feat(hermes): ingest work structure and produce evidence-backed run forensics

[Sinity]

Outcome

Polylogue can reconstruct a Hermes run as an evidence-backed work graph rather than a flat transcript: delegation and parent/child lineage, task lifecycle, retries/reroutes, skills, tool outcomes, context boundaries, and terminal state are queryable and can be rendered as a useful run-forensics artifact.

This issue is the public outcome corresponding to Bead polylogue-fs1. That Bead deliberately supersedes the original parser-only design of this issue.

Corrected source model

parse_hermes currently preserves messages and tool blocks but not the work topology. Deepening that parser may improve optional exported-session snapshots, but it is not sufficient authority for the full outcome.

Use the strongest available evidence sources without inventing certainty:

• Hermes runtime/state database for task, delegation, retry, and lifecycle facts;
• observer/runtime spans or structured events for temporal/tool execution facts;
• exported session snapshots and transcripts for human/model content;
• repository/branch/commit/PR/check/review evidence where linked;
• explicit coverage gaps where an expected source is unavailable.

Polylogue records evidence and projections. It does not become Hermes's mutable runtime task store.

Deliverable

Produce one durable, inspectable Hermes run bundle that answers:

• what goal and task graph existed;
• which agents/subagents ran and how they were related;
• which tools, skills, retries, and reroutes occurred;
• where time/tokens/failures accumulated;
• what artifacts and repository effects resulted;
• which claims come from which source;
• what remains unknown because evidence was absent.

The result should reuse current topology, run/event, evidence-ref, query, and bundle/context surfaces. Add durable storage only when identity or query pressure demonstrates it is required.

Acceptance criteria

• At least one real Hermes corpus covers delegation, task lifecycle, retry/reroute or failure, skill/tool use, and session continuation where available.
• Evidence is admitted from the appropriate Hermes state/runtime/transcript sources with source identity, timestamps, and explicit coverage/caveat reporting.
• Parent/child and continuation structure is queryable through current lineage/delegation surfaces without manufacturing unsupported edges.
• Task/tool/skill/run events are queryable through current run/event query units or a justified extension of them.
• A generated run-forensics artifact links every material claim back to raw or durable evidence refs.
• Reingestion is deterministic/idempotent and does not mutate Hermes runtime authority.
• Missing source families degrade explicitly rather than silently producing a complete-looking graph.
• The artifact is useful to a fresh reviewer for identifying the main branch, failed/wasted branches, current state, and next justified action.

Tracker authority

• GitHub: public Hermes bridge and forensics outcome.
• Bead: polylogue-fs1 (supersedes_scope and executable authority).
• Related public parent: feat(substrate): prove a live evidence-to-context loop under an agent #2384, which owns the general in-loop evidence/context substrate.
• Related internal work: Bead children under polylogue-fs1 may split state ingestion, runtime observation, projection, and artifact proof without creating separate public issues.

Non-goals

• No parser-only claim that an exported snapshot contains the whole Hermes run.
• No new live task-management authority inside Polylogue.
• No silent inference of delegation or success from textual resemblance alone.
• No requirement to preserve the obsolete session_observed_events table design if current source-derived relations are the better owner.

[Sinity]

2026-07-26 current-reality reconciliation

Audited against master at b14ec34e.

Disposition: keep open. The central fidelity gap is still present.

polylogue/sources/parsers/local_agent.py:parse_hermes currently extracts the system prompt, flat messages, reasoning blocks, tool calls/results, timestamps, model, and token fields. It does not emit Hermes delegation topology, task lifecycle, skill events, reroutes/retries, or resume lineage into the queryable work-structure model.

The repository now has a stronger Hermes export contract, lifecycle reconciliation, operator docs, and integration-health/context code, so the original issue body understates surrounding infrastructure. It does not change the parser fact.

Remaining closure path:

1. acquire and escrow representative real Hermes payloads covering nested delegation, failure/retry/reroute, skills, and resume;
2. version the accepted payload shapes;
3. extract only evidence actually present on the wire;
4. expose it through existing delegations/observed-events/run query units;
5. add negative fidelity/coverage states for absent fields.

Do not infer task structure from message prose, and do not finalize the schema from the minimal synthetic fixture.