Skip to content

Version Packages#29

Merged
steflsd merged 1 commit into
mainfrom
changeset-release/main
Jul 17, 2026
Merged

Version Packages#29
steflsd merged 1 commit into
mainfrom
changeset-release/main

Conversation

@steflsd

@steflsd steflsd commented Jul 9, 2026 •
edited by github-actions Bot
Loading

Copy link
Copy Markdown
Contributor

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@serovaai/ficta@0.2.0

Minor Changes

  • #60 083e548 Thanks @steflsd! - Enable context-bound entity-family surrogates for structured entities in trusted keyed request bodies and document their metadata boundary.

  • 2dbfc1a Thanks @steflsd! - Add live managed-registry reloads and Markdown-aware PII case coverage.

  • #40 726f504 Thanks @steflsd! - Enforce strict managed-registry contracts and separate engine, proxy, protocol, and Gateway responsibilities.

  • #58 b6c2cdf Thanks @steflsd! - Link unique high-confidence organization aliases to registered entity anchors and report values-free ambiguity counts in protection stats and egress proofs.

  • 7c8272c Thanks @steflsd! - Standardize validation commands on check, replacing the verify and verify:* scripts.

  • #50 60cc3a5 Thanks @steflsd! - Move legal identity recognition into the Presidio sidecar, preserve structured regex protection, and add values-free trace diagnostics.

  • #59 9031dad Thanks @steflsd! - Add gated context-bound entity-family surrogate rendering and exact restoration across buffered, streaming, and tool-call transports.

  • #56 e465e5a Thanks @steflsd! - Introduce the entity-aware managed registry v1 contract with explicit form boundaries, strict whole-registry validation, and restart-safe live reloads.

  • #66 efe1203 Thanks @steflsd! - Observe residual surrogate tokens that survive restore. A surrogate-shaped token with no dictionary mapping — mutated, truncated, or invented by the model (e.g. a wildcard entity-family reference like FICTA_ORG_<entityTag>_*) — is now counted per response and surfaced as a values-free total in the proxy log (⚠️ N unrestored surrogate token(s)), protection-stats.json, and the stats summary. Detection covers opaque, typed, and entity-family token shapes plus entity-family prefix fragments, across buffered, streamed, and SSE restore paths. Observe-only: response bytes are unchanged, and restore remains exact-match — unknown tokens are never fuzzily recovered.

  • #79 b96f1b0 Thanks @steflsd! - Country-scoped Presidio registry. The shipped recognizer config (renamed
    presidio/default_recognizers.za.yaml → presidio/default_recognizers.yaml) is now fully
    country-tagged, and the sidecar's FICTA_PRESIDIO_SUPPORTED_COUNTRIES env var (default za,us,mu,
    the SA-legal reference profile) decides at load time which country-specific recognizers run;
    locale-agnostic recognizers always load. Notably, the UK NHS recognizer no longer runs on default
    traffic — add uk to the country scope for UK-matter deployments. FICTA_PII_PRESIDIO_ENTITIES
    is now a pure optional narrowing allowlist: when unset, /analyze requests omit the entities
    field and the deployment's loaded registry is the detection surface. The proxy also now strips
    every inbound x-ficta-* header before forwarding upstream, instead of enumerating known ones.

  • #42 35b2e33 Thanks @steflsd! - Add a runtime-admin grant for per-chat raw trace capture and decouple capture from log verbosity.

Patch Changes

  • b19cf46 Thanks @steflsd! - Align source-checkout Presidio startup, logging-root configuration, and documented runtime defaults.

  • #41 1bd5178 Thanks @steflsd! - Allow attachment-only Gateway drafts to enter protection review or send with the existing generic review instruction.

  • 2f3bf7f Thanks @steflsd! - Carry transient UTF-16 source spans through the PII detector contract for offset-based redaction.

  • 737e967 Thanks @steflsd! - Consolidate CI, secret scanning, and release automation into one gated workflow run.

  • 0b26f15 Thanks @steflsd! - Detect organization names via Presidio NER. Ship an NLP-engine config (presidio/nlp_engine.za.yaml, mounted as NLP_CONF_FILE) that un-suppresses ORGANIZATION — upstream ignores it as "many false positives" — so unregistered client/counterparty/company names get a best-effort catch from spaCy en_core_web_lg. This is probabilistic and over-redacts (headings, common nouns); exact confidentiality still comes from the registered-value registry.

  • af9dea2 Thanks @steflsd! - Document routing ficta claude through a local Anthropic-compatible proxy via FICTA_ANTHROPIC_UPSTREAM, so an alt model (e.g. GPT‑5.6 "sol" behind CLIProxyAPI on a ChatGPT subscription) can run with redaction intact. Loopback upstreams need no FICTA_ALLOW_CUSTOM_UPSTREAM.

  • #48 6dd6b82 Thanks @steflsd! - Explain restored-value underlines with accessible origin-specific tooltips.

  • e2eb4a4 Thanks @steflsd! - Redact embedded case variants of registered and detected entities and add explicit boundary policy for future aliases.

  • bd8448e Thanks @steflsd! - Improve Presidio organization coverage for accounting tables and same-document business variants while filtering structural NER false positives.

  • f3cbc49 Thanks @steflsd! - Live protected registry: values published from the gateway admin UI take effect in the running proxy without a restart. New POST /__ficta/registry/reload (loopback-gated, request body ignored, counts-only response including skippedTooShort for values below FICTA_REGISTRY_MIN_LEN), ProtectionEngine.reloadRegistryValues() registering new managed-file values into the live vault, and a stat-based cache key for the managed-registry plugin (a rewritten file is actually re-read — also fixes stale registry counts in per-request log metadata and ficta doctor). Additions are live; deletions still apply on restart (removing a value mid-process would break restore of surrogates already in transcripts). Protocol gains FICTA_REGISTRY_RELOAD_PATH, RegistryReloadOk/Error, and isRegistryReloadOk.

  • ee44595 Thanks @steflsd! - Map compact Markdown detection offsets back to raw request ranges so internally formatted entities are fully redacted.

  • f3cbc49 Thanks @steflsd! - Close two systemic PII-miss classes observed on markitdown-converted documents:

    • Markdown-aware NER input: NLP backends (Presidio/OpenMed) now detect over compact Markdown-normalized text with exact raw-offset mapping (**bold**, \_ escapes, ### headings, list markers, ~~strike~~), fixing contaminated spans, internal-formatting gaps, and recall lost to formatting; regex recognizers keep the raw text.
    • Case-variant coverage: an entity detected in one casing is redacted in every casing actually present in the request (title-case prose vs ALL-CAPS headings/signature blocks), for detected PII and word/name-like registry values alike (digit-bearing secrets are never case-folded). Registry-derived variants carry permanent provenance, so the detected restore-into-tools policy withholds a secret's case twin from tool-call arguments exactly like the canonical form.

  • #70 64335d9 Thanks @steflsd! - Neutralize caller-provenance labels: explicit pre-send selections are now recorded with
    source: "user-selected" and plugin: "protection-preview" (previously gateway-user /
    gateway-preview). The strings appear in protection-preview findings, traces, and stats labels;
    no behavior depends on them.

  • #34 50a49f7 Thanks @steflsd! - Add a gateway trace-capture header so raw proxy trace/audit capture can be scoped per chat thread.

  • 2d55f15 Thanks @steflsd! - Add a loopback-only pre-send protection preview with resolver-authored spans and content-bound, single-use tickets for user-selected chat protections.

  • 0b26f15 Thanks @steflsd! - Add an opt-in FICTA_PRESERVE_LITERALS mode that injects a system/developer instruction carrying the exact surrogate tokens present in each outbound request, telling the model to reproduce them verbatim. This improves restore reliability: models otherwise truncate or editorialise long opaque tokens (FICTA_62a02923…), which leaves them unrestorable. The instruction only ever adds surrogate tokens the proxy already minted (never raw values) and runs after the fail-closed leak gate.

  • f195c54 Thanks @steflsd! - Record fail-closed detector outages across body, query, and header surfaces as explicit values-free blocked-request proof events.

  • aacf45d Thanks @steflsd! - Keep published documentation focused on shipped behavior and public threat-model boundaries, with private strategy and future-roadmap material maintained separately.

  • #39 b60f2bf Thanks @steflsd! - Move request captures into lazy runs/run-* directories and write current protection stats to protection-stats.json.

  • 89da819 Thanks @steflsd! - Enforce required-registry readiness on standalone provider traffic and expose the blocking state to Gateway.

  • #74 f640551 Thanks @steflsd! - Skip Doppler registry API and secret-loading commands when the launch directory has no resolvable Doppler scope.

  • f3cbc49 Thanks @steflsd! - Audit attribution now prefers the registry's identity when a value is both a registered secret and a probabilistic detection: the trace/stats report env-file / secret / exact (e.g. CLIENT_CFO) instead of the detector's guess (person / pii / high). Reporting only — span selection, surrogates, restore, and restore-into-tools provenance are unchanged.

  • 7202694 Thanks @steflsd! - Resolve overlapping body entities by source spans so registered values win exact boundaries without partial leaks.

  • f3cbc49 Thanks @steflsd! - Scope trace-mode restore-highlight markers to assistant output. The highlight triple (the gateway's show/hide toggle format) now rides only on streamed text fragments and their sibling fields; metadata/replay events that echo the request back (response.created / response.in_progress instructions) restore plainly — surrogates still never reach the client, but the echoed preamble is no longer littered with marker sentinels.

  • 7625390 Thanks @steflsd! - Restore repeated document-linked short organization aliases in Presidio identity detection.

  • #47 2f799b0 Thanks @steflsd! - Highlight locally restored response values without requiring sensitive trace capture.

  • e79624e Thanks @steflsd! - Separate capture logs by proxy role so concurrent instances no longer share one directory. The
    standalone/web server now writes under ~/.ficta/logs/gateway/, and each ficta <agent> shim under
    ~/.ficta/logs/agents/<agent>/<instance>/ (one subtree per process, so two ficta claude sessions
    never interleave runs/ or race protection-stats.json). Set FICTA_LOG_ROOT to relocate the root;
    FICTA_LOG_DIR still fully overrides the exact path. Existing config.toml files whose
    [logging].log_dir equals the default root are treated as neutral so the split applies.

  • 1ae6497 Thanks @steflsd! - Preserve contained PII candidates for occurrence resolution and add legal-document ID, Mauritius phone, and ordinal-date Presidio coverage.

  • 8c09e18 Thanks @steflsd! - Add values-free, per-thread provider-egress evidence for Gateway chats.

  • 89da819 Thanks @steflsd! - Simplify the documented POC configuration and have setup emit only the canonical multi-backend PII setting.

  • 2dbfc1a Thanks @steflsd! - Harden live Protected Registry publication: Gateway now writes private registry files atomically, serializes publish transactions, and verifies a per-generation revision echoed by the proxy together with managed-source health counts. Hosted Gateway deployments are explicitly bound to one WorkOS organization via FICTA_GATEWAY_ORG_ID, matching the proxy's process-global permanent registry.

  • Updated dependencies [083e548, 1bd5178, 2dbfc1a, 726f504, 737e967, af9dea2, b6c2cdf, f3cbc49, 50a49f7, 2d55f15, f195c54, aacf45d, 7c8272c, 89da819, 9031dad, e465e5a, efe1203, 2f799b0, b96f1b0, 8c09e18, 35b2e33, 89da819, 2dbfc1a]:

    • @serovaai/ficta-protocol@0.2.0

@serovaai/ficta-protocol@0.2.0

Minor Changes

  • #60 083e548 Thanks @steflsd! - Enable context-bound entity-family surrogates for structured entities in trusted keyed request bodies and document their metadata boundary.

  • 2dbfc1a Thanks @steflsd! - Add live managed-registry reloads and Markdown-aware PII case coverage.

  • #40 726f504 Thanks @steflsd! - Enforce strict managed-registry contracts and separate engine, proxy, protocol, and Gateway responsibilities.

  • #58 b6c2cdf Thanks @steflsd! - Link unique high-confidence organization aliases to registered entity anchors and report values-free ambiguity counts in protection stats and egress proofs.

  • 7c8272c Thanks @steflsd! - Standardize validation commands on check, replacing the verify and verify:* scripts.

  • #59 9031dad Thanks @steflsd! - Add gated context-bound entity-family surrogate rendering and exact restoration across buffered, streaming, and tool-call transports.

  • #56 e465e5a Thanks @steflsd! - Introduce the entity-aware managed registry v1 contract with explicit form boundaries, strict whole-registry validation, and restart-safe live reloads.

  • #79 b96f1b0 Thanks @steflsd! - Country-scoped Presidio registry. The shipped recognizer config (renamed
    presidio/default_recognizers.za.yaml → presidio/default_recognizers.yaml) is now fully
    country-tagged, and the sidecar's FICTA_PRESIDIO_SUPPORTED_COUNTRIES env var (default za,us,mu,
    the SA-legal reference profile) decides at load time which country-specific recognizers run;
    locale-agnostic recognizers always load. Notably, the UK NHS recognizer no longer runs on default
    traffic — add uk to the country scope for UK-matter deployments. FICTA_PII_PRESIDIO_ENTITIES
    is now a pure optional narrowing allowlist: when unset, /analyze requests omit the entities
    field and the deployment's loaded registry is the detection surface. The proxy also now strips
    every inbound x-ficta-* header before forwarding upstream, instead of enumerating known ones.

  • #42 35b2e33 Thanks @steflsd! - Add a runtime-admin grant for per-chat raw trace capture and decouple capture from log verbosity.

Patch Changes

  • #41 1bd5178 Thanks @steflsd! - Allow attachment-only Gateway drafts to enter protection review or send with the existing generic review instruction.

  • 737e967 Thanks @steflsd! - Consolidate CI, secret scanning, and release automation into one gated workflow run.

  • af9dea2 Thanks @steflsd! - Document routing ficta claude through a local Anthropic-compatible proxy via FICTA_ANTHROPIC_UPSTREAM, so an alt model (e.g. GPT‑5.6 "sol" behind CLIProxyAPI on a ChatGPT subscription) can run with redaction intact. Loopback upstreams need no FICTA_ALLOW_CUSTOM_UPSTREAM.

  • f3cbc49 Thanks @steflsd! - Live protected registry: values published from the gateway admin UI take effect in the running proxy without a restart. New POST /__ficta/registry/reload (loopback-gated, request body ignored, counts-only response including skippedTooShort for values below FICTA_REGISTRY_MIN_LEN), ProtectionEngine.reloadRegistryValues() registering new managed-file values into the live vault, and a stat-based cache key for the managed-registry plugin (a rewritten file is actually re-read — also fixes stale registry counts in per-request log metadata and ficta doctor). Additions are live; deletions still apply on restart (removing a value mid-process would break restore of surrogates already in transcripts). Protocol gains FICTA_REGISTRY_RELOAD_PATH, RegistryReloadOk/Error, and isRegistryReloadOk.

  • #34 50a49f7 Thanks @steflsd! - Add a gateway trace-capture header so raw proxy trace/audit capture can be scoped per chat thread.

  • 2d55f15 Thanks @steflsd! - Add a loopback-only pre-send protection preview with resolver-authored spans and content-bound, single-use tickets for user-selected chat protections.

  • f195c54 Thanks @steflsd! - Record fail-closed detector outages across body, query, and header surfaces as explicit values-free blocked-request proof events.

  • aacf45d Thanks @steflsd! - Keep published documentation focused on shipped behavior and public threat-model boundaries, with private strategy and future-roadmap material maintained separately.

  • 89da819 Thanks @steflsd! - Enforce required-registry readiness on standalone provider traffic and expose the blocking state to Gateway.

  • #66 efe1203 Thanks @steflsd! - Observe residual surrogate tokens that survive restore. A surrogate-shaped token with no dictionary mapping — mutated, truncated, or invented by the model (e.g. a wildcard entity-family reference like FICTA_ORG_<entityTag>_*) — is now counted per response and surfaced as a values-free total in the proxy log (⚠️ N unrestored surrogate token(s)), protection-stats.json, and the stats summary. Detection covers opaque, typed, and entity-family token shapes plus entity-family prefix fragments, across buffered, streamed, and SSE restore paths. Observe-only: response bytes are unchanged, and restore remains exact-match — unknown tokens are never fuzzily recovered.

  • #47 2f799b0 Thanks @steflsd! - Highlight locally restored response values without requiring sensitive trace capture.

  • 8c09e18 Thanks @steflsd! - Add values-free, per-thread provider-egress evidence for Gateway chats.

  • 89da819 Thanks @steflsd! - Simplify the documented POC configuration and have setup emit only the canonical multi-backend PII setting.

  • 2dbfc1a Thanks @steflsd! - Harden live Protected Registry publication: Gateway now writes private registry files atomically, serializes publish transactions, and verifies a per-generation revision echoed by the proxy together with managed-source health counts. Hosted Gateway deployments are explicitly bound to one WorkOS organization via FICTA_GATEWAY_ORG_ID, matching the proxy's process-global permanent registry.

@steflsd
steflsd force-pushed the changeset-release/main branch 8 times, most recently from af59189 to a1a2577 Compare July 9, 2026 08:59
@steflsd
steflsd force-pushed the changeset-release/main branch 21 times, most recently from ee22e8c to 3eb1a8c Compare July 11, 2026 10:43
coderabbitai[bot]
coderabbitai Bot requested changes Jul 14, 2026
View reviewed changes

@coderabbitai coderabbitai Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@packages/ficta/CHANGELOG.md`:
- Around line 79-81: Remove the duplicate 89da819 and 2dbfc1a commit links from
the Updated dependencies entry, keeping each dependency reference listed exactly
once before publishing.
- Around line 39-42: Update the changelog entry heading text to say “two
systemic classes of PII misses” instead of “two systemic PII-miss classes,”
leaving the surrounding bullets and formatting unchanged.
🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

  • Push a commit to this branch (recommended)
  • Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: ddd7a3e9-892b-44d1-b9f3-99ebf98088cb

📥 Commits

Reviewing files that changed from the base of the PR and between 5282173 and 2da68d3.

📒 Files selected for processing (33)
  • .changeset/align-presidio-demo-config.md
  • .changeset/allow-attachment-only-review.md
  • .changeset/brave-registry-reload.md
  • .changeset/carry-detector-spans.md
  • .changeset/clarify-engine-protocol-boundaries.md
  • .changeset/consolidate-release-workflow.md
  • .changeset/detect-organization-ner.md
  • .changeset/document-anthropic-upstream-proxy.md
  • .changeset/explain-restored-underlines.md
  • .changeset/harden-entity-form-boundaries.md
  • .changeset/improve-accounting-organization-detection.md
  • .changeset/live-registry-reload.md
  • .changeset/map-markdown-offsets.md
  • .changeset/markdown-case-pii-coverage.md
  • .changeset/per-thread-trace-capture.md
  • .changeset/pre-send-protection-review.md
  • .changeset/preserve-surrogate-literals.md
  • .changeset/prove-detector-outages.md
  • .changeset/quiet-log-captures.md
  • .changeset/quiet-pandas-protect.md
  • .changeset/registry-authoritative-attribution.md
  • .changeset/resolve-entity-occurrences.md
  • .changeset/restore-highlight-scope.md
  • .changeset/show-restored-values.md
  • .changeset/split-role-log-dirs.md
  • .changeset/strengthen-legal-pii-detection.md
  • .changeset/thread-egress-evidence.md
  • .changeset/tidy-otters-trace.md
  • .changeset/verify-registry-publication.md
  • packages/ficta/CHANGELOG.md
  • packages/ficta/package.json
  • packages/protocol/CHANGELOG.md
  • packages/protocol/package.json
💤 Files with no reviewable changes (29)
  • .changeset/split-role-log-dirs.md
  • .changeset/consolidate-release-workflow.md
  • .changeset/carry-detector-spans.md
  • .changeset/resolve-entity-occurrences.md
  • .changeset/tidy-otters-trace.md
  • .changeset/allow-attachment-only-review.md
  • .changeset/map-markdown-offsets.md
  • .changeset/preserve-surrogate-literals.md
  • .changeset/markdown-case-pii-coverage.md
  • .changeset/live-registry-reload.md
  • .changeset/restore-highlight-scope.md
  • .changeset/per-thread-trace-capture.md
  • .changeset/align-presidio-demo-config.md
  • .changeset/show-restored-values.md
  • .changeset/quiet-pandas-protect.md
  • .changeset/document-anthropic-upstream-proxy.md
  • .changeset/prove-detector-outages.md
  • .changeset/detect-organization-ner.md
  • .changeset/improve-accounting-organization-detection.md
  • .changeset/strengthen-legal-pii-detection.md
  • .changeset/clarify-engine-protocol-boundaries.md
  • .changeset/pre-send-protection-review.md
  • .changeset/quiet-log-captures.md
  • .changeset/verify-registry-publication.md
  • .changeset/explain-restored-underlines.md
  • .changeset/registry-authoritative-attribution.md
  • .changeset/brave-registry-reload.md
  • .changeset/thread-egress-evidence.md
  • .changeset/harden-entity-form-boundaries.md
🚧 Files skipped from review as they are similar to previous changes (3)
  • packages/ficta/package.json
  • packages/protocol/package.json
  • packages/protocol/CHANGELOG.md

Comment thread packages/ficta/CHANGELOG.md
Comment on lines +39 to +42
- [`f3cbc49`](https://github.com/SerovaAI/ficta/commit/f3cbc4949d9bc28f622a187e5c251d2b1a1f8cdd) Thanks [@steflsd](https://github.com/steflsd)! - Close two systemic PII-miss classes observed on markitdown-converted documents:

- **Markdown-aware NER input**: NLP backends (Presidio/OpenMed) now detect over compact Markdown-normalized text with exact raw-offset mapping (`**bold**`, `\_` escapes, `###` headings, list markers, `~~strike~~`), fixing contaminated spans, internal-formatting gaps, and recall lost to formatting; regex recognizers keep the raw text.
- **Case-variant coverage**: an entity detected in one casing is redacted in every casing actually present in the request (title-case prose vs ALL-CAPS headings/signature blocks), for detected PII and word/name-like registry values alike (digit-bearing secrets are never case-folded). Registry-derived variants carry `permanent` provenance, so the `detected` restore-into-tools policy withholds a secret's case twin from tool-call arguments exactly like the canonical form.

@coderabbitai coderabbitai Bot Jul 14, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Fix the “PII-miss” wording.

Use “two systemic classes of PII misses” instead; the current phrasing is awkward and triggers the spelling check in this user-facing changelog.

Suggested wording 
- Close two systemic PII-miss classes observed on markitdown-converted documents:
+ Close two systemic classes of PII misses observed on markitdown-converted documents:
📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change
- [`f3cbc49`](https://github.com/SerovaAI/ficta/commit/f3cbc4949d9bc28f622a187e5c251d2b1a1f8cdd) Thanks [@steflsd](https://github.com/steflsd)! - Close two systemic PII-miss classes observed on markitdown-converted documents:
- **Markdown-aware NER input**: NLP backends (Presidio/OpenMed) now detect over compact Markdown-normalized text with exact raw-offset mapping (`**bold**`, `\_` escapes, `###` headings, list markers, `~~strike~~`), fixing contaminated spans, internal-formatting gaps, and recall lost to formatting; regex recognizers keep the raw text.
- **Case-variant coverage**: an entity detected in one casing is redacted in every casing actually present in the request (title-case prose vs ALL-CAPS headings/signature blocks), for detected PII and word/name-like registry values alike (digit-bearing secrets are never case-folded). Registry-derived variants carry `permanent` provenance, so the `detected` restore-into-tools policy withholds a secret's case twin from tool-call arguments exactly like the canonical form.
- [`f3cbc49`](https://github.com/SerovaAI/ficta/commit/f3cbc4949d9bc28f622a187e5c251d2b1a1f8cdd) Thanks [`@steflsd`](https://github.com/steflsd)! - Close two systemic classes of PII misses observed on markitdown-converted documents:
- **Markdown-aware NER input**: NLP backends (Presidio/OpenMed) now detect over compact Markdown-normalized text with exact raw-offset mapping (`**bold**`, `\_` escapes, `###` headings, list markers, `~~strike~~`), fixing contaminated spans, internal-formatting gaps, and recall lost to formatting; regex recognizers keep the raw text.
- **Case-variant coverage**: an entity detected in one casing is redacted in every casing actually present in the request (title-case prose vs ALL-CAPS headings/signature blocks), for detected PII and word/name-like registry values alike (digit-bearing secrets are never case-folded). Registry-derived variants carry `permanent` provenance, so the `detected` restore-into-tools policy withholds a secret's case twin from tool-call arguments exactly like the canonical form.
🧰 Tools
🪛 LanguageTool

[grammar] ~39-~39: Ensure spelling is correct
Context: ...o systemic PII-miss classes observed on markitdown-converted documents: - **Markdown-aw...

(QB_NEW_EN_ORTHOGRAPHY_ERROR_IDS_1)

🤖 Prompt for AI Agents 
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@packages/ficta/CHANGELOG.md` around lines 39 - 42, Update the changelog entry
heading text to say “two systemic classes of PII misses” instead of “two
systemic PII-miss classes,” leaving the surrounding bullets and formatting
unchanged.

Source: Linters/SAST tools

Comment thread packages/ficta/CHANGELOG.md Outdated
@github-actions
github-actions Bot force-pushed the changeset-release/main branch 25 times, most recently from 1080f68 to 616d16e Compare July 15, 2026 11:01
@github-actions
github-actions Bot force-pushed the changeset-release/main branch from 616d16e to f117c42 Compare July 17, 2026 06:39
@steflsd
steflsd merged commit a4e7a94 into main Jul 17, 2026
4 checks passed
@steflsd
steflsd deleted the changeset-release/main branch July 17, 2026 06:43
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@coderabbitai coderabbitai[bot] coderabbitai[bot] requested changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@steflsd