m149 c: bench joins the operator surface
c promotes bench into the cli help and the coverage registries, so the command is discoverable and its tests count.
loswfx
m149 b: apply shows its work before it writes
b makes bench report --apply preview the team.roles diff and only write with --yes.
m149 a: a bench that runs with nothing configured
a adds bench run --offline: a deterministic no-config, no-network default for ci and demos.
m149 start: harden the bench, then take it live
m149 hardens the model bench for routine use, then runs the cohort live for real per-role numbers.
m149 plan-ready: take the bench live
m149 hardens the model bench for routine use and runs the candidate cohort live for real per-role numbers.
m148 closeout: the model panel is a harness now
m148 turns model selection into a reproducible role-keyed benchmark scored on read-only-tool reliability.
m148 a: the cheap screen before the spend
a adds a promptfoo wide-net screen over the openrouter cohort, dry-validated in ci without spend.
m148 b3: the scorecard picks the team
b3 turns scored rows into a per-role recommendation and an applicable mixed-model team config.
m148 b2: the bench runs the real loop
b2 lands bench run: bind a model to a role, run the real agent loop, harvest the ledger into a scored row.
m148 b1: harvesting refusal from the ledger
b1 lands the harvest that reads one run's ledger and scores tool-call reliability, with a relevance/irrelevance metric.
m148: a benchmark that measures whether a model will actually read the file
m148 turns model selection into a role-keyed benchmark scored on whether a model uses read-only tools without refusing.
m121: shadow holds the keys, engagement carries the source
per-engagement provider secrets land in the shadow, never reach git history. cost source auto-derives from the engagement identity. m120's visibility surface is now operational.
m119: events tail closes the last sut gap
gap 12 closed minimally. all thirteen sut gaps from rounds 4 and 5 now durable scenarios. flywheel delivered seven closure milestones in two days.
m118: agent run learns to refuse over-budget dispatches
gap 11 was the highest-severity round-4 surprise. m118 closes it pre-dispatch: cap exceeded, no run, event recorded.
m116: engagement replay activated against synthetic ledger
gap 18 closed with one fixture step. read-and-print is what m94 actually ships — the verify flag was sut-doc imagination, not kernel.
m115: shadow surface activated against fixture tree
gap 16 closed with one fixture step. m86 query + m88 search + m91 drift all read from the workspace-local fallback tree — faithful exercise, not a mock.
m114: four round-4 gaps become durable scenarios
the capability shipped milestones ago; the scenarios never landed. m114 fixes that for four gaps in one cycle — no new features, only the coverage.
m99 lands: loswfx cost — summary, by-role, recommend
loswfx cost summary rolls up workspace usage × catalog cost; cost by-role breaks down per-role per-model; cost recommend surfaces cheaper alternatives. m96-m99 cluster closes.
m98 lands: loswfx roles configure — per-role provider/model
loswfx roles list renders the workspace roles table; roles configure <role> --provider --model writes the per-role override; loswf-roles-v1 envelope; 3 @m98 scenarios green.
m97 lands: provider + model catalog with cost metadata
loswf-catalog-v1 schema lands; per-model context window + cost per million input/output tokens + best-for hint; loswfx providers/models commands surface the catalog.
m96 lands: loswfx auth — credential store + resolution chain
loswf-auth-v1 store at ~/.loswfx/auth.json; add/list/remove CLI; ResolveAPIKey resolution chain — flag > auth.json > env var; 4 @m96 scenarios green.
m95 lands: engagement init --interactive — guided onboarding
loswfx engagement init --interactive prompts through engagement repo, shadow path/repo, skills path, roles; --answers <json-file> script-friendly path; engagement.init.interactive event records the flow.
m94 lands: loswfx engagement replay — deterministic decision-sequence playback
loswfx engagement replay --log <path> [--json] reads a ledger and replays the decision sequence; 2 @m94 scenarios cover fixture playback + missing-log error path.
m93 lands: loswfx milestone run --live — chained orchestration
loswfx milestone run --live flips the existing dry-run path into chained orchestration; helpful-next-action paths preserved; the kernel can now execute a milestone end-to-end without per-phase operator gating.
m112: live-engagement gaps become permanent kernel use cases
three M111 gaps fixed structurally + frozen as durable Gherkin scenarios. the live engagement is now a permanent SUT surface — each cycle feeds the flywheel.
m92 lands: SUT round-3 — gaps 7-10 closed (proposals quarantine, fake-pass, config preservation, skill-load)
round-3 closes the four round-2 carry-forward gaps: proposals/ quarantine, fake-provider silent-pass downgrades, legacy config preserved across quarantine, skill.load.succeeded/failed events made explicit.
m111: first live-engagement milestone — Phase 1 deliverables ship + SUT round 5 catches five gaps
real client work as SUT discipline. Phase 1 deliverables ship to iksnae/gitugh, shadow notes push to loswf/shadow-iksnae, five live gaps land in the kernel gap log.
m91 lands: shadow drift visibility — green/amber/red severity
shadow.drift reads commits-ahead + last-push-at against the configured remote; severity classifier maps to green/amber/red; loswfx shadow drift CLI surfaces the state.
m90 lands: side: dual frontmatter + atomic dual-commit primitive
side: dual joins the skill frontmatter validator as a third value; CommitDualArtifact runs the client + shadow commits as an atomic pair with recovered-state on partial failure.
loswf agency ships its first live client response — gitugh product design rfp
the kernel that wrote 110 milestones today just wrote a real client response. audit → research → product brief → engagement plan → response, surfaced as a PR.
m89 lands: promote proposal — the brain-to-body bridge (dry-run)
loswfx promote proposal --file <path> --dry-run reads a shadow proposal artifact and renders the would-be PR body with shadow citations as provenance; real PR creation is the m89-followup.
m110 lands: engagement init emits schema.legacy_read on normalized files
Gap 14a from SUT round 4 closes — init and the M84 probe path now emit the same schema.legacy_read event on legacy reads. operators audit both surfaces uniformly.
m88 lands: shadow search — case-insensitive substring + ranked hits
loswfx shadow search <query> case-insensitive substring scan across artifact bodies; sort.SliceStable ranks by match count; filter combines with m86 frontmatter k=v.
m109 lands: cost-allocation tags + the m100-m109 cloud arc closes
ten milestones in one grinding session. substrate-and-glue every step, boundaries defined behind m100's interface refactor. the kernel can deploy.
m87 lands: shadow branches — experiments without polluting main
loswfx shadow branch list/create/switch/merge/abandon — agents get full git semantics on shadow; lazy git init turns any shadow path into a real repo; ledger captures every branch lifecycle event.
m108 lands: OpenTelemetry-shaped tracing + JSON spans to stdout
Tracer + Span boundary lands. JSONTracer emits structured spans to stdout — CloudWatch Logs Insights queryable, OTEL SDK swap is a follow-up.
m86 lands: loswfx shadow query — cross-engagement read layer
loswfx shadow query [k=v ...] walks shadow across engagements and filters by frontmatter fields; the cross-engagement substrate becomes queryable; 3 @m86 scenarios green.
m107 lands: POST /api/milestone/run write endpoint
operator API gains a milestone-dispatch write endpoint. CLI + API now produce identical audit trails; only data.source differs.
m85 lands: schema migrate command — workspace modernization arc closes
loswfx schema migrate command + migration framework lands; M77 quarantine becomes the fallback path when migrate can't map; engagement.workspace.migrated event records the plan.
m106 lands: SSM Parameter Store joins the M96 auth chain
credential precedence is now flag > auth.json > SSM > env. cloud-mode kernels read per-engagement params via IAM-scoped roles; local mode unchanged.
m84 lands: tolerant config reader + schema.legacy_read event
config readers tolerate missing schemaVersion; loswfx schema legacy-probe reports legacy state; workspace.RecordSchemaLegacyRead emits an audit event so tolerant reads leave a trail.
m105 lands: Step Functions ASL for chained milestone execution
EmitMilestoneStateMachine — each scope item becomes a sequential Task state invoking the kernel Lambda. Retry policy + 4h timeout. Halt-and-resume deferred.
m83 lands: loswfx workspace unify — legacy .loswf/ migrates into .loswfx/
loswfx workspace unify migrates legacy .loswf/tools and .loswf/config.yaml into the unified .loswfx/ tree; engagement.workspace.unified ledger event records movedFileCount and configAbsorbed.
m104 lands: SNS fanout for ledger events
Emitter interface + EmittingStore decorator. SNS filter policies move fanout cost to the right place; ledger stays source-of-truth.
m82 lands: shadow-first harvest — audit/brief/plan move from proposals/ to shadow
repo-audit / product-brief / engagement-plan / build deliverable paths move from proposals/ to .loswfx/agency-shadow/; new @m82 workflow-coverage scenario asserts proposals/ stays empty post-build.
m103 lands: loswfx deploy template — SAM generator with cost defaults baked in
engagement → SAM template in one command. arm64, no VPC, S3 lifecycle, log retention, per-agent IAM. cost-optimization pillar is non-negotiable at the generator level.
m81 lands: workspace spec + drift validator — modernization arc begins
docs/loswfx-workspace-spec.md lands as the canonical on-disk shape; workspace-spec-current cardinal practice enforces drift detection; M82-M85 planning drafts queue the arc.
m102 lands: Lambda handler shim + Graviton container image
the kernel now ships as a Lambda function — Handler wraps RunKernelWorkflow, three event envelopes, arm64 container image. runtime identical to local; only the entry point changes.
m80 lands: readiness skills.path + work list partial-build signal
loswfx readiness Readiness panel + JSON gain skills.path coverage; work list annotates items with their most-recent agent.build.partial signal so silent dead-ends surface.
m101 lands: s3Store + CompositeStore behind the M100 interface
cloud ledger substrate — s3Store + composite primary/mirror — slots in behind m100's interface without touching a single call site.
m79 lands: agent run --skill resolves capability; mismatch refused
loswfx agent run --skill resolves the capability; CLI rejects conflicting --skill/--capability pairs; RunAgentCapability refuses on work-item skill mismatch.
m100 lands: ledger.Store is now an interface
first of the m100-m109 aws cloud-deployment arc — ledger.Store is now an interface; jsonl + mirrored-sqlite live behind it; 37 call sites compile unchanged.
m78 lands: kernel refuses workflow.Run on engagement-seeded work items
RunKernelWorkflow refuses engagement-seeded work items via skill-routing; SuggestedCapabilityForSkill maps skill→capability; engagement.workflow.refused event captures the refusal.
m77 lands: legacy-quarantine on engagement init — SUT round-2 gap 1 closed
engagement init quarantines pre-existing .loswfx state into a timestamped .loswfx.previous-<utc>/ directory; engagement.init.legacy.quarantined event lands; SUT round-2 gap 1 closes.
m76 lands: chained end-to-end engagement journey against gitugh
the engagement harvest closes: intake → audit → research → brief → plan → iteration → build → ship runs end-to-end against the gitugh fixture as a single chained godog journey.
m75 lands: ship phase activated — review + retro close the harvest pipeline
ship phase activates: scripted review capability gates ready-to-merge signal; retrospective runs against the engagement with What worked / What broke / Carry forward sections; phase ordering enforced.
m74 lands: build phase activated — verify capability + verdict artifact
build artifact records ≥1 intended SUT change + a verification command; the verify capability runs against the build's named SUT and writes a verdict artifact with PASS/FAIL status.
m73 lands: iteration-plan phase activated — child-promotion + Skill on WorkItem
iteration-plan promotes 3-8 child work items with effort estimates; canonical child-promotion write path lands; Skill persists on workflow.WorkItem; titles reference deliverables from the prior phase.
m72 lands: engagement-plan phase activated — composition gating + relative weeks
engagement-plan composes 3-7 numbered phases with Risks (≥2) and Out-of-scope sections; phase composition gates on prerequisites; timelines use relative weeks not calendar dates.
m71 lands: product-brief phase activated — frontmatter + required sections
product-brief joins simulateBuildArtifact; brief shape mandates frontmatter + Product / Customer goal / Source documents sections + README evidence row; open questions reference missing evidence.
m70 lands: research phase activated — skill-aware build simulation + brief shape
research phase activates: skill-aware build simulation routes per-skill output; brief shape mandates frontmatter + evidence table + 3 numbered findings; the brief lands in shadow not the tracked tree.
m69 lands: audit phase activated — skill metadata + 8 step-defs
audit phase activates with skill metadata on seeded work items, the SUT git-status cleanup, and 8 new step-defs (scripted-build-runs, file-exists-in-engagement-workspace, audit-frontmatter, contract-inputs-warns).
m68 lands: engagement intake closes — all four scenarios green
engagement.init.failed event lands; intake decomposes the engagement into 2-6 work items (audit, brief, research, plan, iteration, build, ship); all four intake scenarios green.
m67 lands: engagement intake activation begins
engagement init honors a fake-gh env override for fixtures, emits engagement.init.applied, gets new activation step-defs including aGhFixtureForGitugh — phase 1 begins.
m66 lands: engagement-pipeline gherkin harvest — sketches
eight engagement-phase feature files land as sketches: intake, audit, research, product-brief, engagement-plan, iteration-plan, build, ship — the spec layer for the M67-M75 activation arc.
m55 lands: dashboard arc closes — practices history + cost + receipt mix
practices-history.jsonl writer + dashboard practices/cost/mix subcommands close the M52-M55 dashboard arc; cross-engagement health, activity, cost, and receipt mix all answerable from one cli surface.
m65 lands: practices-history rotation + dashboard SKIP column
practices.RotateHistory keeps the history file bounded; dashboard practices learns a SKIP column so skipped checks become visible instead of folding into the OK count.
m64 lands: loswfx milestone grind — multi-cycle driver
loswfx milestone grind --cycles N enumerates N consecutive M<N> dispatches; halt-on-closed-highest detection so the driver stops when the queue is drained.
m54 lands: activity sparkline column on the cross-engagement table
ActivitySeries readmodel + Sparkline/BarChart primitives + the sparkline column lands on the M52 multi-engagement table; dashboard learns to draw.
m63 lands: milestone-run dispatcher — RunAgentCapability + --dry-run
loswfx milestone run dispatches the m62 milestone-execute capability via RunAgentCapability; --dry-run prints the would-be dispatch; --surface and --role overrides for testing.
m62 lands: milestone-execute capability — full interface + dispatcher wiring
milestone-execute joins the capability registry with Name/SystemPrompt/UserPrompt/Tools all wired; dispatcher recognizes it as a first-class capability.
m53 lands: foreman chat — transcript schema + bubbletea panel + dispatch wiring
loswf-foreman-chat-transcript-v1 schema lands in shadow; ForemanChatModel renders the bubbletea panel; ChatService wires prompt → existing M34-M40 dispatch path; foreman finally answers.
m61 lands: milestone-execute system prompt — the grinder loop in code
MilestoneExecuteSystemPromptBody constant encodes the operator's TDD-iterate-until-DoD loop as a capability system prompt; halt conditions named explicitly.
m60 lands: loswfx milestone — the grinder loop becomes a cli surface
loswfx milestone active surfaces the highest unclosed plan; loswfx milestone run [N] resolves and dispatches; resolveActiveMilestone helper lands; the grinder pattern becomes a kernel command.
m59 lands: contract-inputs-present — the 14th cardinal practice
VerifySkillContractInputs joins the practices catalog as the 14th cardinal practice; cli handler renders results; skips without SkillsPath; passes without warnings for non-contract skills.
m52 lands: MultiEngagementHealthView — the cross-engagement dashboard
MultiEngagementHealthView readmodel surfaces every engagement on disk with name/status/artifacts/receipts/failed/lastActivity columns; status derives mechanically from receipt + ledger state.
m58 lands: edit-source detector reads the ledger channel
the m50 edit-source detector grows a ledger-reader; rows tagged source: ledger merge with source: artifact rows so feedback cascade events flow into the same analysis.
m57 lands: loswfx flags discovered — registry snapshot + drift cli
loswfx flags discovered walks internal/cli for per-group flag discovery; --src override; --json stable map; --diff compares against the default registry.
m51 lands: loswfx health + practices.RunAll summary enrichment
loswfx health ships as the single-pane operator view: practices total/passed/warned/failed counts, 7d receipts grouped, recent activity top 5, structured verdict — OK / OK WITH WARNINGS / FAILING.
cadence pass: m60 → m65
six more milestones — the milestone-grinder loop becomes a loswfx capability: cli surface, prompt body, full Capability interface, --dry-run dispatch, multi-cycle driver, plus practices-history rotation + dashboard SKIP column.
m56 lands: VerifySkillContractInputs — the m50 contract data gets a verifier
VerifySkillContractInputs lands as the verifier any caller can gate on; required-with-zero-matches fails; optional inputs absent without failing; {repo} substitution honors empty placeholder.
m50 lands: contract migration complete + edit-source detector
the remaining 19 skills migrate to contract: frontmatter (27 total); schema gains a kind: marker (deliverable/methodology/verdict/criteria); the edit-source detector lands as the cross-engagement learning surface.
cadence pass: m56 → m59
four milestones in one cadence pass — contract data goes load-bearing, flag-registry gets a cli, edit-source detector grows a ledger channel, contract-inputs-present joins the catalog as the 14th validator.
m49 lands: skill contract frontmatter pilot + role-contract validator
the inputs/process/outputs/verification contract that lived as freeform markdown in 14 of 19 skills moves into machine-parseable yaml frontmatter; role-contract validator pilots against 8 skills.
m48 lands: trustworthiness arc closeout — owed tests + flag discovery + tracking-doc canonicalization
four M43-M47 carry-forwards land in one pass: agent-loop SubmittedTool gate gets a direct unit test, flag-registry gains per-group auto-discovery, 9 M4-M9 tracking docs canonicalized, contributor guide ships.
m47 lands: proof_shadow_repo activated + scripted-adapter key-precedence pinned
three @pending tags inside proof_shadow_repo cleared via system-prompt assertions; scripted-adapter cache reset centralized through a godog hook; five direct unit tests pin the matchKeys precedence owed since M45.
m46 lands: remaining proof_* activations + schema-deviation soft warning
two pre-existing proof_* features activated, the last @pending tag cleared in proof_triage_decomposition, scriptedtest gains typed submit-tool builders, and validators learn the soft-warning dimension.
loswfx v0.4.0 — orchestrator trustworthiness + native skill support
114 commits since v0.3.3 across two arcs — m41 through m48 trustworthiness + the parallel native-skill-support work.
m45 closeout: operator-facing scripted provider + phase scripts
phase script bundle (3 tandem + 8 onboarding), `loswfx config use fake-scripted` operator wiring, milestone-template canonization.
m40 closeout: foreman console lists and confirmations
bubbles v1 proposal selection + huh v1 confirmation. x is confirmation-first; cancel returns cleanly without executing.
m39 closeout: proper charm foreman console
native tui upgraded to bubbletea v2 + lipgloss v1; explicit foreman panel with keyboard-driven proposal execution.
m38 closeout: foreman proposal native actions
by-id shortcut for foreman proposal resume — `--approval-request-id <id>` resolves the command automatically.
m37 closeout: foreman proposal console read model
foreman session view now carries proposal lifecycle counts and renders exact resume commands inline.
m36 closeout: foreman proposal resume
approved foreman proposals can be resumed by approval-request id with exact-match consumption semantics.
m35 closeout: foreman action proposal
executable foreman previews can now be recorded as durable pending proposals, reusing the existing approval-request evidence shape.
m34 closeout: foreman action preview
foreman prompt responses now carry a typed action preview — label, command, safety status, source, work item, trace.
research note: icm folder-as-agent-architecture overlap
arxiv 2603.16021 surfaces a five-layer folder-and-markdown agent architecture. loswfx independently shipped most of it.
m44 closeout: scripted dialogues end-to-end
the scripted provider drives a full tandem phase end to end. four tracks landed; the eleventh cardinal practice locks acceptance-driven closeout.
m43 closeout: orchestrator trustworthiness
three drift validators land alongside bail visibility and a scripted-provider foundation. the orchestrator stops drifting silently.
m42 closeout: engagement as system under test
loswfx exercises its own pipeline as the system under test. practices enforced by code; gherkin coverage of operator surfaces broadened.
m41 closeout: native tui and cli surface overhaul
the operator console grew up — bubbletea snapshots, foreman sessions, capability catalogs, and trace correlation all rendered inline.
m33: foreman prompt interface
added the first explicit foreman prompt interface with deterministic intent classification and evidence context.
m32: foreman contextual next actions
added deterministic next-action selection to foreman sessions.
m31: foreman session threads and native console
derived foreman thread history from durable dispatch and queue evidence.
m30: foreman action dispatch and session persistence
promoted the first foreman-dispatched action with readiness gating and durable evidence.
m29: team bootstrap and foreman readiness
turned provider and team setup into a coordinated readiness surface — foreman.
m28: provider readiness and diagnostics persistence
hardened provider readiness with recorded probe results, custom endpoint probes, and gherkin use-case coverage.
m27: loop trust and ledger and api error semantics
addressed trust items from issue 10 with orphan transition diagnostics and 404 semantics for missing work items.
m26: guided local loop execution and operator ux correctness
made local loops scriptable and operator-friendly with setup gating and corrected policy-denial semantics.
m25: setup facts api contract alignment
promoted setup facts from m24 tui behavior into a read model and `/api/status` contract.
m24: tui readiness and guided local operations
made the native tui more useful for local bootstrap with a read-only readiness section.
m23: capability templates and project configuration bootstrap
removed the highest-friction manual configuration step with a helper command.
m22: local usage packaging and operator bootstrap
added the first supported local install path without a full release channel.
m21: sample project loop hardening
converted the manual sample-project smoke into repeatable product guidance and test coverage.
m20: external client consumption and control panel readiness
validated current api consumption from the external-client point of view with a deterministic client smoke and openapi fixes.
m19: agent harness expansion
added the first loswfx-owned agent harness slice with named capability runs.
m18: durable proposed-action execution
promoted durable tool_run proposed-action execution through the cli with exact command matching and approval consumption.
m17: local usage hardening and install readiness
turned the local smoke path into repeatable operator guidance with a readiness command and comprehensive docs.
m16: surface diagnostics and proposed-action semantics
turned m15 observable surface facts into a concrete diagnostics contract with an action eligibility vocabulary.
m15: multi-surface operational scale and proposed-action readiness
moved from single-surface contracts to multi-surface operator readiness with surface-scoped fixtures and operator journeys.
m14: openapi contract seed and client handoff readiness
turned the m13 external-client decision into a validated contract handoff with an openapi seed, error catalog, and fixture validation.
m13: external client readiness and use-case contracts
established the external-client readiness boundary with a contract audit, live-response-backed fixtures, and use-case coverage.
m12: operator service hardening and projection durability
hardened the operator service foundation with a durable jsonl queue projection artifact, freshness diagnostics, and high test coverage.
m11: bounded operator queue and approval execution resumption
converted m10 follow-ups into working foundations with bounded queue reads and approval-gated test-evidence resumption.
m10: durable approvals and client contract hardening
turned the m9 approval affordance into durable approval request and decision evidence with live-response-backed fixture tests.
m9: tui workflow console and external client readiness
turned the m8 tui foundation into a richer workflow console and proved an end-to-end operator loop with contract fixtures.
m8: native tui and service contracts
landed the tui-first native ui foundation and proved service contracts for future external clients.
m7: agent developer workflows and operator ux
turned the m6 provider and tool foundation into operator-ready workflows with adrs and a capability catalog.
m6: local provider harness and tool policy
moved from fake-backed provider execution to a real local-first openai-compatible harness with named tool policy.
m5: multi-agent factory loop
proved the first auditable multi-agent factory loop with planner, builder, and reviewer roles and bounded retry.
m4: provider runtime and execution hardening
hardened the provider runtime with a fake-backed execution adapter, normalized result shapes, and execution health.
m3: live projection and command center hardening
hardened the projection adapter boundary and added browser automation for command-center validation.
m2: operator command center and projection boundary
turned the local kernel into an operator-facing foundation with typed read models and a static command center.
m1: local kernel proof
proved durable local workflow state, routes, and bounded phase attempts with audit evidence and causal tracing.
loswfx, day one
the rfp went out may 22; the kernel-versus-host boundary was asserted in code 24 hours later.