fix(agents): decouple the delegation duty from hard rule 1's write exception

[prajwal-395]

Why

On 2026-08-19 the primary firstmate personally executed a branch prune across ten repositories instead of delegating it. The investigation traced the cause to upstream commit a24eac1 (2026-07-29, PR kunchenguid#1272), which rewrote AGENTS.md section 1's delegation sentence from an absolute rule into one gated on hard rule 1's captain-approved project-operation exception.

That commit's own message is entirely about write permission and never mentions delegation. The effect was that a rule about whether firstmate may write became the only boundary on whether firstmate must delegate, with no bound on magnitude. This is inherited from upstream rather than fork-local, and is fixed here in the fork.

Two factors compounded it. No rule anywhere in AGENTS.md stated what delegating is for, and a rule with no stated cost loses to a rule that has one. Section 7's "simplest direct end-to-end path" instruction has one, and although it means "do not build machinery" it reads as "do not add indirection" - and delegation is indirection.

Change 1 - decouple the delegation duty from the write exception

Section 1, before:

You are the captain's only point of contact for all software work across all of their projects.
Outside hard rule 1's concrete captain-approved project operation exception, you do not do project-specific work yourself.
For all other project-specific work, delegate coding, investigation, planning, bug reproduction, and audits to a crewmate you spawn and supervise, or to a secondmate whose registered scope fits.

After:

You are the captain's only point of contact for all software work across all of their projects.
Delegating is what keeps you reachable: work you do yourself blocks that only point of contact for as long as it runs, while work a crewmate does leaves you free to keep talking.
You do not do project-specific work yourself.
Delegate coding, investigation, planning, bug reproduction, and audits to a crewmate you spawn and supervise, or to a secondmate whose registered scope fits.
Hard rule 1's concrete captain-approved project operation exception lets you write to a project directly; it does not relieve you of delegating.
When an approved operation is large enough that running it yourself would keep you from answering the captain, brief and spawn a worker for it and tell the captain that is what you are doing.

The duty to delegate is now unconditional. The exception licenses writing, not not-delegating.

The one-project bound was omitted at the captain's instruction

The investigation's proposed wording also included the clause "and it never covers work spanning more than one project or more than a few minutes". The captain reviewed both variants and explicitly chose "decouple delegation from the exception, but do not add the one-project bound".

That clause is therefore deliberately absent from this change, at the captain's instruction. The change carries no project-count and no duration bound. The escalation trigger is grounded in the availability rationale instead: an approved operation gets briefed and spawned when running it would keep firstmate from answering the captain.

Change 2 - state why the rule exists

The rationale is now the second line of section 1, immediately after the "only point of contact" line:

Delegating is what keeps you reachable: work you do yourself blocks that only point of contact for as long as it runs, while work a crewmate does leaves you free to keep talking.

The substance of the proposal, drawn from firstmate's own Grok Bot doctrine, is preserved: availability is the reason, and self-executed work is what costs it. The two halves are joined so the sentence carries the cost in the same breath as the rule.

Change 3 - disambiguate the direct-path sentence

Section 7, before:

For one-off or infrequent operational work, start with the simplest direct end-to-end path.
Do not build wrappers, control planes, policy layers, custom verifiers, or automation unless the direct path exposes a concrete blocker or repeated need that justifies the added machinery.

After:

For one-off or infrequent operational work, start with the simplest direct end-to-end path, meaning do not build machinery around it.
Do not build wrappers, control planes, policy layers, custom verifiers, or automation unless the direct path exposes a concrete blocker or repeated need that justifies the added machinery.
Delegating that work to a crewmate is not machinery and is still the default.

Sites changed: 1. A repo-wide search for the sentence and its close variants (simplest direct, control planes, policy layers, added machinery) across AGENTS.md, .agents/skills/, public skills/, bin/, and docs/ returns this one location only. The near-identical line in the captain's personal global instructions lives outside this repository and is out of scope here.

Out of scope, verified unchanged

• Hard rule 1's own text, including its exception sentence. Byte-identical to the base, proven below.
• The other four hard rules.
• Section 9, the escalation and etiquette contract.
• The captain's personal global instructions file, which is not in this repository.

Only AGENTS.md is touched, 7 insertions and 3 deletions, so the change is close to net-neutral as the file's size discipline requires.

Verification

Hard rule 1 hashed at the merge base and at HEAD, plus the acceptance greps:

--- hard rule 1 must be byte-identical to the base ---
c282a98a5bde78e94dddfbf711ecd8e56247f718  -
c282a98a5bde78e94dddfbf711ecd8e56247f718  -
--- the exception no longer gates the delegation duty ---
0
--- the one-project bound must NOT be present ---
0
--- rationale present ---
1
--- direct-path sentence disambiguated everywhere ---
       1
       1

git diff --stat fork/main HEAD is a single file:

 AGENTS.md | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

Test suite

Full local suite run on the change:

FM_TEST_SUMMARY total=157 failed=4 skipped_gate=24 duration_ms=4025944

No test needed modification. This change edits prose only and alters no structure or documentation-audience mapping that the suite asserts on. Four scripts failed, and each was individually investigated and attributed to a cause other than this change:

1 and 2 - tests/fm-lint-workflows.test.sh and tests/fm-lint.test.sh: missing local actionlint. Both fail with exit 127, fm-lint-workflows.sh: actionlint not found; install actionlint 1.7.12 for CI parity. The pinned actionlint is not installed on this machine; ShellCheck 0.11.0 is present and passed. Nothing was installed system-wide to chase them. These pass in CI, which installs actionlint. This change touches no shell script and no workflow file, so it cannot affect either.

3 - tests/fm-backend-herdr-presentation-e2e.test.sh: load-induced, not merge-introduced. The failing assertion is "concurrent secondmate recovery failed ... error: herdr presentation recovery could not acquire its session lock; refusing a concurrent resume". Re-run standalone on this branch it passes: FM_TEST_SUMMARY total=1 failed=0 duration_ms=285069. The cause is a timing bound rather than a defect in the change under test: spawn_herdr_presentation_order_lock_acquire in bin/fm-spawn.sh retries 50 times at 0.1s, so it gives up after 5 seconds, while the concurrent peer holds the lock across a full recovery. Under whole-suite load that bound is exceeded.

It is also not a regression introduced by the upstream sync. The only change to bin/fm-spawn.sh between the pre-merge base 6a90309 and this branch is PR #23's relaunch-placement block, which is gated on RELAUNCH_ENDPOINT_MISSING and does not touch the presentation lock or the concurrent-recovery path that failed. The failing code path is identical on both sides of the merge. The herdr identity scrub is not implicated either: the test sets its own HERDR_SESSION from the lab helper and calls herdr_forget_inherited_pane first, so it is self-isolating.

4 - tests/fm-wake-queue.test.sh: pre-existing, reproduces without this commit. It fails on "the oversized unread status line was truncated or omitted", preceded by grep: out of memory. This one is not a load artifact - it reproduces standalone at low load. It was then run on fork/main with this commit absent and failed identically, same assertion and same grep: out of memory. It is therefore pre-existing on main and unrelated to this change, and is left for separate triage rather than widened into this PR.

Base

Rebased onto current fork/main (699b14a), which carries the upstream sync from PR #21 and PR #22, so the diff against main is this one commit.