<?xml version="1.0" encoding="utf-8"?>
<entry xmlns="http://www.w3.org/2005/Atom" xml:base="https://docs.spec-kitty.ai/architecture/profile-load-reliability/">
  <id>https://docs.spec-kitty.ai/architecture/profile-load-reliability/</id>
  <title>Profile-Load Reliability (Squads &amp; WP Prompts)</title>
  <updated>2026-10-08T00:00:00.000Z</updated>
  <author><name>Spec Kitty</name></author>
  <link rel="alternate" type="text/html" href="https://docs.spec-kitty.ai/architecture/profile-load-reliability/"/>
  <summary>Why adversarial and research squads stopped loading charter agent profiles, and the stabilization design: resolve-then-inject, fail-loud dispatch, and /spk-load-profile.</summary>
  <content type="html">&lt;span id=&quot;profile-load-reliability-squads--wp-prompts&quot;&gt;&lt;/span&gt;
&lt;p&gt;This document records a corroborated investigation into why &lt;strong&gt;research / adversarial
squads recently stopped loading charter agent profiles&lt;/strong&gt;, and defines the &lt;strong&gt;near-term
stabilization design&lt;/strong&gt;. The deployable charter-backend evolution that this
work makes possible is a &lt;strong&gt;separate, backlog-scoped&lt;/strong&gt; design — see
&lt;a href=&quot;/architecture/charter-backend-service-future/&quot;&gt;Charter Backend Service (Future)&lt;/a&gt;.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Status (2026-09-30).&lt;/strong&gt; Partly shipped. The activation-allowlist fix (#3810, §4.1)
and the &lt;code dir=&quot;auto&quot;&gt;directive:&amp;#x3C;id&gt;&lt;/code&gt; selector fix (#3816, §6 D1) shipped in 3.2.6. #3814 was closed.
The rest is open on the &lt;strong&gt;4.x Work&lt;/strong&gt; milestone: the orchestrator-injects contract
(#3811, §4.2), &lt;code dir=&quot;auto&quot;&gt;/spk-load-profile&lt;/code&gt; (#3812, §4.3) and WP-prompt hygiene (#3813, §4.4).
The charter-backend spike (#3815) stays in the Product backlog. Where the sections below
say “3.2.6”, read it as the cycle the design was written for.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Provenance.&lt;/strong&gt; Root cause established 2026-08-30 by five agents across four lenses
(researcher, debugger, reviewer, architect), each profile-loaded through the
charter API. The squad overturned an initial mis-attribution (see §2.4) and converged
on the activation-allowlist root cause with live-command and git evidence.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Tracking.&lt;/strong&gt; Epic &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/3809&quot;&gt;#3809&lt;/a&gt;
(“squads must never dispatch unprofiled”). Children originally scoped to &lt;strong&gt;3.2.6&lt;/strong&gt;: #3810 (activation-
allowlist bug, §4.1), #3811 (orchestrator-injects fail-loud contract, §4.2), #3812
(&lt;code dir=&quot;auto&quot;&gt;/spk-load-profile&lt;/code&gt;, §4.3), #3813 (WP-prompt hygiene, §4.4), #3816 (&lt;code dir=&quot;auto&quot;&gt;directive:&amp;#x3C;id&gt;&lt;/code&gt;
selector bug, §6 D1). In &lt;strong&gt;Product backlog&lt;/strong&gt;: #3815 (charter-backend design spike) and
#3814 (research/doc template parity — demoted post scope-review as purely additive).
Hard dependency edges are encoded on the issues: #3810→#3811→#3812→#3813 (#3814 is a latent backlog edge — applies only if scheduled)
(#3816 is a &lt;em&gt;soft&lt;/em&gt; compaction edge on #3811, not a hard block).&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;1-symptoms&quot;&gt;1. Symptoms&lt;/h2&gt;
&lt;p&gt;Squad delegates dispatched at a point-cut (per the &lt;code dir=&quot;auto&quot;&gt;adversarial-squad&lt;/code&gt; skill) were
expected to load a doctrine lens via the charter API and act under it. Recently, delegates
assigned certain lenses acted &lt;strong&gt;unprofiled&lt;/strong&gt; — no initialization, boundaries, directives,
or tactics applied — while nothing errored visibly at the orchestrator.&lt;/p&gt;
&lt;h2 id=&quot;2-root-cause--a-charter-activation-allowlist-gap-issue-1&quot;&gt;2. Root cause — a charter activation-allowlist gap (Issue 1)&lt;/h2&gt;
&lt;p&gt;The loading &lt;strong&gt;mechanism is healthy&lt;/strong&gt;. In live testing, 7/7 &lt;em&gt;activated&lt;/em&gt; profiles resolved
&lt;code dir=&quot;auto&quot;&gt;EXIT 0&lt;/code&gt; through &lt;code dir=&quot;auto&quot;&gt;spec-kitty agent profile show &amp;#x3C;id&gt;&lt;/code&gt;, and &lt;code dir=&quot;auto&quot;&gt;spec-kitty charter context --action &amp;#x3C;x&gt; --json&lt;/code&gt; never failed. The failure is a &lt;strong&gt;data + fail-open&lt;/strong&gt; defect, not a
code or template regression.&lt;/p&gt;
&lt;h3 id=&quot;21-the-gate&quot;&gt;2.1 The gate&lt;/h3&gt;
&lt;p&gt;&lt;code dir=&quot;auto&quot;&gt;src/specify_cli/cli/commands/profiles_cmd.py:337&lt;/code&gt; (FR-014, introduced in #1636
&lt;code dir=&quot;auto&quot;&gt;488fe34e0c&lt;/code&gt;) refuses any profile absent from the charter’s &lt;code dir=&quot;auto&quot;&gt;activated_agent_profiles&lt;/code&gt;
allowlist:&lt;/p&gt;
&lt;div class=&quot;expressive-code&quot;&gt;&lt;figure class=&quot;frame not-content&quot;&gt;&lt;figcaption class=&quot;header&quot;&gt;&lt;/figcaption&gt;&lt;pre data-language=&quot;plaintext&quot;&gt;&lt;code&gt;&lt;div class=&quot;ec-line&quot;&gt;&lt;div class=&quot;code&quot;&gt;&lt;span style=&quot;--0:#d6deeb;--1:#403f53&quot;&gt;Error: profile &apos;doctrine-daphne&apos; is not activated.&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;The two source files exist (&lt;code dir=&quot;auto&quot;&gt;packs/built-in/agent_profiles/doctrine-daphne.agent.yaml&lt;/code&gt;,
&lt;code dir=&quot;auto&quot;&gt;…/randy-reducer.agent.yaml&lt;/code&gt;) and are declared DRG nodes — they are &lt;strong&gt;de-activated, not
missing&lt;/strong&gt;. Count check: &lt;strong&gt;25 source profiles, 23 activated → the 2 omitted are exactly the
2 the squad skill recommends.&lt;/strong&gt;&lt;/p&gt;
&lt;h3 id=&quot;22-the-regression-commit&quot;&gt;2.2 The regression commit&lt;/h3&gt;
&lt;p&gt;&lt;code dir=&quot;auto&quot;&gt;9a99801f1b&lt;/code&gt; (2026-08-08, &lt;em&gt;“feat(charter): activate writing-comms &amp;#x26; diagramming doctrine
set”&lt;/em&gt;) first &lt;strong&gt;materialized&lt;/strong&gt; the allowlist (23 entries) in
&lt;code dir=&quot;auto&quot;&gt;.kittify/charter/charter.yaml:1744&lt;/code&gt;, silently omitting &lt;code dir=&quot;auto&quot;&gt;doctrine-daphne&lt;/code&gt; and
&lt;code dir=&quot;auto&quot;&gt;randy-reducer&lt;/code&gt;. The parent commit had &lt;strong&gt;no allowlist key&lt;/strong&gt; → three-state &lt;code dir=&quot;auto&quot;&gt;None&lt;/code&gt; → &lt;em&gt;all&lt;/em&gt;
built-in profiles resolved (NFR-001 back-compat), so squads worked before that date. This
matches “recently stopped.”&lt;/p&gt;
&lt;h3 id=&quot;23-it-ships-upstream&quot;&gt;2.3 It ships upstream&lt;/h3&gt;
&lt;p&gt;&lt;code dir=&quot;auto&quot;&gt;src/charter/activation/packs/default.yaml:187&lt;/code&gt; — the built-in &lt;code dir=&quot;auto&quot;&gt;default&lt;/code&gt; preset as it then stood — carried its own
narrower (16-entry) allowlist that &lt;strong&gt;also omits both profiles&lt;/strong&gt;. &lt;strong&gt;Every new project inherits the gap&lt;/strong&gt;,
not just this dogfooding checkout.&lt;/p&gt;
&lt;h3 id=&quot;24-the-seam-gap-why-its-silent&quot;&gt;2.4 The seam gap (why it’s silent)&lt;/h3&gt;
&lt;p&gt;The &lt;code dir=&quot;auto&quot;&gt;adversarial-squad&lt;/code&gt; skill
(&lt;code dir=&quot;auto&quot;&gt;src/charter/offering/skills/adversarial-squad/SKILL.md&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;randy-reducer&lt;/code&gt; at line 43,
&lt;code dir=&quot;auto&quot;&gt;doctrine-daphne&lt;/code&gt; at line 47) &lt;strong&gt;hardcodes&lt;/strong&gt; those persona names with no reconciliation
against the activated set.
Its fallback clause sanctions reading the raw YAML &lt;strong&gt;only for “a read-only harness that
cannot invoke the CLI.”&lt;/strong&gt; A &lt;em&gt;de-activation&lt;/em&gt; &lt;code dir=&quot;auto&quot;&gt;EXIT 1&lt;/code&gt; is a &lt;strong&gt;different failure mode the
clause does not cover&lt;/strong&gt; — so a compliant delegate has no sanctioned recovery and proceeds
&lt;strong&gt;unprofiled&lt;/strong&gt;. There is no canonical “squad-eligible profiles” query; the skill roster
and the activation set drifted apart the moment the allowlist was compiled.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Adjudicated divergence.&lt;/strong&gt; An initial trace blamed #1840 &lt;code dir=&quot;auto&quot;&gt;60f038cee2&lt;/code&gt; (2026-07-28) for
flipping the load instruction from a YAML read to a CLI call. The debugger and architect
lenses refuted this: #1840 &lt;strong&gt;strengthened&lt;/strong&gt; the instruction (it never removed loading);
the CLI works for any &lt;em&gt;activated&lt;/em&gt; profile. The true cause is the 2026-08-08 allowlist.
A secondary, narrower contributor remains for genuinely CLI-less harnesses: #1840’s
fallback YAML path has since drifted across two relocations.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;3-secondary-finding--wp-prompt-naming-hygiene-issue-2-low&quot;&gt;3. Secondary finding — WP-prompt naming hygiene (Issue 2, LOW)&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;ad-hoc-load&lt;/code&gt; does not exist anywhere&lt;/strong&gt; (0 hits). It was a conflation of the
then-current compatibility alias with its canonical profile-load skill.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Resolved by #3732 (FR-008):&lt;/strong&gt; the alias and the canonical skill were folded into one
skill, &lt;code dir=&quot;auto&quot;&gt;spk-charter-profile-load&lt;/code&gt;; the alias name was retired, not redirected. WP source
templates (&lt;code dir=&quot;auto&quot;&gt;packs/built-in/missions/*/templates/task-prompt-template.md&lt;/code&gt;, the
&lt;code dir=&quot;auto&quot;&gt;implement&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;review&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;tasks(-packages)&lt;/code&gt; prompts, &lt;code dir=&quot;auto&quot;&gt;reviewer-implementer-role-separation. tactic.yaml&lt;/code&gt;, and the rc35 handoff migration’s inserted text) now name
&lt;code dir=&quot;auto&quot;&gt;/spk-charter-profile-load&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;research&lt;/code&gt; and &lt;code dir=&quot;auto&quot;&gt;documentation&lt;/code&gt; task-prompt templates carry no profile-load section at
all&lt;/strong&gt; — a real cross-mission-type inconsistency, but &lt;em&gt;additive&lt;/em&gt;, not a rename target.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;4-near-term-stabilization-design&quot;&gt;4. Near-term stabilization design&lt;/h2&gt;
&lt;p&gt;Three coordinated changes. The activation-data fix restores squads immediately; the
orchestrator-injects contract closes the failure &lt;strong&gt;class&lt;/strong&gt;.&lt;/p&gt;
&lt;h3 id=&quot;41-fix-the-activation-data-close-by-construction&quot;&gt;4.1 Fix the activation data (close by construction)&lt;/h3&gt;
&lt;p&gt;Activate the two doctrine lenses everywhere the allowlist is authored:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;This project: &lt;code dir=&quot;auto&quot;&gt;spec-kitty charter activate agent-profile doctrine-daphne randy-reducer&lt;/code&gt;
(or edit &lt;code dir=&quot;auto&quot;&gt;charter.yaml&lt;/code&gt;; edits are read live).&lt;/li&gt;
&lt;li&gt;Upstream default preset: none needed now. The registry file &lt;code dir=&quot;auto&quot;&gt;src/charter/activation/packs/default.yaml&lt;/code&gt;
was retired by #3732; its successor &lt;code dir=&quot;auto&quot;&gt;packs/built-in/presets/default.yaml&lt;/code&gt; carries no agent-profile
allowlist.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Regression guard:&lt;/strong&gt; a test asserting &lt;em&gt;every lens the &lt;code dir=&quot;auto&quot;&gt;adversarial-squad&lt;/code&gt; skill names
resolves &lt;code dir=&quot;auto&quot;&gt;EXIT 0&lt;/code&gt;&lt;/em&gt; (or, equivalently, source-profile-count parity for squad-eligible
lenses). This discharges directive 043 (close-by-construction) for these two instances; the failure &lt;strong&gt;class&lt;/strong&gt; closes on §4.2 — see §6 D4.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;42-orchestrator-resolves-then-injects-the-durable-seam-fix&quot;&gt;4.2 Orchestrator-resolves-then-injects (the durable seam fix)&lt;/h3&gt;
&lt;p&gt;Make the &lt;strong&gt;orchestrator&lt;/strong&gt; — the one context that reliably &lt;em&gt;can&lt;/em&gt; resolve — load each lens
&lt;strong&gt;once&lt;/strong&gt;, with overlays / &lt;code dir=&quot;auto&quot;&gt;specializes_from&lt;/code&gt; lineage / &lt;code dir=&quot;auto&quot;&gt;enhances&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;overrides&lt;/code&gt; applied, and
&lt;strong&gt;inject the resolved profile + compact action context inline&lt;/strong&gt; into each delegate prompt.
Consequences:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Removes the subagent’s &lt;strong&gt;mandatory&lt;/strong&gt; dependency on a runtime call (CLI &lt;em&gt;or&lt;/em&gt; backend) — the
injected floor is self-contained; on-demand pulls remain &lt;code dir=&quot;auto&quot;&gt;required-capable&lt;/code&gt; per §6 D1. Works for
read-only, shell-less, headless, and stale-cached-copy harnesses alike.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Fail-loud by construction:&lt;/strong&gt; an unresolved / de-activated lens errors &lt;em&gt;at the
orchestrator&lt;/em&gt;, which substitutes, activates, or aborts — a delegate can &lt;strong&gt;never&lt;/strong&gt; be
dispatched unprofiled. This is what converts Issue 1’s class from “silent unprofiled” to
“loud at dispatch.”&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Provenance:&lt;/strong&gt; injected delegates run as sub-invocations under the orchestrator’s own Op
trail (&lt;a href=&quot;/architecture/governed-profile-invocation/&quot;&gt;governed profile invocation&lt;/a&gt;) — they open no
separate Op, keeping the audit record uniform with the &lt;code dir=&quot;auto&quot;&gt;dispatch --profile&lt;/code&gt; path (§4.3).&lt;/li&gt;
&lt;li&gt;Use &lt;code dir=&quot;auto&quot;&gt;charter context&lt;/code&gt;’s existing &lt;code dir=&quot;auto&quot;&gt;mode: compact&lt;/code&gt;; inject profile-init + boundaries +
directive &lt;strong&gt;IDs&lt;/strong&gt;, not full bodies. Keep a &lt;em&gt;hybrid&lt;/em&gt;: a runtime pull for an on-demand
tactic stays &lt;strong&gt;allowed where available&lt;/strong&gt; (sharpened to &lt;code dir=&quot;auto&quot;&gt;required-capable&lt;/code&gt; in §6 D1).&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;43-the-spk-load-profile-id-instructions-primitive&quot;&gt;4.3 The &lt;code dir=&quot;auto&quot;&gt;/spk-load-profile &amp;#x3C;id&gt; &amp;#x3C;instructions&gt;&lt;/code&gt; primitive&lt;/h3&gt;
&lt;p&gt;Build on &lt;code dir=&quot;auto&quot;&gt;spk-charter-profile-load&lt;/code&gt; (the single profile-load skill since #3732) a
dispatch primitive &lt;code dir=&quot;auto&quot;&gt;/spk-load-profile &amp;#x3C;name|id&gt; &amp;#x3C;instructions&gt;&lt;/code&gt; that both &lt;strong&gt;resolves&lt;/strong&gt; the
profile and &lt;strong&gt;carries the task&lt;/strong&gt; the profiled agent runs — the surface the orchestrator
emits under §4.2. &lt;em&gt;(New surface: &lt;code dir=&quot;auto&quot;&gt;/spk-load-profile&lt;/code&gt; does not exist today.)&lt;/em&gt; Per §6 D3 this is a redirect/emitter over &lt;code dir=&quot;auto&quot;&gt;dispatch --profile&lt;/code&gt;, &lt;strong&gt;not&lt;/strong&gt; a third
resolution engine (directive 044).&lt;/p&gt;
&lt;h3 id=&quot;44-wp-prompt-hygiene-low-optional&quot;&gt;4.4 WP-prompt hygiene (LOW, optional)&lt;/h3&gt;
&lt;p&gt;If pursued, rename WP-template references &lt;code dir=&quot;auto&quot;&gt;/spk-charter-profile-load → /spk-load-profile&lt;/code&gt;,
&lt;strong&gt;keeping the leading slash&lt;/strong&gt;, and:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Do it via a &lt;strong&gt;new forward migration&lt;/strong&gt;, not by editing the shipped rc35 migration in
place (mutating emitted text diverges already-migrated installs).&lt;/li&gt;
&lt;li&gt;Gate on &lt;code dir=&quot;auto&quot;&gt;pytest tests/charter_offering/test_spk_skill_pack.py&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;The research/documentation missing-section gap is a &lt;strong&gt;separate additive&lt;/strong&gt; item.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;45-a-read-only-render-reads-only-the-profiles-it-names-6917&quot;&gt;4.5 A read-only render reads only the profiles it names (#6917)&lt;/h3&gt;
&lt;p&gt;The work-package board (&lt;code dir=&quot;auto&quot;&gt;spec-kitty agent tasks status&lt;/code&gt;) renders one structural boolean
per claimed row: &lt;code dir=&quot;auto&quot;&gt;profile.sentinel&lt;/code&gt;, the 👤 human-in-charge marker. Until #6917 that
boolean cost a whole &lt;code dir=&quot;auto&quot;&gt;ActiveCharterService&lt;/code&gt; build — pack context, activation ledger,
offering service — and then the complete &lt;code dir=&quot;auto&quot;&gt;AgentProfileRepository&lt;/code&gt;: every shipped profile
document plus the 358-node lineage graph. Measured on this repository: 1.14 s for the
service, 0.93 s for the catalogue read inside a ~4 s command.&lt;/p&gt;
&lt;p&gt;Two current facts replace that cost:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Id-scoped read.&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;ActiveCharterService.agent_profile_sentinels(ids)&lt;/code&gt; (and the
&lt;code dir=&quot;auto&quot;&gt;build_agent_profile_sentinels(repo_root, ids)&lt;/code&gt; builder the CLI calls) resolves exactly
the ids the rows name. It reuses &lt;code dir=&quot;auto&quot;&gt;AgentProfileRepository&lt;/code&gt;’s own loader — built-in → org →
project layer order, overlay field merge, schema/skip handling, &lt;code dir=&quot;auto&quot;&gt;applies_to_languages&lt;/code&gt;
gate — through an opt-in &lt;code dir=&quot;auto&quot;&gt;only_ids&lt;/code&gt; scope, so answers agree with a full load. It skips
&lt;code dir=&quot;auto&quot;&gt;PackContext&lt;/code&gt; because &lt;code dir=&quot;auto&quot;&gt;sentinel&lt;/code&gt; is structural, like &lt;code dir=&quot;auto&quot;&gt;get_provenance()&lt;/code&gt;; a narrowed or
denied &lt;code dir=&quot;auto&quot;&gt;activated_agent_profiles&lt;/code&gt; therefore cannot hide the marker.
Contract: &lt;code dir=&quot;auto&quot;&gt;kitty-specs/issue-6917-sentinel-scoped-profile-read-01M6BV17/contracts/agent-profile-sentinels.md&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Lazy lineage.&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;AgentProfileRepository&lt;/code&gt; builds the shipped DRG on first lineage use
(&lt;code dir=&quot;auto&quot;&gt;_graph()&lt;/code&gt;, restored from #6462 after a later refactor re-froze it in the constructor),
so profile-map readers — &lt;code dir=&quot;auto&quot;&gt;get&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;list_all&lt;/code&gt;, provenance, the sentinel read — never parse it.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Guarded by &lt;code dir=&quot;auto&quot;&gt;tests/charter/test_tasks_status_sentinel_scope_6917.py&lt;/code&gt; (a census on the
production CLI entry point: a two-profile board parses two documents, an empty board
parses none, a positive control proves the counter sees the whole catalogue) and
&lt;code dir=&quot;auto&quot;&gt;tests/charter/test_agent_profile_scoped_load.py&lt;/code&gt; (layer/merge/language parity with the
full repository).&lt;/p&gt;
&lt;h2 id=&quot;5-scope-boundary&quot;&gt;5. Scope boundary&lt;/h2&gt;

























&lt;table&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;Near-term (shipped or queued)&lt;/th&gt;&lt;th&gt;Backlog&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;§4.1 activation-data fix + regression guard (shipped, #3810)&lt;/td&gt;&lt;td&gt;Deployable charter backend service (API/MCP)&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;§4.2 orchestrator-injects contract + fail-loud dispatch (#3811, 4.x Work)&lt;/td&gt;&lt;td&gt;Remote compute / shared cache / out-of-cycle context precompute&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;§4.3 &lt;code dir=&quot;auto&quot;&gt;/spk-load-profile&lt;/code&gt; consolidation (#3812, 4.x Work)&lt;/td&gt;&lt;td&gt;—&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;§4.4 WP-prompt hygiene (optional, LOW; #3813, 4.x Work)&lt;/td&gt;&lt;td&gt;—&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;
&lt;p&gt;The backend evolution is deliberately &lt;strong&gt;out of the near-term scope&lt;/strong&gt;. §4.2 is intentionally a transport
change the &lt;em&gt;orchestrator&lt;/em&gt; owns, so a future backend slots in behind the &lt;strong&gt;same resolution
contract&lt;/strong&gt; without touching delegates. See
&lt;a href=&quot;/architecture/charter-backend-service-future/&quot;&gt;Charter Backend Service (Future)&lt;/a&gt;.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Freeze note.&lt;/strong&gt; The &lt;code dir=&quot;auto&quot;&gt;adversarial-squad&lt;/code&gt; skill and charter packs are doctrine surfaces;
during the 3.2.x cycle some changes fell under the doctrine-surface freeze (&lt;code dir=&quot;auto&quot;&gt;pr:deferred&lt;/code&gt;). Sequence
the data fix (§4.1) first — it restores squads with the least surface.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;h2 id=&quot;6-dialectic-review-outcomes-2026-08-30&quot;&gt;6. Dialectic review outcomes (2026-08-30)&lt;/h2&gt;
&lt;p&gt;Four dialectic squads (thesis ↔ antithesis, 8 profile-loaded delegates) stress-tested the
§4/§5 design. All four converged; none required synthesizer escalation. This section
&lt;strong&gt;amends&lt;/strong&gt; the sections named. Confidence figures are the delegates’ own.&lt;/p&gt;
&lt;h3 id=&quot;d1--resolution-locus-42-hybrid-sharpened&quot;&gt;D1 — Resolution locus (§4.2): &lt;strong&gt;hybrid, sharpened&lt;/strong&gt;&lt;/h3&gt;
&lt;p&gt;Thesis (inject, 0.8) and antithesis (live-resolve, HIGH-on-facts) agree the answer is the
hybrid §4.2 already names, but both narrow it:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Injection carries a fail-loud FLOOR&lt;/strong&gt; — resolved profile identity + boundaries + proof
of successful resolution, &lt;strong&gt;pinned at dispatch&lt;/strong&gt;. This is what makes “never dispatch
unprofiled” true and is the &lt;em&gt;only&lt;/em&gt; path for CLI-less/shell-less/headless harnesses.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;On-demand pulls stay &lt;code dir=&quot;auto&quot;&gt;required-capable&lt;/code&gt;, not merely “allowed where available”&lt;/strong&gt; — a
delegate that needs an unanticipated tactic/body must be able to pull it live where the
harness permits; otherwise the on-demand path silently degrades on exactly the harnesses
that need it. Do &lt;strong&gt;not&lt;/strong&gt; freeze &lt;em&gt;everything&lt;/em&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Two new must-fix findings (live-verified by the antithesis):&lt;/strong&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;spec-kitty charter context --include directive:&amp;#x3C;id&gt;&lt;/code&gt; returns &lt;code dir=&quot;auto&quot;&gt;EXIT 1&lt;/code&gt; “No directive
found”&lt;/strong&gt; for every directive ID tested (agent-profile and tactic selectors
work; the &lt;em&gt;directive&lt;/em&gt; selector does not). This makes “inject IDs, not bodies”
(§4.2) a &lt;strong&gt;dead end&lt;/strong&gt; — delegates cannot expand injected directive IDs. Fix the
selector, or inject directive bodies. &lt;strong&gt;Filed as #3816 (P1, Bug); fixed in 3.2.6.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Sequencing dependency: §4.1 must land before any §4.2 injection rollout.&lt;/strong&gt; Injecting
a resolved roster &lt;em&gt;today&lt;/em&gt; would freeze the current 23-of-25 allowlist (daphne/randy
de-activated) into every delegate with &lt;strong&gt;no runtime recovery&lt;/strong&gt; — converting Issue 1
from “silent unprofiled” into “silently &lt;em&gt;wrong&lt;/em&gt; roster, permanently.” A live delegate
self-heals once §4.1 lands; an injected one cannot.&lt;/li&gt;
&lt;/ol&gt;
&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;d2--backend-service-42-future--charter-backend-service-futuremd-keep-backlog-08--085-convergent&quot;&gt;D2 — Backend service (§4.2-future / charter-backend-service-future.md): &lt;strong&gt;keep backlog&lt;/strong&gt; (0.8 / 0.85, convergent)&lt;/h3&gt;
&lt;p&gt;Both sides independently affirm the existing scope filing. The backend is the &lt;em&gt;right&lt;/em&gt;
boundary and &lt;em&gt;right to defer&lt;/em&gt;. Hard gates before it leaves backlog:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Land §4.1 + §4.2 &lt;strong&gt;first&lt;/strong&gt; — the backend fixes zero Issue-1 defects; a warm cache over
the same omitted allowlist is “faster and more consistently wrong.”&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Design cache-invalidation before any cache is trusted&lt;/strong&gt; — a warm authority serving
stale doctrine violates the fail-loud invariant. This must land &lt;em&gt;on paper first&lt;/em&gt;.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Never a hard dependency&lt;/strong&gt; — degradation always falls back to in-process resolution,
never to unprofiled dispatch.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Promotion triggers (a number, not an argument):&lt;/strong&gt; (1) measured resolution latency
materially on the dispatch critical path; (2) ≥2 repos in one program sharing a
centrally-refreshed corpus; (3) a real shell-less fleet dispatched by a non-orchestrator
surface. &lt;em&gt;(Live corroboration: a delegate’s own CLI resolved doctrine from a sibling fork
checkout, not CWD — the environment-drift motivation is real but not yet urgent.)&lt;/em&gt;&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;d3--spk-load-profile-shape-43-consolidate-but-reconcile-with-spec-kitty-dispatch&quot;&gt;D3 — &lt;code dir=&quot;auto&quot;&gt;/spk-load-profile&lt;/code&gt; shape (§4.3): &lt;strong&gt;consolidate, but reconcile with &lt;code dir=&quot;auto&quot;&gt;spec-kitty dispatch&lt;/code&gt;&lt;/strong&gt;&lt;/h3&gt;
&lt;p&gt;Thesis (0.72) and antithesis (moderate-high) converge: consolidation’s atomicity win is
real &lt;strong&gt;for the orchestrator-emitted dispatch case&lt;/strong&gt;, but a naive &lt;code dir=&quot;auto&quot;&gt;&amp;#x3C;id&gt; &amp;#x3C;instructions&gt;&lt;/code&gt;
free-text positional is the wrong shape. &lt;strong&gt;§4.3 is amended:&lt;/strong&gt;&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;spec-kitty dispatch&lt;/code&gt; already carries load+task&lt;/strong&gt; (task positional, &lt;code dir=&quot;auto&quot;&gt;--profile&lt;/code&gt; flag —
&lt;code dir=&quot;auto&quot;&gt;dispatch.py:225-232&lt;/code&gt;). &lt;code dir=&quot;auto&quot;&gt;/spk-load-profile&lt;/code&gt; must be a &lt;strong&gt;redirect/emitter over
&lt;code dir=&quot;auto&quot;&gt;dispatch --profile&lt;/code&gt; + the §4.2 inject contract, not a third resolution engine&lt;/strong&gt;
(directive 044 — else it duplicates &lt;code dir=&quot;auto&quot;&gt;dispatch&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;&amp;#x3C;instructions&gt;&lt;/code&gt; must be passed structurally, not as a bare positional tail&lt;/strong&gt; — mirror
&lt;code dir=&quot;auto&quot;&gt;dispatch&lt;/code&gt;’s inversion, so there is no &lt;code dir=&quot;auto&quot;&gt;&amp;#x3C;id&gt; &amp;#x3C;instructions&gt;&lt;/code&gt; whitespace-split ambiguity.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Keep a load-only interactive path&lt;/strong&gt; — the “adopt a profile” trigger has no single task;
&lt;code dir=&quot;auto&quot;&gt;&amp;#x3C;instructions&gt;&lt;/code&gt; is optional, and pure-load must stay first-class and testable
independently of task text.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;d4--issue-1-fix-41-vs-42-data-first-seam-closes-085--085-convergent-on-sequencing&quot;&gt;D4 — Issue 1 fix (§4.1 vs §4.2): &lt;strong&gt;data-first, seam-closes&lt;/strong&gt; (0.85 / 0.85, convergent on sequencing)&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;§4.1 data fix is the correct primary &lt;em&gt;first&lt;/em&gt; move&lt;/strong&gt; — smallest surface, clears the
freeze, restores squads, and (with the guard) discharges directive 043 for these two
instances. Apply at &lt;strong&gt;both&lt;/strong&gt; authoring sites (&lt;code dir=&quot;auto&quot;&gt;charter.yaml&lt;/code&gt; + &lt;code dir=&quot;auto&quot;&gt;default.yaml:187&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;§4.2 is the actual &lt;em&gt;class&lt;/em&gt; close&lt;/strong&gt; — data alone leaves the fail-open fallback and the
two-list drift intact; the class recurs on the next de-activation.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Critical guard-design refinement (both delegates):&lt;/strong&gt; the §4.1 regression guard &lt;strong&gt;must
derive its lens list from a single canonical “squad-eligible profiles” query&lt;/strong&gt;, not a
hand-copied list — otherwise it becomes a &lt;em&gt;third&lt;/em&gt; drift-prone copy. That query is the
&lt;strong&gt;seed of §4.2&lt;/strong&gt; (one source consumed by the skill roster, the guard, and the injector).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Do not mark Issue 1 closed on §4.1 alone&lt;/strong&gt; — close it on §4.2; track §4.1 as mitigation.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;new-tickets-surfaced-by-the-dialectic-filed-post-review&quot;&gt;New tickets surfaced by the dialectic (filed post-review)&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;[Bug]&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;charter context --include directive:&amp;#x3C;id&gt;&lt;/code&gt; selector returns &lt;code dir=&quot;auto&quot;&gt;EXIT 1&lt;/code&gt; for valid
directive IDs (D1.i above) — blocks the inject-IDs compaction strategy. &lt;strong&gt;Filed #3816.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;[Constraint]&lt;/strong&gt; the fail-loud contract’s guard and injector must consume one canonical
squad-eligible query (D4) — &lt;strong&gt;folded into #3810/#3811 acceptance criteria.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;[Constraint]&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;/spk-load-profile&lt;/code&gt; reconciles with &lt;code dir=&quot;auto&quot;&gt;dispatch --profile&lt;/code&gt; and keeps a
load-only path (D3) — &lt;strong&gt;folded into #3812 acceptance criteria.&lt;/strong&gt;&lt;/li&gt;
&lt;/ul&gt;</content>
</entry>
