097 — Agentic analysis drops decisions that change no result

Type: reference/types/adr.md · Status: accepted

Status: accepted Date: 2026-09-29

Context

After ADR 096, an agentic-system analysis still asked workers to make decisions whose alternatives changed nothing in the result, and then asked later steps to accept, restate or check them:

  • A scoping job chose brief or full depth for each lens. Both lenses always ran, no code read the depth, and each lens member stated its scope again.
  • The memory specialist followed a private handoff protocol: an input bundle, memory-input.md, hashed by hand into canonical-register-sha256, plus method-sha256 and worker-model. Only the first was checked, against gitignored run state, so no clean checkout could verify it.
  • A review job wrote the public review's prose after verification. No check covered that prose, and it restated the Bounded synthesis.
  • The caller's source identity and the boundary worker's source.identity were authored independently and never compared. Publication compared identities by exact string, so a difference of form failed after the whole analysis ran.
  • The workflow continued when the Blockers text, loosely normalized, equalled none. "None found" counted as a blocker.

Decision

Delete the scoping job, the brief/full depth, the overview's ## Lens scoping section, memory-input.md and its generator, and the memory handoff protocol. The memory and epistemic jobs read boundary.md and output/runtime.md under the same input and completion contract as every other job. Drop canonical-register-sha256, method-sha256 and worker-model from the memory report type. Model identity stays in the workflow's operational state, and inputs-commit pins the method.

Delete the review job. Code renders the public review from the verified overview: # <System>, an Evidence basis: line built from the boundary fields, the Bounded synthesis, and ## Limitations. Relative links are rewritten to resolve from the review into the retained set. The reconciliation writes a one-sentence ## Description of 50 to 250 characters. Code uses it as the description of a complete run's overview and of the review, so verification reads it. The overview type's Bounded synthesis contract now says the synthesis must read without the members' context.

Normalize the source identity once, when the workflow is constructed. The rule strips surrounding whitespace, a trailing / and a trailing .git, and lowercases a URL's scheme and host. The run slug, destination inspection, the boundary job's prompt and publication use that form. The boundary validator refuses a frozen source.identity that differs.

The verify job's validator accepts ### Blockers only as exactly none or a Markdown list of - entries. The workflow continues only on exactly none.

Considered alternatives

Keep scoping with a separate effort budget. Advance control over inspection effort would need a mechanism that consumes the depth; none existed. If a budget becomes necessary, it should be specified on its own.

Export the memory provenance into the retained set. Moving memory-input.md into the set would make the pin checkable, but the file only restated the boundary and runtime member the set already holds.

Keep a separately written review for a different audience or length. One account serving both readers costs length. On the PageIndex set, before ADR 096 deleted it, the review body had 464 words, the Bounded synthesis 611 and the Limitations 220. The operator accepted the length in exchange for publishing only verified prose.

Keep the description outside the set. A side file from the reconciliation would keep the overview's code-written description, but nothing would check the published description.

Structured decision fields for blockers and returns. Workflow decisions could be represented as structured fields, from which headings are derived. Fixing the form of the one prose section code reads removes the misreading without a second representation.

Remove the return route to the memory specialist. Deferred to the second real run. The option under test is no returns, with amendments and limitations as the only correction.

Consequences

A complete run now has four job kinds: boundary, runtime, the two lenses, and rounds of reconcile and verify. Public reviews get longer and carry only prose that verification read. A reconciliation must write a Bounded synthesis that stands alone. Run directories opened before this change hold ## Lens scoping and the removed memory fields, and no longer validate. The workflow's normalization does not extend to the publication CLI's inspect-destination --source-identity argument, which is compared exactly. The decision is untested by a real run; the second real run tests it together with ADR 096 and cannot isolate either.

This ADR adopts the design proposal Fewer authored decisions in agentic-system analysis.