Skip to content

fix: clarify agent execution and delegation boundaries - #448

Merged
Brian Krabach (bkrabach) merged 2 commits into
mainfrom
fix/explicit-agent-execution-guidance
Oct 21, 2026
Merged

Brian Krabach (bkrabach) merged 2 commits into
mainfrom
fix/explicit-agent-execution-guidance

Conversation

@bkrabach

@bkrabach Brian Krabach (bkrabach) commented Oct 21, 2026 •

Copy link
Copy Markdown
Collaborator

What changed

Clarifies Anchors root execution and delegation contracts, adds a mounted-tool/session-permission boundary to the shared agent baseline, and tests root/child prompt construction. This PR now also applies the repository's documented lean-head growth process to account for the required root guidance.

Why

Make tool selection, permitted scope, evidence, and delegation handoffs explicit while keeping unavailable capabilities and authority with the parent. Preserve those instructions rather than remove them to fit the previous size ceiling.

CI correction history

The initial PR head 3982e224 failed CI run 39691151047. All six Ubuntu/Windows Python 3.11–3.13 jobs failed the same existing lean-head assertion: system.md measured 1,599 characters against the 1,154-character ceiling. The ten Anchors contract tests passed in each inspected job. The failure was a real guardrail regression, not an environment failure; this PR does not waive it.

The follow-up changes only tests/test_lean_head_guardrail.py and the existing docs/lanes/zc6t-lean-head-ship/DONE-NOTE.md. They pin the system prompt ceiling to 1,599 characters, update the aggregate minimum budgeted saving from 588 to 143 characters, and document the reason, measured source counts, and CI failure. They clarify that the ceilings are budgets, not a new live census. Byte-pinned source files, fixtures, historical experiments, and frozen quality bounds are unchanged.

How to verify

  • Focused Anchors tests previously passed: 10 passed. The original three candidate files still match their recorded validation hashes.
  • Cross-repository isolated validation: 395 unique passed, 0 failed, 2 skipped because Unicode spellings remained distinct in that environment. Ordinary/managed schemas and a bounded exact-call CLI/provider smoke passed.
  • Prompt construction passed at root, named-child, and self-child seams. Child activation/execution was mocked; initialized provider-driven child behavior and freeform model efficacy are not claimed.
  • Astra/Fable review corrections were incorporated; final independent static review passed.
  • Full CI for this updated head is pending; no new-head CI pass is claimed yet.

Breaking changes

None. No runtime enforcement, permission, default, or retry-policy change. No new cost, cache, or task-success measurement is claimed.

Generated with Amplifier

Co-Authored-By: Amplifier <240397093+microsoft-amplifier@users.noreply.github.com>
Keep the reviewed execution and handoff rules. Apply the documented lean-head growth process: pin the 1,599-character source ceiling, account for the resulting 143-character minimum budgeted saving, and append the rationale and failed-CI record. Preserve byte pins, fixtures, historical measurements, and frozen quality bounds. This changes accounting and documentation only; it makes no cost or efficacy claim.

Generated with Amplifier

Co-Authored-By: Amplifier <240397093+microsoft-amplifier@users.noreply.github.com>
@bkrabach
Brian Krabach (bkrabach) merged commit 81b9bb0 into main Oct 21, 2026
7 checks passed
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.

2 participants