examples, §6 auto_approve schema, §7 meta-cmd table, §12 plan all updated. Original "." references preserved in commit history. Constraint: aliases must not themselves contain "__" so the parse stays unambiguous. Tool names from MCP servers may have underscores freely. Second fix bundled — uninformative broker error: Previously "broker error: transport: HTTP response code said error" Now "broker error: transport: HTTP 400: {full body snippet}" ffi/curl.lua M.post_sse changes: - FAILONERROR no longer set (was hiding the response body). - raw_body accumulator added alongside the SSE buffer; captures every byte regardless of SSE shape. - After perform, check status_code via curl_easy_getinfo. On >=400, return (nil, "HTTP <code>: <body[:400]>"). 2xx unchanged. - End-of-stream SSE flush only runs on 2xx (no false event on error bodies that aren't SSE-shaped). - Phase 1 callers reading just first return slot stay correct. End-to-end verified: - :model cloud + tools=[boltzmann__read_file ...] + "Use boltzmann__read_file with path=/etc/hostname" → Claude emits tool_call with name="boltzmann__read_file", args='{"path": "/etc/hostname"}'. ok=true, transport clean. - Force-bad tool name "bad.name.with.dots" → err string carries the full bedrock 400 with the regex-pattern message visible. TC #26 (sub-loop end-to-end) is now testable against cloud — the error that blocked it is resolved. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|
marfrit
|
0fde77fe35 |
safety: confirm_tool_call gate with auto-approve policy
Phase 2 commit #2 per docs/PHASE2.md §12. Implements just the per-call confirm-gate surface; Phase 3 stubs (is_destructive, norris_step) stay unimplemented with their error() bodies. M.confirm_tool_call(name, args, cfg) checks cfg.mcp.auto_approve for: - exact match on "<alias>.<tool>" - "<alias>.*" glob covering a whole server Miss falls back to a [y/N] readline prompt. Empty or non-"y" answer rejects (matches the existing confirm_cmd UX from PHASE0 §10). Pretty-printing renders args as compact JSON, truncated at 80 chars with "..." suffix so one-line prompts stay readable. Smoke-test passes all eight cases per §12 verify-row #2: exact match / alias glob → auto-approve, no prompt miss + y / n / empty / nil-cfg → prompt shown, expected verdict empty args / long args → clean rendering, truncation works Note: PHASE0 §4 module-layout had a "lands in Phase 2" hint on the norris_step stub; the actual landing is Phase 3 per PHASE0 §11 row 3. Comment in safety.lua updated to clarify. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
 claude-noether
|
4310207738 |
Phase 0: scaffold tree + manifest
- README, .gitignore, CLAUDE.md (project conventions) - docs/PHASE0.md — full Phase 0 manifest (locked substrate) - 10 root .lua modules + 4 ffi/ bindings, all stubs raising NotImplemented with module-scoped responsibilities matching the manifest - config.lua wired to current dirac/hossenfelder endpoints (qwen-coder-7b snappy/32k + cloud via OpenRouter through hossenfelder) File names match docs/PHASE0.md §4 exactly. Module bodies fill in across later phases; the tree shape is locked. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |