feat(hermes): ingest NeMo Relay ATOF streams#3087
Merged
Merged
Conversation
Problem: real Hermes NeMo Relay evidence includes append-only ATOF JSONL that was watched and retained but not admitted or normalized. What changed: structurally detect and group real ATOF records by Hermes session, retain observer-only LLM/tool/approval/error/subagent/context evidence, and use UUID-plus-scope-phase replay protection. Add a privacy-redacted, provenance-checked v0.1 fixture and real-route tests. Compatibility: ATOF remains an observer session correlated by the Hermes session id; it does not merge into the conversational transcript or create session links. Ref polylogue-fs1.2.1 Co-Authored-By: Codex <noreply@openai.com>
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
|
Warning Review limit reached
Next review available in: 46 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: ASSERTIVE Plan: Pro Run ID: 📒 Files selected for processing (9)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Sinity
changed the title
Sinity
changed the title
Sinity
added a commit
that referenced
this pull request
Jul 18, 2026
…rification ledger (#3092) ## Summary Continuation of PR #3087 (merged as fee0799 with only its first commit — this branch grew four more commits afterward that never made it into that squash-merge, since #3087 was merged mid-session while this work was still landing). Admits real ATOF lifecycle events, proves ATIF tool_calls fidelity against real bytes, closes a real state.db v16→v19 schema-drift gap, and imports Hermes's own verification_evidence.db claim-vs-evidence ledger as structured outcomes (Ref polylogue-fs1.2.1, polylogue-wj25). ## Problem - Two real, well-defined ATOF lifecycle event shapes (the outer session-scope boundary `kind=scope,category=agent`, and `hermes.session.end` marks) fell into the generic unrecognized-event bucket despite being present, unused, in the already-committed real ATOF fixture. - The real ATIF fixture used to originally admit fs1.2.1 was a minimal message-only session, so `tool_execution_spans` fidelity stayed "inferred" despite real trajectories with the `tool_calls`/`observation` shape sitting on disk. - The live `~/.hermes/state.db` schema is at `SCHEMA_VERSION=19`, nine versions past the v16 baseline previously verified — nine real columns (gateway-routing identity from Hermes's #9006 chat-platform bridging, compression-retry state) were silently dropped by every import. - `verification_evidence.db` — Hermes's own command/exit-code/status verification ledger — was retained as a watched artifact (#3084) but never parsed, so Hermes sessions carried none of the structured verification evidence the archive's claim-vs-evidence lanes are built around. ## Solution Four commits (`b96cc97d1`, `c8db6706e`, `1d716e985`, `18f326632`): 1. **ATOF lifecycle events** (`hermes_spans.py`): `kind=scope,category=agent` and `hermes.session.*` marks now normalize to typed `hermes_context_span` evidence. `context_events` fidelity moves to exact; `unrecognized_atof_events` drops to zero on the real fixture. 2. **ATIF tool_calls proof**: appended one real, privacy-redacted step (4 parallel `tool_calls` + correlated `observation.results`, drawn from a separate live trajectory) to the checked-in ATIF fixture. `tool_execution_spans` now reports exact fidelity like `llm_request_spans`. Added `tests/fixtures/hermes/atif/README.md` documenting provenance (the ATOF fixture already had one). 3. **state.db v19 drift** (`hermes_state.py`): added `gateway_identity` and `compression_recovery` capability groups, extended `lifecycle` with `expiry_finalized`, wired all fields into the existing `hermes_session_metadata` event. Added a schema-drift canary test pinning the live-verified v19 column set against the parser's capability maps, and fixed a pre-existing stale contract file (`state_db_v16.contract.json`) that testmon hadn't caught. 4. **Verification-ledger import** (new `hermes_verification.py`): parses `verification_events`/`verification_state` into a `verification:<hermes_session_id>` observer session — same deferred-physical-merge pattern as ATOF/ATIF. `command`/`canonical_command`/`output_summary` round-trip verbatim (this is Hermes's own recorded evidence, not conversational content, so ATOF/ATIF's payload-hygiene bound doesn't apply — documented in the module docstring). Wired through every route `state.db` already uses (dispatch, source_walk/source_parsing live ingest, live/batch full-batch ingest, import-explain CLI). `session_id="default"` (the producer's own unknown-identity fallback) is marked `ambiguous_correlation`, never trusted. Artifact-taxonomy split follows state.db's own precedent exactly: raw-path classification stays non-session (protects schema sampling from raw SQLite bytes), marker-payload classification is session-eligible. ## Verification - `devtools test tests/unit/sources/parsers/test_hermes_spans.py` — 17/17. - `devtools test tests/unit/sources/parsers/test_hermes_verification.py` — 9/9. - `devtools test tests/unit/sources/parsers/test_hermes_state_schema_canary.py` — 2/2. - Full hermes-touching regression sweep (dispatch/sampling/watcher/artifact-inspection/import/lineage/import-explain, 236 tests) — all green. - `devtools verify --quick` green after every commit. Known honest gaps, not silently assumed: ATIF `subagent_delegation` stays "inferred" (no real export sampled has a non-empty `subagent_trajectories`); the verification-ledger session id has no profile-qualification (same accepted gap ATOF/ATIF's `observer:` id already has); ATOF partial-final-line/rotation-truncation cursor behavior and unpaired-event debt remain deferred (fs1.2.1's remaining AC). Ref polylogue-fs1.2.1 Ref polylogue-wj25 --------- Co-authored-by: Claude <noreply@anthropic.com>
Sinity
added a commit
that referenced
this pull request
Jul 21, 2026
## Summary Closes the remaining conformance-law gaps in the OriginSpec admission kernel (`polylogue/sources/origin_specs.py`) against every clause of `polylogue-2qx.1.1`'s acceptance criteria: a public-origin/Provider-token leak guard, a stream/parser binding parity check against dispatch, and a live proof of the `lifecycle="reserved"` admission path. ## Problem `polylogue-2qx.1.1` asked for the OriginSpec admission kernel (built on the `polylogue/declarations` DeclarationSpec kernel from `polylogue-o21.1`) plus a conformance law where missing parser, stream/parser conflict, ambiguous detector order, absent fixture, undeclared coverage, leaked Provider token, and non-injective Provider-to-Origin collision each yield a source-locatable diagnostic. Auditing the live repo found the kernel **already exists on master** — `polylogue/sources/origin_specs.py` landed incrementally via unrelated feature PRs (#3051, #3087, #3088, #3092, #3201, #3228: Claude Workflow admission, Hermes ATOF/ATIF, the Grok parser, provider_completeness migration) without ever citing this bead, so `bd`'s tracking never reflected it — the 2026-07-18 warroom sweep found no bead-tagged commits and reset the bead to open even though the code was there. It already covers all 11 `Origin` tokens, consumes the declarations kernel correctly, and is wired into `provider_completeness.py`, the artifact taxonomy, the live watcher, and Claude Code orchestration. Auditing it clause-by-clause against the AC found three real gaps: 1. No diagnostic caught a public origin `public_name` colliding with a raw `Provider`-wire token — the doctrine in `docs/provider-origin-identity.md` ("no provider wording on public origin surfaces") was enforced only indirectly (public_name must equal `origin.value`, and no current `Origin`/`Provider` pair happens to collide), with no explicit guard or diagnostic for a future collision. 2. No check parity-verified declared `stream_parser_path` presence against `sources/dispatch.py:STREAM_RECORD_PROVIDERS` — a stream/parser binding could silently drift from dispatch's actual streaming behavior with no diagnostic. 3. The module docstring still claimed a live "reserved origin" pilot, but Grok shipped a real parser (`polylogue-611`, #3201) after this bead was written, so no `OriginSpec` is currently `lifecycle="reserved"` and that admission path had no live proof. ## Solution - Added `_PROVIDER_TOKEN_VALUES` and a guard in `OriginSpecRegistry.register()` that rejects a declaration whose `public_name` collides with any `Provider` enum value, ordered before the existing public-name/origin equality check so it is independently testable (a synthetic declaration can trip it without also tripping the equality check). - Added `validate_stream_parser_parity()`, parity-checked against `sources/dispatch.py:STREAM_RECORD_PROVIDERS` the same way `validate_dispatch_precedence()` already parity-checks detector tightness order. **Zero edits to `dispatch.py`** — only a new read-only import of its existing `STREAM_RECORD_PROVIDERS` constant, respecting this slice's ownership boundary with the concurrent `ingest_worker.py`/`decoders.py` lane. - Updated the module docstring to honestly state Grok's promotion from reserved to executable, and point at the new synthetic reserved-lifecycle test as the live proof of that admission path (since no real production `Origin` is currently reserved). - Added five tests to `tests/unit/sources/test_origin_specs.py`: undeclared- coverage rejection, provider-token-leak rejection, a synthetic `lifecycle="reserved"` registration (built via `replace()` of the real Grok spec) proving the reserved admission contract requires no parser/detector tightness while still requiring coverage/fixture evidence, and positive plus negative stream/parser parity checks. ## Verification - `devtools test tests/unit/sources/test_origin_specs.py` → 8 passed - `devtools test tests/unit/sources/test_dispatch_ordering.py tests/unit/sources/test_dispatch_payloads.py` → 21 passed - `devtools test -k "provider_completeness or origin_specs or dispatch_ordering"` → 26 passed - `mypy --strict polylogue/sources/origin_specs.py tests/unit/sources/test_origin_specs.py` → Success, no issues found - `devtools verify --quick` → exit 0 (all 16 steps ok: format, lint, mypy, render all, topology, layering, closure-matrix, schema roundtrip, manifests, ci-workflows, doc-commands, docs-coverage, test-infra-currency, test-clock-hygiene, pytest-timeout-overrides, degrade-loudly) - No new module/file added, so no topology projection regen was required. ## AC matrix | # | AC | Status | | --- | --- | --- | | 1 | `origin_specs.py` defines complete typed fields/lifecycle states, consumes `polylogue/declarations` without adding source semantics to that kernel | Satisfied (pre-existing on master; verified by reading `polylogue/declarations/*` and confirming `origin_specs.py` only wraps `DeclarationSpec`/`DeclarationRegistry`, adding no kernel-layer semantics) | | 2 | Claude Code / ChatGPT / reserved Grok pilots derive/validate deterministic tightness, parser/assembly registration, public origin values, coverage/completeness rows, fixtures, docs, reparse consequences from one declaration each | Satisfied for Claude Code (executable, tightness 60, stream+artifact-rule bindings) and ChatGPT (executable, tightness 70). Grok itself is now executable (real parser shipped after this bead was written, #3201) rather than reserved; the reserved-lifecycle admission contract is proven live by this PR's new synthetic test rather than by a real reserved origin, since every current public `Origin` token is admitted as executable or compatibility-only | | 3 | `dispatch.py` uses derived registration/order or is parity-checked by it; adding a synthetic executable origin requires one OriginSpec + adapter + fixtures | Satisfied — `validate_dispatch_precedence()` (pre-existing) plus this PR's new `validate_stream_parser_parity()` parity-check `dispatch.py`'s `RECORD_DETECTOR_PROVIDER_ORDER`/`STREAM_RECORD_PROVIDERS` without restructuring dispatch.py | | 4 | Missing parser / stream-parser conflict / ambiguous or cyclic detector order / absent fixture / undeclared coverage / leaked Provider token / non-injective collision each yield a source-locatable diagnostic + repair | Satisfied — missing parser, absent fixture, and non-injective collision without policy were pre-existing `ValueError` raises (now also covered by a dedicated undeclared-coverage test in this PR); ambiguous detector order was pre-existing (`ambiguous_detector_tightness`); this PR adds the previously-missing stream/parser-conflict diagnostic (`stream_parser_parity_mismatch`) and leaked-Provider-token guard. "Cyclic" detector order has no distinct failure mode beyond ambiguous tightness, since tightness is a total order, not a graph | | 5 | Provider stays in adapters, Origin is the public vocabulary, Gemini/Drive-style collisions require an explicit collision policy | Satisfied (pre-existing `collision_policy` field + registration guard, exercised by `aistudio-drive`'s `Provider.GEMINI`/`Provider.DRIVE` fiber) | | 6 | Focused declaration/dispatch/public-schema/completeness/fixture/mutation/render/quick verification passes | Satisfied — see Verification section above | `polylogue-2qx.1.2` (migrate every current origin onto OriginSpec, delete parallel inventories) is explicitly out of this slice's scope per the bead's own design text — it is largely already superseded in practice too (all 11 origins are already declared here and `provider_completeness.py` already projects from `ORIGIN_SPECS`), but a formal audit of that bead's own AC (exact-eleven-token enumeration, parallel-inventory deletion, dispatch golden fixture) is left to whoever picks it up. Ref polylogue-2qx.1.1 <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **Bug Fixes** * Prevented origin names from exposing internal provider terminology. * Added validation to ensure origin evidence is complete before registration. * Improved consistency checks for origins that support streaming data. * **New Features** * Added support for reserved-lifecycle origins used in fixtures, including required supporting evidence without requiring parser configuration. * Added diagnostics for mismatches between streaming capabilities and origin configuration. <!-- end of auto-generated comment: release notes by coderabbit.ai --> Co-authored-by: Claude <noreply@anthropic.com>
Sinity
added a commit
that referenced
this pull request
Jul 21, 2026
…gue-2qx.1.2) (#3250) ## Summary Audits and closes the remaining gap in polylogue-2qx.1.2 ("migrate every current origin onto OriginSpec"). Most of this bead's scope was already shipped to `master` untagged across #3051/#3087/#3088/#3092/#3201/#3228/#3246 (closing `polylogue-2qx.1.1`'s notes confirm this). This PR verifies that prior work end-to-end, proves the parity checks actually catch regressions (not just pass today), and closes one concrete residual gap: a second, un-declared provider-to-assembly-hook dispatch that Claude Code, Codex, and AI Studio/Drive depend on. ## Problem `polylogue/sources/origin_specs.py` already declares all eleven `Origin` tokens with lifecycle/detector-tightness/parser/fixture/coverage bindings, and `polylogue/sources/provider_completeness.py` already derives its rows from `ORIGIN_SPECS` (`PACKAGE_MODE_SPECS` is a tuple comprehension over it, parity-proven by `test_provider_completeness_is_a_projection_of_every_origin_spec`). Dispatch-order and stream-parser-binding parity are also already mechanically checked (`validate_dispatch_precedence`, `validate_stream_parser_parity`) against the real production values in `sources/dispatch.py` (`RECORD_DETECTOR_PROVIDER_ORDER`, `STREAM_RECORD_PROVIDERS`). What was genuinely missing: the bead's design explicitly calls for "Claude Code and Codex specs [to] expose typed assembly/orchestration/title/action extension hooks (consumed later by 2qx.2, j2zz, ih67) — a typed field on the spec, not another registry." `polylogue.sources.assembly.get_assembly_spec()` is exactly that second, un-declared provider dispatch (`ProviderAssemblySpec` factory for `Provider.CLAUDE_CODE`/`Provider.CODEX`/ `Provider.GEMINI`) — it existed in production but had no OriginSpec declaration or parity check, so a caller had no way to discover the hook from `origin_specs.py` and no test would catch it drifting from the real registry. Also: `docs/provider-origin-identity.md` documented Provider/Origin/Source doctrine but never pointed at `origin_specs.py` as the per-origin admission authority a reader should trust over any hand-maintained list. ## Solution - `polylogue/sources/origin_specs.py`: added `OriginSpec.assembly_spec_path` (`"module/path.py:ClassName"` or `None`) declaring which `ProviderAssemblySpec` a given origin's provider wire binds. Populated for the three origins with a real live hook: `claude-code-session` → `assembly_claude_code.py:ClaudeCodeAssemblySpec`, `codex-session` → `assembly_codex.py:CodexAssemblySpec`, `aistudio-drive` → `assembly_gemini.py:GeminiAssemblySpec` (its `Provider.GEMINI` wire is the one `get_assembly_spec` keys on). Added `validate_assembly_spec_parity()`, following the exact idiom of `validate_dispatch_precedence`/ `validate_stream_parser_parity`: it checks the declared field against the live `get_assembly_spec(provider)` return value rather than replacing that registry with a second one in `origin_specs.py`. - `tests/unit/sources/test_origin_specs.py`: two new tests — one asserting the three declared hook paths (and `None` for two origins without a live hook), one asserting `validate_assembly_spec_parity(get_assembly_spec) == ()` plus the mutation shape (a resolver that always returns `None` makes every origin that declares a hook path report `assembly_spec_parity_mismatch`). - `docs/provider-origin-identity.md`: added an "Origin admission" section naming `origin_specs.py` as the per-origin declaration authority and pointing at both live-production parity checks. - No parallel hand-maintained inventory was deleted, because the audit found none left inside my ownership (`polylogue/sources/**`) — see the residual below for the one still outstanding elsewhere. ### Residual found outside owned scope `polylogue/cli/shell_completion_values.py:_ORIGIN_DESCRIPTIONS` is a hand-maintained `dict[str, str]` of 8 of the 11 origin tokens (missing `beads-issue`, `grok-export`, `unknown-export`) used for shell-completion help text, and it is **not** parity-checked against `Origin`/`ORIGIN_SPECS` the way `polylogue/agent_integration/spec.py:ORIGIN_MEANINGS` already is (that one has a `RuntimeError` guard requiring exact `Origin`-enum order match — already satisfies AC4, no action needed there). `cli/**` is out of this lane's ownership (per dispatch instructions, another lane owns it concurrently); flagging as a follow-up rather than touching it here. ## Verification - `devtools test tests/unit/sources/test_origin_specs.py tests/unit/sources/test_provider_completeness.py` → **15 passed** - `devtools test tests/unit/sources/ -k "origin_specs or provider_completeness or dispatch or assembly"` → **144 passed** - `devtools test tests/unit/sources/test_parsers_props.py` (protected file) → **43 passed** - `devtools verify --quick` → all 16 steps `ok`, `"exit_code": 0` (ruff format/check, mypy --strict, render all, topology/layering/closure-matrix, manifests, ci-workflows, doc-commands, docs-coverage, test-infra-currency, test-clock-hygiene, pytest-timeout-overrides, degrade-loudly) - Anti-vacuity mutations, run manually then reverted (`git checkout --`), each confirmed the *existing* production dependency actually fails on the mutation before I added anything: - Removing `_grok_spec()` from the registration loop → `ORIGIN_SPEC_REGISTRY.diagnostics()` reports `missing_origin_spec` for `grok-export`. - Dropping Claude Code's `detector_tightness` below `gemini_cli_session`'s (contradicting its real last-in-branch-order position) → `validate_dispatch_precedence(RECORD_DETECTOR_PROVIDER_ORDER)` reports `dispatch_tightness_mismatch`. - Deleting Codex's new `assembly_spec_path` declaration → `validate_assembly_spec_parity(get_assembly_spec)` reports `assembly_spec_parity_mismatch` for `codex-session` (this is the mutation the new code in this PR is proven against). ## AC matrix (polylogue-2qx.1.2) 1. **Satisfied (pre-existing).** All eleven `Origin` tokens are declared exactly once in `ORIGIN_SPECS`; `set(by_origin) == set(Origin)` is asserted in `test_origin_specs_cover_the_public_enum_and_admission_lifecycles`, and `OriginSpecRegistry.diagnostics()` emits `missing_origin_spec` for any omission (proven above). 2. **Satisfied (pre-existing + this PR).** Every executable origin declares acquisition modes, detector/parser/stream-parser/fixture/coverage/fidelity/reparse fields; `claude-code-session` additionally declares 6 typed `OriginArtifactRule`s for its workflow/journal/transcript/sidecar/adopt-manifest/coordinator-stream family. `grok-export` is `lifecycle="executable"` today (it shipped a real parser via #3201/polylogue-611 after 2qx.1.1's pilot design assumed it would stay reserved — documented at the top of `origin_specs.py`); the `reserved` lifecycle contract itself is proven against a synthetic variant since no current `Origin` is genuinely reserved. `unknown-export` is `lifecycle="compatibility-only"` with explicit fallback semantics; `beads-issue` is executable with `issue-jsonl` acquisition and its own non-`accepted` completeness maturity caveat (no harvested schema sample yet). This PR adds the Claude/Codex/AI-Studio assembly hook declaration. 3. **Satisfied (pre-existing).** `validate_dispatch_precedence`/`validate_stream_parser_parity` parity-check tightness and stream-binding against `sources/dispatch.py`'s real `RECORD_DETECTOR_PROVIDER_ORDER`/`STREAM_RECORD_PROVIDERS`; mutation-proven above (deleting/reordering a declaration fails). 4. **Satisfied (pre-existing) for `provider_completeness.py`; this PR extends it to the assembly-hook axis.** `PACKAGE_MODE_SPECS` is a direct tuple comprehension over `ORIGIN_SPECS`, parity-proven by `test_provider_completeness_is_a_projection_of_every_origin_spec`. `agent_integration/spec.py:ORIGIN_MEANINGS` is separately parity-guarded against `Origin` at import time. **Deferred**: `cli/shell_completion_values.py:_ORIGIN_DESCRIPTIONS` remains a stale, un-parity-checked hand-maintained inventory — out of this lane's ownership (`cli/**`), flagged as a residual above rather than a named follow-up bead (small, mechanical, single-file fix for whichever lane next touches CLI completions). 5. **Satisfied (pre-existing).** `aistudio-drive`'s `provider_wires=(Provider.GEMINI, Provider.DRIVE)` with `collision_policy` is registration-enforced (`OriginSpecRegistry.register` raises `ValueError` if `len(provider_wires) > 1` without a `collision_policy`, tested in `test_origin_spec_rejects_missing_fixture_and_noninjective_collision_without_policy`); `docs/provider-origin-identity.md` already documented the non-reversal doctrine, and this PR adds the pointer from that doc to the enforcing code. 6. **Misframed — recorded honestly rather than claimed satisfied.** `polylogue-2qx.2` (Claude Code orchestration artifact admission) already closed via PR #3088, and its own close reason states it shipped as a `claude_workflow_materializer` convergence stage *without* the OriginSpec migration this bead was meant to gate — the blocking edge was explicitly force-closed as "over-blocking" because the merged implementation disproves that ordering. So AC6 as written ("Claude/Codex extension hooks are the only admission path used by 2qx.2...") cannot be retroactively true for `2qx.2`. `j2zz` (Codex child-action lowering) and `ih67` (Codex title enrichment) are still open; this PR's new `assembly_spec_path`/`validate_assembly_spec_parity` gives them a real typed admission point to consume going forward, but I did not modify `assembly_codex.py`/`ih67`/`j2zz` scope — those remain their own beads' work, not migrated by me here. ## Files touched - `polylogue/sources/origin_specs.py` - `tests/unit/sources/test_origin_specs.py` - `docs/provider-origin-identity.md` Ref polylogue-2qx.1.2 <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit - **Documentation** - Added guidance on how origin metadata is declared, validated, and kept aligned with runtime behavior. - Documented checks for provider completeness and assembly-related configuration. - **Improvements** - Added validation to detect mismatches between declared origin capabilities and available assembly behavior. - Added assembly configuration metadata for Claude Code, Codex, and AI Studio origins. - **Tests** - Added coverage confirming assembly declarations and runtime parity checks behave as expected. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Admits Hermes NeMo Relay ATOF JSONL streams (already watched since #3084) into normalized parsing: structural detection, per-Hermes-session grouping, observer-only LLM/tool/approval/error/subagent/context evidence, and UUID+scope-phase replay protection.
Problem
Real Hermes runtime evidence includes append-only ATOF observability streams that the daemon watched and retained but never admitted or normalized — a fidelity gap in the maximalist Hermes support program (
Ref polylogue-fs1.2.1).Solution
sources/dispatch.pystructural detector +origin_specs.pyregistration;parsers/hermes_spans.pygrows ATOF record normalization (observer session correlated by Hermes session id — deliberately NOT merged into the conversational transcript, no session links); privacy-redacted real-bytes v0.1 fixture with provenance notes.Verification
tests/unit/sources/parsers/test_hermes_spans.pyagainst the redacted real fixture (per the lane-B session's commit; see commit body).Note: authored by the lane-B Codex session, which was cut off by quota exhaustion after pushing this commit; PR opened by the war-room coordinator. Lane B resumed on the same branch afterward with several more fixes (real ATOF lifecycle events, ATIF tool_calls fidelity, state.db v19 schema drift, verification_evidence.db import) -- since this PR merged before those additional commits landed, they are tracked separately in #3092.