fix(mcp): railway_service_redeploy multi-account dispatch (Bug #15) by gHashTag · Pull Request #110 · gHashTag/trios-railway

gHashTag · 2026-05-01T18:12:54Z

Summary

Fixes MCP Bug #15 — railway_service_redeploy now works across all Railway accounts.

Problem

railway_service_redeploy failed with RAILWAY_TOKEN not set or invalid for any non-IGLA service because it used build_client() fallback which requires user-scoped RAILWAY_TOKEN (not available with project-scoped tokens).

Solution

Added find_project_for_service() helper that:
- Queries all configured accounts
- Finds which project contains the target service
- Returns correct project_id
Updated railway_service_redeploy to:
- Auto-resolve project if not provided
- Always use build_client_for_project() (never build_client())

Impact

✅ Works for all acc0/acc1/acc2/acc3/acc4 scarabs (18 services)
✅ Maintains backward compatibility (explicit project param still works)
✅ No more "frozen" MCP connector after repeated failures

Files changed

crates/trios-railway-mcp/src/tools.rs: +29 -3 lines

Verification needed

After merge:

Railway auto-rebuilds MCP service db786a4b (~3 min)
Re-run 18 scarab redeploys via railway_service_redeploy
Verify Bug A: bpb_samples writes flowing again

🌻 phi² + phi⁻² = 3 · TRINITY · NEVER STOP

Extracts 5 stable library crates from trios-railway: - tri-core: deploy(), kill(), rotate(), snapshot(), fleet_list() - tri-hunt: seed_hunter_status(), smoke_race(), rung_schedule(), prune_diverging(), mirror_siblings() - tri-exp: next_exp_id(), claim_exp_ids() via Neon sequence - tri-canon: validate(), validate_for_deploy(), tripwires #97-108 - tri-ledger: append(), DDL migration, append-only enforcement Creates bin/tri and bin/tri-gardener as thin shim CLIs that delegate to the public crate APIs. All crates compile with zero clippy warnings. Closes #69. Part of #68. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…+E.5) Phase E.2: 4-layer seed policy enforcement - forbidden_seeds table (42/43/44/45 banned for quorum) - sanctioned_seeds table (F17-F21 Fibonacci, Lucas-closed) - seed_policy_violations table (R5-honest tripwire) - enforce_seed_policy() trigger (priority=0 checks, fresh validation) - Smoke tests: forbidden rejected, sanctioned allowed, replay allowed Phase E.5: 10-min smoke-first experiment configs - E1: Champion reproduce (seed=42, anchor for all) - E2-E3: Quorum-3 candidates (seeds 43/44, σ² validation) - E4: Capacity push (h=1536 ctx=16, breach <1.85?) - E5: GF16 storage test (L-R9 guard, TRAIN-001 prep) - E6: Hybrid-001 (3T+15GF16, 18.4 GOPS target) - E7: LR φ-optimal (lr=αφ/φ³=0.004, INV-8 verification) DB-level protection: parallel agents now hard-rejected from inserting forbidden seeds with priority=0 (quorum violation). Single source of truth. Closes #81 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…eproduction Golden Float Family (GF8/GF16/GF32/GF64/GFTernary): - G1: GF8 (8-bit) - ultra-low-power speculative - G2: GF16 (16-bit) - production baseline, BENCH-004b ready - G3: GF32 (32-bit) - FP32 drop-in replacement - G4: GF64 (64-bit) - double-precision scientific - G5: GFTernary (2-bit) - bulk quantized for HYBRID-001 Champion Exact Reproduction: - train_v2 h=1024 ctx=12 WT+resid (no attn) - Exact BPB=1.8921 target, Δ≤0.005 tolerance - Full 120K steps budget (not smoke test) All configs follow Golden Float whitepaper φ-constants: - GF8: φ⁴+φ⁻⁴ = 7 (L₄) - GF16: 6/9 ≈ 1/φ, L-R9 safe (d_model≥256) - GF32/GF64: Lucas-closed mantissa (13/18, 21/42) - GFTernary: {-φ, 0, +φ} Trinity basis Closes #81 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

@50

Phase E.GF: 8 formats × 10-min budget sweep per zig-golden-float Account distribution: acc0(6 lanes @50), acc1(14 lanes @50,80,90) Experiment matrix (20 total): - acc0: GF8, GFTERN, GF32, GF64, FP16, BF16, FP32 (priority 50, 90) - acc1: GF16 variants, GF32, BF16, FP32 (priority 50, 80, 90) All use Fibonacci seeds (1597, 2584, 4181): - 3× extreme-low-power (GF8) - 4× extreme-low-power + bulk ternary (GFTERN) - 3× 16-bit baseline (GF16) - 3× IEEE half (FP16) - 3× 32-bit baseline (FP32) - 3× Google brain-float (BF16) - 1× IEEE single (FP32) - priority 80 champion replay - 1× IEEE single (FP32) - priority 90 - 2× IEEE single (FP32) - priority 90 SQL artifact: .trinity/phase_e_gf_sweep.sql L7 audit: gardener_decisions row enqueued Closes #81 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Two GF32 entries with wrong seeds deleted. 8 valid experiments remain. Untracked files added (.swarm/) for consistency. Closes #81 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…ments 8 correct experiments already exist in experiment_queue: - 2× GF8 (acc0) - 2× GF16 (acc0) - 2× GF32 (acc0) - 2× FP32 (acc0) - 1× BF16 (acc0) Let Railway workers run and verify if duplicates occur. Will re-address constraint after initial results. Closes #81 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

… fix - Add ALLOWED_PROJECT_IDS whitelist + build_client_for_project() to MCP gateway - MCP tools now route to correct per-account token based on project ID - Add AccountConfig + load_accounts() reading RAILWAY_TOKEN_ACC{0..3} env vars - Add env_for_project() for per-account environment ID resolution - Add batch-deploy subcommand with TOML config, multi-account, bounded concurrency - Add variables_upsert_parallel() for faster deploys - Fix snapshot fleet auth mode (was hardcoded team, now auto-detects per account) - Export is_uuid_like() from trios-railway-core - Extract snapshot_one_account() to fix clippy too_many_lines - Remove hardcoded IGLA_PROJECT_ID/IGLA_PROD_ENV_ID from tri-railway CLI - Add .env to .gitignore Closes #81 Agent: GENERAL

- Add railway-template.json: 8 formats × 4 accounts deployment config - Update disaster-recovery/fleet-snapshot.json: add acc0 project - Add Dockerfile.igla-gf: GF format training container - Add format_benchmark.zig: CPU format performance benchmark - Add railway-service.json: service config reference - Add Phase E/F SQL scripts for experiment tracking Formats: GF8/GF16/GF32/GF64/GFTernary/FP32/FP16/BF16 Champion: GF32 fastest (29s vs 39.6s baseline) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Fixed f32ToBf16: correct bf16 constants (128/0x007F instead of 256/0x00FF) - Fixed format_results: changed from const to var for mutability - Updated header to include GF8/16/32/64 formats Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…oring - fleet_health: queries all accounts, returns service counts + connectivity - seed_list: lists all seed/training services across accounts - Both tools use multi-account token routing via AccountConfig Agent: GENERAL

- Fixed Unicode/std.debug.print issues by using std.log.warn - Fixed f32ToBf16 constant errors (128/0x007F instead of 256/0x00FF) - Fixed format_results const/var for mutability - Added GF8/GF32/GF64 formats to benchmark Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Release build: trios-railway-mcp 7.8MB (release profile) Experiments: 80 new across seeds 100-129, formats GF16/FP32/BF16/GF8/GFTernary Queue: 132 pending (43 acc0, 49 acc1, 20 acc2, 20 acc3) Workers: 64 alive across acc0-acc2 Agent: ReleaseCannon

…redeploy, experiment_queue_insert tools Adds 4 new MCP tools for full operational control via the gateway: - experiment_queue_status: Neon DB queue breakdown by status/account - worker_status: alive/stale/dead worker counts per account - service_batch_redeploy: bulk redeploy services on an account - experiment_queue_insert: insert experiments into queue Dependencies: tokio-postgres, rustls, tokio-postgres-rustls, webpki-roots Closes #11 Agent: GENERAL

The #[tool] macro only registers tools when they're inside the #[tool_router] impl block. experiment_queue_status, worker_status, service_batch_redeploy, and experiment_queue_insert were in a separate impl block and invisible to the tool router. Agent: GENERAL

db_connect() was using tokio_postgres::NoTls but Neon requires TLS (sslmode=require). Now builds a rustls ClientConfig with webpki-roots Mozilla CA bundle for proper TLS handshake. Agent: GENERAL

…ect timeout tokio-postgres doesn't understand channel_binding=require and sslmode=require libpq params, causing the connection to hang indefinitely. Now strips these params before connecting and adds a 10s timeout. Agent: GENERAL

Root cause: rustls 0.23 requires an explicit CryptoProvider. Without it, the TLS handshake panics at runtime with 'Could not automatically determine the process-level CryptoProvider'. Now calls install_default() before creating the TLS config. Also: keep sslmode=require in URL (only strip channel_binding). Agent: GENERAL

- Added Gaussian weight distribution (σ=0.1) for realistic testing - Added 3-layer MLP inference benchmark (10→8→4→1) - GF16 outperforms fp16 by ~47x in MLP inference MSE - bf16 shows same accuracy as GF16 in inference scenario Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Fixed u32 overflow in Gaussian weight generation (use u64) - Added UNIFORM [-100, 100] distribution test - Results confirm whitepaper: GF16 wins on large dynamic range - GF16: 0.0198 MSE (best) - fp16: 184.2 MSE (~93× worse) - bf16: 335.9 MSE (~170× worse) Key findings: - GF16 φ-distance (6:9) provides superior dynamic range - fp16/bf16 collapse on large values due to smaller mantissa - MLP test with small weights showed GF16≈bf16 due to limited range Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

7-phase decomposed roadmap covering: - Phase 1: Critical fixes (connection pooling, CryptoProvider, tests, dead code) - Phase 2: 5 missing tools (logs, queue update, var upsert, batch deploy, bpb samples) - Phase 3: Architecture (split tools.rs, bearer auth, rate limiting, kill switch) - Phase 4: Observability (tracing, health check, metrics) - Phase 5: Code quality (constants, error types, seed validation, concurrency) - Phase 6: CI/CD (GitHub Actions, auto Docker build+push) - Phase 7: Documentation (tool catalog, architecture, runbook) Also includes fleet-wide NEON_DATABASE_URL injection experience log. Agent: GENERAL Soul: RailRangerOne

Cherry-picked SSE transport onto ring-69 branch for Railway deploy. Adds GET /sse + POST /message routes for legacy MCP clients (Roo/Cline). Changes: - Cargo.toml: add transport-sse-server feature + tokio-util - main.rs: dual transport (SSE + Streamable HTTP) - Cargo.lock: +2 lines (tokio-util was already transitive) Agent: GENERAL Soul: SSEntry

…to allowlist

…DRY token-kind handling - Replace const ALLOWED_PROJECT_IDS with OnceLock<Vec<String>> loaded from ALLOWED_PROJECT_IDS env var (comma-separated). Falls back to hardcoded DEFAULT_ALLOWED_PROJECT_IDS when env var is absent. - Fix default whitelist to 6 correct project IDs: abdf752c (acc0), e4fe33bb (acc1/IGLA), 12c508c7 (acc2), 8ab06401 (acc3), 0247abaa (acc4), 475a2290 (acc5/acc6). Removes stale da1fb0c7 and f3350520. - Extract resolve_auth_mode() method on AccountConfig to eliminate 4 duplicated token-kind match blocks across fleet_health, seed_list, service_batch_redeploy, and build_client_for_project. - Update build_client_for_project doc comment (0..3 → 0..7). Improvements #1, #2, #3 from fleet audit.

experiment_queue → strategy_queue, workers → scarabs. Also fix created_by from 'mcp-gateway' to 'human' (must match CHECK constraint). Fixes #9 — experiment_queue_status, worker_status, experiment_queue_insert all now reference the correct post-migration table names.

… insert #13: railway_service_redeploy and railway_service_delete now accept optional 'project' parameter for multi-account token dispatch. When provided, uses build_client_for_project() instead of global RAILWAY_TOKEN. Backward-compatible — falls back to build_client(). #14: experiment_queue_insert now uses $2::jsonb cast so postgres handles text→jsonb conversion, fixing 'error serializing parameter 1' when tokio-postgres passes a String for a jsonb column.

#14 real fix: add postgres-types with with-serde_json-1 feature so serde_json::Value implements ToSql for jsonb columns. Pass params.config_json directly instead of serializing to String first. Also amends #13 (multi-account dispatch for redeploy/delete).

Replace build_client() fallback with auto project resolution: - Add find_project_for_service() helper - Uses build_client_for_project() instead of build_client() - Auto-detects project by querying all accounts for service_id - Fixes "RAILWAY_TOKEN not set or invalid" for non-IGLA services Before: failed on acc0/acc3/acc4 scarabs (no user-level token) After: works for all configured multi-account services Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

gitguardian · 2026-05-01T18:12:59Z

⚠️ GitGuardian has uncovered 1 secret following the scan of your pull request.

Please consider investigating the findings and remediating the incidents. Failure to do so may lead to compromising the associated services or software components.

🔎 Detected hardcoded secret in your pull request

GitGuardian id	GitGuardian status	Secret	Commit	Filename
31559428	Triggered	PostgreSQL Credentials	`23f1535`	Dockerfile.igla-gf	View secret

🛠 Guidelines to remediate hardcoded secrets

Understand the implications of revoking this secret by investigating where it is used in your code.
Replace and store your secret safely. Learn here the best practices.
Revoke and rotate this secret.
If possible, rewrite git history. Rewriting git history is not a trivial act. You might completely break other contributing developers' workflow and you risk accidentally deleting legitimate data.

To avoid such incidents in the future consider

following these best practices for managing and storing secrets including API keys and other credentials
install secret detection on pre-commit to catch secret before it leaves your machine and ease remediation.

^{🦉 GitGuardian detects secrets in your source code to help developers and security teams secure the modern development process. You are seeing this because you or someone else with access to this repository has authorized GitGuardian to scan your pull request.}

Dmitrii Vasilev and others added 28 commits April 28, 2026 16:53

chore(ring-69): cleanup duplicate experiments (IDs 114, 115 removed)

2166fa8

Two GF32 entries with wrong seeds deleted. 8 valid experiments remain. Untracked files added (.swarm/) for consistency. Closes #81 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

fix(mcp): use tokio-postgres-rustls for Neon TLS connection

209bd09

db_connect() was using tokio_postgres::NoTls but Neon requires TLS (sslmode=require). Now builds a rustls ClientConfig with webpki-roots Mozilla CA bundle for proper TLS handshake. Agent: GENERAL

feat(mcp): extend account support to ACC0-ACC7, add project 475a2290 …

c08762b

…to allowlist

chore: update Cargo.lock for postgres-types with-serde_json-1 feature

175557d

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(mcp): railway_service_redeploy multi-account dispatch (Bug #15)#110

fix(mcp): railway_service_redeploy multi-account dispatch (Bug #15)#110
gHashTag wants to merge 28 commits into
mainfrom
fix/mcp-redeploy-multiaccount

gHashTag commented May 1, 2026

Uh oh!

gitguardian Bot commented May 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

gHashTag commented May 1, 2026

Summary

Problem

Solution

Impact

Files changed

Verification needed

Uh oh!

gitguardian Bot commented May 1, 2026

⚠️ GitGuardian has uncovered 1 secret following the scan of your pull request.

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant