Skip to content

feat(agent): blue_cell graph, langgraph server and compose wiring (Phase 2b-ii) - #6

Merged
PurpleCHOIms merged 1 commit into
mainfrom
feat/agent-graph-wiring
Oct 26, 2026
Merged

PurpleCHOIms merged 1 commit into
mainfrom
feat/agent-graph-wiring

Conversation

@PurpleCHOIms

Copy link
Copy Markdown
Member

The final implementation PR of the agent runtime design — AutoBot no longer borrows a Decepticon LangGraph server.

  • autobot/agent/blue_cell.py: the investigation agent assembled from AutoBot's own runtime — thin LiteLLM factory, directly-assembled middleware in the design's canonical order (UntrustedOutput → PromptInjectionShield → TokenBudget → Summarization(StateBackend) → PromptCaching → PatchToolCalls), the five read-only sensor tools, prompt from the packaged prompts dir. Factory-style graph export (model env read at serve time).
  • autobot/agent/middleware/budget.py: thin TokenBudgetMiddleware (AUTOBOT_BUDGET_TOKENS, default 400k, Command(goto=END) with an explicit budget note; 0 disables) — AutoBot's own BUDGET slot implementation (Decepticon's was proxy-coupled USD budgeting).
  • langgraph.json (blue_cell graph) + containers/langgraph.Dockerfile + the autobot-agent compose service in examples/local-sensor (health-checked; the monitor's BLUE_AGENT_URL contract unchanged — flip its default to the in-stack service in a follow-up after the live harness pass).
  • Assembly wiring tests (default read-only tool set, middleware order matches the design, injection respected) + budget tests. Suite: 240 passed; ci-lint clean.

Remaining after this PR: boot the in-stack server and run the full live harness (monitor watch/analysis against it), then stage 2c in Decepticon.

…ase 2b-ii)

The blue-only agent runtime — AutoBot no longer borrows a Decepticon
LangGraph server:

- autobot/agent/blue_cell.py: the investigation agent assembled from AutoBot's
  own runtime — thin LiteLLM factory, directly-assembled middleware in the
  design's canonical order (UntrustedOutput → PromptInjectionShield →
  TokenBudget → Summarization(StateBackend) → PromptCaching → PatchToolCalls),
  the five read-only sensor tools, prompt loaded from the packaged prompts
  dir. Factory-style graph export (model env read at serve time).
- autobot/agent/middleware/budget.py: thin TokenBudgetMiddleware
  (AUTOBOT_BUDGET_TOKENS, default 400k, Command(goto=END) with an explicit
  budget note — 0 disables). AutoBot's own BUDGET slot implementation
  (Decepticon's was proxy-coupled engagement USD budgeting).
- langgraph.json (blue_cell graph, ./packages/autobot dependency) +
  containers/langgraph.Dockerfile + the autobot-agent compose service in
  examples/local-sensor (health-checked, BLUE_AGENT_URL contract unchanged).
- Assembly wiring tests: default tool set (read-only structural guarantee),
  middleware order matches the design, injection respected; budget tests.
  Suite: 240 passed; ci-lint clean.
@PurpleCHOIms
PurpleCHOIms merged commit 11a5937 into main Oct 26, 2026
1 check failed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant