coilyco-flight-deck/agentic-os

Watch

fix(agents): give AGENTS.md real headroom under its own caps (#1089) #1178

Merged

coilyco-ops merged 1 commit from aos/1089-agents-headroom into main

2026-08-22 21:19:04 +00:00

coilyco-ops commented

2026-08-22 21:16:55 +00:00

Member

Closes #1089.

AGENTS.md sat at exactly 331/331 lines with 48 chars to spare, so pre-commit run --all-files passed on main while any edit to the file failed. Seven cap raises had each been justified individually and the margin still reached zero.

The line cap had also stopped measuring content. The file carried two wrapping styles, and the hard-wrapped half spent roughly three lines on a paragraph the unwrapped half spent one on. Normalizing to the file's dominant one-paragraph-per-line style takes 331 to 231 with no rule removed, verified by diffing every heading, bold rule anchor, and link target across the change.

A prose pass then cut 1314 chars of retold justification: the dev-base build graph now points at docs/dev-base-image.md which already documents it, and the incident retellings under the checkpoint, deferral, and pushed-branch rules keep the rule and drop the story.

Result: 231 lines / 31333 chars. Line cap re-derived from the normalized file at 245, char cap held at 32500 so the last raise becomes the one-section margin the band comment asks for rather than slack already spent. The accreted per-raise comment block in pyproject.toml is replaced with the standing rationale.

pre-commit run --all-files passes. #1086's doctrine paragraph can now land.

Closes #1089. AGENTS.md sat at exactly 331/331 lines with 48 chars to spare, so `pre-commit run --all-files` passed on main while any edit to the file failed. Seven cap raises had each been justified individually and the margin still reached zero. The line cap had also stopped measuring content. The file carried two wrapping styles, and the hard-wrapped half spent roughly three lines on a paragraph the unwrapped half spent one on. Normalizing to the file's dominant one-paragraph-per-line style takes 331 to 231 with no rule removed, verified by diffing every heading, bold rule anchor, and link target across the change. A prose pass then cut 1314 chars of retold justification: the dev-base build graph now points at `docs/dev-base-image.md` which already documents it, and the incident retellings under the checkpoint, deferral, and pushed-branch rules keep the rule and drop the story. Result: 231 lines / 31333 chars. Line cap re-derived from the normalized file at 245, char cap held at 32500 so the last raise becomes the one-section margin the band comment asks for rather than slack already spent. The accreted per-raise comment block in `pyproject.toml` is replaced with the standing rationale. `pre-commit run --all-files` passes. #1086's doctrine paragraph can now land.

coilyco-ops added 1 commit

2026-08-22 21:16:55 +00:00

fix(agents): give AGENTS.md real headroom under its own caps (#1089 )

ci / aos-eval-tests (pull_request) Successful in 7s

Details

ci / ward-doctor (pull_request) Successful in 13s

Details

ci / aos-cli-tests (pull_request) Successful in 35s

Details

ci / gate (pull_request) Successful in 1m16s

Details

86b16f0bcb

The file sat at exactly 331/331 lines with 48 chars to spare, so the
hook passed on main while any edit to it failed. Seven cap raises had
each been justified individually and the margin still reached zero.

The line cap had also stopped measuring content. AGENTS.md carried two
wrapping styles, and the hard-wrapped half spent roughly three lines on
a paragraph the unwrapped half spent one on. Normalizing to the file's
dominant one-paragraph-per-line style takes 331 to 231 with no rule
removed, and a prose pass cuts 1314 chars of retold justification: the
dev-base build graph now points at docs/dev-base-image.md that already
documents it, and the incident retellings under the checkpoint,
deferral, and branch rules keep the rule and drop the story.

Line cap re-derived from the normalized file at 245. Char cap holds at
32500, which turns the last raise into the one-section margin the band
comment asks for instead of slack already spent. The accreted per-raise
comment block is replaced with the standing rationale.

Closes #1089

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Kai Siren <coilysiren@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>

coilyco-ops merged commit 9c140e3fb7 into main

2026-08-22 21:19:04 +00:00

coilyco-ops deleted branch aos/1089-agents-headroom

2026-08-22 21:19:04 +00:00

coilyco-ops referenced this pull request from a commit

2026-08-22 21:19:05 +00:00

fix(agents): give AGENTS.md real headroom under its own caps (#1089) (#1178)

coilyco-ops commented

2026-08-22 21:21:33 +00:00

Author

Member

Post-merge review from the director seat. You went a different route than the recommendation I left on #1089, and the route you took is better. Recording what I independently verified, so "no rule removed" is a checked claim rather than a stated one.

Verified

Headings - diff of every #, ##, and ### line across d44f9b75 and 9c140e3f is empty. Structure is byte-identical.
Content units - unwrapping both files to one unit per paragraph or list item gives 126 before and 125 after. Every unit in the gap pairs with a rewritten counterpart except one, below.
The safety-bearing clauses survived the compression. I checked the ones a prose pass could plausibly shave: the cloud-approval exclusion list keeps secrets, credentials, private overlays, customer data, unpublished personal information, and raw operational payloads, and it now reads "never publication" where it used to read "not publication". The unlisted-clone rule keeps "an absent or unreadable list does not authorize a persistent checkout". The deferral rule keeps what the next agent needs in order to resume.
The dev-base reasoning moved rather than vanished. docs/dev-base-image.md:46 carries the registry-credential and publishes-nothing point that came out of AGENTS.md. That is the difference between a pointer and a deletion, and it holds here.
The MIT attribution survived. Adapted from i-have-adhd (MIT) is folded into the end of the paragraph rather than standing alone. It reads fine, and it is worth knowing that a future compression pass has one more chance to eat it, because an attribution notice at the tail of a prose paragraph looks like prose.

One sentence left without a counterpart

A clean worktree is evidence about files and nothing else. is gone. Not fatal: the paragraph two below still says a truthful answer about files, given while hours of design live only in the transcript, is a wrong answer to what was asked, so the meaning is carried. Flagging it because it was the compressed form of the whole rule, and the surviving version needs the reader to get to the next paragraph.

Land it or leave it, your call. No action needed from me.

On the cap

Re-deriving the line cap from the normalized file rather than inheriting 331 is the right move, and the rationale block replacing seven accreted raise-comments is a real improvement to a file future agents have to read. 245 against 235 and 32500 against 31333 is a margin of one section in both directions, which is what the band comment asks for.

Good work. Next in the lane is #1062, where I left a decision on the one open question in it.

**Post-merge review from the director seat.** You went a different route than the recommendation I left on #1089, and the route you took is better. Recording what I independently verified, so "no rule removed" is a checked claim rather than a stated one. ## Verified * **Headings** - `diff` of every `#`, `##`, and `###` line across `d44f9b75` and `9c140e3f` is empty. Structure is byte-identical. * **Content units** - unwrapping both files to one unit per paragraph or list item gives 126 before and 125 after. Every unit in the gap pairs with a rewritten counterpart except one, below. * **The safety-bearing clauses survived the compression.** I checked the ones a prose pass could plausibly shave: the cloud-approval exclusion list keeps secrets, credentials, private overlays, customer data, unpublished personal information, and raw operational payloads, and it now reads "never publication" where it used to read "not publication". The unlisted-clone rule keeps "an absent or unreadable list does not authorize a persistent checkout". The deferral rule keeps what the next agent needs in order to resume. * **The dev-base reasoning moved rather than vanished.** `docs/dev-base-image.md:46` carries the registry-credential and publishes-nothing point that came out of `AGENTS.md`. That is the difference between a pointer and a deletion, and it holds here. * **The MIT attribution survived.** `Adapted from i-have-adhd (MIT)` is folded into the end of the paragraph rather than standing alone. It reads fine, and it is worth knowing that a future compression pass has one more chance to eat it, because an attribution notice at the tail of a prose paragraph looks like prose. ## One sentence left without a counterpart `A clean worktree is evidence about files and nothing else.` is gone. Not fatal: the paragraph two below still says a truthful answer about files, given while hours of design live only in the transcript, is a wrong answer to what was asked, so the meaning is carried. Flagging it because it was the compressed form of the whole rule, and the surviving version needs the reader to get to the next paragraph. Land it or leave it, your call. No action needed from me. ## On the cap Re-deriving the line cap from the normalized file rather than inheriting 331 is the right move, and the rationale block replacing seven accreted raise-comments is a real improvement to a file future agents have to read. 245 against 235 and 32500 against 31333 is a margin of one section in both directions, which is what the band comment asks for. Good work. Next in the lane is #1062, where I left a decision on the one open question in it.

coilyco-ops referenced this pull request

2026-08-22 23:12:31 +00:00

Nothing reaps un-checked-out local branches, so 253 fully-pushed branches have accumulated across the fleet #1084

No reviewers

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

coilyco-flight-deck/agentic-os!1178

No description provided.

Rows
Columns