Why: The user asked for a researched design covering subagent use,
tiered model routing (Claude via dmxapi for complex work, MiniMax for
the rest), a millisecond/sub-second task sandbox (record-only), and a
verdict on why the agent still feels weaker than codex/claude-code/kimi.
Why: Four research threads (external frontier survey, subagent/routing
evidence, harness-vs-model attribution, local code audit and probes)
converge on role/stage routing instead of whole-task routing, and on
three corrections to the sandbox assumptions (overlay is not a snapshot;
snapshots defer conflicts to merge time; destroy must follow sealing).
Evidence: Local measurements are SELF receipts in the Dev task card
(bwrap namespace ~7ms, overlay same, docker ~142ms; MiniMax reasoning
split; dmxapi /v1/messages cache_read; EXT-16 sandbox missing the
python-docx layer). External claims carry per-item source URLs in the
proposal's section 8.
Impact: Documents-only addition: new proposal 2026-10-10 (E1-E5/S0
increments, D/R workstreams, sandbox tiers L0-L2) plus INDEX rows for it
and the previously unlisted G1 package. No product code changes; all
implementation increments stay unauthorized until approved.
Refs: proposals/2026-10-09-agent-quality-improvement-plan.md (P4/P5,
model A/B), 2026-10-09-ext-g1-b1-b3-exec-backend.md (EXT-16),
Dev task feature/agent-exec-routing-sandbox-plan.
Why: RDM G1/G2 are blocked on four base capabilities - the member agent
cannot see published project knowledge inside a run, plugin commands are
unreachable as tools, members cannot see or open conversations in shared
projects, and the isolated exec tier needs a backend seam.
Evidence: SELF - code citations verified at Dev main (runmgr.go:726-736
owner-only Retrieve; hostcaps.go:64-66 owner-only search; server.go:438-441
owner-only listProjects; store/projects.go:124-130 owner-scoped conversation
creation); bwrap availability tested on this host (2026-10-09).
Impact: Docs only. New proposal docs/proposals/2026-10-09-ext-g1-b1-b3-exec-backend.md
registering EXT-13..EXT-16 per the compat-policy template; no code change.
Refs: RDM plan V1.1 sections 3.3/3.4/5.3; owner Q1-Q8 (Q4=option A, Q2)
Why: The 2026-10-09 diagnosis on real run transcripts identified the direct
conversation path's failure chain and a phased plan was approved with six
explicit user rulings (open up tools, kernel-transcript-authoritative context,
markdown-source deliverables with all file types retained, gate demotions,
uncapped stronger-model evals, real-model eval required for behavior changes).
The plan needs a durable main record per the long-term-direction policy, and
the 2026-09-30 preview proposal's "fs tools hidden by default" ruling is
superseded by ruling 1, which must be discoverable from the index.
Evidence: SELF - the record cites the diagnosis session, the transcript
sources (pg1 + legacy SQLite, 62 assistant turns) and the before-baseline
report location on the task card; no code or pointers changed in this repo.
Impact: New proposal 2026-10-09-agent-quality-improvement-plan.md plus one
index row. Subsequent phase tasks reference this record instead of restating
the plan.
Refs: Dev task feature/agent-quality-phase01; supersedes the fs-tool default
in 2026-09-30-file-preview-and-reserved-capabilities.md §2.3.
Why: v0.2.0 is the hard prerequisite for RDM shared work, and the release
checklist referenced a stale SurfaceVersion and tag message.
Evidence: SELF - scope lines are tied to merged Backend PRs (#66-#110);
Backend #110 CI success; UI composite run 1369 shows 14 files/103 tests.
Impact: Adds docs/records/2026-10-09-base-v0.2.0-release.md and points the
compat policy and the EXT-1 proposal at the actual 2.5.0-experimental line.
Tag creation and push stay with the owner (dual-key rule).
Refs: Dev feature/base-v020-release; Backend #110
Why: U0 (merged in RDM) rules that domain logic stays downstream and the base needs seven minimal generic capabilities; the formal four-element applications are the only legal path to base changes.
Evidence: All current-state citations re-verified at Dev 34a2bbe (backend fe7c245); RDM requirement source pinned at RDM main ca72cfc.
Impact: New proposal doc + one INDEX row; no code, no surface change in this PR.
Refs: RDM U0 proposal section 5 / ruling D1; extension-compat-policy; phase-plan section 3.3.
Why: The RDM plan's 2026-10-07 snapshot (downstream at zero, N1 in
flight, O1/O3/O7 blocked) was contradicted by the RDM repo's 103 merged
PRs; the planning index registered this as an open reconciliation item
and named it the preferred follow-up to the stage-record task.
Evidence: isolated clone replay at RDM 4feb2f0 (pins=v0.1.0) - make
test exit 0, -ms1 PASS (same-source gold caveat recorded), ms0
acceptance PASS=9/SKIP=0 on a disposable postgres container, five
upload probes, and entrance probes (chat 400 vs commands 200); card/
PR snapshot JSON; all receipts SELF (RDM has no CI), stored under
docs/records/2026-10-08-rdm-status-*.json. The private minutes probe
receipt stays local by decision; only hashes and metrics are recorded.
Impact: adds plan SS2.3 (row-by-row re-verdict, MS1 seven-clause audit,
six registered gaps), marks SS2 as a historical snapshot, updates the
index milestone table and closes the RDM-snapshot open item, registers
the dual-main-record ruling and the missing-CI gap.
Refs: Dev task feature/rdm-status-review; Documents index next-step
(2026-10-08).
Why: The planning index listed the stage-record reconciliation as the
preferred next task: N1 WP6b/6c transfers, the Q1 platform-task vs
squad-run boundary, I1 reserved command consumption, and the I2 card's
dangling reference all feed N2/Q3 and RDM dependency judgments.
Evidence: Dev merges #230/#231/#234/#235/#237/#240/#242 verified in
main via git log --merges on 2026-10-08; stage claims are SELF-level
card statements, explicitly not replayed; intent command.go:18-25/78-84
checked for the Cancel/Handoff ruling.
Impact: Adds a status-backfill section (SS7) to the N1/Q1/I1 plan with
per-item done/transferred/todo ownership, an I2 status note in the
stack completion plan (the sole I2 main record), and INDEX navigation
updates; registers the truncated Q2/Q3 archive lines as a new open
item.
Refs: Dev task feature/agent-stack-stage-records; planning index
"next-step" recommendation (2026-10-08).
Why: documents main gained two plans after the index snapshot, and Dev has since merged N2, N3, Q2, and Q3.
Evidence: 21 markdown proposals are present at HEAD; merge SHAs were read from Dev origin/main git log --merges.
Impact: navigation only; proposal bodies and acceptance claims are unchanged.
Refs: Documents #43#44; Dev #244#247#248#249#250#251#253
Why: the index branch was cut before N3 and Q2/Q3 plans landed on main.
Evidence: merge-tree reported no content conflicts against origin/main 16795ed598.
Impact: bring those two proposal files into the index branch without rewriting history.
Refs: documents main
Why: N1-N3/Q1/I1-I2 are closed; per the completion doc dependency
order Q2 then Q3 are next. Both need plans grounded in current code:
squad has no topology abstraction (two modes, dormant Cond
combinators, dead FactItem.Verified), RespondDecision has no actor
attribution (the mailbox acceptance attack surface), WorkflowDef
execution compresses steps into SpawnSubagent, and adapters/a2a does
not exist.
Evidence: three-area survey with file:line citations at the pinned
baselines (agent-squad c462bc4, backend 6e02a1f); the review
equivalence baseline is quoted from backend runReview with its test
anchors; the P4 ruling is quoted verbatim from agent-harness.md.
Impact: documentation only; Q2 implementation starts after merge,
Q3 after Q2 closes.
Refs: 2026-10-02-agent-stack-completion.md Q2/Q3 sections
Why: N2 is closed; per the completion doc's dependency order N3 is
next and needs a plan grounded in current code (the crash window
between persist and intent commit is unprotected, the product
conversation path has zero durability, span names/attributes do not
follow the GenAI convention, the memory port does not exist).
Evidence: five-area survey with file:line citations at the pinned
baselines (agent-native 74aa60b, backend 7980a8a, core 0b5c774);
audit N1/N21 verified fixed, N6/N20 verified still open.
Impact: documentation only; work starts after this plan merges.
Refs: 2026-10-02-agent-stack-completion.md N3 section
Why: expose existing milestone records and distinguish source observations from combination tests.
Evidence: linked records exist at the fixed source version; proposal blobs and the observation receipt remain unchanged.
Impact: index wording and navigation only; no plan or acceptance decision changes.
Refs: docs/proposals/INDEX.md
Why: author and task-account names do not establish ownership of a plan.
Evidence: independent review found four missing owner fields in the adopted set and N2 also has none; fourteen source headers explicitly name weichen.
Impact: correct the index attribution and clarify the local N2 document adoption boundary; source proposals remain unchanged.
Refs: planning-index review P2; docs/proposals/INDEX.md
Why: make existing directions, milestones and tasks discoverable without treating document claims as acceptance.
Evidence: fixed file blobs, 55 PR API records and Git ancestry are preserved as SELF observations; document links and all 19 proposals are checked.
Impact: navigation and evidence only; original plans, approval scope and business code are unchanged.
Refs: Dev feature/planning-index; docs/proposals/INDEX.md
Why: N1/Q1/I1/I2 are closed; per the completion doc's dependency order
N2 is next and needs a plan grounded in current code (the audit's N4 is
still open; the product-side ctxasm trim is the real gap for acceptance
clause 2; MCP 2026-07-28 features are absent).
Evidence: six-area survey with file:line citations at the pinned baselines
(agent-native ead690b, backend 68addfd); audit items N2/N3/N18/N19/N22
verified fixed against current code, N4 verified still open.
Impact: documentation only; work starts after this plan merges.
Refs: 2026-10-02-agent-stack-completion.md N2 section
Why: The further-objective docs are now the system's next big-phase
target; the plan pins the two-tree-one-net product on top of the
generic platform with two hard invariants - downstream-only domain
logic and always-degradable-to-generic-agent when the plugin is off.
Evidence: docs-only change; milestone mapping verified against
further-objective 10 (W1-W16), readiness claims verified against the
pinned worktrees (EXT-1 symbols in surface.txt, v0.1.0 tags, RDM repo
seed facts as recorded in vision doc 13).
Impact: Adds one proposal; no code changes.
Refs: docs/proposals/2026-10-07-agent-stack-n1q1i1-plans.md
Why: The completion proposal defines the N1/Q1/I1 phases but not their
work packages; D1 has landed, so the three phases are now unblocked and
need file-level plans grounded in the current pins.
Evidence: docs-only change; file:line claims verified against the pinned
worktrees (native 3a110a7, squad 5a57636, intent 65aaebc, backend
e7fb276).
Impact: Adds one proposal and one link line; no code changes.
Refs: docs/proposals/2026-10-02-agent-stack-completion.md
Why: downstream parsers are unreachable through the Host handle and the base lacks xlsx/pptx parsers; MS1 needs both. The snapshot also missed interface methods, so adding one was undetectable.
Evidence: request follows compat-policy sect.6 four elements; behavior contract and limits in the proposal body (documents has no CI; verified via git diff --check recorded on the Dev task card).
Impact: docs only; compat-policy sect.3 gains one sentence, usage.md format list gains xlsx/pptx.
Refs: docs/proposals/2026-10-06-ext-5-host-parse-document.md
Why: AGENTS.md requires PROJECT.md but the referenced project facts and verification entry did not exist.
Evidence: SELF git diff --check and checked the README reference and Dev verification targets.
Impact: adds documentation-specific validation and evidence limits; no runtime change.
Refs: Agent-System-Dev feature/governance-reliability audit remediation
Why: The comparison has to say what was borrowed, what was refused, and which product behaviors changed, without copying upstream text.
Evidence: The record lists the verified security gaps and the features implemented in this task. Proposal section 2.5 links to it.
Impact: Deferred and rejected items stay deferred and rejected. Adopted context and secret behavior is no longer described as a later batch.
Refs: feature/sillytavern-review
Why: The October 3 durable note described Pico5 before packages 1-23 existed, and the stack proposal had no place for the v1.0.1 and ZCode v3.14.3 findings.
Evidence: The new record, the correction at the top of the durable note, and proposal section 2.4 link to each other. No code changes.
Impact: Documents only. Later batches can cite the adopt, hold, and reject rows.
Refs: Dev feature/pi-zcode-second-pass
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The review record still said the proposal was unmerged, but Documents PR #34 is on main and this branch is cut from that commit.
Evidence: origin/main is 76044b5, the merge of feature/agent-stack-blueprint.
Impact: One sentence in the review record. The proposal text is unchanged.
Refs: Dev feature/pi-durable-borrow
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Pico5's package is a contract sketch, but its commit and recovery rules match four verified gaps in how runs survive a hard kill.
Evidence: Reviewed earendil-works/pi @ c7cdb460a, packages/durable and docs/pico-v5.md, against the skeleton section 4.2 and the current backend store, run manager, and SSE handler.
Impact: Documents only. The record rejects taking the package as a dependency and lists what the matching backend task changes.
Refs: Dev feature/pi-durable-borrow
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The three libraries cannot be reused or recovered independently, and the tested dependency combination is not the one the product builds.
Evidence: Read-only audit at native c6d867c, squad 590fd86, intent 1dbbe81, recorded with file references in docs/records/2026-10-02-agent-stack-audit.md.
Impact: Documents only. No library code or gitlinks change in this commit. Later stages stay separate tasks.
Refs: Dev feature/agent-stack-blueprint
Co-authored-by: Cursor <cursoragent@cursor.com>
weichen
