Skip to content

feat(pr-description): record each decision as a question, answer and who decided it - #9

Merged
Silthus merged 4 commits into
mainfrom
feat/decision-provenance
Oct 17, 2026
Merged

Silthus merged 4 commits into
mainfrom
feat/decision-provenance

Conversation

@Silthus

@Silthus Silthus commented Oct 17, 2026 •

Copy link
Copy Markdown
Owner

Problem

The Decisions block of a PR could not say who really decided. Two things went wrong:

  • An accepted recommendation and an override both showed as "👤 Human".
  • Whole tickets showed as "🤖 Agent", even when the user had settled what was inside them.

On PostHog/posthog#118924 this credited the agent with a send route the user picked over its recommendation. It also listed an unanswered question as a finished call.

Merge Danger

🟢 Low, two-way door. The change is to skill text only. pr-description and orchestration-wayfinder rely on the new fleet decisions output (Silthus/fleet 4025fde and 706b16c, on main and rolled out), which reads the decision lines these skills now ask for. Maps that have no decision lines yet print No load-bearing decisions., so the block for them is written by hand from the transcript.

Summary

  • pr-description: the Decisions table has the columns Question, Answer, Who and Where, with one row per decision, never one per ticket. Who has three classes:

    • 👤 Human: the user chose, changed the recommendation, or a standing instruction decides it.
    • 👍 Accepted default: the user accepted the recommendation as written.
    • 🤖 Agent: decided without asking.

    Rows sort in that order. An unanswered question goes on an Open line and blocks the PR. Decisions are recorded where they are made as decision lines: - 👤 **<question>** <answer>.

  • orchestration-wayfinder:

    • HITL parks and resolutions record decision lines.
    • The conductor puts its own charting decisions on the map.
    • Every resolution comment records the ticket's load-bearing decisions as decision lines. That covers the human's answers and the worker's own calls.
    • Questions a ship worker leaves for the user go on Open.
  • architecture: the resolution comment lists locked decisions as decision lines.

Evidence

Decisions

Question Answer Who Where
Which classes does Who have? Human, human with the agent's default, and agent only 👤 Human retro thread
What is a row? A question with a brief answer or a link to details, kept concise 👤 Human retro thread
Which fixes ship? All four proposed: record at grilling time, rebuild fleet decisions, three-class schema, Open line 👍 Accepted default retro thread
How is the middle class labelled? 👍 Accepted default 🤖 Agent this PR
Where does an unanswered question go? An Open line under the table that blocks the PR 🤖 Agent this PR
How are rows ordered? 👤, then 👍, then 🤖, and by time within each 🤖 Agent this PR
Which line wins when a question repeats? The newest by comment time 🤖 Agent this PR
Does grilling write decision lines? No. It is a third-party snapshot in Fleet, so orchestration-wayfinder carries the rule 🤖 Agent this PR

Involvement: 1 review round and 3 final reads (Opus 5.5); fleet proof gate passes on HEAD.

🤖 Generated with Claude Code

Silthus and others added 4 commits October 17, 2026 04:26
…who decided it

The Decisions table credited the agent with calls the user made and could
not tell an override from an accepted recommendation. Rows now record the
question and its answer, Who has three classes (👤 Human, 👍 Accepted
default, 🤖 Agent), unanswered questions go on an Open line, and decisions
are kept as decision lines where they are made.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… on the map

Address review findings: decision lines live under Who with a generic mark,
a later line for the same question wins, feedback that changed a draft makes
it 👤, charting answers land on the map, and ship and implementation calls
are written as decision lines.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ines

Final-read findings: AFK task, research, prototype and spec workers now
record their own calls as decision lines too, the charting rule covers
every decision locked while charting, and review rows take their mark from
pr-description.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ines

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Silthus
Silthus marked this pull request as ready for review October 17, 2026 05:09
@Silthus
Silthus merged commit 81167ae into main Oct 17, 2026
2 checks passed
@Silthus
Silthus deleted the feat/decision-provenance branch October 17, 2026 05:09
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