Repository navigation
Releases
Mythify v7.0.0
Mythify 7.0.0 narrows the tool to its job: evidence for agent work. Product
planning is removed, which is the breaking change behind the major version.
route can also show what your own routing policy file says about a task,
and the README is now a short product page with the full feature guide in
docs/guide.md. Existing .mythify/ state keeps working; leftover product
data is ignored and never deleted.
Added
routeshows what the user's own routing policy file says about the task. With
MYTHIFY_AGENT_POLICY(a path) or~/.agents/policy.jsonpresent, the text
output ends with aDelegationsection and--jsongains adelegation
object: the matched policy step and action, the role with its options and
effort (the role'strickyeffort on high risk, the tier's reviewer effort for
the reviewer role), and the review tier for the task's risk with every extra
tier key printed as written. It is advisory: options named in the file's
banned-model list are withheld, an effort above its ceiling or in its
never-list is lowered to the highest allowed level (or withheld when none
is), banned names and forbidden effort words in any printed policy text are
replaced with[withheld](a name containing one is partly withheld), and
Mythify still never picks, ranks, or switches a model. Without a file the
output is byte-for-byte unchanged; an unreadable or invalid file prints one
[WARN]line and is ignored. Step 1 (production, deploys, servers, or money)
prefers recall: a clause whose operative verb is an operation verb (deploy,
ship, release, publish, promote, restart, roll back, stop or start a service),
ssh, scp, or rsync as the main verb, a migration against production, or an explicit
money phrase is step 1, with or without an environment word. Only clear code
work stays build: a coding verb whose object is a code noun that the operation
word modifies (a deploy script, an ssh client); an ambiguous prompt goes to
step 1. The delegation review tier is at least high for step 1, and also when
the wording names a risky-code term from the newhigh_tier_termslist (auth,
billing, checkout, race, migration, secrets, and similar, whole words only),
with the reason in the basis; route's own classification is unchanged.
ssh, scp, and rsync are operation verbs whatever follows them, stop, start,
restart, reload, or kill plus a service, server, container, or worker is step
1, run or apply a migration with an environment word is step 1, "and" before
an operation verb starts a new clause, and documentation words (docs, typo,
notes) cancel step 1 only together with a coding verb. The image step also takes icons
and favicons. The task-to-step terms for production, image,
bulk, and browser_check live in a newpolicy_stepssection of
protocol/classification-rules.json(schema 6) and do not change the
classificationblock. The protocol gains one sentence in "Delegation is your
choice": follow the user's or project's rules on who does what, whichroute
shows when a policy file is present.
Changed
routesends requests to write a product roadmap, PRD, RFC, specification,
requirements document, or design document to planning (plan); without the
product task type they had classified as trivial. These terms count only
near a creation verb (write, draft, create, prepare, author, produce,
outline, plan), so "What is an RFC?" still routes as a question and "inspect
the product requirements document" as a review.- The README is a short product page: the problem, how it works, who it is
for, and a quickstart that ends at a first verified step. The full feature
tour moved to the newdocs/guide.md. - Maintainer files moved off the top level:
MAINTAINING.md,
RELEASE-CHECKLIST.md, andROADMAP.mdnow live indocs/maintainers/.
User-facing docs no longer carry version numbers or release-history notes;
the CHANGELOG and the architecture decision log keep that history.
Removed
- Breaking: product planning is removed.
mythify product ...(create,outcome,non-goal,bet,risk,
check,approve,decide,promote,measure,show,list) is an
unknown command and exits 64, like any other removed command. Also removed:
the elevenproduct_*MCP typed tools (themythifytool no longer lists
producteither), theproductroute and theprompt productpacket
(routeno longer returnsproduct, and prompts about roadmaps, priorities,
or product strategy now route by the remaining rules, typicallymapor
plan), theproducttask type and its classification terms, the
productlineage kind (--parent product:NAMEandlineage attach product
are refused), the product lines instatus,summary, andreport
(summary --jsonloses itsproductskey andstatus --jsonits
active_productkey),plan show's product line, the "Product planning"
section and product lines in the protocol (AGENTS.mdfollows), the product
parts of the skills and docs, anddocs/product-planning.md. Why: Mythify's
job is evidence for agent work; product strategy belongs in your own tools.
Leftover state is ignored and left in place: a.mythify/products/folder,
anactive_productpointer, and a plan'sproduct_refare never read,
written, or deleted, and ledger records with aproductkey stay scoped
records that never count for a step.ROADMAP.md, whichproduct show --markdownrendered, is now a static file of outcomes, non-goals, and bets
atdocs/maintainers/ROADMAP.md.
An agent following the old protocol text that runsproduct createnow gets a
usage error and should usemap createfor open decisions orplan create
for work. docs/DRIFT.mdanddocs/evidence/(efficacy reproduction notes and smoke
results), which described past releases rather than the current tool.
Full changelog: v6.2.0...v7.0.0
Mythify v6.2.0
Mythify 6.2.0 makes verification faster and more selective without trusting
self-report. It can reuse an identical passing run, flag a check that ran
fewer tests, name project checks, gate host hooks on evidence, bound review
loops, judge risk from the changed paths, and measure the flow. It adds
commands, flags, and record fields only; the new completion checks warn by
default and refuse only under strict context (plans started by
campaign start).
Added
report --metrics [--json] [--plan NAME]prints read-only flow metrics:
per step, verify runs, failures, reused runs and seconds saved, executed
seconds, and time fromstarted_atto the first failing and first passing
run (n/awithoutstarted_at); totals add test-count drops and findings
review rounds with open, fixed, and dismissed findings. It never moves the
report cursor and writes nothing.- CI cancels superseded pull request runs through a workflow
concurrency
group; pushes tomainare never cancelled. verify run --reuseandplan verify --reuserecord an earlier passing run
as reused, without executing the command again, when the command, git
commit, worktree digest, Mythify version, and environment digest all match.
The new record carriesreused_fromandreused_duration_seconds, counts
for the step gate like a fresh run, and shows as reused inhistory,
status, andreport. Zero-test and no-op passes and failed runs are never
reused. Every executed record gainsprovenance.environment_digest;
MYTHIFY_REUSE_ENVnames extra environment variables to include. Reuse is
opt-in because the digest does not cover files ignored by.gitignore,
services, or the network.- A test-count ratchet. A passing run that executed fewer tests than the
previous passing run of the same command recordstest_count_dropand
warns.step ID completedcompares the satisfying run with the step's
baseline run and warns on a lower count; on astrict_contextplan it
refuses unless--accept-test-drop REASONis given, which is stamped on the
step astest_drop_accepted.campaign advancetakes the same flag, and
statuslists each command whose latest passing run dropped until a later
run no longer does. - Findings reviews with bounded rounds:
review round NAMEopens a round
stamped with the commit and worktree digest it saw (or records a clean pass
with--clean),review findingrecords a finding on the latest round, and
review resolvemarks it fixed or dismissed (a dismissal needs a reason).
A new round is refused while an earlier finding is open, and round 4 needs
--rethink TEXTexplaining why the approach still stands after three
rounds. A finding that repeats a dismissed one is refused; one that repeats
a fixed one is recorded withrepeat_ofand a warning.review show
reports whether the latest clean round still covers the current source,
statuslists open findings, a round-3 review without a clean pass, and a
stale clean round (newest review only), andprompt reviewcarries the
re-review scope and the dismissed findings.review provestill accepts only blast-radius reviews. - Risk from the changed paths.
protocol/classification-rules.json(schema 5)
gainspath_riskglobs, andclassify_pathsrates a list of paths high,
medium, or low. A step stampsstarted_commitandstarted_atwhen it goes
in progress. Completing a step whose changed paths are high risk needs a
review that covers the current source and--human-input TEXT, the person's
own words approving the change: it warns by default and refuses on a
strict_contextplan. The step recordsrisk,human_input, and, when
MYTHIFY_RISK_GATE=0waived the check,risk_gate_waived.stepand
campaign advancetake--human-input,plan showprints a step's risk,
androute --jsonaddsclassification.diff_risk, which makes the review
advisory independent when the changed paths are riskier than the task
wording. - Named checks.
verify check NAME --set COMMANDstores a command in
.mythify/checks.json(names are[a-z0-9-]+; a command that cannot fail is
refused),verify checklists them, andverify check NAMEruns one and
records an executed verification stamped withcheckandcheck_template.
{files}expands to the changed files, or to--files, passed to the
command as positional parameters ("$@") so a file name is never parsed as
shell text; with none, nothing runs and nothing is recorded.--reuse,--timeout,
--claim, and--jsonare supported, andMYTHIFY_DISABLE_RUN=1refuses
with exit 2. A step'sverify_commandmay be@NAME:plan verifyruns the
check, a run of that check made while the step is in progress counts for it,
and the step gate accepts any passing run of the check whatever its expanded
text.plan createandplan add-stepwarn when@NAMEis not defined yet.
Prompt packets (prompt next,prompt handoff,campaign prompt) and
statuslist the defined checks. verify gate [--plan NAME] [--json], a read-only check for host stop or
pre-commit hooks. It judges each in-progress step with the same rules as
step ID completed(including stale evidence, and a test-count drop or a
high-risk change under strict context), runs nothing, and writes nothing.
Exit 0 means nothing is in flight without evidence; exit 2 means unverified
work is in flight, with the step and the next command on stderr; exit 1
means the active or named plan is missing or unreadable. See
docs/checks-and-gate.md.
Changed
- The step evidence gate moved out of
scripts/mythify.pyinto
scripts/mythify_step_gate.py, with no change in behavior.
Protocol
An agent following the 6.1 text now also: passes --reuse only when an
identical passing run proves the claim; notices a test-count drop and gives
--accept-test-drop REASON under strict context; runs named checks with
verify check NAME; lets verify gate block unverified work in a host hook;
records review rounds, fixes or dismisses each finding with a reason, never
raises a dismissed finding again, and rethinks the approach after three
rounds (--rethink); and asks the person for --human-input before
completing a step that changed high-risk paths, never writing those words
itself. MYTHIFY_RISK_GATE=0 waives the risk check with a stamp.
Also in this release: 6.1.0
6.1.0 was never published on its own, so its changes ship here for anyone upgrading from 6.0.0.
Added
- Plan-backed campaigns:
campaign start|status|prompt|watch|advance|learn
andprompt campaign. Starting a campaign reuses the current plan unless
a name is specified, preserving steps, results, and verification anchors.
Advancement uses the plan's strict evidence gate. Prompts and watchers
only steer the host; they do not launch another executor. - Campaign continuation in
prompt next, resume routing, typed MCP tools,
and the installed chat skills. - A campaign guide covering existing-plan preservation, evidence-gated
advancement, learning, prompt watching, and installed-copy upgrade checks.
Fixed
- Crafted active-plan pointers and plan-store symlinks can no longer direct
plan or campaign reads and writes outside the plan directory. - Campaign inspection leaves corrupt JSON unchanged and refuses to proceed,
preserving the read-only behavior of status, prompt, and watch commands.
Full Changelog: v6.0.0...v6.2.0
Mythify v6.0.0
Summary
Major release: one runtime, model agnostic, smaller. Mythify is now standard
library Python 3.9+ end to end: the Node MCP server is gone, and mythify mcp
serves MCP over stdio by running the CLI for every tool call. Mythify no
longer names, ranks, routes to, or spawns any model, provider, or vendor CLI;
the host picks any subagent it wants, and delegated output stays material
until a verify run on the integrated result passes. The CLI goes from 36 to
22 top-level commands: 16 are removed, seven of them views folded into
status, and product (product planning with human gates and measured
outcomes) and mcp are added. The generated AGENTS.md shrinks from 36,510
bytes to 7,777. The 13 bugs reproduced in the 5.8 review, and the gate and
installer bugs found during the release reviews, are fixed with regression
tests. scripts/lint.py checks the repository for drift.
The full list of breaking changes, additions, fixes, removals, and migration notes is in CHANGELOG.md.
Upgrading from 5.x
- Run
scripts/install_user.shfrom the 6.0.0 checkout or the CLI tarball below. Node is no longer needed, and the installer removes a verified 5.x install. - Copy the new
AGENTS.mdinto your project, andCLAUDE.mdonly if your host reads it and you have no CLAUDE.md of your own; runmythify protocol check. Delete.cursorrules. - Update saved MCP tool names (for example
workflow_routeis nowroute,memory_storeis nowmemory_set); the table is in the changelog.
Assets
mythify-cli-6.0.0.tar.gz: standalone CLI, deterministic archive.mythify.skill: the chat skill bundle.SHA256SUMS: checksums for both.
Mythify v5.8.0
Full Changelog: v5.7.0...v5.8.0
Mythify v5.7.0
Full Changelog: v5.6.0...v5.7.0
Mythify v5.6.0
Full Changelog: v5.5.0...v5.6.0
Mythify v5.5.0
Full Changelog: v5.4.0...v5.5.0
Mythify v5.4.0
Full Changelog: v5.2.0...v5.4.0
Mythify v5.3.0
Full Changelog: v5.2.0...v5.3.0
Mythify v5.2.0
Mythify makes an AI coding agent prove its claims. Before any piece of work can
be marked finished, the agent has to run a real command and show what came back.
What is in 5.2.0
Wayfinding maps. Work that is too big for one session, where the route to
the finish is not visible yet, now gets a decision map instead of a premature
plan. A ticket is a question, not a slice of the build, and its type decides who
is allowed to answer it. A grilling or prototype ticket will not close
without --human-input, because an agent that answers its own question has
proved nothing. When no ticket and no fog remain, map promote hands the
destination, its decisions, and its scope boundary to a plan.
Graph-engineering hardening across both runtimes.
Evidence and provenance fixes, including reading the expected provenance
version from package.json rather than a duplicated constant.
Dependency audit fixes for fast-uri, hono, and ip-address.
The MCP surface is 47 tools: 44 core plus 3 fanout, mirrored between the CLI and
the packaged server and checked by protocol/surface-manifest.json.
Documentation
The project front door has been rewritten for a broader audience, including
people shipping software with AI assistants who are not engineers themselves.
The README now leads with the concrete failure Mythify prevents, start-here.md
states the one habit before the command surface, and the docs index routes by
intent instead of listing files. This landed after the tag, so it is on main
rather than inside the archives attached here.
Two documentation drift bugs were fixed in the same pass: docs/README.md and
docs/design.md both claimed a 38-core MCP surface, against an authoritative 44.
Install
git clone https://github.com/hannsxpeter/mythify.git
cd mythify
./scripts/install_user.sh --project /path/to/your/projectNo account, no API key, no npm install. Zero-dependency Python 3.9+ for the CLI,
with an optional Node 20+ MCP server.
To install without a clone, download mythify-cli-5.2.0.tar.gz below, extract
it, and run ./mythify-cli-5.2.0/scripts/install_user.sh --skip-mcp --project /path/to/your/project.
Assets
| Asset | What it is |
|---|---|
mythify-cli-5.2.0.tar.gz |
Standalone CLI: Python runtime, protocol manifests, chat skills, installer |
mythify-mcp-5.2.0.tgz |
MCP server package for a local npm install |
mythify.skill |
Chat skill archive |
SHA256SUMS |
Checksums for the three assets above |
Verify before installing:
shasum -a 256 -c SHA256SUMSEvidence, honestly
The efficacy evidence for this project is one small reproducible smoke
comparison, documented in docs/evidence/efficacy-reproduction.md. It confirms
that the evidence mechanism works in that run. It is not a demonstrated
improvement in task success or speed, and the project does not claim one.
Full Changelog: v5.1.0...v5.2.0