pod
Other formats agents might prefer:
markdownjsonllms.txt

Agent? You probably want markdown or json, or Pod over MCP.

Reported issues for Perseus — Live Context Engine for AI Agents

Pod holds 20 of 108 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to Perseus — Live Context Engine for AI Agents.

Most discussed

Document protocol-style context contracts with capability manifests (PDD alignment)

Body:

Motivation

PDD (arXiv:2605.12981) argues the durable artifact should be a machine-enforceable protocol of structural, behavioral, and operational invariants rather than prose. Perseus already has schema= validation and perseus validate; this issue documents how a .perseus/context.md block can be authored as a small protocol: typed handshake (schema), behavioral assertions (idempotency/safe-failure properties), and a capability manifest (side-effect, latency, resource bounds) — an

Read the thread · 2026-08-03 · closed · 4 comments

Expose Vault memory origin and entity-reference metadata in Perseus rendering

Summary

Expose Vault memory origin and external-entity reference metadata in Perseus memory rendering surfaces.

Why

Perseus already treats provenance and memory recall as first-class concerns at the render/orchestration layer, including @memory, @mimir, automatic memory injection, and provenance-aware documentation. Meanwhile, Perseus Vault is the persistent memory backend and now has follow-up issues proposed for:

  • first-class external entity references on memories
  • first-class ori

Read the thread · 2026-07-20 · closed · 4 comments

Reflective queries as a first-class operation in the context engine (question-driven read-time synthesis)

Origin: design discussion with @sowerkoku on sowerkoku/knowledge-kernel#2 (comment, reply). Vault-side mechanics: Perseus-Computing-LLC/perseus-vault#741 (and the abductive-context dependency, perseus-vault#740).

Problem

Memory retrieval today is context-driven: what's relevant to the current turn gets injected. The discussion surfac

Read the thread · 2026-07-21 · closed · 3 comments

session_start hook emits correct but fresh session does not receive it

Summary

The session-start hook emits the correct <memory-prep> block, but a fresh Rovo CLI session does not expose that recalled memory to the assistant.

What happened

The hook:

  • ~/.rovo/hooks/perseus-vault/session_start.sh

returns the expected memory-prep output when run directly, including the correct always-on rules and conventions.

Example manual verification:

printf '%s' '{"title":"default call prep format first names full names atlassian rovodev boundary 50+ skills ca

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/790) · 2026-07-15 · closed · 2 comments

### [patent] Claim element: deterministic reproducibility guarantee (same template + same state => byte-identical context)

**Epic:** Patent: Non-Provisional Conversion

Agentic systems are probabilistic and **cannot** claim reproducible context. Perseus can. Make it a positive, tested limitation.

### Acceptance criteria
- [ ] Property test: identical template + identical source state ⇒ byte-identical assembled context across repeated runs
- [ ] Document the determinism contract (ordering rules, tie-breaking, budget allocation) in disclosures
- [ ] Exhibit showing two runs hashing to the same digest
- [ ] Identify a

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/489) · 2026-06-28 · closed · 2 comments

### feat(skills): make procedural-memory trust receipt-backed and coverage-aware

## Summary

Extend the outcome-verified procedural-memory work from #948 and the evidence-linked lesson lifecycle in #926 so that lesson trust is based on receipt-backed verifier coverage as well as known-outcome win rate.

## Problem

The current procedural-memory ledger accepts a boolean `success` outcome and derives attempts, successes, failures, recent-tail win rate, and attribution-aware retirement. It does not yet represent an inconclusive verifier result as a first-class outcome, bind eac

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/1017) · 2026-08-27 · closed · 1 comment

### test(deployment): prove disconnected Perseus + Vault + Ledger operation and claim boundary

## Context

Hypothesis 3 is not closed by saying that the products are local-first. We need a reproducible acceptance run showing that the Perseus stack can be installed, operated, observed, recovered, and updated inside a disconnected/on-prem deployment boundary.

Iron Bank provides hardening/scanning evidence, but its own documentation says hardened containers do not receive an ATO or CtF; the downstream program remains responsible for authorization. This issue is therefore an evidence and cla

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/997) · 2026-08-19 · closed · 1 comment

### feat(supply-chain): ingest SPDX/CycloneDX and build queryable software lineage

## Context

The Kessel Run DevSecOps CSO is an active umbrella vehicle for Software and DevSecOps offerings, with focused Areas of Interest and submission instructions released as specific requirements arise:

- SAM.gov: https://sam.gov/workspace/contract/opp/2dbc9f187ca543398f9f1c6e7601fde9/view
- CISA SBOM overview: https://www.cisa.gov/topics/information-communications-technology-supply-chain-security/sbom
- CISA 2025 SBOM minimum-elements guidance: https://www.cisa.gov/resources-tools/resour

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/995) · 2026-08-19 · closed · 1 comment

## Most recent

### bench(agent-utility): add a paired coding-agent memory lane with delivery, use, and outcome evidence

## Source signal

Stele Bench publishes an open apparatus for paired coding-agent project-memory evaluations:

- article: https://stele-ai.dev/blog/anatomy-of-an-agent-memory-benchmark
- apparatus: https://github.com/Stele-Dev/stele-bench
- preliminary evidence boundary: https://raw.githubusercontent.com/Stele-Dev/stele-bench/main/eval/stele-bench/results/2026-08-preliminary/README.md
- methodology: https://raw.githubusercontent.com/Stele-Dev/stele-bench/main/eval/stele-bench/METHODOLOGY.md

Its

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/992) · 2026-08-19 · closed · 0 comments

### feat(context): commitment-preserving verifiable compression — Context Codec borrow (2605.17304)

## Summary
Adopt Context Codec's commitment-level compression contract: before any compaction, extract typed, source-grounded semantic atoms (goals, constraints, decisions, preferences, tool results, evidence, safety boundaries) and verify their preservation after compression — with conservative fail-closed fallback for low-confidence and safety-critical atoms.

## Problem
"LLM context is not just tokens; it is a set of commitments" (arXiv:2605.17304). Existing compaction methods shorten length 

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/971) · 2026-08-16 · closed · 0 comments

### feat(context): trajectory-mined context-source failure attribution (TRACE 2608.09153)

## Summary
Borrow TRACE's trajectory-mining + multi-component causal attribution loop to give the context engine a debugging layer: mine historical agent trajectories for implicit dissatisfaction signals (user corrections, rephrasing, abandonment) and attribute failures to the specific context source (system prompt, knowledge base, tool schema, skill) that caused them, then drive CREATE/UPDATE remediation of that source.

## Problem
Perseus compiles prompts from heterogeneous context sources. Wh

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/968) · 2026-08-16 · closed · 0 comments

### feat(context): auditable context-compilation DAG (AGoT 2502.05078)

## Summary
Represent context compilation as a typed, versioned DAG (nodes = requirement / retrieved record / summary / contradiction / policy constraint / tool output / decision; edges = supports / depends_on / contradicts / invalidates / selected_for), with selective expansion under hard budgets and a sufficient/abstain/escalate terminal evaluator. Optional CISC-style confidence-weighted candidate prioritization, strictly behind evidence gates.

## Problem
The context engine assembles packets, 

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/962) · 2026-08-15 · closed · 1 comment

### fix(skill-mining): review-follow-up — redaction retry fail-closed + empty-render telemetry (#932 review)

## Context
Independent adversarial review of #932 (deleg_aba71539, Aug 9 2026) passed all security gates (path traversal rejected at every name join, no auto-injection paths, redaction fail-closed on the initial render) but surfaced two MEDIUM findings. Both are fixed and verified locally on branch `fix/perseus-932-review-findings` (commit 11a4fce, 26 focused tests green, build --check clean) — this issue tracks landing them.

## Findings
1. **Redaction fail-closed covered only the first redacti

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/934) · 2026-08-09 · closed · 1 comment

### Add a first-party local runtime evaluation protocol with capability and auth provenance

## Problem

Helix Agent Lab uses an asynchronous local MCP server, deterministic prompt-case scoring, hidden simulator state, runtime capability descriptors, bounded logs, reproducible seeds, and explicit separation between stateful system benchmarks and prompt-only evaluations. It also keeps provider authentication inside official installed CLIs instead of turning consumer credentials into a general API.

Perseus has strong memory/context/evidence benchmarks, but its cross-runtime evaluation co

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/927) · 2026-08-06 · closed · 1 comment

### Adopt exact external-artifact prior-action checks as a first-class context/action integration

## Problem

Helix-AGI separates semantic familiarity from exact prior-action state: its interaction ledger records stable artifact identifiers (message, email, post, comment, thread, or notification IDs) when a response action occurs, then annotates later reads when the same artifact has already been handled. Semantic similarity alone cannot answer whether an agent has already acted on this exact artifact.

Perseus has durable memory, external references, context decisions, and Ledger receipts, 

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/925) · 2026-08-06 · closed · 1 comment

### Add a reproducible code-context benchmark that separates retrieval overhead from delivered prompt tokens

## Summary
Add a reproducible, offline benchmark for code-context retrieval that reports both retrieval cost and context delivered to the agent. The benchmark should make Perseus's efficiency claims auditable without conflating local retrieval work with provider-billed prompt usage.

## Motivation
The GrapeRoot thread made several distinct claims that are easy to blur together: graph lookup can avoid model exploration, context can be pre-injected, and users can estimate money/token savings from 

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/922) · 2026-08-06 · closed · 1 comment

### Add an optional code-graph context provider for symbol-aware pre-injection

## Summary
Add an optional, local code-graph provider that can feed Perseus's existing context-routing and context-decision pipeline with symbol-aware, dependency-aware candidates before broad file search.

## Why this is worth doing
A recent review of GrapeRoot's public workflow highlighted a useful narrow idea: build a graph of files, symbols, imports/call relationships, and recently touched files, then select bounded code context before the agent starts exploring. Perseus already has hybrid r

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/921) · 2026-08-06 · closed · 1 comment

### feat(context): add a reviewable sanitized agent projection with per-scope consent and revocation

## Problem

Perseus can select and render context, but it needs a first-class release boundary that makes the difference between internal source/derived state and the exact representation sent to an agent inspectable. Without that boundary, users and operators cannot reliably answer “what did this agent receive, under which policy, and why?”

## Proposed change

Add a versioned `agent_projection` / context-release contract that compiles selected Vault/context records into a task-scoped, privacy-

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/917) · 2026-08-04 · closed · 0 comments

### feat(context): add decision-oriented Rank/Ask MCP operations over scoped context

## Problem

The current context engine has retrieval, routing, budgeting, and explainable decision records, but the user-facing abstraction is still mostly “retrieve or inject context.” Agent products often need a narrower contract: rank a known candidate set for this user/task, or answer one constrained context question without exporting the full profile.

A decision-oriented surface would make context useful at the point where an agent chooses, while reducing unnecessary context-window exposur

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/916) · 2026-08-04 · closed · 0 comments

### Docs/setup: standardize all installer and MCP examples on ~/.local/bin/perseus

## Summary
Standardize docs and setup examples on the stable launcher path `~/.local/bin/perseus` instead of version-pinned Python minor-version paths.

## Why
During a local upgrade from v1.0.24 to v1.0.26, one remaining setup example still pointed to a version-specific macOS path (`.../Library/Python/3.13/bin/perseus`). It was locally corrected to the more durable `~/.local/bin/perseus` pattern.

The repo already appears to prefer the stable launcher in several places, but the convention is no

[Read the thread](https://github.com/Perseus-Computing-LLC/perseus/issues/909) · 2026-08-03 · closed · 0 comments

The remaining reports are on [the project's issue tracker](https://github.com/Perseus-Computing-LLC/perseus/issues).