chore(#470): C10 — the flip: delete the Dev Platform from core (~49k LOC)

[Weegy]

epic #470 C10 — the flip: delete the Dev Platform from core

The Dev Platform now lives in byte5ai/omadia-dev-platform and installs via the Hub (or a ZIP upload). This is core's side of the move.

One PR by necessity, not preference. A partial delete leaves tsc red — index.ts, config.ts, the routers and the SPA are one dependency knot. It is split into 7 commits by area so it stays reviewable and revertible.

243 files changed · +149 / −59,356 · 217 files deleted.

---

Zone table — ratchet before → after

node scripts/check-core-decoupling.mjs --report

Zone	Before	After	Note
middleware/src	1576	19	6 are C12's publicPaths; rest are comments
middleware/test	1031	62	fixture strings + the legacy-key regression test → C13
middleware/packages	89	17	plugin-api CHANGELOG recording the removal → C13
middleware/scripts	8	CLEAN
middleware/sidecars	195	CLEAN
middleware/migrations	69	69	C11 — 0022–0030 stay, deliberately
middleware/package.json	0	CLEAN
middleware/env-example	19	1	the plugin repo URL, which cannot be reworded
web-ui/app	229	19	nav test fixtures + two comments → C13
web-ui/messages	4	CLEAN
web-ui/config	0	CLEAN
ci-workflows	16	CLEAN
scripts	30	27	the ratchet's own pattern list — see Note on zero
compose	36	CLEAN
TOTAL	3302	214	baseline updated 3300 → 214

Nine of fourteen zones read CLEAN. The baseline is updated consistently in all three places the README requires: decoupling-baseline.json, the README "Baseline 214" line, and the acceptance.md guard row.

main was already +2 over its own baseline when this branch was cut (middleware/test 1031 vs 1030). Pre-existing; the deletion absorbs it.

Note on zero. scripts/check-core-decoupling.mjs counts itself — 27 of the remaining 214 are its own literal pattern list. The total can never reach the 0 the README defines as done until the script excludes its own path. That is a one-line fix and belongs to C13; flagging it because "ratchet reads 0" is the stated definition of done and it is currently unreachable by construction.

---

What was deleted

Area	Files	LOC
middleware/src/devplatform/	62	18,022
middleware/test/devplatform/	58	16,878
middleware/sidecars/ (dev-runner, dev-runner-daemon, dev-dind)	34	12,990
web-ui/app/admin/dev-platform/	29	4,348
middleware/packages/dev-runner-shim/	23	4,077
i18n keys (en + de + allowlist)	3	782
web-ui/app/_components/devjobs/	6	771
middleware/src/{index,config}.ts	2	565
docker-compose.dev-platform.yaml	1	321
CI workflows	3	72

Plus scripts/dev-transcript.ts, web-ui/app/_lib/useDevJobEvents.ts, the /admin grid card, and all 43 DEV_* config keys.

i18n: 299 leaf keys per locale (adminDevPlatform.* 288, chat.devJob.* 9, admin.index.cards.devPlatform.* 2) = 598 across en+de, plus the 4 orphaned i18n-identical-allowlist.json entries the validator flagged. npm run i18n:check → OK, 3560 keys.

---

KEPT — and why

Every item below was traced to a live importer or an explicit spec instruction before being left in place.

Kept	Reason
auth/publicPaths.ts — both exemptions	C12. Deleting them now 401s every in-flight runner phone-home; runners carry no session cookie. Single revertible commit, on its own.
migrations 0022–0030	C11. Contiguous slot range in the core ledger; the handoff seeds by filename against a per-file schema witness and never deletes donor rows. Core still applies them, so no operator data moves in this PR.
services/githubAppJwt.ts	implementation.md §5. It was moved into core to close the one core→devplatform reverse dependency. Sending it out recreates that leak across a repo boundary. The plugin uses a capability-first local RS256 fallback (SEAMS S2).
express, pg, zod	§5. Plugin peerDependencies, resolved through the host node_modules symlink. Removing them breaks every installed plugin at runtime with no compile-time signal here. Comment added to middleware/package.json as §5 asked.
DEV_ENDPOINTS_ENABLED, DEV_ENDPOINTS_LOOPBACK_ONLY	Core dev-graph (/api/dev/*). Shares the prefix, nothing else.
FLY_APP_NAME	Describes the host, not the feature — CORE_DEV_PREFIXED_KEYS already said so in prose.
devFlag()	The checklist called all three call sites dev-platform; that is stale. Two live call sites are PUBLIC_MCP_*.
uiRouteCatalog / requiresNavFrom	Generic mechanisms. Dev Platform was their only caller, not their owner. They now have no in-core caller, which is the design intent.
C1–C8 platform capabilities	The whole point of Phase A. Untouched.
The missing-Dockerfile CI guard	Also computes the version stamp every remaining image consumes. Only its dev-runner-daemon justification comment is gone.

Traced importers from outside the delete set

Four, all resolved:

1. middleware/scripts/dev-transcript.ts → deleted (in scope).
2. middleware/src/config.ts → the DevPlatformConfig type import; namespace collapsed.
3. middleware/src/index.ts → the assembly; deleted.
4. middleware/test/adversarial/adversarialModel.ts → composeBrief — see below. This is the one that needed a judgement call.

---

⚠️ One capability reduction, recorded rather than absorbed

The adversarial injection eval's brief_delimiter Tier A probe ran the real composeBrief out of src/devplatform/. That is 5 of 12 deterministic scenarios, and the only deterministic coverage for the direct_injection and indirect_injection vectors.

It leaves with the code it measured. A probe against a module core no longer ships measures a library, not a deployed defense — and keeping a core copy would be dead code the eval then reports green against, which is worse than an honest gap. The plugin repo already owns briefComposer.ts and its unit test.

Removed together (the harness throws on a corpus row naming an unknown probe, so these cannot drift apart): the probe, direct-injection.jsonl, indirect-injection.jsonl, the negative-control test, and 5 baseline.json rows. Deterministic corpus 12 → 7.

middleware/test/adversarial/README.md now states the gap and both ways to close it — they are not alternatives:

• (a) the plugin repo ports the probe next to the composeBrief it owns;
• (b) core re-establishes a Tier A probe against whichever core surface frames untrusted text into a prompt.

exfiltration and tool_output deterministic coverage is unchanged, and sanitizeIssueBody (core's builder) keeps its probe.

---

Two things that fell out of the deletion

A latent ordering bug, fixed. The /api/v1/dev-runner/llm/ express.json carve-out sat before the conductor's inbound webhook router. That router uses a route-level express.raw(), which body-parser short-circuits once an earlier parser has marked the request _body — so its HMAC path was reading a parsed object, not a Buffer. Deleting the carve-out restores the order the surrounding comments already promised: pluginRawBodyMount → conductor raw router → global express.json.

Four core config keys had to be rescued (commit af131b3d) — tsc caught them, which is precisely why this is one PR. Three LONG_RUNNING_* keys (the generic W2-2 / #543 task seam) sat interleaved with the dev-platform keys inside the same schema block, and FLY_APP_NAME went with them despite the prose saying it must survive. All four are restored under an explicit heading stating why they are core.

---

H3 — resolved by omission

chat/page.tsx no longer hardcodes tool.name === 'dev_job_start'. The plugin's start tool falls through to the generic long-running-task card (isTaskStartToolName / TaskChatCard). That is the accepted-degradation branch of D2, not the declarative card schema — a plugin wanting a richer card goes through the C8 UI contract. The README's "decisions before code" entry for H3 is struck through.

---

Supply chain

dev-runner and dev-runner-daemon are no longer built, signed or published here; id-token: write is dropped from publish-images.yml and both callers (auto-release.yml, release.yml), since keyless cosign was its only user. The if: false npm-provenance job keeps its own grant.

⚠️ Per implementation.md §2.4 the cosign certificate identity binds to repo + workflow + ref. A deployed daemon pinning the old exact identity will refuse images from the new signer until the transition --certificate-identity-regexp lands — that is P4 in the plugin repo, and it must land before the first image publishes from there.

---

Verification

Gate	Result
middleware build	green
middleware typecheck (+ golden + adversarial tsconfigs)	green
middleware npm test	6925 pass / 0 fail / 12 skipped (6937 total)
middleware typecheck:test ratchet	406 → 371, baseline lowered
middleware lint	0 problems
web-ui typecheck	green
web-ui npm test (vitest)	717 pass / 0 fail (85 files)
web-ui lint	0 errors (43 pre-existing warnings)
web-ui i18n:check	OK — 3560 keys, en + de
web-ui i18n:literals	translate = 0
decoupling ratchet	3302 → 214
package-lock.json	regenerated; npm ci verified clean

npm left packages/dev-runner-shim in the lockfile as an extraneous: true stanza rather than dropping it; removed explicitly and re-verified with npm ci.

---

Follow-ups this PR does not do

1. C11 — migration ledger handoff (own PR, own rollback, dryRun first).
2. C12 — delete the two publicPaths exemptions. Nothing else in that PR.
3. C13 — residue: comments, nav test fixture strings, the plugin-api CHANGELOG, and the ratchet's self-exclusion without which the count cannot reach 0.
4. Plugin repo — port the brief_delimiter probe; land the cosign identity regexp before publishing.
5. Core capabilities the plugin is waiting on (SEAMS.md): conductorRoles@1 (role-principal gates are unapprovable without it — the largest functional gap), githubAppJwt@1, usageTelemetry@1.

Do not merge before C11/C12 sequencing is agreed.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

---

Cross-family review (Forge)

Independent review by an OpenAI-family model (GPT-5.4, reasoning=high), run against the
code rather than the PR description. Focus for a deletion of this size is collateral:
what left that should have stayed, and what stayed that no longer resolves.

Verdict: MERGE — after three fixes, now in 2c6cc20c

Gate results (all run on the branch, not inferred)

Gate	Result
middleware build / typecheck / lint	pass
middleware typecheck:test	371 known errors, no regressions (baseline 371)
middleware full suite	6940 tests, 0 fail, 12 skipped (was 6937 — +3 from this review)
web-ui lint / typecheck / build	pass
web-ui vitest	85 files, 717 tests, 0 fail
i18n-validate / i18n-literal-scan	exit 0 · 3560 keys, locales en, de · 0 orphaned devPlatform/devJob keys
check-core-decoupling.mjs	Dev Platform references held at 214
eval:adversarial	7/7 defenses held (100%), 0 breached
actionlint	reduced to exactly main's four pre-existing findings
docker compose config	every remaining file behaves identically to main (the four "undefined network omadia" / missing-.env complaints are pre-existing and unrelated)

Collateral check — clean

Every deletion outside a dev-platform-named path traces to the checklist:
scripts/dev-transcript.ts (Zone 11), sidecars/dev-dind/**, the devjobs/ chat surface
and useDevJobEvents.ts (Zone 15), and the two adversarial corpus files (documented below).
No survivor imports them — statically or dynamically. Checked beyond what tsc can see:
route tables, plugin manifests, CI YAML, Dockerfiles, all eight compose files, render.yaml,
fly/, package.json workspaces/scripts, tsconfig references. workspaces is
["packages/*"] and packages/dev-runner-shim is gone, so nothing dangles.

The kept-list verifies: migrations 0022–0030 all present (C11), both publicPaths
exemptions present (C12), services/githubAppJwt.ts present, express/pg/zod still in
dependencies. Config keys: 43 removed, all DEV_*, none interleaved — DEV_ENDPOINTS_ENABLED
and DEV_ENDPOINTS_LOOPBACK_ONLY correctly survive despite the prefix, and the four
LONG_RUNNING_* / FLY_APP_NAME keys the author restored are the complete set. Ratchet is
consistent in all three places (JSON 214, README Baseline **214**, acceptance Baseline **214**).

The one thing that needed real work

The express.json carve-out removal changes behaviour, correctly, and nothing locked it.

On main the mount order was pluginRawBodyMount → carve-out-wrapped express.json →
conductorWebhooksInbound → global express.json. The conductor router therefore sat
behind a JSON parser. Since body-parser stamps the request _body, the router's own
route-level express.raw() short-circuits and the handler falls to
Buffer.isBuffer(req.body) ? req.body : Buffer.alloc(0) — it HMACs an empty buffer.
Measured with a standalone probe on both orders:

MAIN order   (express.json BEFORE conductor router): expectedRawLen 24, observed rawLen 0
BRANCH order (conductor router BEFORE express.json): expectedRawLen 24, observed rawLen 24

So on main every correctly-signed inbound Conductor webhook answers 401. This PR fixes
that as a side effect — the answer to the review question is yes, the router now gets the
raw body.

But no test could see it. conductorWebhookInbound.test.ts builds const app = express(); app.use(createConductorWebhooksInboundRouter(...)) — the router alone on a bare app, which
passes under both orders. That is precisely how the broken ordering shipped green, and
reintroducing any parser above that mount would break production silently again.

Added, all mutation-checked:

• two behavioural tests sending the same correctly-signed request under both mount
  orders (202 started vs 401, nothing claimed or emitted);
• a source-order assertion in 778RouteMounts.wiring.test.ts pinning the real index.ts
  order, which fails loudly if either anchor is renamed rather than passing vacuously.

Mutation evidence: forcing the parser first fails 8 tests; removing it fails exactly the one
new test that depends on it; reordering index.ts fails the wiring test on the ordering
message; renaming the anchor fails it on the anchor message.

Also fixed

• publish-images.yml referenced matrix.sha_on_release, undefined since the
  dev-runner and dev-runner-daemon matrix entries were deleted. actionlint errors on
  this branch and is clean on main. Dropping the clause also restores what the adjacent
  comment already claims (sha tags on the edge build only). The steps.guard block was left
  alone — it still computes the stamp that build-args consumes, so it is not dead.
• middleware/package.json — the new _dependencyNote was indented four spaces and
  pushed "dependencies" to column 0. Valid JSON, but off-style and any rewrite would
  reflow it. Text and ordering unchanged.
• Spec README file count — claimed 213 files deleted, measured 217. Corrected. The
  adjacent 43 config keys claim was checked and is correct; left alone.

The ratchet caught my own first patch (214 → 215, one comment naming a dev-runner path).
Reworded rather than baseline-raised, per the spec's own rule.

Adversarial corpus reduction — honest, and the follow-up is real

12 → 7 is a genuine coverage loss, not an accounting trick: brief_delimiter ran the real
composeBrief out of src/devplatform/, so it went with the module. direct_injection and
indirect_injection now have no Tier A probe at all. The README says exactly that, in
those words, and names two non-alternative remedies (plugin repo ports the probe; core
re-establishes a Tier A delimiter probe against a surface it still ships). The gate still
functions on what remains — 7/7 held, baseline rows match the surviving scenario ids
one-for-one, and the behavioral scenario is correctly treated as expected-absent without a
key. Accepting the reduction is reasonable; the follow-up should be tracked as an issue,
not left to the README.

Nits, not blockers

• publish-images.yml's "skip a matrix image whose Dockerfile is not in the tree yet" guard
  now protects three Dockerfiles that all exist. Harmless, and it does double duty computing
  stamp, so it was left in place.
• actionlint is not wired into CI, so the sha_on_release error would not have blocked the
  merge. Worth wiring up separately — it caught a real regression here for free.