Z.1: KB conflict counts were reported as ROW counts (kb_knowledge rows), not DISTINCT GTIN counts - a single conflicted product with 2 competing NCM candidates produces 2 rows, doubling the apparent conflict count. Real number, measured directly in production: 262 distinct GTINs in CONFLICT (all under knowledge_key=ncm), matching the ~264 baseline closely - no real growth, just a row-vs-GTIN counting mismatch in the earlier report. Z.3: corrected the Missão 4 claim that NFCE_ITENS was the "fonte encontrada" for what PIS/COFINS was actually transmitted - that overstated certainty. Relabeled PISCOFINS_FONTE_TRANSMITIDA = NÃO_CONCLUSIVO in docs/PISCOFINS_TRANSMITTED_SOURCE_V1.md and both dossiê copies: a blank NFCE_ITENS column is equally consistent with "nothing was transmitted" and "the SGV transmitted a value in the XML but never wrote it back to that column" - we cannot tell without the actual XML. Z.4: docs/AUTOFILL_SAFE_3_CELULAS_V1.md - table of the 3 real AUTO_FILL_SAFE cells (000801/000802 CFOP, 000697 IPPT) with proposed value, field_contract rule and concrete evidence for each. Z.5: docs/NCM_27_BLOCKED_PRODUCTS_V1.md now marks the 15 test/garbage-named products as "produto de teste" in the kb.attest queue (still counted in coverage, just lower priority); the 3 legacy + 5 out-of-table NCMs are flagged for reclassification once the official NCM catalog (Fase 2 of this mission) is loaded and approved. Z.6: CLAUDE.md records independent Hub/Agent versioning (docs/VERSIONING.md) as the adopted standard going forward. No schema/migration change. No Firebird write anywhere. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
6.6 KiB
GTI Discovery
This project is strictly read-only. Discovery code must use a database-level read-only account or transaction/driver mode; lexical SQL validation is only defense-in-depth. Discovery code must not perform DML or DDL, modify database objects or permissions, or store credentials and other secrets.
Run the offline tests with:
python -m pytest tests -q
Firebird write policy
Default policy:
- All client/production Firebird databases are STRICTLY READ-ONLY.
- Never perform INSERT, UPDATE, DELETE, DDL, permission changes, or destructive operations against any client/production ERP database unless explicitly authorized in a future policy change.
Explicit development exception:
The following database is a disposable TEST DATABASE and is explicitly authorized for writes:
D:\PROJETOS\GTISISTEMAS\Projeto\TESTESOFICIAIS.FDB
For this exact file only, the agent MAY perform:
- SELECT
- INSERT
- UPDATE
- DELETE
- DDL
- transactional tests
- rollback/recovery tests
- destructive development experiments when necessary
This database has external intact backups and can be restored.
Safety requirement:
Before ANY Firebird write, the code must verify that the canonical absolute database path is exactly:
D:\PROJETOS\GTISISTEMAS\Projeto\TESTESOFICIAIS.FDB
If the path differs in any way, writes must fail closed.
Never infer write authorization from filename alone. Never enable writes to other Firebird databases.
Secrets:
- credentials remain local in ignored environment files;
- never log, commit, or expose passwords.
Development priority:
- TESTESOFICIAIS.FDB may be modified freely as needed to validate the MVP.
- Do not over-engineer rollback protections for this disposable test database unless they are part of the product behavior being validated.
GTI Fiscal write-target rule
The SGV keeps its fiscal defaults in the canonical product record (C000025); every other routine consumes them.
GTI Fiscal therefore writes ONLY to the canonical product record, and only fields whose mapping is proven
(docs/SGV_PRODUCT_FISCAL_UI_MAP_V1.json, write_target=true). It never writes to NFCE, NFCE_ITENS, item tables,
issued documents or any other transactional/historical table; those are read-only evidence for validation.
Company context (UF, CRT) is read from C000004 at run time — never hardcoded.
Local completion mission — exact write allowlist (operator-authorized)
For the "GTI FISCAL LOCAL COMPLETION V1" mission the operator explicitly authorized writes to the disposable COPIES listed, by exact canonical path, in
docs/LOCAL_WRITE_ALLOWLIST_V1.json (CONSTRUCAO, MMMODAS, RINCAO, UNIAO under Projeto\Banco\, plus TESTESOFICIAIS.FDB). Authorization is by exact
path only — never by directory or .FDB extension. Any other Firebird under D:\PROJETOS\GTISISTEMAS stays strictly READ-ONLY. Writes go through the
guarded executor (path allowlist, company-context guard, schema/type/length guard, compare-before-write, rowcount, transaction, audit) and only to
C000025 fields whose contract in docs/FISCAL_FIELD_CONTRACT_V1.json has WRITE_ALLOWED=true. This authorization does not extend to production or client databases.
Production deploy rule (Hub)
Normal deploy = powershell -ExecutionPolicy Bypass -File .\scripts\deploy_hub.ps1 [-PublishAgent -Version X] (docs/DEPLOY_HUB.md). Do not recreate ad-hoc deploy scripts (deploy115.py etc.) and do not improvise SSH/docker against 10.0.10.117. The script handles VPN/SSH retry, backup, upload, migrations, rebuild, health, release publish, download verification, and automatic rollback; the real error is in logs/last_deploy_result.json and the log.
Versioning (adopted standard, see docs/VERSIONING.md)
Hub and Agent are versioned and released independently — never assume they share a version
number, and never make the Hub's GTI_FISCAL_VERSION bump touch the Agent's own version files
(agent/pyproject.toml, the AGENT_VERSION defaults) or vice versa. The panel's "Atualização
disponível" already compares an Agent's reported version against the highest PUBLISHED
Windows AgentRelease, never against the Hub's version — see docs/VERSIONING.md for the exact
mechanism and the current (informal) compatibility matrix.
Permanent mission rules (fiscal completion missions, recorded 2026-09-28)
- No AI/LLM/embeddings, no paid API, no scraping — every rule/decision is derived from real ERP
evidence or an explicit, audited operator/ADMIN decision (
FiscalContextEntry), never inferred or guessed. - No Firebird triggers or schema changes, ever, for any discovery/completion mission. All queries against any Firebird database (client, lab copy, or disposable test DB) are read-only unless the exact path is on an explicit, mission-scoped write allowlist (see above) — writes still go only through the guarded executor.
- Deploy only via
scripts\deploy_hub.ps1(see "Production deploy rule" above) — never ad-hoc SSH/docker/psql against 10.0.10.117. - If the Agent (agent/) changes in a way that affects its runtime behavior: bump the single-source
version, build a new installer, validate it in a fresh Windows Sandbox (install, upgrade in-place,
identity preserved), and never overwrite an already-published release. A pure version-label sync
bump (release/VERSION, pyproject.toml files, the
AGENT_VERSIONdefault string) that changes no Agent behavior is not itself an "Agent change" requiring a new installer. - Never reenroll an existing installation (never mint a new
installation_id/token for a machine that already has one) — an in-place upgrade always preserves identity and local SQLite state. - No remote-execution channel to any client's Agent host, ever (by design — see
docs/rtc_auto_apply.md). An Agent version upgrade on a real client machine is always performed by a human operator on-site; the assistant never attempts RDP/WinRM/SSH/PowerShell-remoting into a client host, and never asks for or accepts a productionAGENT_TOKEN/admin session credential pasted in chat to work around that — live validation of a real installation, when needed, comes from the operator (a panel screenshot, or a value they read and paste themselves). - No real client/production data (CNPJ, razão social, product names/descriptions, per-product NCM)
is ever published in a location a non-admin can reach (e.g.
central/static/). Any evidence artifact referenced from a UI card must either be served through an authenticated ADMIN-only route, or be anonymized (generic client labels, no names/documents/descriptions) before being published as a static file.