vietreport-admin/solution-erp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[CLAUDE] Docs: S129 disposition owner-gated đợt 16-07 — (ii) bỏ trần giữ 1 số 100K hub-shape + (iv) provenance /3.0 + email hub a7aff375; (i)/(iii) CHỜ-HUB

Browse Source 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
pqhuy1987
2026-07-16 22:03:42 +07:00
parent 521565897b
commit 949ec33ee7
7 changed files with 72 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
.claude/agent-memory/memory-budget.json
View File

@ -62,7 +62,7 @@
"sample_questions": "eval/mfe/sample-questions.json",
"state_file": ".claude/agent-memory/.mfe-state.json",
"honest_caveats": {
"token_estimate": "char/4 is NOT real tokenization; VN-diacritic hot-memory ~3.0-3.5 byte/tok; report a RANGE [bytes/4 .. bytes/3.0]; FIT verdict uses bytes/3.0 (worst-case). Real tokenizer count lies inside the band -- never a single false-precise number.",
"token_estimate": "char/4 is NOT real tokenization; VN-diacritic hot-memory ~3.0-3.5 byte/tok (convention -- chua do tokenizer that tai SE; dinh-chinh 58f5afd8; real-N when it matters = harness Read-truncation notice per hub 'do token that' method, draft 2026-07-15, adopt-on-broadcast); report a RANGE [bytes/4 .. bytes/3.0]; FIT verdict uses bytes/3.0 (worst-case). Real tokenizer count lies inside the band -- never a single false-precise number.",
"judge_layer": "Branch-A recall/apply scorer is SCAFFOLD-only (empty). Self-grade same-session = meaningless (just read the answers). Numbers meaningful ONLY after (a) sample-questions mature in age AND (b) an independent cross-session/different-model judge scores.",
"age_band": "age is a FLAG, never a cut (mark RC-...10-29-11 age=false-proxy). An item drops only on status-change (mark->Disabled, guard->retired, AS row deleted), NOT by age.",
"read_only_scope": "MFE WRITES exactly one file: .claude/agent-memory/.mfe-state.json (its own strikes-state for cross-run Goodhart compare). 'READ-ONLY' is scoped to token_governor / the budget caps (owner-authority) which it NEVER writes -- it is NOT the claim 'writes nothing'."
@ -86,11 +86,9 @@
},
"harness_floor": {
"_note": "Harness-15 A1/A3 (S81, 2026-06-20): SAN-harness = fixed per-spawn cost (NOT tunable) = tool-schema + framing + own persona/role file + lead-pasted base-doc slice + task prompt. SEPARATE HOUSE (A3 anti-double-count): persona + lead-pasted-docs belong HERE (floor), NOT counted in token_governor.l1_always (which = own agent-memory + archive index + work-state block only). MEASURED-ESTIMATE not exact (H15 honest-note b): persona = directly measured bytes (.claude/agents/<name>.md 4.3KB-13.3KB => ~1.3K-4.0K tok via /3.3); tool-schema + framing = harness-injected (cannot byte-count locally), estimated comparable to AI_INFRA same-toolset-family (Read/Write/Edit/Bash/Grep/Glob/Skill/RAG).",
"measured_token_estimate": 55000,
"_s112_remeasure_note": "S112 (2026-07-12) RE-MEASURE — supersedes the S81 ~21K figure, which was a UNDER-COUNT (owner flagged 'why is SE so little vs sisters 60K+'). Byte-measured session-start DISK-inject floor = 57,595B (root CLAUDE.md 13,638 + docs/CLAUDE.md 10,222 + commands/session-start.md 23,669 + user-memory MEMORY.md 10,066) = ~14.4-19.2K tok ALONE. PLUS harness-INJECTED tool-schema + framing + MCP-instructions + deferred-tools + skills-catalog (~50 tool defs + 4 MCP blocks + ~90 deferred + ~60 skills), NOT byte-countable locally, ~30-50K tok. => FULL session-start floor ~45-70K tok, matching the fleet's measured ~60-71K (sister harness_floor_est @S103=71,364). The old 21K was the anomaly: S81 GUESSED the harness block at only ~5K(tool)+2K(framing) AND counted a small 'lead-pasted slice ~5-9K' instead of the full dual-CLAUDE.md + 23.7KB session-start body + user-memory inject. 55000 = honest mid estimate (~19K byte-measured + ~36K harness-injected-est). measured_token_estimate stays an ESTIMATE (harness block not locally byte-countable), not a false-precise number.",
"rounded_up_tokens": 100000,
"_owner_override_note": "S112 (2026-07-12) owner-directive: rounded_up_tokens raised 30000->100000 = OWNER-SET RESERVE (project-owner authority per token_governor.role_boundary_note + mark RC-pqhuy1987-21-06-2026-01-58-01), NOT a re-measurement (100K != a measured floor). NOTE: measured_token_estimate was subsequently re-measured 21000->55000 THIS SAME SESSION (S112, see _s112_remeasure_note) after the owner asked why SE's floor read so low vs sisters -- this override_note predates that re-measure; the live field is 55000. Rationale: reserve headroom for the grown fixed per-spawn overhead (roster 11->12 agents, +office-document, +2 monitor, more MCP tool-schemas, richer lead-pasted slices at the 380K read-set). CAVEAT (honest): harness_floor is an ACCOUNTING/AWARENESS figure (separate house, A3 anti-double-count) documented here; it is NOT actively subtracted by the current scripts (mfe-eval.ps1 / crystallized-backfill.ps1 govern via token_governor caps, not this floor). If owner wants 100K wired as an ACTIVE reserve subtracted from headroom, that is a separate change. Re-measure available on request to check the true current floor.",
"components_note": "persona(measured 1.3K-4.0K, DIRECTLY via Get-ChildItem byte-count of .claude/agents/*.md, NOT via measure-agent-memory.ps1 which only sizes agent-memory tiers) + tool-schema(est ~5K) + framing(est ~2K) + lead-pasted base-slice(est ~5-9K; em-main injects CLAUDE.md/README-slice + task ctx) + prompt(est ~1K). README decision-tree corpus = 32704B but em-main pastes a SLICE, not whole. The ~21K sum is SE's OWN reasoning over SE's OWN agents (persona byte-measured + each harness-injected sub-component estimated independently); it lands NEAR AI_INFRA's ~21K because the toolset family is identical (Read/Write/Edit/Bash/Grep/Glob/Skill/RAG), NOT because borrowed. SE's governing cap = 100K (owner-set reserve S112 2026-07-12; was 30K round-up), independent of AI_INFRA's figure."
"_owner_override_note": "S112 (2026-07-12) owner-directive: rounded_up_tokens raised 30000->100000 = OWNER-SET RESERVE (project-owner authority per token_governor.role_boundary_note + mark RC-pqhuy1987-21-06-2026-01-58-01), NOT a re-measurement (100K != a measured floor). Rationale: reserve headroom for the grown fixed per-spawn overhead (roster 11->12 agents, +office-document, +2 monitor, more MCP tool-schemas, richer lead-pasted slices at the 380K read-set). S129 (2026-07-16) owner-directive 'bo luon cac tran, chi giu tran cao nhat, thong nhat 1 cach voi hub': the parallel field measured_token_estimate RETIRED -- ONE floor number only, mirroring hub shape (hub _floor_reserve_tok = single 100000 + lineage-in-note, no stored parallel estimate; hub lineage ~21K measured -> anh-round 30K -> 60K s78 -> 100K s90). Retired-field history (verbatim archived at runs/2026-07-16-S127-adap-dot-16-07/owner-gated-proposals-dot-16-07.md disposition): S81 ~21K UNDER-COUNT -> S112 re-measure 55000 honest mid-estimate (~19K byte-measured disk-inject 57,595B + ~30-50K harness-injected est; full session-start floor band ~45-70K tok, fleet-parity ~60-71K). WHY retired: a stored 'measured/estimate' number invites the mislabel class (san-su-that 190c11ba; SE dinh-chinh 58f5afd8) -- when a real figure is needed, MEASURE FRESH via harness Read-truncation notice N (hub 'do token that khong can API key' method, draft 2026-07-15, adopt-on-broadcast), do NOT store-and-trust. CAVEAT (honest, unchanged): harness_floor is an ACCOUNTING/AWARENESS figure (separate house, A3 anti-double-count); it is NOT actively subtracted by the current scripts (mfe-eval.ps1 / crystallized-backfill.ps1 govern via token_governor caps, not this floor). If owner wants 100K wired as an ACTIVE reserve subtracted from headroom, that is a separate change. 0 script consumers of the retired field (grep-verified S127 + S129).",
"components_note": "floor COMPONENT LIST (what the fixed per-spawn cost consists of): persona (directly byte-measurable via Get-ChildItem .claude/agents/<name>.md 4.3KB-13.3KB => ~1.3K-4.0K tok; NOT via measure-agent-memory.ps1 which only sizes agent-memory tiers) + tool-schema + framing + lead-pasted base-slice (README decision-tree corpus = 32704B but em-main pastes a SLICE, not whole) + prompt -- the harness-injected parts are NOT byte-countable locally. Per-component S81 estimates and their ~21K sum = RETIRED history with the estimate field (owner 2026-07-16, see _owner_override_note). SE's governing number = rounded_up_tokens 100K (owner-set reserve S112), independent of AI_INFRA's figure (hub's 100K = their own measure+round, toolset family identical Read/Write/Edit/Bash/Grep/Glob/Skill/RAG)."
},
"token_governor": {
"_note": "Harness-15-v2 (S82, 2026-06-21): UPDATED by delta broadcast 2026-06-20-Governance-harness-15-v2-hot-feed-update (supersedes_scope = tier-1-sizing + L2/L3-caps ONLY; rest of H15 unchanged). TWO CHANGES vs S81: (1) Tier-1 = HOT-FEED LARGE per-role (was flat 12K -- too thin, caused lead to forget work across sessions); (2) L2/L3 caps REMOVED (on-demand, no artificial tier-limit, bounded only by model context window). Still the SECOND governor (token) ORTHOGONAL to the BYTE governor (tiers/archive_gate above) -- keep BOTH (B(e)); byte measures file-size-on-disk, token measures context-loaded; VN text ~3.0-3.5 byte/tok so byte/4 = upper bound => real headroom LARGER. Budget = MINIMUM-to-USE floor (FILL Tier-1 with real work-state up to the number; under-fill ONLY when high-value content exhausted; NEVER garbage-stuff -- token-saving = forgetting work).",