Next Steps¶
Released baseline: v1.39.36
Next implementation candidate: v1.40.0
This is Primr's canonical executable queue. It states the next bounded change, the order in which later release gates unlock, and the evidence required to advance. It does not assign delivery dates or effort estimates. Version labels are candidate ship points, not calendar commitments.
Use one source for each kind of information:
| Question | Source of truth |
|---|---|
| What should be implemented next? | This file |
| What must unlock before a later release band? | ROADMAP.md |
| Why is a workstream designed this way? | docs/design/ |
| What has already shipped? | CHANGELOG.md |
| What must the product remain? | Company Analyst Product Contract |
When these sources disagree, correct the stale source in the same pull request. Do not create another active queue in a design document or issue description.
Product filter¶
Every release must strengthen the same product: a bare company-and-website invocation produces an evidence-grounded long-form Strategic Overview and selected YAML strategy documents as reliable Markdown and Word artifacts. Primr is a research and consulting-prep tool whose reports mix cited facts with reasoned strategy that is usually not public. It is not only a fact checker: quality means grounded research plus sharper, more specific, more useful strategic thinking, never just more verifiable statements. Primr Zero stays the free-first agent-host path. Provider-backed execution stays estimate-first, explicitly approved, and budget-governed.
Architecture, providers, ledgers, memory, and agent protocols are supporting capabilities. They advance only when they improve report quality, execution choice, safety, reliability, or maintainability without creating a second product.
Current executable card¶
v1.39.36 is the released baseline. It keeps doctor, cancellation, and queue
accept honest, updates the optional Anthropic, A2A, and Google GenAI SDKs,
and adds an opt-in leadership brief from sources already in hand. That brief
does not search the web or change the Strategic Overview. v1.40 still waits
on the side-by-side read. v1.39.35 raises the soupsieve and dnspython
floors, installs CPython 3.15.2 in CI, and makes new requests use current
model ids. The Strategic Overview writing steps on this line are included.
v1.39.34 ordered the
roadmap around Primr's outputs: Overview depth, the strategy library, skill
packs, and Primr Zero parity. Model defaults and prices were rechecked
against every provider on 2026-10-29 with no catalog change. Longer Strategic Overview sections move
the default full run from $1.24 to about $1.26 ($1.08 without the AI Strategy).
Catalogs were rechecked on 2026-11-02. No newer generally available flagship
and no price change. Scheduled Gemini steps are unchanged.
A later owner decision, recorded 2026-11-14 and rechecked for Gemini 2.5 on
2026-11-16, moves the Pro default from Gemini 4 Argon to gemini-4-carbon
at the same $2/$10 rate through 2026-11-30 and $4/$20 from 2026-12-01.
Argon stays registered. Every xAI tier, including writing, uses grok-4.7.
Direct gemini-2.5-pro and gemini-2.5-flash shut down on 2026-12-01.
v1.39.33 states plainly that Primr's reports mix
cited facts with reasoned strategy and that Primr is not only a fact checker.
v1.39.32 refocuses v1.40 on making the
Strategic Overview, AI Strategy, and skill pack better, checked by side by
side reads instead of human fact adjudication. In v1.39.31, every default route uses the current
provider flagship (Grok 4.7, Gemini 3.8 Flash and 4 Argon, Gemini 3.5
Flash-Lite, GPT-6 Luna and 6.1 Terra, Claude Haiku and Sonnet 5.5) by owner
decision rather than a measured promotion, and estimates disclose scheduled
price changes. In v1.39.30 the evaluator harness adds an
authority_without_support variant, so v1.40 judges are also tested on
whether a trust cue substitutes for support. In v1.39.29, AGENTS.md is the single agent guide and
CLAUDE.md imports it; the operator guide is clients/primr-operator.md.
In v1.39.28, Word reports open with the title and
Executive Summary; recorded collection counts move to one closing note, and
writer section envelopes are parsed at the section boundary instead of leaking
as prose. Ambiguous envelopes withhold the polished DOCX. v1.39.27 makes
primr doctor warn when OpenAI's
free model list reports a shutdown date the registry does not record; the
check is report-only. v1.39.26 adds the library-only evaluator
validation harness described below and takes reviewed dependency updates; no
judge has been measured with it and it arms no gate. v1.39.25 raises the transitive filelock,
virtualenv, and dnspython security floors, routes OpenAI-only reasoning to
GPT-5.6 Terra after the o4-mini shutdown, and replaces retired OpenAI and xAI
image models. It does not advance the report-quality gate. v1.39.24 refreshes
current provider catalogs and standard/cache prices, fixes prompt-tier accounting and gateway ceilings, and
retains evaluated report defaults. It binds judge comparisons to the original
claims, verdicts, and source captures, rejects incomplete local judge plans, and
excludes failed or changed artifacts from agreement counts. It retains bounded
ingestion, native PDF containment, explicit browser sandbox defaults, and the
patched telemetry/environment dependency floors. It retains provider-specific
model retirement checks, explicit spend gates, the MCP session-resource security
floor, maintained Python 3.11-3.15 support, and the updated direct Gemini catalog.
Keyless startup, diagnostics, streams, collection signatures, and Word evidence
labels now follow the interface rules. Research stages have typed ownership
boundaries and an acyclic first-party import graph. Dependency proposals remain
reviewed and maintainer-authored. These maintenance changes do not advance the
report-quality or backend-promotion gate. The additional security review confines
agent artifact reuse by current-job ownership, bounds HTTP body expansion,
protects Windows scrape paths, and repairs QA and cloud-runner integration.
v1.40.0 candidate: Strategic Overview depth and quality¶
Objective: make the Strategic Overview the best consultant-prep document Primr can produce: a long-form report with clear sections, typically 50 or more pages, written the way a large consulting firm would prepare to understand a company, guide it, and hit the ground running. It serves any role, not only a CIO. It is not a financial analyst's report and not a fact lookup.
Why this is next: the Overview is the primary product and the foundation every strategy document and skill pack builds on. Reviews of real reports found genuine insight held back by fixable prompt rules.
What good looks like:
- A clear governing view of the company that the whole report develops.
- Every section earns its place and adds depth. When a theme recurs, later sections add a new angle, evidence, consequence, or question instead of restating it. Repetition is fixed by deepening sections, never by cutting length or dropping sections.
- Specific to this company: its names, numbers, mechanisms, and tensions, not consultant boilerplate that would fit any firm. Real financial and operating anchors where the company or its filings disclose them.
- Facts cited where sources exist, following good company-research practice; strategic reasoning written with confidence, without a hedge on every sentence.
How changes are checked: read before and after reports for the same companies side by side, blind to which is which, and keep the change when it reads better without losing depth or length. Model judges may help as a second opinion, not a verdict. Hard floors stay automated and blocking: valid citations, valid Markdown and Word artifacts, and spend and egress limits. Any paid run needs a fresh estimate, explicit approval, and a cost cap.
Explicitly outside this card: shortening the report, provider promotion, unapproved paid runs, a deterministic prose-quality gate, memory, and a new public command.
Implementation status (2026-11-16): steps 1 to 4 below shipped in v1.39.35 for the standard writer, the fast-mode system prompt, and the deep-research overview prompts (higher section floors, build-don't-restate, disclosed financial anchors, load-bearing labels, no stray scaffolding, and one analytical voice). The composed overview selects its own persona, and the section briefs use that same voice. Strategy prompts are unchanged. A follow-through on 2026-11-03 stops the composed overview, the fast trust polish, and the company-brief example from replacing a disclosed or reported figure with a wide band. v1.40 still waits on step 5, which compares outputs side by side. That comparison needs an approved, capped paid run because the Standard writer is provider-backed. Frozen evidence can avoid a new collection, and it does not make the rewrite free.
The intervention distinction in the AI Strategy card is later lane work. It does not finish step 5, and v1.40 still waits on the side-by-side read.
Next, in order (2026-10-29):
- Fix the prompt rules that cause restatement across sections: give each section ownership of the ideas it should develop and let same-part sections see the frameworks they build on, while keeping word floors and section count.
- Anchor financial and operating sections on disclosed figures where they exist, with ranges only for true estimates.
- Replace blanket per-claim hedging with confident analysis; keep citations and labels where they add value.
- Remove leftover scaffolding in rendered text (stray labels, duplicated validation lines).
- Compare before and after on the same companies, using frozen evidence where possible. Frozen evidence skips a new collection. The rewrite still needs an approved, capped paid run.
Follow-through after v1.40¶
- Then the strategy library, skill packs, and Primr Zero parity (v1.41 to v1.43), each checked by side by side reads. The AI Strategy card now asks for the smallest useful intervention. A job ad, a DNS signal, or a public complaint stays too weak to justify a skill or an agent. The role plan can now store an explicit link or an explicit refusal. Pack authoring still does not turn a weak signal into a skill. The paid side-by-side read remains later work on this lane.
- Then compare backends. Use frozen evidence and output contracts, record quality, reliability, latency, and measured cost, and make an explicit promotion decision. Preserve preview status when evidence is inconclusive.
Maintenance that may ship as patch releases meanwhile: client-facing report polish found in real output, a model-currency recheck against the provider catalogs before each release, and reviewed dependency updates.
Read-only UTCP discovery for cloud job status is not planned. No named
consumer is blocked on a UTCP manual. GET /status/{job_id} and the MCP job
check already return the owner-checked status envelope. A UTCP runtime would
be a new dependency, so it stays out of the package.
The evaluation plan owns measurement design. These steps do not authorize paid runs or create another active release card.
Parallel readiness lanes¶
These lanes may produce independently safe patch releases, but they do not reorder the current executable card or bypass their promotion evidence.
OpenRouter preview¶
The optional paid route is wired through CLI, MCP, A2A, supervised workers,
estimation, request-level price ceilings, privacy controls, exact-cost capture,
and runtime budget accounting. v1.39.13 makes dry-run budget decisions explicit:
--budget 10 reports a $10 per-run ceiling, whether the estimate fits, and
whether launch is executable. A smaller ceiling reports launch as blocked.
Remaining promotion gate:
- Keep hermetic routing, accounting, and budget regressions green.
- Run representative report-quality and estimate-versus-actual comparisons only after an exact estimate, explicit approval, and a cost cap.
- Retain preview status unless quality, reliability, and accounting evidence support promotion. A configured key or successful one-off run is not enough.
The Primr ceiling is per run. An OpenRouter account or key spending limit is a separate defense-in-depth control and does not replace Primr approval.
Maintained Python support¶
The maintenance lane covers Python 3.11 through 3.15, with newest stable 3.15 for development, release tooling, containers, and primary platform checks. Native dependencies must install across the claimed matrix. Browser sync runtime checks retain a fail-closed boundary for unverified greenlet builds; 3.15 needs greenlet 3.5.6 or newer. The isolated native fuzz engine remains on its hash-pinned CPython 3.13 wheel, with its own audit. These runtime changes do not promote a provider or report-quality gate.
Documentation integrity¶
The explicit primr draft --local BUNDLE path produces an unvalidated draft
from frozen prep text using an installed model. Keep runtime isolation, source
excerpt review, transactional MD/DOCX artifacts, and truthful coverage in the
free regression gate. Local draft and self-comparison eval rows remain
review-only; full-report or backend promotion still depends on the quality
instruments and independent comparisons above.
Every user-facing patch updates the smallest appropriate surface:
- Keep the root README as the front door.
- Put operator procedures in task guides.
- Keep only the current executable card here.
- Keep dependency order and long-range gates in the roadmap.
- Move completed implementation detail to the changelog.
- Update examples whenever the CLI or machine-readable contract changes.
Protocol and plugin currency¶
Track specification revisions, SDK releases, and host-client acceptance separately. Current MCP core compatibility does not imply Tasks, OAuth discovery, or every client extension. Agent Plugins schema validity does not prove executable availability, writable state, host authorization, or runtime support. Use the compatibility review as a dated evidence snapshot, then recheck primary sources when upgrading. Keep transport and installed-layout regressions in the normal free release gate.
Version gates after v1.39.36¶
Later bands advance in dependency order. A version is cut when its exit criteria hold, not because a date or effort estimate was written down.
| Candidate | Capability gate | Depends on |
|---|---|---|
| v1.40 | Strategic Overview visibly deeper and better (50+ pages, every section adding depth), confirmed by side by side reads | Stable report contract |
| v1.41 | Strategy library complete and role-agnostic, AI Strategy first, 30+ pages per document | v1.40 writing improvements |
| v1.42 | Skill packs tied to the company's roles and strategy | v1.41 strategy library |
| v1.43 | Primr Zero reaches the same artifact quality with no Primr API spend | v1.40 to v1.42 artifact contracts |
| v1.44 | Measured host-versus-cloud promotion decision and honest single-provider full-report execution | v1.43 parity reads |
| v1.45 | Durable agent job lifecycle and remaining MCP/A2A control-plane parity | Stable provider execution and existing authorization contract |
| v1.46+ | Governed run history, retention/deletion/export, progressive artifacts, and further cost levers | v1.45 consumption boundaries |
| v2.0 | Measured analyst quality, backend freedom, and safe delegation hold together | Completion of the required 1.x pillars |
| v2.x | Strategy Delta and measured repeat-engagement continuity | Proven report lift without weaker freshness, privacy, uncertainty, or cost |
| v3.0 | VLM-first extraction, vertical compounding, and stable post-artifact handoff | Mature 2.x evidence and contracts |
Release and version update protocol¶
Use this order for every release. It keeps planned versions distinct from published versions and prevents package, docs, tag, and registry drift.
- Name the candidate. Put one candidate version and one executable card in
this file. Keep
ROADMAP.mdCurrent State on the latest published version. - Implement atomically. Record user-facing changes under Changelog
[Unreleased]. Do not pre-claim completion in the roadmap. - Prove the exit criteria. Run focused regressions first, then the complete release gate. Record paid measurements only when separately approved.
- Choose the final semantic version. Use a patch for compatible fixes and polish, a minor for a compatible capability, and a major only for the stated product-contract gate.
- Synchronize version surfaces in one release change. Update
pyproject.toml,src/primr/__init__.py,uv.lock,CITATION.cff, plugin manifests, the OpenClaw image default, roadmap Current State and release ledger, the docs index currency line, and the dated changelog heading. - Advance this queue. Mark the released baseline and name the next candidate card before merging the release change.
- Merge through a green pull request. Delete the feature branch and verify
the exact merge commit through the independent
mainworkflows. - Publish from the green commit. Tag that exact commit, let trusted publishing create GitHub and PyPI artifacts, and verify their hashes.
- Verify the installed product. Upgrade the supported local launcher,
confirm
primr --version, and run a nonbillable smoke check.
If any publication or verification step fails, do not advance Current State or the next executable card. Correct the release record before starting new work.
Explicitly not next¶
- Another README expansion or competing roadmap summary.
- A provider promotion based on connectivity alone.
- A deterministic prose-quality gate or a lone LLM judge.
- Shortening a report or dropping sections to remove repetition.
- A claim-aware rewrite before the representative ledger evaluation.
- Memory before retention, deletion, provenance, and freshness rules.
- A generic orchestration framework, daemon, SaaS surface, or new runtime without a measured product bottleneck.
Standing validation policy¶
Use free local validation first. Deterministic checks guard spend, egress, filesystem writes, schemas, citation resolution, version integrity, and other prose-invariant structure. Content quality is judged by reading whole outputs side by side, blind where practical; model judges are a second opinion, not a gate. Every paid evaluation or run requires an exact estimate, explicit approval, and a cost cap.