or.2: orchestrator-agent design correction — no nested subagent dispatch

Claude Code categorically forbids subagents from spawning other
subagents (code.claude.com/docs/en/sub-agents and the Agent SDK
subagents page). The `or.1` architecture (and `pr.1` that ran on
top of it) presumed a named-exception for `ailang-implement-
orchestrator` to dispatch implementer / spec-reviewer / quality-
reviewer / tester per task. That exception never existed at the
platform level; the `Agent` tool was silently dropped from the
orchestrator-agent's tool set at dispatch time.

Architecture revised, doc-only:

- Per-task phases (implementer → spec-compliance check → quality
  check) now run as sequential role-switches in the orchestrator-
  agent's own context, not as nested subagents.
- The four role-files (implementer / spec-reviewer / quality-
  reviewer / tester) become phase reference files the
  orchestrator-agent consults at role-switch boundaries.
- Boss-context offload preserved; fresh-per-phase context given
  up (the property that drove or.1's nested-dispatch design is
  not buildable in Claude Code).

Files touched:
- skills/implement/agents/ailang-implement-orchestrator.md
  (frontmatter, Iron Law, Phase 2/3 rewrite, rationalisations,
  red flags)
- skills/implement/SKILL.md (frontmatter, Iron Law, sub-status
  vocabulary note, cross-references)
- skills/README.md (Conventions: no more named exception; agent
  roster: role-files recast as phase references)
- docs/journals/INDEX.md (or.2 entry appended)
- docs/journals/2026-05-11-iter-or.2.md (new — full rationale)
- docs/roadmap.md (P1 tool-wiring item removed; resolved as
  categorically-not-fixable)
- docs/WhatsNew.md (user-facing correction entry appended)

No code changes; no bench impact; CLAUDE.md untouched (no stale
references to fix there).
This commit is contained in:
2026-05-11 13:01:38 +02:00
parent bd801b979c
commit 5c3bd9ab24
7 changed files with 340 additions and 97 deletions
+189
View File
@@ -0,0 +1,189 @@
# iter or.2 — orchestrator-agent design correction
**Date:** 2026-05-11
**Branch:** none (trivial-mechanical carve-out — doc-only revision, single commit on main)
**Status:** DONE
**Tasks completed:** N/A (doc revision, not a plan-driven iter)
## Summary
This entry corrects a factual error in the `or.1` architecture that
surfaced during the first real dispatch (pr.1, same day). The
revision lands as one direct-on-main commit because the change is
doc-only: no code, no tests, no behaviour shift in any compiler
crate. The Iron-Law mini-fix from earlier today (`origin/main`
`main`) is preserved unchanged — that fix remains correct.
### What was wrong
`or.1` and `pr.1` shipped the `ailang-implement-orchestrator` agent
with frontmatter `tools: Read, Edit, Write, Bash, Glob, Grep, Agent`
and an architecture that depended on the orchestrator-agent
spawning further subagents per task (implementer → spec-reviewer →
quality-reviewer → tester, each in fresh context). The architecture
was framed in `skills/README.md` as "agents do not call other
agents — one named exception" (the orchestrator-agent being the
exception).
The exception never existed in Claude Code. The platform forbids
**all** nested-subagent dispatch, categorically:
> "Subagents cannot spawn other subagents, so `Agent(agent_type)`
> has no effect in subagent definitions."
> — https://code.claude.com/docs/en/sub-agents
> "Subagents cannot spawn their own subagents. Don't include
> `Agent` in a subagent's `tools` array."
> — https://code.claude.com/docs/en/agent-sdk/subagents
The `Agent` tool was silently dropped from the orchestrator-agent's
tool set at dispatch time (no error, no warning). The pr.1 run
caught this empirically; the diagnosis closed the question today.
### Why the failure took the form it did
The Boss-Orchestrator (me) designed `or.1` on the false premise
that the project's house rule "agents do not call other agents"
was a local convention with a documented exception. In reality the
house rule was a *restatement of an external constraint* — and the
"named exception" was incompatible with the constraint itself.
This is a verification-of-capability failure: the architecture was
locked in before checking the platform docs for what subagents can
do. The right order is the other way around. (Memorised as a
feedback memory for future architecture decisions.)
### The corrected architecture
Same skill body, same skill triggers, same branch-and-merge
discipline, same Boss-side end-report contract. What changes is
*inside* the orchestrator-agent's run:
- Per-task phases (implementer → spec-compliance check → quality
check) run sequentially as **role-switches in the orchestrator-
agent's own context**, not as nested subagents.
- The `ailang-implementer.md` / `ailang-spec-reviewer.md` /
`ailang-quality-reviewer.md` / `ailang-tester.md` files now serve
as **phase reference files** — the orchestrator-agent consults
them at each role-switch to inhale the discipline of that role,
but does not dispatch them as separate subagents.
- The Boss-context offload property is preserved (the Boss still
sees only the ≤500-token end-report; per-task chatter stays
inside the orchestrator-agent's context).
- The fresh-per-phase context property — which was the design goal
that drove `or.1` to attempt nested dispatch — is **given up**.
Pure role-switching inside one context is a weaker discipline
than fresh-context-per-phase. The trade-off is named openly: the
cost is the loss of fresh-eyes review per phase; the benefit is
that the architecture is buildable in Claude Code at all.
- The pr.1 "degraded mode" was not actually degraded — it was the
only mode this agent can run in. The pr.1 journal's degraded-
mode caveat is preserved as-is (historical record of how the
constraint was discovered).
### Why option (A), not (B) or (C)
Three options were on the table:
- **(A)** Single-context inline phases (what landed) — preserves
Boss-context offload, gives up fresh-per-phase context.
- **(B)** Roll back to pre-or.1: Boss dispatches each per-task
subagent itself. Preserves fresh-per-phase context, gives up
Boss-context offload.
- **(C)** Hybrid: orchestrator-agent for setup / branch / journal,
Boss for per-task dispatches in between. Preserves both
theoretically but multiplies moving parts.
(A) was chosen because the Boss-context offload is the scarce
resource (user chat budget). Empirically, the pr.1 run with the
plan template's verbatim per-task content blocks made fresh-per-
phase context contribute little additional signal beyond a careful
single-context phase loop. The decision was made by the user
2026-05-11 after the diagnosis was presented; the rationale belongs
in the user's hands, not the agent's.
### Files changed
- `skills/implement/agents/ailang-implement-orchestrator.md`
frontmatter (`Agent` removed from tools, description rewritten),
Iron Law (sequential role-switches instead of fresh subagents),
Phase 2.1 / 2.2 / 2.3 rewritten as inline phases, Phase 3 likewise,
common rationalisations and red flags updated.
- `skills/implement/SKILL.md` — frontmatter description, Iron Law,
per-task sub-status mechanics note (vocabulary is preserved but
now describes inline role-phase status), one new
Common-Rationalisation row, Cross-references rewritten (role-files
recast as phase references, with a "Why inline phases" subsection
pointing here).
- `skills/README.md` — Conventions: "Agents do not call other
agents — one named exception" → "Agents do not call other agents"
(the exception is gone, with an explanation of why). Agent roster
rows for implementer / spec-reviewer / quality-reviewer / tester
recast as phase references for the orchestrator-agent.
- `docs/roadmap.md` — P1 entry "`Agent` tool wiring for the
dispatched orchestrator-agent" removed (resolved as
categorically-not-fixable).
- `docs/journals/INDEX.md` — this entry appended.
- `docs/WhatsNew.md` — user-facing correction entry appended.
### Files NOT changed (and why)
- `CLAUDE.md` — no stale references; the file does not encode the
named-exception story.
- `skills/implement/agents/ailang-implementer.md`,
`ailang-spec-reviewer.md`, `ailang-quality-reviewer.md`,
`ailang-tester.md` — these remain unchanged. Their content
(discipline, reading list, output format) is correct in the new
architecture; their *role* in the system has shifted from
"dispatched subagent" to "phase reference for the orchestrator-
agent". The shift is documented in the README roster row and the
SKILL.md Cross-references; no in-file edits are required for
correctness. A future cleanup iter may rewrite their frontmatter
to drop the "subagent" framing if it begins to mislead — for now
the cost of touching four files outweighs the marginal clarity
gain.
- `docs/specs/2026-05-11-implement-orchestrator-agent.md` and
`docs/specs/2026-05-11-plan-recon-subagent.md` — specs are
historical records of the design as conceived; they remain
accurate to that moment. Drift between spec and the corrected
architecture is acknowledged here, not by editing the specs.
## Concerns
None. The revision is deliberate and self-contained.
## Known debt
- **Phase reference files frame themselves as dispatched subagents.**
`ailang-implementer.md` and its three siblings still read like
files describing a subagent receiving a carrier from a dispatcher.
In the corrected architecture, the orchestrator-agent IS the
dispatcher and the worker for those phases. If the next
`/implement` run shows the orchestrator-agent confused by this
framing during phase reads, a follow-up cleanup iter rewrites
the four files. Not pre-emptive — wait for an empirical signal.
- **No verification ladder for the corrected design.** `or.1`'s
verification ladder (items 15 in
`docs/specs/2026-05-11-implement-orchestrator-agent.md`) was
written for the nested-dispatch architecture. Items 13
(Boss-side dispatch, branch creation, journal write) carry over
unchanged. Items 45 (intentional BLOCKED, intentional re-loop)
are weaker in the inline-phase world — re-loop is a same-context
retry, not a fresh-context re-dispatch — and a new ladder entry
for the inline phase-switch should be added the next time
`/implement` runs against a non-trivial plan. Tracked
informally; no ticket.
## Blocked detail
N/A — DONE.
## Commits
This entry's revision lands as a single trivial-mechanical commit
on `main` (no iter branch, no plan template). The commit subject is
`or.2: orchestrator-agent design correction — no nested subagent dispatch`.
## Stats
N/A — no orchestrator-agent dispatch (this is a doc-only revision
made directly by the Boss-Orchestrator).
+1
View File
@@ -8,3 +8,4 @@
- 2026-05-11 — iter or.1: orchestrator-refactor → 2026-05-11-iter-or.1.md
- 2026-05-11 — fieldtest canonical-type-names → 2026-05-11-fieldtest-canonical-type-names.md
- 2026-05-11 — iter pr.1: plan-recon subagent → 2026-05-11-iter-pr.1.md
- 2026-05-11 — iter or.2: orchestrator-agent design correction (no nested subagent dispatch) → 2026-05-11-iter-or.2.md