ALZ-F01 Discovery sub-agent pattern: report phase and fetch in code
Orchestrator report phase (ADR 0013) and fetching in code (ADR 0014), with the report renderer. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
@@ -3,7 +3,7 @@
|
||||
| | |
|
||||
|---|---|
|
||||
| Status | Proposed |
|
||||
| Date | 2026-10-29 |
|
||||
| Date | 2026-10-29, changed 2026-11-01 |
|
||||
| Decided by | Built as a default; awaiting product owner and platform confirmation |
|
||||
| Requirements changed | Earlier draft settings `claude-sonnet-4-5` / `gpt-5.4`; DSC-08's note "state if the critic shares a model with any executor" |
|
||||
|
||||
@@ -13,8 +13,8 @@ The Azure AI resource behind the gateway deploys `gpt-4o`, `gpt-5.6-sol`, `claud
|
||||
|
||||
## Decision
|
||||
|
||||
The agent and seven sub-agents use `gpt-5.6-sol` (temperature 1, max tokens 32000), as the CP4 agents do. The critic uses `claude-sonnet-5-5` (`DSC_08_CRITIC_LLM_MODEL`), a different model family from the agents it judges. Eval cases wait 330 seconds, because `gpt-5.6-sol` is a reasoning model.
|
||||
The agent and all eight sub-agents use `gpt-5.6-sol` (temperature 1, max tokens 32000), as the CP4 agents do. The critic was to use `claude-sonnet-5-5`, but the cluster's LiteLLM doesn't serve it (checked 2026-11-01), so the product owner moved it to `gpt-5.6-sol`.
|
||||
|
||||
## Consequences
|
||||
|
||||
Confirm both names on the cluster's LiteLLM gateway before the first run. If `claude-sonnet-5-5` isn't served, the critic's calls fail and the critique phase reports `failed`. Point `DSC_08_CRITIC_LLM_MODEL` at another model the gateway serves. Removing the variable instead would put the critic on the same model as the agents it judges.
|
||||
The critic can share blind spots with the agents it judges. It states this in its `note`, as DSC-08 requires. Its verdict is still enforced in code. A different model for the critic remains the better design, so revisit this when one is on the gateway.
|
||||
|
||||
@@ -0,0 +1,15 @@
|
||||
# ADR 0013: a report phase, written in code
|
||||
|
||||
Status: Proposed. Date: 2026-11-01.
|
||||
|
||||
## Context
|
||||
|
||||
The run ended at the verdict, with nothing an Engagement Lead could hand to a customer. The individual-agents pattern has a report step (its ADR 0012).
|
||||
|
||||
## Decision
|
||||
|
||||
Add a fifth phase, `report`, answered in `pre_process` with no model. It reads the workspace, takes the verdict from `_check_critique`, and stores the report as hash-chained records. `app/report.py` is this pattern's own copy: the two patterns share no code. The approval gate moves from the critique step to this one.
|
||||
|
||||
## Consequences
|
||||
|
||||
The report can't state more than the workspace holds, and it can't overrule the enforced verdict. A change to the report must be made in both patterns. The step's copy is a draft: `scripts/render_discovery_report.py` writes the clean copy from `report_data` after approval.
|
||||
@@ -0,0 +1,15 @@
|
||||
# ADR 0014: MCP results are fetched and stored in code
|
||||
|
||||
Status: Proposed. Date: 2026-11-01. Changes how ADR 0003 works.
|
||||
|
||||
## Context
|
||||
|
||||
The orchestrator's model copied each MCP result into `workspace_put`. On `SUB-002` it stored only the first 7 entries of the call log, despite being told to store results unchanged. The critic then reported five real scans as unmatched. The hash chain protects a record once stored, but can't show that it matches what the tool returned.
|
||||
|
||||
## Decision
|
||||
|
||||
The model names the calls through `fetch`. Code makes them, stores the results exactly as they came back, and returns a receipt. `agent.yaml` still declares the MCP tools so that code can call them, and `before_model` withholds them from the model. `workspace_put` is no longer offered.
|
||||
|
||||
## Consequences
|
||||
|
||||
A stored result is what the tool returned, and a refused call is stored and reported. The model reads a record only when a step needs it, which saves tokens and suits large estates. It has less freedom, and every call must be one `fetch` allows. The sub-agents are unchanged.
|
||||
Reference in New Issue
Block a user