nobrainer-tech-flow 0.0.0-stage → 2.4.0

package/docs/releases/v1.6.0.md

@@ -0,0 +1,64 @@
+# Version 1.6.0 — portable autopilot
+
+This record covers the reviewed source and local verification. The
+[GitHub release](https://github.com/nobrainer-tech/nobrainer-tech-skills/releases/tag/v1.6.0)
+records publication and its tag commit. A local version or passing test is not
+itself publication evidence.
+
+## What changed
+
+Clear coding and everyday tasks can execute directly, including chat-only and
+non-Git work. Larger tasks retain one clarification round when needed, concise
+TODO progress, existing authorization and a bounded corrective loop. Markdown
+state suffices for resume without native goals or telemetry. Optional subagents
+do not require persistent sessions. Existing project specs, docs and diagrams
+are reused; new scaffolding requires an actual acceptance need.
+
+The review records five baseline contract defects,
+their fixes and dated ecosystem research. The fifteen-skill inventory is unchanged.
+
+## Evidence and limits
+
+- Baseline `d94dd05`: 121 deterministic tests passed despite the source contradictions.
+- Final source and release documentation: 125 deterministic tests passed.
+- Codex CLI `0.153.1`: one four-case batch each for baseline/candidate with
+  requested `gpt-6-astra` and `gpt-5.6-luna`, low effort. Behavior was acceptable;
+  this establishes no improvement percentage or universal model claim.
+- A separate frozen four-case holdout passed all 16 prewritten criteria after
+  good/bad evaluator controls. One Astra batch, no retries, explicit source
+  injection, no tool calls; this is simulation evidence.
+  [Sealed cases, calibration, output and verdict](../evals/artifacts/v1.6.0/holdout/receipt.json).
+- Claude Code `2.1.241`: blocked before model execution by organization subscription
+  access. No API credential substitution; current Claude behavior remains unverified.
+- Actual non-Git artifact execution on Codex/Astra passed: a 43-word invitation,
+  three-item organiser checklist and completed existing TODO. Independent readback
+  checked every brief fact, unchanged skill hashes and absence of new Git/workflow
+  scaffolding. Artifacts and receipt.
+- Final deterministic, syntax and public-clean release checks are recorded in the
+  GitHub release and CI run for the reviewed commit.
+
+The CLI accepted the requested models but did not separately expose backend
+model identity. Global skill metadata affected input usage; the holdout emitted
+a description-truncation warning. These probes do not prove clean automatic
+discovery, alias selection, all client integrations or token-cost savings.
+
+Reproducible public artifacts:
+
+- Frozen four-case prompt
+- [Source SHA-256 manifest](../evals/artifacts/v1.6.0/candidate-source-hashes.json)
+- [Model receipts, usage and read commands](../evals/artifacts/v1.6.0/smoke-receipts.json)
+- Astra candidate output
+- Luna candidate output
+
+The source manifest is checked against the [frozen source snapshot](../evals/artifacts/v1.6.0/source-snapshot.zip)
+by the test suite. Its 31 files were recovered from release commit `d51d1e10db881182601afef80df9e3196d4779ec`
+and verified against the original manifest. Later source changes do not inherit
+these runtime results. Raw local CLI logs are not published because host metadata is outside the public package.
+Outputs and selected read receipts preserve the relevant bounded evidence.
+
+## Recovery
+
+Pin the reviewed commit or release. Version `v1.5.0` remains available as a source
+rollback; a focused revert restores prior behavior without changing private user
+skills or unrelated projects. Website publication uses its own scoped source,
+preimage backup and public readback.

package/docs/releases/v1.6.1-publication-readback.md

@@ -0,0 +1,53 @@
+# Version 1.6.1 publication readback — 2026-09-05
+
+The source release and matching website are published. This records delivery;
+it does not change the failed behavioral verdicts in [release evidence](v1.6.1.md).
+
+## Source
+
+- [PR 44](https://github.com/nobrainer-tech/nobrainer-tech-skills/pull/44) merged
+  at `2026-09-05T05:01:23Z`.
+- Merge and peeled `v1.6.1` tag: `807b6d7bb7e0ecccc074581c85d7679d44428ced`.
+- Annotated tag object: `02ea6bb0345d10a66c65a890136f9441bff3afc2`.
+- [Release](https://github.com/nobrainer-tech/nobrainer-tech-skills/releases/tag/v1.6.1)
+  published `2026-09-05T05:02:46Z`; not draft or prerelease.
+- Reviewed source commit `7ae0e95054bf29abb02d9ec4fcedb2947b1d3079` is an
+  ancestor of the merge with an identical tree. Both Linux and macOS
+  [release CI jobs](https://github.com/nobrainer-tech/nobrainer-tech-skills/actions/runs/33946027681)
+  passed; local validation, 126 tests and staged/scoped secret scans passed.
+
+## Website
+
+The website source merged at `c52b0160abc687a9ecd1ca6b274f86c71b1b13c5` on
+its existing production-source branch. Six site tests, static validation and
+independent review passed before deployment. A preimage backup preceded an
+additive transfer; unrelated commercial-site work was not deployed.
+
+Public readback verified 35 files with HTTP 200 and exact staged bytes,
+including the root sitemap and LLM index. The existing server-only route
+remained byte-identical. The changed public surfaces are:
+
+- [Skills 1.6.1](https://nobrainer.tech/skills/)
+- [Comparison discovery](https://nobrainer.tech/skills/compare/)
+- [NoBrainer in Codex](https://nobrainer.tech/skills/integrations/codex/)
+
+After the owner corrected the source classification, Codex moved from the
+comparison category to host integrations; its former URL redirects to the
+integration page. The original publication receipts above describe the initial
+deployment, not this later correction. The classification correction merged
+in website source commit `11e0fe716cbb50d5b3555d29d7d8d71a503e5e76`; its
+public readback verified all 36 staged files, including the integration, legacy
+redirect, sitemap and LLM index. Ten website regression tests passed.
+
+The source-to-change manifest and website AGENTS rule require an appropriately
+classified, discoverable destination: workflow comparison, host integration or
+attributed inspiration. Empty or missing coverage fails validation. Nineteen
+historical landscape entries retain their original dates.
+
+## Limits and recovery
+
+No fresh client-discovery, full runtime-compliance, quality superiority or
+resource-savings claim is made. The recorded Luna completion-label and scope
+failures remain failures after publication. Source rollback is `v1.6.0`;
+website recovery uses the retained preimage and an additive restore. New routes
+can be removed only as an explicitly scoped rollback action.

package/docs/releases/v1.6.1.md

@@ -0,0 +1,82 @@
+# Version 1.6.1 — clearer authority, sufficient checks
+
+This patch clarifies the portable instruction contract. It is **not a measured
+model-quality upgrade**. The [GitHub release](https://github.com/nobrainer-tech/nobrainer-tech-skills/releases/tag/v1.6.1)
+records publication and the exact tag; a manifest alone does not prove release.
+
+## Changes
+
+- Ultra explicitly respects user instructions over skill guidelines within the
+  host hierarchy and explains the exact source of a skill-caused pause.
+- Build stops expanding verification after acceptance and required checks pass,
+  unless a new change, failure or unresolved concern justifies more.
+- Setup reconciles stale or conflicting project rules before adding instructions.
+- Historical v1.6 runtime receipts verify their original frozen source archive;
+  later edits cannot inherit that runtime proof.
+
+The fifteen-skill portfolio, model selection policy and prompt-size caps remain
+unchanged. No standing agent team, new runtime, provider fork or extra skill was
+added. Existing flow diagrams remain accurate: `NOT_NEEDED`, no stages or edges
+changed. See the source-contract review
+and public-source research.
+
+## Fresh model evidence, including failures
+
+Requested model: `gpt-5.6-luna`, medium effort, Codex CLI `0.149.1`.
+Backend model identity was not separately exposed. The two injected skill bodies
+were frozen before testing; this does not verify automatic client discovery.
+
+| Test | v1.6.0 baseline | v1.6.1 candidate |
+|---|---|---|
+| Three paired four-case response simulations | 95/96; 1/3 runs pass all gates | 94/96; 1/3 runs pass all gates |
+| One sealed two-case response simulation | 15/16, gates pass | 14/16, completion-label gate fails |
+
+The candidate used `FINISHED` while describing only what it would do. No actual
+execution occurred in those simulations. Development failures include omitted
+explicit simulation/publication state. Scores and hard gates are separate;
+full category points do not override a missing required fact.
+
+**Measured promotion is rejected.** No further behavioral rule was added or
+tuned against the holdout. The existing, independently justified contract
+clarifications use their separate source-contract acceptance; these results
+support no quality, token-savings or model-equivalence claim.
+
+One separate actual local CSV repair produced a passing focused test and changed
+only the authorized implementation file, preserving the test and sentinel. Its
+**full scope acceptance failed**: the model ran read-only `git status` despite
+an explicit ban on Git operations. That command exited 128 outside a repository;
+no Git write or publication occurred. Artifact success is not a complete runtime
+pass. The one-case fixture does not cover arbitrary CSV behavior or other hosts.
+
+Cases, rubric, exact outputs, receipts and independent verdict
+are public, alongside [runtime artifact and scope results](../evals/artifacts/v1.6.1/runtime/verdict.json).
+The artifact hash manifest freezes
+these records; tests verify integrity and score arithmetic.
+
+Accounting: eight valid simulation calls, two excluded calls from an incomplete
+initial fixture, one pre-request configuration rejection, and one actual-execution
+call. No quality retries. Development input was 11,586 tokens per baseline call
+and 11,646 per candidate call. The execution turn reported 64,216 cumulative
+input tokens, including 48,128 cached, across tool turns. These are usage receipts,
+not peak context size, RAM measurements or a cost comparison. Raw host logs stay
+private; the published execution response has its temporary path redacted.
+
+## Deterministic and publication gates
+
+Both skill validators and all 126 unit tests pass on the reviewed source.
+Independent review accepts the contract patch and archive repair. Ultra remains
+1600 words, bootstrap 190 words, and identical AGENTS/CLAUDE 200 lines. Syntax,
+public-clean and remote publication checks are separate release gates. Passing these checks does not
+turn the failed behavioral gates above into passes. Exact release CI and
+publication state are available from the release and associated pull request.
+
+Website source coverage distinguishes workflow comparisons, host integrations
+and attributed inspiration. [Using NoBrainer in Codex](https://nobrainer.tech/skills/integrations/codex/)
+explains their complementary roles and cites official OpenAI guidance. Community
+prompt captures are qualified research leads, not authenticated vendor prompts.
+
+## Recovery
+
+Pin `v1.6.0` for the preceding source release. Use a focused revert for these
+instruction changes; preserve unrelated work. Website deployment requires a
+preimage backup and additive transfer, with public route and version readback.

package/docs/releases/v1.7.0.md

@@ -0,0 +1,42 @@
+# v1.7.0 — bounded commands and first use
+
+Date: 2026-09-05.
+
+Adds an optional executable helper inside the existing Sessions skill. The
+fifteen-skill portfolio and client-neutral routing remain intact. Installation
+now appears before architecture, with coding and everyday-task trials.
+
+## Verification
+
+- 140 deterministic tests pass, including 14 runner regression tests.
+- Canonical structure and suite validators pass.
+- An isolated copy installation includes the executable helper. Three actual
+  runs verify completion, timeout (exit 124), and output limit (exit 125).
+  [Machine-readable smoke receipt](artifacts/v1.7.0-copy-install.json) records
+  the exact helper hash, platform, limits and observed results.
+- Independent implementation and integration review covered process-group
+  cleanup, closed pipes, signal-resistant children, fresh receipt publication,
+  and fast-command deadline accounting.
+
+The smoke ran on macOS with Python 3.14.5. CI separately verifies its advertised
+matrix; local checks alone do not establish other client or platform support.
+The timeout includes a finite cleanup grace, so elapsed duration can exceed the
+configured execution deadline. Output forwarding occurs after child cleanup.
+
+## Scope and limits
+
+[Runner boundary](../BOUNDED_RUNNER.md), [first task](../TRY_IT.md), and
+Trending decisions describe usage.
+The research screened 435 discoveries across the collected views, then inspected
+selected sources; it was not a source-code audit of every repository.
+
+The helper requires POSIX and Python 3.11+. It bounds command execution time and
+captured output, not model tokens, billing, memory, filesystem or network access.
+Processes escaping the owned process group are outside its cleanup guarantee.
+A blocked downstream output consumer can delay receipt publication.
+No comparative model-quality, token-savings, universal client compatibility or
+security-sandbox claim follows from these checks. Previous model evaluations
+remain historical evidence and were not rerun or relabeled as v1.7.0 results.
+
+The GitHub release record identifies the published source commit. Website and
+scheduled-post publication require their own live readback after release.

package/docs/releases/v1.7.1-publication-readback.md

@@ -0,0 +1,64 @@
+# v1.7.1 publication readback — 2026-09-05
+
+NoBrainer Tech Flow v1.7.1 is published at
+<https://github.com/nobrainer-tech/nobrainer-tech-flow/releases/tag/v1.7.1>.
+The release tag resolves to `a5b2dcccae3aeb8e64f19f121f98f74f59719d82`.
+The old and new GitHub repository addresses resolve to the same release commit.
+
+## Product and installation
+
+- Public product: NoBrainer Tech Flow, “From task to done.”
+- Canonical site: <https://nobrainer.tech/flow/>.
+- The 15 module names, `nb-ultra`, package/plugin installation identities and
+  bootstrap marker remain compatible. This is a branding release; the bounded
+  runner and protected runtime files retain their v1.7.0 behavior.
+- Downloaded public v1.7.1 archive: suite validator passed; isolated copy
+  installation passed; the installed bounded runner returned `completed`.
+- Linux and macOS post-merge checks passed:
+  <https://github.com/nobrainer-tech/nobrainer-tech-flow/actions/runs/33951044769>.
+
+## Website and coverage
+
+The collected Trending snapshot covers 435 discoveries across repository and
+featured-developer views with daily, weekly and monthly ranges and selected
+language categories. Reconciled metadata screening and targeted pinned source
+review produced 40 canonical comparison pages. This does not mean 435 runtime
+audits or exhaustive coverage of all languages and future Trending results.
+
+The initial Flow deployment matched all 58 staged public files. Nine redirect
+checks covered root, deep comparison with query and asset paths on the two
+subdomains and former `/skills/` prefix. Both subdomains have working HTTPS.
+The pre-existing server-only Runpane legacy page remains available.
+
+The owner-requested comparison clarity update is also published. Website PR 8
+merged as `61bbda65f968368a59616f4edbbee40e6ef0d001`: all 40 source pages
+now explain concrete source strengths, both task routes, additional Flow value
+and practical fit, with details below. All 56 current Flow files matched the
+public readback; nine redirect cases passed again. The site validator, ten
+contract tests, ten desktop/mobile route checks and live mobile ECC readback
+passed. The library version stays v1.7.1.
+
+Verification digest and scope: [publication artifact](artifacts/v1.7.1-publication.json).
+
+## Social and weekly research
+
+Authenticated readback confirmed exactly seven existing scheduled X posts on
+@nobrainer_tech, September 5–11 at 18:00 Europe/Warsaw, with the updated brand,
+current version where mentioned and `/flow` links. Their dates were preserved;
+no duplicate posts were created. The saved public bio identifies the account as
+creator of NoBrainer Tech Flow. Historical published posts were not rewritten.
+
+The weekly ecosystem task is active for Friday 09:00 Europe/Warsaw, next run
+September 11, using Sol low. It reads the current published release and existing
+queue, limits deeper source reads, and prepares drafts without auto-publishing.
+
+## Proof limits and rollback
+
+No new controlled model-behavior benchmark was run for this branding release.
+The optional runner bounds one process group by time and captured output; it is
+not an account quota, token billing limiter, sandbox or persistent agent engine.
+
+Previous release tags remain intact. Website preimages and checksum manifests
+were retained before each transfer; deployment used checksum comparison and no
+remote deletion. Existing domains, installation identifiers and dated historical
+evidence remain available for rollback and compatibility.

package/docs/releases/v1.7.1.md

@@ -0,0 +1,49 @@
+# Version 1.7.1 — NoBrainer Tech Flow
+
+**From task to done.** Tell it what you need. Flow clarifies the goal, does the
+work, and checks the result.
+
+This is a product naming and documentation release. Maintained display labels
+use **NoBrainer Tech Flow**, repository links use
+[`nobrainer-tech/nobrainer-tech-flow`](https://github.com/nobrainer-tech/nobrainer-tech-flow),
+and the product website is [nobrainer.tech/flow](https://nobrainer.tech/flow/).
+
+Package and plugin identities remain `nobrainer-tech-skills`; marketplace IDs,
+all fifteen skill names, `nb-ultra`, adapter module paths and installer targets
+are preserved. The [migration guide](../MIGRATION_TO_FLOW.md) records each exact
+identifier and how to update an existing source without adding a duplicate
+installation.
+
+## Verification scope
+
+The skill bodies, client adapter code and bounded command runner are unchanged
+from [v1.7.0](v1.7.0.md). Their existing evidence remains scoped to the tested
+source and host. Historical release, review, specification and evaluation files
+retain their original bytes and provenance.
+
+Both skill validators and all **140 deterministic tests passed locally**,
+including the source check without Git metadata. Readback confirmed consistent
+1.7.1 manifests, preserved installation identities, identical 200-line AGENTS
+and CLAUDE instructions, unchanged runtime and historical files, and a passing
+whitespace check and scoped secret scan. Commit CI and publication readback remain separate;
+this document alone is not proof of publication, repository redirection or a
+client update.
+
+The workflow diagram still describes the unchanged `nobrainer-ultra` behavior;
+no new workflow diagram or runtime probe is required by this naming change.
+
+The completed local run passed **140/140 tests** and both validators. The
+Trending reconciliation records
+435 screened discoveries, resolved metadata gaps and the expanded public
+comparison coverage; it is not an external-runtime benchmark.
+
+## Publication and recovery
+
+The [GitHub release](https://github.com/nobrainer-tech/nobrainer-tech-flow/releases/tag/v1.7.1)
+and its reviewed tag commit establish source publication after release. Website
+deployment and the old repository/site redirects require separate live readback.
+No marketplace, model-performance or token-cost claim follows from the rename.
+
+Before updating, retain the installed source pin and any local configuration.
+To roll back, use the prior reviewed source with the same technical package and
+skill IDs; do not delete unrelated installations or change their paths.

package/docs/releases/v1.8.0.md

@@ -0,0 +1,63 @@
+# Version 1.8.0 — adaptive session restart and portable refinements
+
+This source record covers the v1.8.0 candidate prepared on 2026-09-06 from
+`origin/main` at `50de71c0c32f500da181f214062cc405b904a506`. It is a handoff for
+owner audit. It does not claim a merge, tag, GitHub release, marketplace
+distribution, client loading or clean-session runtime behavior.
+
+## Scope
+
+- The existing `nobrainer-sessions` portfolio now ships the adaptive
+  `session-restart` protocol and its deterministic Python decision helper. The
+  protocol requires a durable checkpoint, exact identity, verified takeover and
+  controlled retirement before the source session is archived. It does not add a
+  sixteenth skill, watcher daemon, scheduler, global hook or native session API.
+- The source-backed OpenAI Codex refinements already present in the base are
+  included: bound restart context fragments, preserve reusable input order and
+  bytes where a host supports reuse, bind review monitoring to the exact head,
+  exclude pending draft feedback, keep retries finite and diagnosed, and select
+  only review perspectives justified by the changed boundary. The adoption map
+  remains source-contract evidence,
+  not model-quality, token-cost or provider-runtime evidence.
+- Active package and adapter metadata is consistently `1.8.0` across the npm
+  package, lockfile, root plugin, Claude, Codex, Cursor, Kimi, Gemini and
+  marketplace manifests. The fifteen canonical skills, package/plugin IDs and
+  client-neutral adapter paths remain unchanged.
+- The active README, release notes, compatibility statement and session-restart
+  guide now describe the v1.8.0 source and keep the published v1.7.1 source as
+  the rollback anchor. Dated v1.7.1 release, publication, review and evaluation
+  records retain their original bytes.
+
+## Verification boundary
+
+The release-preparation checks cover portable source and repository behavior:
+
+- `python3 scripts/validate_skills.py` and
+  `python3 scripts/validate_skills.py --suite` validate the active portfolio,
+  manifests, links and public-clean rules.
+- `python3 -m unittest discover -s tests -q` covers the deterministic adapter,
+  installer, security, lifecycle, restart and suite gates.
+- The restart helper is invoked through its real Python CLI with a normalized
+  observation and must return its deterministic `CREATE_FRESH` decision; the
+  helper never creates, sends, archives or locks a session.
+- `git diff --check`, the active-manifest equality readback and the staged
+  gitleaks policy scan cover the handoff bytes before publication.
+
+The fresh run returned `OK` from both validators, `154 tests ... OK` from the
+full unittest suite, `CREATE_FRESH` from the helper smoke input and no leaks from
+the staged gitleaks scan.
+
+These checks establish source and repository proof only. No client was installed
+or exercised for this candidate, and no claim is made about native transport,
+all-client restart support, model quality, token savings, billing or marketplace
+availability. The owner must perform the merge, tag, release and any external
+readback separately.
+
+## Rollback
+
+The rollback anchor is the published `v1.7.1` tag at commit
+`a5b2dcccae3aeb8e64f19f121f98f74f59719d82`. Pin that exact commit (or the tag
+after verifying it resolves to the same commit), restore the v1.7.1 active
+metadata and verify the source with the checks belonging to that tree. Keep the
+v1.8.0 checkout and any audit evidence until the rollback is accepted; do not
+delete unrelated installations or rewrite historical release records.

package/docs/releases/v1.8.1.md

@@ -0,0 +1,32 @@
+# Version 1.8.1 — dated fresh-session titles
+
+This patch keeps the adaptive restart policy from v1.8.0 and adds a portable
+human-readable naming rule for clients that can create or rename conversations.
+
+## Behavior
+
+- There is no fixed restart interval. Adaptive mode assesses after accepted
+  milestones and context-pressure signals. Two compactions trigger assessment;
+  daily mode adds a 24-hour assessment trigger. Neither forces a restart.
+- Every fresh conversation keeps one stable task title and appends its own
+  ` | started DD-MM` date. The source conversation receives its own verified
+  start date when available.
+- Formatting removes an existing start suffix before adding the current one, so
+  retries and later handoffs do not stack dates.
+- The registry preserves the full timestamp and timezone. Session ID, goal ID and
+  checkpoint digest remain the identity and transfer evidence.
+- Unsupported title creation or rename is reported but does not block an
+  otherwise safe handoff. No universal title, session or archive API is claimed.
+
+## Verification boundary
+
+Repository validators and deterministic contract tests verify the documented
+fields, idempotent suffix rule, identity boundary, README diagram and synchronized
+manifests. They do not create, rename or archive a real client conversation.
+Native behavior still requires adapter and client readback.
+
+## Rollback
+
+The rollback anchor is the published `v1.8.0` tag at commit
+`13a5a6842678fff39b9c9f35040ddb9ba9d1ea8f`. Pin that tag and use its existing
+session-restart contract. Do not rewrite the historical release record.

package/docs/releases/v1.9.0.md

@@ -0,0 +1,33 @@
+# Version 1.9.0 - decision depth and feasibility gates
+
+This release adds one canonical decision skill contract with three explicit
+depths while preserving the fifteen-skill portfolio and existing technical
+package identities.
+
+## Scope
+
+- `just decide` and `just-decide` select a short LIGHT decision.
+- `decide`, `nb-decide` and `nobrainer-decide` select STANDARD analysis.
+- `deep decide` and `deep-decide` select HIGH analysis.
+- Ranking first checks feasibility, deadline, budget and minimum observable
+  reliability. It then compares total cost, time to useful output, growth,
+  reversibility and evidence-backed robustness or antifragility.
+- Panels, model changes and artificial score gaps are optional. A missing
+  decisive fact produces a bounded experiment or explicit unknown.
+
+## Verification boundary
+
+The release candidate passed both skill validators, the deterministic package
+suite, dedicated decision-contract tests, syntax checks, whitespace checks and a
+staged secret scan. A three-case MAIN holdout exercised the LIGHT, STANDARD and
+HIGH invocation routes and checked the expected output constraints.
+
+These checks validate the portable instruction contract. They do not prove
+automatic discovery in every client, model quality, decision quality in live
+production or any financial outcome.
+
+## Rollback
+
+The rollback anchor is the published `v1.8.1` tag. Revert to that exact source
+after verifying the tag and rerun its validators. Existing package/plugin IDs,
+skill paths and compatibility aliases remain unchanged.

package/docs/releases/v2.0.0-candidate.md

@@ -0,0 +1,28 @@
+# nobrainer-tech-flow 2.0.0 candidate
+
+Status: **owner review required**. This document is a proposed release note,
+not proof of a published tag, client upgrade or runtime behavior.
+
+nobrainer-tech-flow now uses `nobrainer-tech-flow` as both its repository ID
+and its explicit entry skill. The former `nobrainer-ultra` entry is recognized
+as a migration source, not a new invocation. Public package and plugin IDs
+also use `nobrainer-tech-flow`; existing client registrations need a reviewed
+one-time migration to avoid loading old and new adapters together.
+
+The eighteen-skill portfolio adds `nobrainer-auto-fine-tune`. It selects
+worker routes from the models and limits the active client actually exposes,
+benchmarks only when a bounded comparison is worth its cost, and preserves
+the user's selected MAIN model and effort. Project start now connects a
+short-term goal to the owner's longer-term direction, checks the existing
+layers and wiki, and keeps session takeover and archival evidence separate.
+
+The installer previews and applies a concise global instruction block for
+Codex, Claude Code, OpenCode or Copilot CLI. An installed Flow skill checks
+the latest stable release once per active day, with honest offline states.
+Safe auto-upgrade and session restart are explicit personalization options;
+an inactive client still needs an external scheduler for unattended checks.
+
+Rollback: retain the last verified 1.14.1 source, the exact prior
+client instruction files and installed skill locations. Do not remove the
+old package registration or entry skill until the new client loads one
+canonical Flow path and the owner accepts the migration readback.

package/docs/releases/v2.0.0.md

@@ -0,0 +1,53 @@
+# nobrainer-tech-flow 2.0.0 verification record
+
+Source version: 2.0.0. Publication is a separate check against the public tag
+and release archive; this document alone does not establish distribution.
+
+## What is included
+
+- One workflow entrypoint and 18 modules, including standard first-setup Auto Fine Tune.
+- Read-only repository/client preflight, stable selection IDs and situational recommendations.
+- Selected installation plus required entry/audit modules, client-specific personalization,
+  idempotent readback and guarded rollback.
+- Task-sized planning, SDD, wiki use, native-worker routing and integration contracts.
+- Optional bounded Jev/Laya decision adapters, explicit data/call limits and fallback.
+- Daily active-use version checking, browser preference guidance and safe worktree cleanup.
+
+## Reproduce the repository checks
+
+From the exact reviewed release checkout, run:
+
+```sh
+python3 scripts/validate_skills.py --suite
+python3 -m unittest discover -s tests -q
+```
+
+The suite includes isolated Codex and Claude homes, partial selection and required
+dependencies, multiple-client state, idempotence, preservation of later edits,
+rollback, offline preflight, update responses, decision-adapter error/fallback
+handling and worktree cleanup refusal/removal cases. It does not write the user's
+actual client configuration. The public tests are executable evidence; older
+internal review journals are not part of the current proof.
+
+For a narrower setup check:
+
+```sh
+python3 -m unittest discover -s tests -p 'test_guided_onboarding.py' -v
+python3 -m unittest discover -s tests -p 'test_personalization.py' -v
+python3 -m unittest discover -s tests -p 'test_typed_decisions.py' -v
+```
+
+## Evidence limits
+
+These checks establish local helper behavior and source contracts. They do not
+prove that every client loads the workflow, that a requested model was honored,
+that all hosts support parallel workers, or that live Jev/Laya inference improves
+quality, latency or cost. Auto Fine Tune records unverified values as UNKNOWN or
+UNMEASURED. Native tools, access and authorization remain host-dependent.
+
+The website's automatically animated portal scenario is a scripted illustration.
+Its worker states, console lines and verification steps are not a live run or a
+benchmark. A real task must supply its own inspected artifacts and readback.
+
+Before upgrading a real client, follow the migration guide, preserve its previous
+instructions and verify the exact new entrypoint in a fresh client session.