SDK context budgets, cursor-chain efficiency, and MCP mutation parity

.agents/pm/features/pm-nilh.toon

@@ -6,7 +6,7 @@ status: closed
 priority: 2
 tags[4]: agent-ux,"area:cli",gh-issue,token-efficiency
 created_at: "2026-07-06T08:03:03.701Z"
-updated_at: "2026-07-31T14:16:08.723Z"
+updated_at: "2026-08-01T00:35:36.460Z"
 closed_at: "2026-07-20T10:01:06.896Z"
 author: maintainer-agent
 estimated_minutes: 120
@@ -23,9 +23,10 @@ dependencies[4]{id,kind,created_at}:
   pm-g9xk,related,"2026-07-20T08:49:49.562Z"
   pm-xm7c,related,"2026-07-20T08:49:49.562Z"
   pm-mpbb,implements,"2026-07-26T17:21:13.002Z"
-comments[2]{created_at,author,text}:
+comments[3]{created_at,author,text}:
   "2026-07-20T09:59:55.002Z",codex-20260720,"Evidence: mutation commands now default to id/status/changed_field_count, --no-changed-fields preserves the compact legacy item envelope, --full-changed-fields restores the full delta, and --id-only retains precedence. Isolated packed creates measured 71-byte compact vs 583-byte full output (87.82% reduction). Token-budget consumer and manifest were updated intentionally."
   "2026-07-31T14:16:08.723Z","harness:claude-code","Closed-item conformance check 2026-07-31, filed as pm-awe3t6. No reopen proposed - the CLI half of this item shipped and holds - but the parity claim in the title is inverted at HEAD.\n\nMeasured in one sandbox tracker, same type, same fields, same second. CLI pm create --json returns 72 bytes: {id, status, changed_field_count}. MCP pm_create with identical arguments returns 373 bytes: a result wrapper containing the full ten-field item echo, an empty warnings array, a next_transition object, and changed_field_count. That is 5.2x, and it is the MCP default; the lean form on MCP requires idOnly:true, which returns 45 bytes, while the CLI needs no flag.\n\nSo this item made the CLI lean and the direction of parity reversed rather than converged: the surface that is now verbose is the one where tokens are unavoidable, because every MCP tool result enters the model context whether the agent needed it or not, whereas a CLI caller can pipe through jq and discard. The result wrapper is a second divergence - the CLI emits no such key - which is the shape difference the agent guidance warns about."
+  "2026-08-01T00:35:36.460Z","harness:codex","Conformance correction 2026-08-01: pm-awe3t6 restores the parity this closed item intended. An isolated identical create now serializes to 57 bytes on both the CLI default and the MCP structured result, with the same id/status/changed_field_count keys. MCP protocol framing still wraps structuredContent.result as documented, while the mutation result itself has one cross-surface accessor shape; fullChangedFields:true remains the explicit full-envelope opt-in."
 files[6]{path,scope,note}:
   scripts/release/token-budget-gate.mjs,project,Compact mutation consumer compatibility
   scripts/release/token-budgets.json,project,Intentional agent output budget update

.agents/pm/history/pm-nilh.jsonl

@@ -16,3 +16,4 @@
 {"ts":"2026-07-26T16:52:59.224Z","author":"harness:claude-code","author_source":"detected","agent_harness":"claude-code","agent_instance":"ad0de54baae9cc12afd5a5a6","op":"update","patch":[{"op":"replace","path":"/metadata/updated_at","value":"2026-07-26T16:52:59.224Z"},{"op":"add","path":"/metadata/release","value":"v2026.7.21"}],"before_hash":"a1417a03edfd1fb52e09c66cca36fe08d5ef2d285f00d7d1d4a739555ab28518","after_hash":"4cbab13de3d0da18dd9e4ae810b8a881ccac584806e1a4fe28c132dd65409303","message":"Historical release attribution backfill (pm-3j6it6): stamp the release tag whose window contains this item close event, derived from its immutable history stream, so changelog attribution stops depending on updated_at"}
 {"ts":"2026-07-26T17:21:13.320Z","author":"harness:claude-code","author_source":"detected","agent_harness":"claude-code","agent_instance":"ad0de54baae9cc12afd5a5a6","op":"update","patch":[{"op":"add","path":"/metadata/dependencies/3","value":{"id":"pm-mpbb","kind":"implements","created_at":"2026-07-26T17:21:13.002Z"}},{"op":"replace","path":"/metadata/updated_at","value":"2026-07-26T17:21:13.320Z"}],"before_hash":"4cbab13de3d0da18dd9e4ae810b8a881ccac584806e1a4fe28c132dd65409303","after_hash":"e422423c03f54290d0352ec9369dbf88216d23fe4c523d3d6c99bb1c819294eb","message":"Terminal graph backfill (pm-lnrk / pm-3j6it6 lane): add 1 evidence-derived relationship edge(s). Evidence classes used: typed hierarchy mirror so semantic traversal reaches the owning lineage (parent alone is not a traversable semantic kind). Release attribution is stamped, so this metadata write cannot re-bucket the item in the generated changelog."}
 {"ts":"2026-07-31T14:16:08.723Z","author":"harness:claude-code","author_source":"detected","agent_harness":"claude-code","agent_instance":"4cc69f82c87e6e8164c8f176","agent_provenance":{"model":null,"effort":{"value":"xhigh","source":"environment"}},"op":"comment_add","patch":[{"op":"add","path":"/metadata/comments/1","value":{"created_at":"2026-07-31T14:16:08.723Z","author":"harness:claude-code","text":"Closed-item conformance check 2026-07-31, filed as pm-awe3t6. No reopen proposed - the CLI half of this item shipped and holds - but the parity claim in the title is inverted at HEAD.\n\nMeasured in one sandbox tracker, same type, same fields, same second. CLI pm create --json returns 72 bytes: {id, status, changed_field_count}. MCP pm_create with identical arguments returns 373 bytes: a result wrapper containing the full ten-field item echo, an empty warnings array, a next_transition object, and changed_field_count. That is 5.2x, and it is the MCP default; the lean form on MCP requires idOnly:true, which returns 45 bytes, while the CLI needs no flag.\n\nSo this item made the CLI lean and the direction of parity reversed rather than converged: the surface that is now verbose is the one where tokens are unavoidable, because every MCP tool result enters the model context whether the agent needed it or not, whereas a CLI caller can pipe through jq and discard. The result wrapper is a second divergence - the CLI emits no such key - which is the shape difference the agent guidance warns about."}},{"op":"replace","path":"/metadata/updated_at","value":"2026-07-31T14:16:08.723Z"}],"before_hash":"e422423c03f54290d0352ec9369dbf88216d23fe4c523d3d6c99bb1c819294eb","after_hash":"0aeb07e988b0e76f0baa62b125d9df196f6a49ef700235d1fc8264e12c47cf9b"}
+{"ts":"2026-08-01T00:35:36.460Z","author":"harness:codex","author_source":"detected","agent_harness":"codex","agent_instance":"264406b88560fcdcbb17052e","agent_provenance":{"model":null,"effort":null,"role":null},"op":"comment_add","patch":[{"op":"add","path":"/metadata/comments/2","value":{"created_at":"2026-08-01T00:35:36.460Z","author":"harness:codex","text":"Conformance correction 2026-08-01: pm-awe3t6 restores the parity this closed item intended. An isolated identical create now serializes to 57 bytes on both the CLI default and the MCP structured result, with the same id/status/changed_field_count keys. MCP protocol framing still wraps structuredContent.result as documented, while the mutation result itself has one cross-surface accessor shape; fullChangedFields:true remains the explicit full-envelope opt-in."}},{"op":"replace","path":"/metadata/updated_at","value":"2026-08-01T00:35:36.460Z"}],"before_hash":"0aeb07e988b0e76f0baa62b125d9df196f6a49ef700235d1fc8264e12c47cf9b","after_hash":"d2a03f14a4ce3a5403c09a9e64aacfa6366087722a642a2d01a00374944a4aa1"}