Skip to content

Latest commit

 

History

History
319 lines (264 loc) · 21.4 KB

SPRINT_BOARD.md

File metadata and controls

319 lines (264 loc) · 21.4 KB

SPRINT_BOARD.md

The ordered queue. Agents pull from the top; they do not reinterpret the order. Hard parts first. When a surface closes: commit + evidence, update CLAIM_LEDGER.md, move to the next row.

Companion files: CLAIM_LEDGER.md · VISION_VS_VERIFIED.md. Last updated 2026-09-03 by CC. Re-measured, not re-asserted — B and C were both QUEUED for four weeks after they had shipped. See the note under the queue.


Execution graph (the loop every agent obeys)

SPRINT_BOARD (ordered)
      │ pull next
      ▼
  IMPLEMENTER ── evidence ──▶ VERIFIER (≥95% of original spec?)
      ▲                          │
      │  <95%: fix loop          │ ≥95%
      └──────────────────────────┤
                                 ▼
                    COMMIT + evidence + ledger update
                                 │
                                 ▼
              next priority  ──or──  PARKING_LOT (if blocked)
                                 │  tokens out / wall-clock end?
                                 ▼
                          TOTAL_RECAP.md

Boundary — stop and wait ONLY for these five live-state gates: merge to main · npm publish · prod deploy / env secrets / Railway infra GO · on-chain txs that spend real funds · domain / DNS ownership actions.

Everything else: log under BLOCKED_FOR_SEAN, pull the next surface, keep building. Waiting for approval on a non-live-state item is a process failure.

Verifier loop (mandatory, per task): implement against the original spec → self-check (tests/smoke/measured output) → separate verifier pass with a different checklist → if <95% match, fix and re-verify → only then advance. Evidence or it didn't happen.


QUEUE (ordered)

# Surface Definition of done (≥95% of spec) State
A 3-file source of truth The three files exist on a branch, populated only with evidenced claims IN PROGRESS (this commit)
B TrustShell presentProof / badge path A portable proof surface a reviewer calls standalone (no engine API key); tests DONE — shipped in @hyperdag/trustshell 1.3.0. client.presentProof (src/lib/trustshell.ts), src/lib/badge.ts, checked-in example examples/proof-badge-trinity-shofet.svg. No API-key gate on the CLI proof path. tests/badge.test.ts + tests/portable.test.ts = 33 tests pass (measured 2026-09-03)
C TrustMarket backend wiring Rating ingestion consuming fold root + dual-auth decision; schema + API + tests; no UI theater DONE — src/services/rating-ingestion.ts (admitRating matches fold root, rejects fold_root_mismatch, and fails closed on an unrecorded dual-auth decision — "an unrecorded ALLOW is not an ALLOW"), route src/routes/v1/ratings.ts, tables repid_ratings + repid_outcomes both exist. tests/rating-ingestion.test.ts + tests/participant-rating.test.ts = 39 tests pass. Both tables hold 0 rows — correct for a pre-user system, not a gap
D TrustTrader backend Fold-root / RepID consumption; schema + API + tests QUEUED
E trustchat + AISocialMirror backend plugs Shared auth/RepID/gate hooks where repos exist; clean stub interfaces where they don't QUEUED
F AITrinitySymphony.com deploy diagnosis Correct Vercel/project mapping as a branch/PR with exact steps for Sean; do NOT flip DNS QUEUED
G Parking lot (when A–F blocked) Ship commits on branches from PARKING_LOT FALLBACK

Why B and C sat QUEUED for a month after shipping

This board is the ordering authority — the loop above says agents "pull from the top and do not reinterpret the order." A stale board is therefore worse than no board: an agent obeying it faithfully rebuilds finished work. B shipped in the SDK's 1.3.0 and C has 39 passing tests behind it; both still read QUEUED on 2026-09-03.

docs/PRIOR-WORK-INDEX.md in trinity-ecosystem records the cost of exactly this — two sprints spent optimising a component already at 97.9% of its bound. The rule that prevents it is cheap: before pulling a surface, measure whether it is already built. Four greps and two jest runs settled B and C here.

One correction worth keeping, because it nearly produced a false finding: checking C's schema by guessing table names (agent_ratings, ratings, marketplace_ratings) returned nothing and read as "the schema was never built." The real tables are repid_ratings and repid_outcomes — this repo names everything repid_*, which CLAUDE-RULE-5 says outright. Trace the route to the table; do not guess the table.

PARKING_LOT (pull when the primary queue is blocked)

  • Searchable encrypted memory cell (agent memory, committed + queryable)
  • Plonky3 recursion stub + measure (proof-of-proof, capture proving time)
  • Family-BFT docs-as-code tests (the quorum-family invariants as executable tests)
  • Canary hardening (fetch-timeout on the HAL harness; the harness lacked one)

BLOCKED_FOR_SEAN (human-gated — parked, not stopping the loop)

Item Gate type Exact action Sean takes Unblocks
Merge PR #364 (Tier-0 harness + MVP packaging) merge to main Review + squash-merge #364 mainline npm run demo:harness
Publish the public trust-harness/CLI package to npm npm publish Decide public package boundary (must NOT be this proprietary repo), then publish reviewers npm install without repo access
Deploy the always-on Railway "manager" worker Railway infra GO Approve service deploy (prepared on branch when ready) laptop-closed overnight loop
TRUSTRAILS_HMAC_SECRET is not armed on either aitrinitysymphony surface — both self-report ready:false [MEASURED 2026-09-03 via GET /api/version] env secret Set a real secret on both the Vercel project (www + apex) and the Railway service (app), then re-check /api/version shows ready:true compliance-receipt audit_hash integrity

OPEN AS OF 2026-09-04 18:00Z (session handoff — do not re-derive)

Everything below was measured this session against the live database, the deployed /health, or the code on main. Dates are attached because a number without one is an assertion. Items closed during the session were deleted rather than ticked — a list that keeps done work trains its reader to skim, which is how the one urgent row gets missed.

Human-gated — nothing below moves without Sean

Item Where exactly Value
⚠ Stranded escrow — decide recovery contract 133a9048…, escrowed 2026-09-04 12:00:13Z Real testnet USDC is committed to a contract that can never be delivered (NULL work_statement_hash, WORK_STATEMENT_REQUIRED at fulfil). #608 stops NEW ones; it does not release this one. Recovering funds is a money-path write and was deliberately not attempted.
TRUSTRAILS_HMAC_SECRET Vercel project trustrails and Railway repid-engine — same value both generate: openssl rand -hex 32. A publicly-known placeholder is in use today, so audit hashes are forgeable. Deleting it from HEAD is not rotation.
SUPABASE_URL + SUPABASE_SECRET_KEY Vercel project trustmarket-landing the sb_secret_… key from Supabase → Settings → API Keys. NOT sb_publishable_… (that one ships in the browser). Route 503s by design until both exist.
trustmarket.dev deploy target Vercel domain is served by an April static upload from trustmarket-coming-soon, not the landing project — merged fixes will not appear until this is repointed
www.trustmarket.dev TLS DNS / Vercel certificate invalid

Verification debts — real work that is NOT_CHECKED until something runs

Neither is a failure. Both are claims with no live witness, and each names the single observation that closes it.

  • The 2026-08-17 HAL orientation fix has never scored a real deliverable. Deliverable traffic stopped 2026-08-17, the same day the fix landed. VERIFIED in simulation (npm run repid:sim: monotone, zero violations), NOT_CHECKED in production. Closes on the first purpose: deliverable event dated after 2026-08-17.

  • The service-quality hook has never executed — but everything it needs is now in place. [MEASURED 2026-09-04 15:56Z, from the public /health, not a dashboard:]

    deployed_commit_short  34e9d28          (#608 is live)
    service_quality_hook   mode=shadow  enrolled_count=2  allowlist=default
    

    So the flag IS set, and allowlist: default means it is using the corrected pair from #608 rather than an env override. Until #608 the allowlist named the most active BUYER, so every fulfilment reported agent_not_enrolled at any flag setting — that is why the absence was not evidence of anything.

    This row is why the observability change was worth making. "Is the flag set?" was unanswerable for a full session and got guessed; it is now one unauthenticated GET, and the answer was yes the whole time. Closes on the first row where service_contracts.metadata ? 'hal_quality_shadow' — expected on the next fulfilment by an enrolled provider (~12:00Z daily). If it does not appear, run npm run verify:quality-hook BEFORE suspecting the flag: it distinguishes "nobody enrolled delivered" from "the hook cannot fire at all".

Found this session, not yet acted on

  • CLAUDE.md is stale and says so nowhere. It records all 12 Trinity agents as offline since ~2026-07-17. FALSE [MEASURED 2026-09-04]: trinity-nexus and trinity-hdm scored on 2026-09-03, trinity-gcm and trinity-veritas on 09-02. Read liveness from recency, never from that line.

  • Live fulfilments still pass task_domain: 'general'. So today's real service work classifies as non-deliverable and earns nothing from HAL. STALE, CORRECTED 2026-09-04 (Beat 99). #605 (merged same day, after this note was written) changed the actual SERVICE_FULFILLED sites (validation-repid-delta.ts:428,466) to default task_domain to 'service_contract', not 'general' — read directly on main, not from a PR description. The "earns nothing from HAL" mechanism was also wrong independent of that: this path (applyValidationEvent) never calls classifyTaskPurpose or isDeliverableDomain — the purpose gate lives in runScoreEvent (the HAL path) and the reward gate in the agents-external score-event route, neither reached from here. task_domain here only feeds the apply_vertical_accuracy trigger's per-domain competence bucket, not a HAL reward decision. The real observation survives (paid contract work was landing in one untyped bucket); the claimed consequence did not.

  • Peer-verification is HALTED BY A FLAG, not broken [corrected 2026-09-04]. Completer last ran 2026-07-04, producer stopped 2026-07-21, 62,841 rows stuck in_review with completed_at NULL — but trinity-task-bridge.ts gates the producer on isProducerHalted('peer_verify'), reading the PRODUCER_HALT_CLASSES env var. That is a deliberate circuit breaker (L2 breaker 2.1, "drain-only, fail-loud"), and the stop date is consistent with the flag being set rather than with code rot. If peer_verify appears in PRODUCER_HALT_CLASSES on the Railway repid-engine service, removing it restarts the producer — one variable, not a rebuild. UNVERIFIED from here: Railway env is not readable from a cloud session. Check the service Variables before planning any repair work. Separately and still true: it has produced zero RepID events in its entire history, so its dormancy cost nothing in scoring, but the anti-gaming consequence path that chronic-flag-accumulator routes to has never once fired.

  • PREDICTION_RESOLVE needs a CALLER, not a fix [MEASURED 2026-09-04]. 2,892 events lifetime, 11 since July — second-largest event type in the ledger. The producer is wired and works: it is emitted from an HTTP route (src/routes/agents-external.ts) and from nowhere else. There is no worker, no cron, no scheduled resolver. So the quiet is absent demand, not a broken path — the same shape as the deliverable-traffic finding. Prediction is also the best-behaved RepID substrate available (objectively resolvable, measures calibration, unfakeable because you commit before the outcome, and proper scoring rules make honest reporting optimal), so pointing traffic at that route is a smaller job than building a new event class.

  • repid_agents.risk_tolerance is read by no code. Dead column. Harmless today, but it is the shape a user-settable risk knob would take, and such a knob is a measured +73 RepID arbitrage on identical work (npm run repid:sim, Part 4). Do not wire it without reading that first.

  • A gate can be added at one end of a pipeline without checking the producers at the other. #607 made work_statement_hash required at fulfil and left contract CREATE permissive, which stranded real money the same day (see the escrow row above; fixed for the cron in #608). src/routes/v1/negotiation.ts, the other writer, was read end to end (Beat 100, 2026-09-04): it does not share the content-validity failure #607 was about. negotiation.ts:1324-1330 calls parseWorkStatement(rfq.scope, ...) and returns 400 on !spec.ok before calling acceptAndAward — an unparseable statement never reaches the money-committing RPC here, unlike the pre-fix cron.

    What the read surfaced instead is narrower and still open. a2a_accept_and_award is documented (a2a-negotiation.ts:57-64) as one atomic Postgres transaction — RFQ CAS + service_contracts insert + a2a_awards insert + losing-bid sweep — specifically so a rejected constraint can't leave "a live, payable, un-provenanced contract" behind. The work_statement column, though, is bound by a separate .update() after that RPC returns success (negotiation.ts:1376-1379), outside the transaction its own neighboring comment describes as the fix for exactly this class of bug. A failure on that update (WORK_STATEMENT_BIND_FAILED, 500) leaves the same shape as the stranded-escrow row above: a live, payable contract with work_statement unbound — reached by a DB-write failure between two calls, not by unparseable content. Real fix is passing the parsed spec into the RPC so binding is atomic with creation; that RPC's signature is in a migration outside this repo (schema managed externally) and was not read this beat, so changing its call contract blind is a money-path guess, not a fix. Not attempted.

  • 106 of 111 behaviour gates are unobservable from outside the process [MEASURED 2026-09-04]. See the flag entry below. The obvious fix — publish them all on /health — is WRONG: /health is public and unauthenticated, so that would hand an attacker a map of which money controls and breakers are currently off. Anything gating funds, chain writes or a breaker needs an authenticated surface. That split is a decision, not a cleanup.

Peer-verification: never switched on, not broken [MEASURED 2026-09-04]

Supersedes the earlier "peer-verification is dead" entry. It is not dead; its consequence path has never executed once, because the only code that computes consensus sits behind a flag that defaults OFF.

Two zeroes are the whole finding, and a zero is a finding rather than an inventory — so these are stated flat, and every other figure below is a query instead of a number (this repo is PUBLIC; live table counts do not belong in it, and a count published here is stale the week after anyway):

peer_verify_queue rows ever reaching `panel_resolved`      0
RepID score events ever attributable to peer-verify        0

Against those zeroes: tens of thousands of votes were cast, and six figures of verifier tasks were dispatched and fully drained — nothing is pending. Work was done. None of it was ever tallied. Re-derive the exact shape yourself with scripts/sql/peer-verify-audit.sql rather than quoting a number from this file.

PEER_VERIFY_PANEL_ENABLED defaults 'false' (src/services/peer-verify-consensus.ts). The panel path is the ONLY writer of panel_resolved and the only caller of the 2-of-3 tally, so every one of those votes went into a system with no enabled tally. That is why chronic-flag-accumulator — which routes gaming patterns here on the promise that "a confirmed peer-verify FAILURE carries the real RepID consequence" — has never delivered a consequence: the promise was never wired to a live path.

Three switches, stacked, all closed:

PEER_VERIFY_PANEL_ENABLED defaults false — no consensus, no resolution, no scoring PRODUCER_HALT_CLASSES contains peer_verify? producer stopped emitting 2026-07-21 (UNVERIFIED — Railway env is not readable from a cloud run) HAL_CHRONIC_FLAG_ENABLED defaults false, and its threshold is a placeholder the code says to calibrate against GA-C's flagged-precision number first

When you do re-run the audit, read the consensus count carefully. Far fewer stranded rows reached genuine consensus than reached two agreeing votes, because computeConsensus returns all_timeout whenever decisive.length === 0 — it tests that BEFORE quorum is ever considered. Two agreeing timeout votes are not a verdict, and a query that counts "two matching votes" overstates the backlog several-fold. That mistake was made here first.

FOUR HYPOTHESES TESTED AND KILLED, so nobody re-runs them. Each looked right:

  1. reader polls a status nobody writes — NO, both writers insert 'pending'
  2. NULL source_response_id from the chronic-flag insert — NO, every row has one
  3. queue_id vs source_response_id key mismatch in the tally — NO, 0 mismatches, 1:1
  4. a large block stranded at consensus — NO; see the all_timeout rule above

Turning the panel on starts scoring real agents on 2-of-3 consensus. That is a decision with a number attached, and the number (the calibrated threshold) does not exist yet. Sean's, not an agent's.

Flag observability: 106 of 111 behaviour gates cannot be seen from outside [MEASURED 2026-09-04]

Generalises the quality-hook defect. That flag gated a live scoring path, defaulted off, and its state was unreadable outside the Railway dashboard — so the question "is it on?" got guessed instead of answered, three times in this project's history. GET /health now reports it (service_quality_hook), on the same contract operator_pager keeps: whether vars are SET and what they resolved to structurally, never a value.

The audit that followed: 426 distinct env names in src/, 111 of them real behaviour gates (a mode switch or early return, not a URL or a tuning constant). Five are observable. Four keyless surfaces report any env state at all.

More interesting than the count: a number of these default ON, which is the more dangerous shape. A default-off gate that nobody knows about is dormant; a default-on gate that nobody knows about is live behaviour nobody chose. Several score-affecting and money-affecting gates are in that group.

DO NOT just add them all to /health. /health is PUBLIC and unauthenticated (publicPaths in src/middleware/auth.ts). Publishing which money controls, circuit breakers and auth requirements are currently off hands an attacker the map. The quality-hook entry is safe there because it gates a scoring shadow observation — no access, no funds. Anything gating money, chain writes or a breaker belongs on an authenticated surface. That split is a decision, not a cleanup.

One caveat any such reporter must not miss. Several HAL gates resolve DB → env → default (src/hal/config.ts), reading repid_config. A reporter that reads process.env would publish the env value while the gate used the DB row — a reporter that disagrees with the thing it reports, which is worse than reporting nothing. getHalConfig() already computes a per-key source; publish the resolved value WITH its source.

Two claims from that audit were REFUTED on re-check, and the shape is worth keeping. OBSERVABILITY_REQUIRE_AUTH and RESILIENCE_REQUIRE_AUTH both read === 'true' ? authMiddleware : passthrough, which looks like "defaults to no auth". It is not: both routers mount AFTER the global app.use(authMiddleware), and neither path is in publicPaths. They are a redundant second layer that is off by default, not an open door. Reading a router-local ternary without checking mount order gets this exactly backwards — verify the mount, not the ternary.

MOCK_FACILITATOR did survive re-check: unset takes the REAL settlement path (the code comment says so) and builds a live wallet. It is guarded by the same amount ceiling and circuit breaker as the mock branch, so this is documented behaviour rather than a trap — but it is three-state, and "unset" is not "false".

Designed, not built

The E2E transaction (x402 + ERC-8004 + RepID + HAL in one contract), a chess match as an honesty-under-stakes test, and a posed-problem event class that pays only when someone else resolves the question. Reasoning for all three is in the session transcript, not here.

DONE (this run)

Surface Evidence
Tier-0 harness built + verified live + packaged as MVP PR #364; reports/2026-08-07/TRUST_HARNESS_MVP_VERIFIED.md
A — 3-file source of truth this branch feat/cc-2026-08-07-orchestration-sot