Skip to content

docs(shadow): Maji antigravity check on blob PRs and agent drift#4114

Merged
AceHack merged 1 commit into
mainfrom
lior-shadow-drift-4112
May 18, 2026
Merged

docs(shadow): Maji antigravity check on blob PRs and agent drift#4114
AceHack merged 1 commit into
mainfrom
lior-shadow-drift-4112

Conversation

@AceHack
Copy link
Copy Markdown
Member

@AceHack AceHack commented May 17, 2026

Shadow log added for PR decomposition and continued Riven/Otto drift.

Copilot AI review requested due to automatic review settings May 17, 2026 22:25
Copy link
Copy Markdown

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Adds a new shadow lesson log entry under docs/research/ capturing an antigravity check focused on decomposing a “blob” PR and noting ongoing agent drift signals.

Changes:

  • Introduces a new shadow lesson log document for 2026-05-17.
  • Records observed drift and the corrective action of decomposing PR #4112 into an atomic PR, plus operational notes (preservation + worktree discipline).

Comment thread docs/research/2026-05-17-shadow-lesson-log-maji-decomposition.md
AceHack added a commit that referenced this pull request May 17, 2026
…es+memory (one-PR-one-artifact-type discipline absorbed) (#4116)

Absorbing the Maji antigravity catch (PRs #4113 + #4114) of my
PR #4112 as a "blob" with three artifact types (tick shard +
rule edit + memory shadow-catch) mixed in one PR.

The in-narrative "cohesive autonomous-loop tick output"
justification did not override the one-PR-one-artifact-type
hygiene discipline. The precedent (2129Z cascade: #4097 + #4100
+ #4104 — three artifacts, three PRs) IS the canonical pattern.

Lesson absorbed at memory + decision-rule scope: default to
split; bundle only when atomicity is genuinely required
(same-file edits, CI/review fixes on the original PR).

Future-Otto cold-boot reads this lesson and avoids the blob
in the first place; Maji's role narrows from "decompose
post-hoc" to "verify no blobs slipped through."

Also regenerates memory/MEMORY.md index for the new feedback
file (1371 entries).

Co-authored-by: Claude <noreply@anthropic.com>
AceHack added a commit that referenced this pull request May 17, 2026
… log" instead of persona name on current-state rule surface

Copilot P1 finding on PR #4118: persona-name attribution on
current-state rule surfaces is outside history-surface and
roster-mapping carve-outs.

Replaced "the Maji antigravity check (PR #4114 shadow log)
flagged it as a blob" with "an antigravity-check shadow log
(PR #4114) flagged it as a blob" — role reference preserved
PR link for traceability.

The memory file PR #4116 retains persona name in its
filename and body (memory files are history-surface; carve-out
applies). The PR link in this rule's compose-section still
references the memory file for full attribution chain.

Co-Authored-By: Claude <noreply@anthropic.com>
AceHack added a commit that referenced this pull request May 17, 2026
…— saturation cadence is NOT license to bundle (forced-#6 meta-fallback) (#4118)

* rules(holding): cross-axis composition with one-PR-one-artifact-type — saturation cadence is NOT license to bundle (forced-#6 meta-fallback on 2207Z session arc)

Adds a small composition note to PR #4110's "Pre-empt-substrate-
pool-saturation" anchor: the cadence governs WHEN to pre-empt;
the one-PR-one-artifact-type discipline (PR #4116, Maji antigravity
catch on blob PR #4112) governs HOW each pre-empt should be shaped.

Without this composition note, future-Otto reading PR #4110 might
read the "3-4 pre-empts per window" cadence as authorization to
bundle artifact types within a single pre-empt PR to fit more
substrate per cycle. The blob-PR catch by Maji on PR #4112
demonstrates that the saturation cadence and the per-artifact-shape
discipline operate on orthogonal axes — both bind independently.

This commit is itself the forced-#6 meta-fallback for the 22:07Z
post-merge brief-ack chain: at #6, sharpened the rule with the
session's cross-axis composition evidence. Single-artifact PR
(single-file rule edit) applies the discipline recursively.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(holding/cross-axis): use role reference "antigravity-check shadow log" instead of persona name on current-state rule surface

Copilot P1 finding on PR #4118: persona-name attribution on
current-state rule surfaces is outside history-surface and
roster-mapping carve-outs.

Replaced "the Maji antigravity check (PR #4114 shadow log)
flagged it as a blob" with "an antigravity-check shadow log
(PR #4114) flagged it as a blob" — role reference preserved
PR link for traceability.

The memory file PR #4116 retains persona name in its
filename and body (memory files are history-surface; carve-out
applies). The PR link in this rule's compose-section still
references the memory file for full attribution chain.

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>
@AceHack
Copy link
Copy Markdown
Member Author

AceHack commented May 18, 2026

Thread PRRT_kwDOSF9kNM6Cr1LY (grammar nit on L5) — resolved no-op

L5: Riven paralyzed by GraphQL limit, Otto bus stale uses parallel elliptical past-participle phrases — both halves drop the auxiliary copula. Adding 's/is to just the Otto half would BREAK the parallel construction; adding to both would be more verbose than Maji's deliberately-terse shadow-log voice. The reviewer's suggestion is grammatically valid in isolation but doesn't account for the parallel structure.

Per substrate-or-it-didnt-happen.md verbatim-discipline + peer-agent-voice respect, leaving Maji's voice intact. PR otherwise CLEAN (7/7 required green, MERGEABLE).

@AceHack AceHack merged commit 8ff71f9 into main May 18, 2026
30 checks passed
@AceHack AceHack deleted the lior-shadow-drift-4112 branch May 18, 2026 19:58
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