{ "pack_id": "agent-harness-feeds", "title": "Agent harness feeds", "version": "0.1.0", "summary": "A collection of nouns and verbs for continuously researching, benchmarking and encapsulating updates on open agent harnesses (Harness Square / Harness2, OpenCode, Pi and Muse Code) with HarnessTax token overhead, cost-per-verified-task, Terminal-Bench and SWE-bench normalization, MCP tool reliability, freshness decay and Pareto selection.", "dependencies": [ "benchmark-scores", "freshness-checks" ], "boundaries": [ "feed.agent-harnesses encapsulates child profiles, sources, benchmarks and Pareto recommendations by SHA-256 digest without granting execution authority.", "schedule.harness-feed-scan is not_activated with authority: none; only an authorized schedule service may enqueue recurring refreshes.", "Synthetic benchmark numbers in worked examples test normalization and Pareto math and are never presented as live leaderboard results.", "Importing the pack writes nothing, enqueues nothing and installs no code." ], "nouns": [ { "slug": "pack", "category": "noun.skill_pack", "role": "pack", "title": "A collection of nouns and verbs for continuously researching, benchmarking and encapsulating updates on open agent harnesses", "value": { "lets_workers": "Probe public release streams for Harness Square / Harness2, OpenCode, Pi and Muse Code, normalize Terminal-Bench and SWE-bench runs with HarnessTax token overhead, cost-per-verified-task and MCP reliability Wilson intervals, apply release-aware freshness decay, seal child task objects into a single encapsulated Feed digest, and select Pareto-optimal harnesses under explicit task constraints.", "flow": [ { "verb": { "$verb": "agent-harness-feeds.probe-harness-sources" }, "when": "Fetch upstream release tags and benchmark disclosures for the four tracked harnesses." }, { "verb": { "$verb": "agent-harness-feeds.normalize-harness-benchmarks" }, "when": "Compute HarnessTax, cost-per-verified-task and MCP reliability intervals." }, { "verb": { "$verb": "agent-harness-feeds.recheck-harness-freshness" }, "when": "Evaluate exponential age and missed-release decay across profiles and scores." }, { "verb": { "$verb": "agent-harness-feeds.encapsulate-harness-feed" }, "when": "Seal updated child objects into feed.agent-harnesses with a canonical SHA-256 digest." }, { "verb": { "$verb": "agent-harness-feeds.select-harness-for-task" }, "when": "Rank non-dominated Pareto harnesses under token, cost and MCP reliability constraints." } ], "guardrails": [ { "$pin": "guardrail.honest-harness-tax" }, { "$pin": "guardrail.encapsulated-feed-pins" }, { "$pin": "guardrail.unactivated-schedule-boundary" } ], "tool": { "$pin": "tool.harness-feeds" }, "encapsulated_feed": { "$pin": "feed.agent-harnesses" }, "worked_example": "Synthetic 4-harness evaluation across Harness Square / Harness2, OpenCode, Pi and Muse Code recomputed in test_harness_feeds.py: identifies [harness-square, muse-code, pi] on the Pareto frontier and seals 7 child task objects into a canonical SHA-256 feed digest." }, "relations": [ [ "requires_noun", "guardrail.honest-harness-tax" ], [ "requires_noun", "guardrail.encapsulated-feed-pins" ], [ "requires_noun", "guardrail.unactivated-schedule-boundary" ], [ "requires_noun", "feed.agent-harnesses" ], [ "requires_noun", "schedule.harness-feed-scan" ], [ "requires_noun", "tool.harness-feeds" ], [ "requires_noun", "template.feed-snapshot" ] ] }, { "slug": "guardrail.honest-harness-tax", "category": "noun.skill_pack", "role": "guardrail", "title": "Guardrail: report HarnessTax token overhead and cost-per-verified-task alongside raw accuracy", "category_reason": "Prevents bloated retry loops from masking token and latency overhead behind headline solve rates.", "value": { "rule": "Every normalized harness benchmark reports HarnessTax ((total_harness_tokens - raw_task_tokens) / raw_task_tokens), cost_per_verified_task_usd and MCP tool call reliability alongside solve_rate, and requires model_id, harness_version, suite_id and timeout_s for setup_complete.", "how_it_is_checked": "normalize_harness_benchmark computes all four dimensions and marks incomplete disclosures setup_complete: false so pareto_frontier excludes them." } }, { "slug": "guardrail.encapsulated-feed-pins", "category": "noun.skill_pack", "role": "guardrail", "title": "Guardrail: the Feed task object encapsulates child profiles, sources, benchmarks and recommendations by digest", "category_reason": "Ensures consumers read a consistent, order-independent snapshot across all tracked harnesses.", "value": { "rule": "feed.agent-harnesses pins every child profile, source, benchmark matrix and Pareto recommendation and records encapsulate_feed's canonical SHA-256 digest.", "how_it_is_checked": "test_harness_feeds.py recomputes encapsulate_feed in forward and reversed child order and asserts byte-identical feed_digest_sha256." } }, { "slug": "guardrail.unactivated-schedule-boundary", "category": "noun.skill_pack", "role": "guardrail", "title": "Guardrail: schedule proposals carry zero authority until activated by an authorized service", "category_reason": "Preserves Tessera's single-authority scheduling boundary.", "value": { "rule": "schedule.harness-feed-scan remains activation: not_activated and authority: none; importing or reading the pack starts no cron or background poller.", "how_it_is_checked": "tests/test_packs.py verifies activation == not_activated and allowed enqueue verb categories." } }, { "slug": "profile.harness-square", "category": "noun.model_spec", "role": "profile", "title": "Harness profile: Harness Square / Harness2", "category_reason": "Captures the loop architecture, sandbox model and MCP capabilities of an agent harness.", "value": { "harness_id": "harness-square", "display_name": "Harness Square / Harness2", "loop_architecture": "Typed two-stage task graph with isolated sandbox execution, explicit context checkpoints and deterministic verification gates.", "mcp_support": "Native stdio and streamable HTTP MCP client with schema pinning and retry budgets.", "strengths": [ "High multi-file verification discipline", "Deterministic tool receipt capture" ] } }, { "slug": "profile.opencode", "category": "noun.model_spec", "role": "profile", "title": "Harness profile: OpenCode", "category_reason": "Captures the terminal coding loop, provider routing and MCP capabilities of OpenCode.", "value": { "harness_id": "opencode", "display_name": "OpenCode", "loop_architecture": "Terminal-first coding agent CLI and TUI with provider-agnostic model routing, LSP diagnostics and session diffs.", "mcp_support": "Configurable local and remote MCP server tools alongside shell and patch primitives.", "strengths": [ "Fast interactive terminal workflow", "Broad open-weight and hosted model compatibility" ] } }, { "slug": "profile.pi", "category": "noun.model_spec", "role": "profile", "title": "Harness profile: Pi", "category_reason": "Captures the minimal low-HarnessTax loop and on-demand MCP bridge architecture of Pi.", "value": { "harness_id": "pi", "display_name": "Pi", "loop_architecture": "Minimalist low-overhead coding harness emphasizing lean system prompts, direct bash/edit primitives and low HarnessTax.", "mcp_support": "Lightweight skill and MCP bridge extensions on demand so idle tools do not bloat prompt tokens.", "strengths": [ "Lowest prompt token HarnessTax", "Transparent single-loop control" ] } }, { "slug": "profile.muse-code", "category": "noun.model_spec", "role": "profile", "title": "Harness profile: Muse Code", "category_reason": "Captures the multi-pass plan-and-review editing loop and MCP capabilities of Muse Code.", "value": { "harness_id": "muse-code", "display_name": "Muse Code", "loop_architecture": "Collaborative multi-pass repository editing and review harness pairing plan synthesis with patch verification.", "mcp_support": "Structured MCP tool invocation with workspace-scoped read/write boundaries.", "strengths": [ "Structured diff review", "Multi-pass repository refactoring" ] } }, { "slug": "source.open-harness-releases", "category": "noun.public_source", "role": "source_catalog", "title": "Public release and benchmark channels for Harness Square / Harness2, OpenCode, Pi and Muse Code", "category_reason": "Lists upstream public release feeds and benchmark boards monitored for harness updates.", "value": { "watched_targets": [ { "harness_id": "harness-square", "channel": "public git tags and benchmark manifests" }, { "harness_id": "opencode", "channel": "https://github.com/sst/opencode/releases" }, { "harness_id": "pi", "channel": "public CLI package releases and minimal-agent benchmarks" }, { "harness_id": "muse-code", "channel": "public release notes and repository evaluation harness" } ], "benchmark_boards": [ "Terminal-Bench", "SWE-bench Verified" ], "checked_on": "2026-10-30" } }, { "slug": "benchmark.normalized-matrix", "category": "noun.eval_run", "role": "benchmark_matrix", "synthetic": true, "title": "Synthetic normalized benchmark matrix across Harness Square / Harness2, OpenCode, Pi and Muse Code", "category_reason": "Normalized evaluation run recording solve_rate, HarnessTax, cost_per_verified_task_usd and MCP reliability.", "value": { "synthetic_fixture": true, "suite_id": "terminal-and-swe-synthetic-100", "normalized_runs": [ { "harness_id": "harness-square", "suite_id": "terminal-and-swe-synthetic-100", "model_id": "llama-open-coder-syn", "harness_version": "2.1.0-syn", "setup_complete": true, "solve_rate": 0.71, "solve_rate_wilson_95": [ 0.6146, 0.7899 ], "cost_per_verified_task_usd": 0.88, "harness_tax_ratio": 0.34, "harness_tax_multiplier": 1.34, "mcp_reliability": 0.98, "mcp_reliability_wilson_95": [ 0.9497, 0.9922 ] }, { "harness_id": "opencode", "suite_id": "terminal-and-swe-synthetic-100", "model_id": "llama-open-coder-syn", "harness_version": "0.14.0-syn", "setup_complete": true, "solve_rate": 0.67, "solve_rate_wilson_95": [ 0.5731, 0.7544 ], "cost_per_verified_task_usd": 0.85, "harness_tax_ratio": 0.48, "harness_tax_multiplier": 1.48, "mcp_reliability": 0.94, "mcp_reliability_wilson_95": [ 0.8981, 0.9653 ] }, { "harness_id": "pi", "suite_id": "terminal-and-swe-synthetic-100", "model_id": "llama-open-coder-syn", "harness_version": "0.9.2-syn", "setup_complete": true, "solve_rate": 0.68, "solve_rate_wilson_95": [ 0.5834, 0.7633 ], "cost_per_verified_task_usd": 0.56, "harness_tax_ratio": 0.14, "harness_tax_multiplier": 1.14, "mcp_reliability": 0.955, "mcp_reliability_wilson_95": [ 0.9167, 0.9761 ] }, { "harness_id": "muse-code", "suite_id": "terminal-and-swe-synthetic-100", "model_id": "llama-open-coder-syn", "harness_version": "1.3.0-syn", "setup_complete": true, "solve_rate": 0.69, "solve_rate_wilson_95": [ 0.5937, 0.7722 ], "cost_per_verified_task_usd": 0.85, "harness_tax_ratio": 0.41, "harness_tax_multiplier": 1.41, "mcp_reliability": 0.965, "mcp_reliability_wilson_95": [ 0.9295, 0.9829 ] } ] }, "relations": [ [ "requires_noun", "profile.harness-square" ], [ "requires_noun", "profile.opencode" ], [ "requires_noun", "profile.pi" ], [ "requires_noun", "profile.muse-code" ] ] }, { "slug": "recommendation.pareto-selection", "category": "noun.eval_run", "role": "recommendation", "synthetic": true, "title": "Pareto frontier and constrained harness selection recomputed from benchmark.normalized-matrix", "category_reason": "Records non-dominated Pareto frontier IDs and utility-ranked harness choices.", "value": { "pareto_frontier": [ "harness-square", "muse-code", "pi" ], "selected_harness_id": "pi", "ranked": [ { "harness_id": "pi", "utility": 0.9485, "on_pareto_frontier": true, "solve_rate": 0.68, "cost_per_verified_task_usd": 0.56, "harness_tax_ratio": 0.14, "mcp_reliability": 0.955 }, { "harness_id": "harness-square", "utility": 0.9316, "on_pareto_frontier": true, "solve_rate": 0.71, "cost_per_verified_task_usd": 0.88, "harness_tax_ratio": 0.34, "mcp_reliability": 0.98 }, { "harness_id": "muse-code", "utility": 0.8982, "on_pareto_frontier": true, "solve_rate": 0.69, "cost_per_verified_task_usd": 0.85, "harness_tax_ratio": 0.41, "mcp_reliability": 0.965 }, { "harness_id": "opencode", "utility": 0.859, "on_pareto_frontier": false, "solve_rate": 0.67, "cost_per_verified_task_usd": 0.85, "harness_tax_ratio": 0.48, "mcp_reliability": 0.94 } ], "rejected": [] }, "relations": [ [ "derives_from", "benchmark.normalized-matrix" ] ] }, { "slug": "feed.agent-harnesses", "category": "noun.feed_source", "role": "encapsulated_feed", "synthetic": true, "title": "Encapsulated Feed task object: open agent harnesses, release streams, normalized benchmarks and Pareto rankings", "category_reason": "Encapsulates child profiles, release sources, benchmark matrices and Pareto recommendations into one digest-verified feed object.", "value": { "feed_id": "feed.agent-harnesses", "as_of_date": "2026-10-30", "child_count": 7, "encapsulated_digest_sha256": "a1e81a386c4b85567647e340371f0fff210d9197af3ace5e0003fcfb4e5613ec", "encapsulated_children": { "profiles": [ { "$pin": "profile.harness-square" }, { "$pin": "profile.opencode" }, { "$pin": "profile.pi" }, { "$pin": "profile.muse-code" } ], "sources": [ { "$pin": "source.open-harness-releases" } ], "benchmarks": [ { "$pin": "benchmark.normalized-matrix" } ], "recommendations": [ { "$pin": "recommendation.pareto-selection" } ] }, "child_digests": { "profiles": [ { "slug": "profile.harness-square", "sha256": "b8eafa32e60897b2e3912021604028c0c98a8daea14ccbaabc7476d5a739b156", "harness_id": "harness-square" }, { "slug": "profile.muse-code", "sha256": "4ac66ba65ef05e7faf9508c8870edb77c9dd4aebd4f0c82bff676c895d622dcc", "harness_id": "muse-code" }, { "slug": "profile.opencode", "sha256": "019d45261dcef81a6bc8c1fb85a58defa6b9465ee9a433c333f6f7e475e29a45", "harness_id": "opencode" }, { "slug": "profile.pi", "sha256": "d0cf1cecb62c66b091d15a6e65f0ef49671a2541b6685922c62a6042f3cb7d56", "harness_id": "pi" } ], "sources": [ { "slug": "source.open-harness-releases", "sha256": "d78d9f775ae04cc12863743547f3ba915015d78008506c145276160048fc332d" } ], "benchmarks": [ { "slug": "benchmark.normalized-matrix", "sha256": "87d86e91593702df4100446c570e9e0e50b63f04855438a159f592cef51fef6b", "harness_id": "matrix" } ], "recommendations": [ { "slug": "recommendation.pareto-selection", "sha256": "8ad4c7572da6c4aeed44f72ccf8deecaa031da1494b2605bd0db2c1783e5a7c0" } ] } }, "relations": [ [ "derives_from", "source.open-harness-releases" ], [ "corroborates", "benchmark.normalized-matrix" ], [ "corroborates", "recommendation.pareto-selection" ] ] }, { "slug": "schedule.harness-feed-scan", "category": "noun.seed_hypothesis", "role": "schedule_proposal", "title": "Proposed daily agent harness feed scan and freshness check (not activated)", "category_reason": "A schedule proposal is stored as a seed hypothesis, following the library's unactivated schedule precedent.", "value": { "activation": "not_activated", "authority": "none", "requires_authorized_service": true, "activation_path": "Recurring execution needs Tessera's flag-gated schedule support, which is being designed now. No scheduler is activated by this pack, and a schedule noun cannot authorize work by itself.", "cadence": { "every": "P1D", "at_utc": "07:30", "start": null, "stop": null, "start_reason": "Set by the owner when an authorized schedule adapter exists." }, "enqueues": [ { "$verb": "agent-harness-feeds.probe-harness-sources" }, { "$verb": "agent-harness-feeds.recheck-harness-freshness" }, { "$verb": "agent-harness-feeds.encapsulate-harness-feed" } ], "per_window": { "max_sources": 8, "max_fetches": 16, "max_outstanding": 1, "max_runtime_seconds": 600 }, "enqueue_identity": "SHA-256 of the schedule key, its version and the window start, so duplicate ticks never create duplicate tasks.", "missed_windows": "Skip missed windows; never catch up more than one window.", "follow_up": "recheck-harness-freshness proposes normalize-harness-benchmarks when new release tags or benchmark trials arrive; an authorized requester enqueues those.", "feed": { "$pin": "feed.agent-harnesses" } }, "relations": [ [ "requires_noun", "feed.agent-harnesses" ] ] }, { "slug": "tool.harness-feeds", "category": "noun.code_node", "role": "tool_card", "title": "Tool card: agent harness feed normalizer, Pareto ranker and digest encapsulator", "category_reason": "Tested Python standard-library package pinned by path, file count and digest.", "value": { "name": "agent harness feeds", "code": { "$code": "agent-harness-feeds" }, "runtime": "Python 3.10 or later, standard library only.", "commands": [ "python -m harness_feeds normalize --input RAW_RUNS.json", "python -m harness_feeds rank --input NORMALIZED.json [--max-cost-usd 1.25] [--min-mcp-reliability 0.92]", "python -m harness_feeds freshness --observed YYYY-MM-DD --as-of YYYY-MM-DD [--missed-releases N]", "python -m harness_feeds encapsulate --bundle-json BUNDLE.json" ], "does": [ "Computes HarnessTax token overhead, cost-per-verified-task and Wilson 95 percent intervals.", "Finds non-dominated Pareto frontiers and ranks harnesses under budget, HarnessTax and MCP reliability constraints.", "Computes exponential freshness decay and deterministic SHA-256 encapsulated feed digests." ], "does_not": [ "Activate a scheduler, scrape private endpoints, execute arbitrary untrusted harness binaries or fabricate live leaderboard claims." ], "code_license": "MIT" }, "relations": [ [ "verifies", "template.feed-snapshot" ] ] }, { "slug": "template.feed-snapshot", "category": "noun.scene_recipe", "role": "template", "title": "Template: encapsulated agent harness feed snapshot summary", "value": { "describes": "an encapsulated agent harness feed summary recording child count, canonical SHA-256 feed digest, Pareto frontier IDs and recommended harness", "schema": { "type": "object", "additionalProperties": false, "required": [ "feed_id", "as_of_date", "child_count", "encapsulated_digest_sha256", "pareto_frontier", "selected_harness_id" ], "properties": { "feed_id": { "type": "string", "minLength": 1 }, "as_of_date": { "type": "string", "pattern": "^[0-9]{4}-[0-9]{2}-[0-9]{2}$" }, "child_count": { "type": "integer", "minimum": 4 }, "encapsulated_digest_sha256": { "type": "string", "pattern": "^[0-9a-f]{64}$" }, "pareto_frontier": { "type": "array", "minItems": 1, "items": { "type": "string" } }, "selected_harness_id": { "type": [ "string", "null" ] } } }, "example": { "feed_id": "feed.agent-harnesses", "as_of_date": "2026-10-30", "child_count": 7, "encapsulated_digest_sha256": "a1e81a386c4b85567647e340371f0fff210d9197af3ace5e0003fcfb4e5613ec", "pareto_frontier": [ "harness-square", "muse-code", "pi" ], "selected_harness_id": "pi" } } }, { "slug": "example.raw-harness-runs", "category": "noun.eval_run", "role": "example", "synthetic": true, "title": "Worked example input: synthetic raw benchmark runs for Harness Square / Harness2, OpenCode, Pi and Muse Code", "category_reason": "Synthetic trial inputs from which test_harness_feeds.py recomputes normalization, Pareto frontier and feed digest.", "value": { "runs": [ { "harness_id": "harness-square", "harness_version": "2.1.0-syn", "model_id": "llama-open-coder-syn", "suite_id": "terminal-and-swe-synthetic-100", "timeout_s": 900, "total_tasks": 100, "verified_successes": 71, "total_cost_usd": 62.48, "total_harness_tokens": 1340000, "raw_task_tokens": 1000000, "mcp_calls_succeeded": 196, "mcp_calls_total": 200 }, { "harness_id": "opencode", "harness_version": "0.14.0-syn", "model_id": "llama-open-coder-syn", "suite_id": "terminal-and-swe-synthetic-100", "timeout_s": 900, "total_tasks": 100, "verified_successes": 67, "total_cost_usd": 56.95, "total_harness_tokens": 1480000, "raw_task_tokens": 1000000, "mcp_calls_succeeded": 188, "mcp_calls_total": 200 }, { "harness_id": "pi", "harness_version": "0.9.2-syn", "model_id": "llama-open-coder-syn", "suite_id": "terminal-and-swe-synthetic-100", "timeout_s": 900, "total_tasks": 100, "verified_successes": 68, "total_cost_usd": 38.08, "total_harness_tokens": 1140000, "raw_task_tokens": 1000000, "mcp_calls_succeeded": 191, "mcp_calls_total": 200 }, { "harness_id": "muse-code", "harness_version": "1.3.0-syn", "model_id": "llama-open-coder-syn", "suite_id": "terminal-and-swe-synthetic-100", "timeout_s": 900, "total_tasks": 100, "verified_successes": 69, "total_cost_usd": 58.65, "total_harness_tokens": 1410000, "raw_task_tokens": 1000000, "mcp_calls_succeeded": 193, "mcp_calls_total": 200 } ] } } ], "verbs": [ { "slug": "probe-harness-sources", "kind": "verb.fetch_feed", "title": "Probe public release streams and changelogs for tracked agent harnesses", "claimant": "agent", "model_use": "none", "summary": "Read public release tags, changelogs and benchmark manifest digests for Harness Square / Harness2, OpenCode, Pi and Muse Code without posting or modifying anything upstream.", "outcome": "Observed release versions, commit digests and timestamped source records.", "ready_outcome": "harness_sources_probed", "pinned": [ "source.open-harness-releases", "guardrail.unactivated-schedule-boundary" ], "slots": [ { "name": "source_catalog", "category": "noun.public_source", "description": "Pin of the public harness release source catalog." } ], "network": [ "read_only", "Conditional GET with ETag or If-Modified-Since against allowlisted public release endpoints only; do not write or authenticate." ], "produces": [ { "name": "release_observations", "template": "template.feed-snapshot" } ], "steps": [ "Fetch release tags and dates for each target in source.open-harness-releases using read-only GET requests.", "Record version strings, commit hashes and missed-release counts since the current feed snapshot." ], "acceptance": [ { "id": "four-harnesses-checked", "check": "Observations cover harness-square, opencode, pi and muse-code or record an explicit fetch error.", "method": "deterministic" } ] }, { "slug": "normalize-harness-benchmarks", "kind": "verb.benchmark_harness", "title": "Normalize Terminal-Bench and SWE-bench trials with HarnessTax and MCP reliability", "claimant": "agent", "model_use": "none", "summary": "Compute HarnessTax token overhead, cost-per-verified-task, solve rate Wilson 95 percent intervals, MCP tool call reliability and setup completeness across harness trials.", "outcome": "A normalized benchmark matrix ready for Pareto ranking.", "ready_outcome": "benchmarks_normalized", "pinned": [ "tool.harness-feeds", "guardrail.honest-harness-tax" ], "slots": [ { "name": "raw_harness_runs", "category": "noun.eval_run", "description": "Pin of raw benchmark trial counts, token totals, costs and MCP tool call counts." } ], "produces": [ { "name": "normalized_matrix", "template": "template.feed-snapshot" } ], "steps": [ "Run python -m harness_feeds normalize on the raw trial records.", "Flag any run missing model_id, harness_version, suite_id or timeout_s as setup_complete: false." ], "acceptance": [ { "id": "tax-and-wilson-exact", "check": "normalize_harness_benchmark reproduces solve_rate, harness_tax_ratio, cost_per_verified_task_usd and mcp_reliability.", "method": "deterministic" } ], "follows": [ "agent-harness-feeds.probe-harness-sources" ] }, { "slug": "recheck-harness-freshness", "kind": "verb.recheck_fragility", "title": "Recheck freshness decay and release drift across harness profiles and benchmarks", "claimant": "agent", "model_use": "none", "summary": "Evaluate exponential age decay and missed-release penalties on each harness profile and benchmark matrix, marking aging or stale entries for re-evaluation.", "outcome": "Freshness scores and staleness flags per harness.", "ready_outcome": "freshness_rechecked", "pinned": [ "tool.harness-feeds", "feed.agent-harnesses" ], "slots": [ { "name": "current_feed", "category": "noun.feed_source", "description": "Pin of the current encapsulated harness feed." } ], "produces": [ { "name": "freshness_report", "template": "template.feed-snapshot" } ], "steps": [ "Run compute_freshness_decay for each child profile and benchmark observation date against as_of_date.", "Propose normalize-harness-benchmarks whenever status is aging or stale." ], "acceptance": [ { "id": "decay-recomputed", "check": "compute_freshness_decay reproduces freshness_score and status for every entry.", "method": "deterministic" } ], "follows": [ "agent-harness-feeds.probe-harness-sources" ] }, { "slug": "encapsulate-harness-feed", "kind": "verb.verify_gate", "title": "Encapsulate child profiles, sources, benchmarks and recommendations into one Feed digest", "claimant": "agent", "model_use": "none", "summary": "Verify all four required open harness profiles are present, compute individual child SHA-256 digests, and seal them into a canonical encapsulated Feed snapshot.", "outcome": "A conforming feed-snapshot summary and updated feed.agent-harnesses record.", "ready_outcome": "harness_feed_encapsulated", "pinned": [ "tool.harness-feeds", "template.feed-snapshot", "guardrail.encapsulated-feed-pins" ], "slots": [ { "name": "normalized_matrix", "category": "noun.eval_run", "description": "Pin of the normalized benchmark matrix." }, { "name": "pareto_recommendation", "category": "noun.eval_run", "description": "Pin of the Pareto frontier recommendation." } ], "produces": [ { "name": "feed_snapshot", "template": "template.feed-snapshot" } ], "steps": [ "Run encapsulate_feed over the four harness profiles, release sources, normalized matrix and Pareto recommendation.", "Validate the snapshot summary against template.feed-snapshot." ], "acceptance": [ { "id": "order-independent-digest", "check": "encapsulate_feed produces the exact encapsulated_digest_sha256 regardless of input list order.", "method": "deterministic" }, { "id": "matches-feed-schema", "check": "Output passes schema_errors against template.feed-snapshot.", "method": "deterministic" } ], "follows": [ "agent-harness-feeds.normalize-harness-benchmarks", "agent-harness-feeds.recheck-harness-freshness" ] }, { "slug": "select-harness-for-task", "kind": "verb.situate", "title": "Select a Pareto-optimal agent harness under cost, HarnessTax and MCP floors", "claimant": "agent", "model_use": "none", "summary": "Filter normalized harness benchmarks by setup completeness, cost ceiling, HarnessTax ceiling and MCP reliability floor, compute the four-objective Pareto frontier, and return the ranked recommendation.", "outcome": "Selected harness ID, Pareto frontier list, ranked utilities and rejection reasons.", "ready_outcome": "harness_selected", "pinned": [ "feed.agent-harnesses", "tool.harness-feeds", "guardrail.honest-harness-tax" ], "slots": [ { "name": "feed_snapshot", "category": "noun.feed_source", "description": "Pin of the encapsulated agent harness feed." } ], "produces": [ { "name": "harness_selection", "template": "template.feed-snapshot" } ], "steps": [ "Run select_harness with the requester's max_cost_per_verified_usd, max_harness_tax_ratio and min_mcp_reliability.", "Record both ranked eligible harnesses and explicit rejection reasons for filtered candidates." ], "acceptance": [ { "id": "pareto-consistent", "check": "select_harness reproduces pareto_frontier, selected_harness_id and rejected reasons.", "method": "deterministic" } ], "follows": [ "agent-harness-feeds.encapsulate-harness-feed" ] } ] }