agentera 3.0.0-dev.4 → 3.0.0-dev.41

package/bundle/references/analysis/evidence-tier-authority.yaml

@@ -0,0 +1,518 @@
+schema_version: agentera.evidenceTierAuthority.v1
+status: active_authority
+purpose: >-
+  Define the authoritative tier model for local agent-history evidence so that
+  bounded consumers receive useful analysis at realistic corpus scale instead
+  of degrading solely because complete local evidence exceeds a single-reader
+  limit. The contract fixes the tiers, supported source families, the consumer
+  map, the required signal semantics, and the deterministic compatibility
+  outcomes for oversized, legacy, missing, corrupt, and incomplete state. It is
+  the single source of truth for what every bounded input contract must supply.
+
+authority:
+  source: references/analysis/evidence-tier-authority.yaml
+  single_source_rule: >-
+    Extraction, readers, validators, tests, help, and capability source
+    contracts may project this authority but must not create a second map of
+    tiers, source families, consumer required fields, signal semantics,
+    compatibility outcomes, or evidence bounds. A consumer that needs a field
+    declares it here; a field declared here is the only name a consumer may
+    require from a bounded tier.
+  implementation_boundary: >-
+    This contract defines and verifies the tier model, source coverage, signal
+    semantics, consumer field supply, and compatibility outcomes. It does not
+    implement the complete extraction writer, migrate consumers to bounded
+    tiers, change profile levels, generate a glossary, or alter suite version
+    metadata. Tasks 2 through 4 of the approved plan implement publication,
+    consumer migration, and bounded profile synthesis against this contract.
+
+decision_55_reconciliation:
+  decision_number: 55
+  date: 2026-05-17
+  preserved_intent: >-
+    Decision 55 established `agentera stats` as the user-facing analytics
+    workflow, forbade a top-level `agentera corpus` command, bound local
+    history to explicit consent, and wrote the internal corpus to a single
+    `$AGENTERA_PROFILE_DIR/intermediate/corpus.json` envelope. The intent
+    survives: analytics remain behind `agentera stats`/`agentera usage`, no
+    `agentera corpus` namespace is introduced, and local-history extraction
+    remains consent-bound and read-only.
+  superseded_surface: >-
+    The single monolithic `corpus.json` envelope is superseded by the tier
+    model below. Complete local evidence is retained as independently bounded
+    shards; a derived bounded signal view replaces the monolithic record array
+    as the analytic input. Decision history itself is not rewritten: the
+    decision record stands as adopted, and this reconciliation is recorded in
+    the tier authority rather than by editing the decision.
+  non_goal: >-
+    Do not rewrite, renumber, or restate Decision 55's adopted text. Do not
+    introduce `agentera corpus` or any new top-level analytics command.
+
+tiers:
+  full_evidence:
+    tier_id: full_evidence
+    rank: 1
+    definition: >-
+      Complete local evidence retained as independently bounded shards. Every
+      record ever extracted from a supported source is addressable here; nothing
+      is dropped to fit a reader limit. Each shard is independently bounded so a
+      single oversized shard cannot poison the whole tier.
+    stored_fields:
+      required:
+        - source_id
+        - source_kind
+        - timestamp
+        - project_id
+        - runtime
+        - source_class
+        - source_product
+        - active_runtime
+        - adapter_version
+        - data
+      optional:
+        - project_path
+        - session_id
+        - conversation_key
+    source_identity_rule: >-
+      `source_id` is the stable, content-addressable identity of one full
+      record. It is derived deterministically from source kind and source parts
+      and must resolve to exactly one retained full-evidence shard. No other
+      tier may invent identities that do not resolve here.
+    completeness_rule: >-
+      The union of full-evidence shards is the lossless local history for the
+      selected source scope. Consumers that need an entire record read it here;
+      consumers that need only a derived view read the signal tier instead.
+  signal:
+    tier_id: signal
+    rank: 2
+    definition: >-
+      A reproducible, bounded, derived view of the full-evidence tier. It
+      carries only the fields each declared consumer needs plus a resolvable
+      evidence anchor back to full evidence. It is small enough to load under
+      the reader bound even when the full-evidence tier is hundreds of
+      megabytes.
+    stored_fields:
+      required:
+        - source_id
+        - source_kind
+        - signal_type
+        - timestamp
+        - project_id
+        - runtime
+        - source_product
+        - evidence_anchor
+      optional:
+        - conversation_key
+        - session_id
+    derivation_rule: >-
+      The signal tier is a pure function of the full-evidence tier: identical
+      inputs produce identical signal membership and ordering. It never stores
+      transcript text, raw tool arguments, raw store paths, or raw session ids
+      beyond what a declared consumer requires with provenance visible.
+    evidence_anchor_rule: >-
+      Every signal record carries `evidence_anchor` resolving to the
+      `source_id` of exactly one retained full-evidence shard. A bounded
+      consumer that needs the full record follows the anchor; it must not
+      re-scan unbounded history.
+
+source_families:
+  coverage_rule: >-
+    The supported source families are fixed. A source family is either supplied
+    by an active runtime or imported as historical Claude history. No supported
+    family may be silently omitted from the contract, the coverage audit, or
+    the compatibility evaluation. An unavailable family reports a deterministic
+    discovery status; it is never absent from the map.
+  families:
+    codex:
+      source_class: active_runtime
+      source_product: codex
+      active_runtime: true
+      accepted_input: JSONL session/turn/response_item and tool_call records
+      store_glob: "*.jsonl"
+    cursor:
+      source_class: active_runtime
+      source_products:
+        - cursor
+        - cursor-agent
+      active_runtime: true
+      accepted_input: >-
+        Cursor project JSONL sessions and Cursor Agent CLI store.db SQLite
+        conversations. cursor-agent is a Cursor source product, never a
+        separate runtime identity.
+      store_glob: "*.jsonl"
+    opencode:
+      source_class: active_runtime
+      source_product: opencode
+      active_runtime: true
+      accepted_input: SQLite opencode.db with session, message, and part tables
+      store_glob: opencode.db
+    copilot:
+      source_class: active_runtime
+      source_product: github-copilot
+      active_runtime: true
+      accepted_input: SQLite session-store.db with conversation message rows
+      store_glob: session-store.db
+    claude-code:
+      source_class: historical_import
+      source_product: claude-code
+      active_runtime: false
+      inclusion_rule: >-
+        Explicit `--import-source claude` only. Excluded from default
+        active-runtime metrics and never described as runtime health, Claude
+        support, or active-runtime coverage.
+      accepted_input: JSONL project session files with user/assistant and tool_use events
+      store_glob: "*.jsonl"
+  filesystem_substrate:
+    note: >-
+      instruction_document and project_config_signal records are extracted
+      from the project filesystem (AGENTS.md, package manifests, config files).
+      They are source kinds, not runtime families, and follow source identity
+      and bounds rules like any other record.
+
+signal_semantics:
+  fixed_taxonomy_rule: >-
+    The signal semantics below are fixed. Current profile and personal-glossary
+    consumers depend on their meaning. A semantic kind present here must remain
+    derivable from a bounded signal record without restoring unbounded
+    transcript content.
+  kinds:
+    correction:
+      meaning: >-
+        A user signal that revises a prior position ("no", "not quite",
+        "actually", "instead"). Used to detect decision drift and overrides.
+      derivable_from: history_prompt and conversation_turn records
+      consumer: glossary
+    decision:
+      meaning: >-
+        A decision-bearing signal (decide, prefer, trade-off, scope, approve,
+        reject). The primary profile-synthesis input.
+      derivable_from: history_prompt and conversation_turn records
+      consumer: glossary
+      current_consumer: profile_synthesis
+    question:
+      meaning: >-
+        An interrogative signal ("why", "how", "should", or a "?"). Surfaces
+        uncertainty and information-gathering intent.
+      derivable_from: history_prompt and conversation_turn records
+      consumer: glossary
+    instruction:
+      meaning: >-
+        Explicit user-authored instructions (AGENTS.md and project memory
+        documents). The highest-signal, most durable evidence.
+      derivable_from: instruction_document records
+      consumer: glossary
+      current_consumer: profile_synthesis
+    configuration:
+      meaning: >-
+        Recurring config patterns across projects (manifest names, scripts,
+        dependencies, build/tool config). Objective evidence of what shipped.
+      derivable_from: project_config_signal records
+      consumer: glossary
+      current_consumer: profile_synthesis
+    record_identity:
+      meaning: >-
+        The stable, addressable identity (`source_id`) of one full record. Lets
+        any consumer cite evidence without retaining its content.
+      derivable_from: every record's source_id
+      consumer: glossary
+    date:
+      meaning: >-
+        The record timestamp anchoring evidence in time. Drives decay,
+        recency, and watermark windows.
+      derivable_from: every record's timestamp field
+      consumer: glossary
+    evidence_anchor:
+      meaning: >-
+        The resolvable link from a bounded signal record back to exactly one
+        retained full-evidence shard. Guarantees that bounded analysis never
+        severs the citation chain.
+      derivable_from: signal record evidence_anchor -> full_evidence source_id
+      consumer: glossary
+  preservation_rule: >-
+    Fixing the contract must not remove the required meaning of any kind above
+    from current consumers. Personal glossary synthesis is named here because it
+    actively consumes these bounded-history fields and semantics.
+
+consumer_map:
+  supply_rule: >-
+    Each supported consumer maps to exactly one bounded input contract that
+    supplies every required field with source identity. The input tier is the
+    smallest sufficient tier; consumers that need full records read
+    full_evidence, others read signal. No consumer may require a field not
+    declared in its input contract.
+  consumers:
+    usage_stats:
+      tier: signal
+      purpose: Skill-invocation analytics and USAGE.md reporting
+      required_fields:
+        - source_kind
+        - timestamp
+        - project_id
+        - runtime
+        - source_product
+      source_identity: source_id with source_product provenance
+      input_contract: signal tier conversation_turn projection
+    report_status:
+      tier: signal
+      purpose: Truthful corpus/coverage status (missing, stale, available)
+      required_fields:
+        - source_kind
+        - source_product
+      source_identity: tier presence and discovery status per source family
+      input_contract: signal tier metadata projection
+    prime_coverage:
+      tier: signal
+      purpose: Prime startup coverage of supported sources and selected scope
+      required_fields:
+        - runtime
+        - source_product
+        - source_class
+        - active_runtime
+      source_identity: available/selected/skipped per source family
+      input_contract: signal tier coverage projection
+    startup_analysis:
+      tier: signal
+      purpose: >-
+        Latent startup-analysis reader measuring state-gathering overhead. Reads
+        conversation_turn and tool_call signals; full tool arguments stay in
+        full_evidence.
+      required_fields:
+        - source_kind
+        - timestamp
+        - runtime
+        - source_class
+        - source_product
+        - evidence_anchor
+      source_identity: source_id and evidence_anchor to full evidence
+      input_contract: signal tier conversation_turn + tool_call projection
+    profile_synthesis:
+      tier: signal
+      purpose: Current profile synthesis (PROFILE.md) from bounded signals
+      required_fields:
+        - source_kind
+        - signal_type
+        - timestamp
+        - project_id
+        - runtime
+        - source_product
+        - evidence_anchor
+      source_identity: source_id and evidence_anchor for every synthesized claim
+      input_contract: signal tier instruction/decision/question/config projection
+  semantic_consumers:
+    glossary:
+      tier: signal
+      status: active
+      input_scope: bounded_personal_history
+      excluded_evidence_classes: [project_file]
+      required_semantics:
+        - correction
+        - decision
+        - question
+        - instruction
+        - configuration
+        - record_identity
+        - date
+        - evidence_anchor
+      rule: >-
+        Personal glossary synthesis actively consumes bounded personal-history
+        signals. Project-file provenance remains repository evidence under the
+        shared glossary ownership contract and is never represented as retained-
+        history evidence. Implementations must not drop a required semantic or
+        remove the field that carries it.
+
+profile_sufficiency:
+  question: >-
+    Does bounded profile synthesis preserve useful evidence distribution
+    compared with the current full-corpus intent? (planning Unknown 2)
+  profile_signal_types:
+    - decision
+    - question
+    - correction
+    - instruction
+    - configuration
+  prior_intended_distribution: >-
+    The prior intended distribution is the complete set of profile-relevant
+    signals (the profile_signal_types above) across all supported source
+    families, mirroring what a full-corpus read would have surfaced before
+    tiering. The signal tier's derivation is total when no cap is applied:
+    every full-evidence record yields exactly one signal, so the bounded and
+    intended distributions are identical. The comparison below is the safety
+    net that detects when signal selection reduces a family below a useful
+    threshold.
+  threshold_rule: >-
+    Profile synthesis is sufficient when the signal tier is uncapped (retained
+    == total) or every profile-relevant source family retains at least the
+    minimum representative fraction of its intended decision-bearing signals.
+    A family with zero intended signals is not a sufficiency violation; a
+    family with intended signals that retains none is.
+  minimum_family_retention: 0.5
+  minimum_family_retention_note: >-
+    The minimum per-family retention is 0.5 (half of the family's intended
+    profile-relevant signals). Below this, a family is underrepresented and
+    profile synthesis must surface the limitation rather than silently reduce
+    confidence for that family's contributions. The threshold is deliberately
+    lenient because proportional-per-family selection already preserves the
+    relative distribution; the check exists to catch a cap severe enough to
+    distort a family's evidence below a useful floor.
+  insufficiency_rule: >-
+    When insufficient, profile synthesis surfaces the underrepresented families
+    with their retained and intended counts, caps confidence for affected
+    categories to the single-signal band (CS4, 30-49) or below, and never
+    fabricates high-confidence claims from sparse or missing evidence. The
+    limitation is reported in the profile output and exit signal, not silently
+    absorbed.
+  derivation: >-
+    Resolves planning Unknown 2. The comparison is reproducible: identical
+    full evidence yields an identical sufficiency assessment because signal
+    derivation and selection are deterministic (Unknown 1). The prior
+    intended distribution is reconstructed from the signal tier's selection
+    report (total per family) and the retained distribution from the published
+    signal records (retained per family), so the comparison needs no separate
+    full-corpus read.
+
+bounds:
+  reader_byte_cap: 67108864
+  reader_byte_cap_note: >-
+    64 MiB. Existing bounded readers cap monolithic reads at this size and
+    safely degrade. The tier model replaces "safely degrade" with "useful
+    analysis" by feeding consumers bounded tiers that fit under this cap even
+    when complete local evidence is hundreds of megabytes.
+  shard_bound_rule: >-
+    Each full-evidence shard is independently bounded. A single oversized shard
+    is reported and recoverable; it cannot make the whole tier unloadable.
+  shard_byte_cap: 67108864
+  shard_byte_cap_note: >-
+    A full-evidence shard encodes no more than `reader_byte_cap` bytes. Shards
+    are grouped by source family; when a family's encoded records exceed the
+    cap the family is split into numbered sub-shards so no shard exceeds it and
+    no record is dropped. This is the measured bound a bounded reader loads.
+  signal_bound_rule: >-
+    The signal tier is bounded to fit under the reader cap. Selection must
+    preserve representative evidence across all supported sources rather than
+    silently truncating one source.
+  signal_byte_cap: 67108864
+  signal_byte_cap_note: >-
+    The whole signal tier encodes no more than `reader_byte_cap` bytes. A signal
+    record carries only the contract's required fields plus `evidence_anchor` and
+    no transcript text, raw tool arguments, or raw store paths; measured
+    per-record size is far below the cap for realistic corpora, so the natural
+    outcome is to retain every derived signal. The selection below is the
+    deterministic safety net that keeps the bound when a real-scale corpus would
+    otherwise exceed it.
+  signal_selection:
+    rule: >-
+      Applied only when the derived signal tier exceeds `signal_byte_cap`. The
+      derivation is otherwise total: identical full evidence yields identical
+      signal membership and ordering.
+    algorithm: >-
+      Group derived signal records by source family. Allocate a per-family record
+      quota proportional to that family's share of the total record count, with a
+      minimum of one record per family so no supported source is silently
+      truncated to zero. Within each family, keep the most recent records first
+      (timestamp descending, then `source_id` ascending as the deterministic
+      tie-break). Reassemble the kept records in the stable tier order
+      (timestamp ascending, then `source_id` ascending). The generation identity
+      is derived from the retained membership so identical input reproduces an
+      identical selected tier.
+    derivation: >-
+      Resolves planning Unknown 1. The bound is the existing reader cap reused as
+      the signal cap because a bounded consumer must load the whole signal tier in
+      one read. Proportional-per-family selection preserves the relative evidence
+      distribution across all supported sources; recency-within-family preserves
+      the most decision-relevant signals; the fixed `(timestamp, source_id)`
+      ordering makes membership and ordering reproducible.
+  real_scale_acknowledgement: >-
+    Real corpora exceed 500 MB. The goal is useful analysis, not merely
+    graceful failure; bounded tiers exist so consumers analyze real scale
+    rather than reporting unavailable or stale results.
+
+compatibility_states:
+  determinism_rule: >-
+    Every compatibility state has a fixed status, reason, recovery action, and
+    outcome. Recovery is actionable: it names the command or correction that
+    restores useful analysis. outcomes are limited to `recover` and `degrade`;
+    a `degrade` outcome must keep provenance visible and never fabricate
+    coverage. These outcomes are covered by passing and failing checks.
+  states:
+    oversized:
+      trigger: A full-evidence shard or legacy monolithic corpus exceeds the reader cap.
+      status: stale
+      reason: oversized
+      outcome: degrade
+      recovery: >-
+        Refresh bounded tiers with `agentera stats refresh --consent
+        local-history` or read the signal tier, which is bounded to fit. Do not
+        attempt to load the oversized artifact whole.
+      determinism: >-
+        Size is computed before read; the artifact is never partially loaded.
+        The signal tier remains usable.
+    legacy:
+      trigger: >-
+        A pre-tiering monolithic corpus.json exists but no bounded tiers have
+        been published.
+      status: stale
+      reason: legacy_monolithic_state
+      outcome: degrade
+      recovery: >-
+        Emit deterministic refresh guidance: run `agentera stats refresh
+        --consent local-history` to publish bounded tiers. Legacy monolithic
+        state receives refresh guidance rather than indefinite read
+        compatibility.
+      determinism: >-
+        Legacy state is detected by envelope shape and missing tier markers; it
+        is never silently treated as current.
+    missing:
+      trigger: No evidence artifact or tier exists at the resolved path.
+      status: missing
+      reason: no_evidence
+      outcome: degrade
+      recovery: >-
+        Report `agentera stats refresh --consent local-history` and the
+        `--corpus PATH` override. Must not auto-build, scan home directories, or
+        run live extraction without consent.
+      determinism: >-
+        Missing is reported as a concrete status with a path, never as an empty
+        success.
+    corrupt:
+      trigger: An artifact exists but is not readable JSON or fails schema validation.
+      status: stale
+      reason: unreadable_or_schema_divergent
+      outcome: degrade
+      recovery: >-
+        Refresh from source history. If a single shard is corrupt, the other
+        shards and the signal tier remain usable; only the corrupt shard is
+        flagged.
+      determinism: >-
+        Parse errors are caught per artifact/shard and reported with the
+        failing reason; one corrupt shard cannot crash tier reads.
+    incomplete:
+      trigger: >-
+        Tiers exist but a supported source family is unavailable, skipped
+        without acceptance, or extracted zero records.
+      status: flagged
+      reason: coverage_gap_or_sparse
+      outcome: degrade
+      recovery: >-
+        Surface the skipped/sparse family with its reason. Available skipped
+        sources require explicit `--accept-coverage-gap`; sparse (zero-record)
+        extraction is distinct from schema divergence.
+      determinism: >-
+        Coverage gaps and sparse extraction are reported per family with
+        discovery status; none is silently omitted.
+
+scope:
+  in_scope:
+    - Tier contract definition, source coverage, consumer field supply.
+    - Signal semantics reservation for current and deferred consumers.
+    - Compatibility state definitions and deterministic recovery outcomes.
+    - Small contract-model code and focused tests making invariants executable.
+  out_of_scope:
+    - Complete extraction writer implementation.
+    - Migration of all consumers to bounded tiers.
+    - Bounded profile synthesis runtime.
+    - Glossary implementation.
+    - Profile level selection.
+    - Suite version bump, release, or publish.
+  version_constraint: >-
+    Suite and version metadata remain at 3.0.0. This contract does not bump,
+    release, or publish anything.

package/bundle/references/analysis/startup-measurement-contract.yaml

@@ -15,7 +15,7 @@ boundary:
   rule: include records with timestamps strictly after committed_at
 
 section_22_compatibility:
-  consumes_adapter_version: agentera-v2-corpus-1
+  consumes_adapter_version: agentera-v3-corpus-2
   required_record_families:
     - conversation_turn
     - tool_call
@@ -39,7 +39,9 @@ benchmark_execution:
   live_host_checks: forbidden_by_default
   runtime_path_approval: >-
     Running mage bench:startupState authorizes inspection of the documented
-    default runtime stores for codex, claude-code, opencode, and github-copilot.
+    default active-runtime stores for codex, cursor, opencode, and copilot.
+    Claude history is excluded unless the caller separately supplies
+    --import-source claude with its concrete local path.
     Custom runtime history sources must be supplied as exact runtime labels and
     concrete filesystem paths. A generic consent flag or runtime label without
     its concrete path is insufficient for custom sources.
@@ -59,10 +61,10 @@ benchmark_execution:
 
 runtime_extraction_contract:
   supported_runtime_order:
-    - codex
-    - claude-code
     - opencode
-    - github-copilot
+    - codex
+    - cursor
+    - copilot
   normalized_record_fields:
     required:
       - source_id
@@ -70,6 +72,9 @@ runtime_extraction_contract:
       - project_id
       - source_kind
       - runtime
+      - source_class
+      - source_product
+      - active_runtime
       - adapter_version
       - data
     optional:
@@ -90,16 +95,16 @@ runtime_extraction_contract:
         parsing encountered schema errors. This is extraction failure evidence,
         not a successful zero-record benchmark window.
       required_counts:
-        - candidate_count when discovery can count candidates
+        - file_count when discovery can count store files
         - record_count, which may be zero or partial
         - error_count greater than zero
       current_examples:
-        - runtime: claude-code
-          candidate_count: 4836
+        - runtime: cursor
+          file_count: 4836
           record_count: 0
           error_count: 2
-        - runtime: github-copilot
-          candidate_count: 1
+        - runtime: copilot
+          file_count: 1
           record_count: 0
           error_count: 1
     no_matching_records:
@@ -110,7 +115,7 @@ runtime_extraction_contract:
         shape, but no supported message or tool records were extracted. This is
         sparse coverage evidence and remains distinct from schema divergence.
       required_counts:
-        - candidate_count when discovery can count candidates
+        - file_count when discovery can count store files
         - "record_count: 0"
         - no error_count greater than zero
     successful_zero_record_window:
@@ -127,7 +132,7 @@ runtime_extraction_contract:
         - runtime: codex
           record_count: 0
           error_count: 0
-  supported_runtimes:
+  source_products:
     codex:
       accepted_input_schema_classes:
         - JSONL session files containing session_meta or turn_context records
@@ -146,6 +151,9 @@ runtime_extraction_contract:
         - Do not retain cwd, project_path, store_path, file_path, raw session ids, or generated salted hashes.
         - Retain only canonical runtime labels, canonical artifact labels, bounded counts, and aggregate status fields.
     claude-code:
+      source_class: historical_import
+      active_runtime: false
+      inclusion_rule: explicit --import-source claude only; excluded from default active-runtime metrics
       accepted_input_schema_classes:
         - JSONL project session files with top-level role or type user/assistant events
         - JSONL message objects with nested message.role plus content/text/message content variants
@@ -180,6 +188,9 @@ runtime_extraction_contract:
         - Do not retain SQLite file paths, project paths, transcript text, raw session ids, or generated salted hashes.
         - Retain only canonical runtime labels, canonical artifact labels, bounded counts, and aggregate status fields.
     github-copilot:
+      runtime: copilot
+      source_class: active_runtime
+      active_runtime: true
       accepted_input_schema_classes:
         - SQLite session-store.db with sessions and turns tables
         - session id columns id, session_id, or sessionID
@@ -312,8 +323,8 @@ state_gathering_sequence:
   start_anchor: first_agentera_state_call_after_capability_related_user_turn
   capability_invocation_markers:
     - slash route such as /agentera plan or /agentera build
-    - bare capability name such as planera, resonera, or orkestrera
-    - natural-language capability mention such as run orkestrera
+    - bare capability name such as plan, discuss, or orchestrate
+    - natural-language capability mention such as run orchestrate
     - assistant capability intro marker
   end_anchors:
     - first_implementation_boundary
@@ -324,7 +335,7 @@ state_gathering_sequence:
 
 counted_event_classes:
   - class: cli_state_call
-    description: Agentera state CLI reads such as hej, todo, docs, plan, progress, decisions, health, and query.
+    description: Agentera state CLI reads such as prime, todo, docs, plan, progress, decisions, health, and query.
   - class: raw_artifact_access
     description: Direct read, grep, or glob access to Agentera state artifacts after CLI state access.
   - class: capability_prose_read
@@ -362,17 +373,17 @@ privacy_boundary:
     - canonical_event_class
     - salted_non_reconstructable_hash
   canonical_artifact_labels:
-    TODO.md: TODO.md
-    CHANGELOG.md: CHANGELOG.md
-    DESIGN.md: DESIGN.md
-    PLAN.md: PLAN.md
-    PROGRESS.md: PROGRESS.md
-    DOCS.md: DOCS.md
-    DECISIONS.md: DECISIONS.md
-    HEALTH.md: HEALTH.md
-    VISION.md: VISION.md
-    OBJECTIVE.md: OBJECTIVE.md
-    EXPERIMENTS.md: EXPERIMENTS.md
+    todo: todo
+    changelog: changelog
+    design: design
+    plan: plan
+    progress: progress
+    docs: docs
+    decisions: decisions
+    health: health
+    vision: vision
+    objective: objective
+    experiments: experiments
     AGENTS.md: AGENTS.md
     SKILL.md: SKILL.md
     AGENTERA_ARTIFACTS: AGENTERA_ARTIFACTS